Datasets:

pszemraj

/

scientific_lay_summarisation-plos-norm

like 6

The MYC genes encode nuclear sequence specific–binding DNA-binding proteins that are pleiotropic regulators of cellular function, and the c-MYC proto-oncogene is deregulated and/or mutated in most human cancers. Experimental studies of MYC binding to the genome are not fully consistent. While many c-MYC recognition sites can be identified in c-MYC responsive genes, other motif matches—even experimentally confirmed sites—are associated with genes showing no c-MYC response. We have developed a computational model that integrates multiple sources of evidence to predict which genes will bind and be regulated by MYC in vivo. First, a Bayesian network classifier is used to predict those c-MYC recognition sites that are most likely to exhibit high-occupancy binding in chromatin immunoprecipitation studies. This classifier incorporates genomic sequence, experimentally determined genomic chromatin acetylation islands, and predicted methylation status from a computational model estimating the likelihood of genomic DNA methylation. We find that the predictions from this classifier are also applicable to other transcription factors, such as cAMP-response element-binding protein, whose binding sites are sensitive to DNA methylation. Second, the MYC binding probability is combined with the gene expression profile data from nine independent microarray datasets in multiple tissues. Finally, we may consider gene function annotations in Gene Ontology to predict the c-MYC targets. We assess the performance of our prediction results by comparing them with the c-myc targets identified in the biomedical literature. In total, we predict 460 likely c-MYC target genes in the human genome, of which 67 have been reported to be both bound and regulated by MYC, 68 are bound by MYC, and another 80 are MYC-regulated. The approach thus successfully identifies many known c-MYC targets and suggests many novel sites. Our findings suggest that to identify c-MYC genomic targets, integration of different data sources helps to improve the accuracy. MYC plays a critical role in regulating cell proliferation, growth, apoptosis, and differentiation. Human malignancies are often associated with aberration of the c-MYC gene [1–3]. The diversity of its functions has been attributed to c-MYC' s ability to activate or repress the transcription of an extensive array of target genes mediating a wide range of cellular activities [4–6]. c-MYC' s actions are mediated by sequence-specific binding of the c-MYC protein, dimerized with its partner MAX, to DNA elements called E-boxes with the core sequence motif 5′-CACGTG-3′ [7–9]. Binding of the MYC–MAX heterodimer to a target gene can directly activate or repress transcription, but many E-boxes do not bind MYC, and in many experimentally confirmed cases, MYC binding is not associated with changes in gene expression. Identifying functional MYC binding sites and target genes is a critical step in understanding both the biological role and molecular mechanism of MYC action. mRNA expression studies have identified many target genes activated or repressed by c-MYC in various animal and human cells or cell lines. The number of experimentally validated c-myc targets are expanding rapidly thanks to the use of high throughput methods [10–13]. The recent studies of Basso et al. [14] and Remondini et al. [15] suggest that the potential list of c-MYC targets could be much larger than what was previously anticipated. However, different experimental and theoretical studies give quite different findings, and much work remains to be done to define the complete set of c-MYC targets. Gene expression studies alone cannot discriminate between direct and indirect targets of c-MYC action, although network-based inference of direct action has been proposed [14]. Recently, chromatin immunoprecipitation (ChIP) and whole genome scale analysis of methylation status have emerged as new sources of relevant data for the analysis of genomic regulatory elements. Theoretical analysis complements and extends experimental study, and several researchers have attempted to predict c-MYC target genes using computational methods. By searching human transcript sequence, analyzing E-box location, and using genomic sequence evolutionary conservation, Schuldiner et al. [16] identified 12 putative targets, two of which were confirmed by subsequent experimental analysis. Zeller et al. [17] built a database of c-Myc responsive genes that have been reported in publications and supported by multiple lines of evidence. They then identified seven out of 12 candidate genes in this database using phylogenetic analysis, and they confirmed six of these predictions using ChIP. The use of evolutionary conservation is also supported by Haggerty [18] who categorized c-MYC targets into two classes: Class I, to which the majority of genes belong, has E-boxes that are evolutionarily conserved; and Class II, which includes genes with no region of homology at or flanking the genomic regions that exhibit MYC binding. The promoter regions of c-MYC–regulated genes often contain E-box sequences that bind MYC with high occupancy. Li et al. [17] analyzed genomic binding sites for c-MYC in Burkitt lymphoma cells and found a strong correlation between MYC DNA binding and gene transcription, strengthening the view that high binding occupancy of a c-MYC site near a gene' s promoter region is, to some extent, a sign of a c-MYC target gene. However, this is not a hard and fast rule, and many sites that bind MYC with high occupancy are not associated with c-MYC target genes. Fernandez et al. [19] performed a large-scale assay for genomic Myc binding sites in vivo by quantitative ChIP. They found that promoter E-boxes are distributed in two groups differing in MYC binding occupancy. The strongest DNA-sequence characteristic of high-affinity/high-occupancy targets was location of the E-box within a CpG island. This observation can be partially explained by the fact that most CpG dinucleotides in the mammalian genome are subject to cytosine methylation [20,21], but the methylation of the CpG dinucleotide in the consensus myc binding site sequence reduces the binding affinity of myc–max dimers for the target DNA [22,23]. Further, DNA methylation is often coupled to and associated with histone methylation and the formation of heterochromatin [24]. Recent work by Ernesto [25] shows that target sites are only recognized MYC by whether they are packaged in chromatin bearing high H3 K4/K79 methylation and H3 acetylation. This is true for both classic E-box (CACGTG) and alternative sequence sites. With the abundant volumes of experimental data on c-MYC target gene expression and in vivo binding of MYC to the genome, there is an increasing need to integrate these information resources and to classify c-MYC responsive genes into direct and indirect targets. In this paper, we identified c-MYC target genes using a computational approach that draws on data from multiple sources including gene expression profiling, gene annotations, ChIP, sequence conservation, and sequence composition. First, we developed a computational model to predict the likelihood of CpG methylation. Next, we developed a computational strategy to predict which E-box sites were likely to be functional MYC binding sites. Finally, the binding predictions were integrated with gene expression and gene annotation data to identify direct and indirect c-MYC target genes. The performance of these tools was validated by comparison with multiple experimental datasets. Our method is able to successfully predict the occupancy of the binding sites as revealed by CHIP. Although this computational method was specifically built on c-MYC data, it also provides useful information on the binding of cAMP-response element-binding protein (CREB) [26], another transcription factor whose binding is sensitive to DNA methylation. After further integration with gene expression data from different tissues and datasets, and the Gene Ontology (GO) annotations for c-MYC targets, we identified 460 likely c-MYC target genes. Of these genes, 215 have been previously identified as MYC bound or regulated and 245 are novel. Our study shows that integrating multiple data sources improved the prediction specificity of MYC binding site prediction. Similarly, using gene expression profiles from several independent studies improved the sensitivity for target gene prediction. Our analysis suggests that much of the variation between microarray-based gene expression assays may be due to limitations of the technology. In addition, there appears to be a significant tissue-specific component to the responses of some c-MYC target genes. The performance of the fifth-order Markov model for genomic methylation status was tested using an independent dataset from the Human Epigenome Project [27]. We found that 86. 4% of these unmethylated sites fell within our predicted hypomethylation regions. In contrast, only 22% of the hypermethylated CpG sites were within the predicted hypomethylated regions (p < 2. 2e−16). We applied this model to predict the CpG islands and hypomethylation regions on the human genome. On the human repeat masked genome NCBI35, we found that 0. 71% of the human genome sequences were CpG islands and 1. 07% of the genome sequences were predicted to be hypomethylated regions, with 82% of the CpG islands falling within hypomethylated regions. As anticipated, promoter regions had much higher percentages of CpG islands and hypomethylated regions than the whole genome. Based on our prediction results, 43% of the human genes had CpG islands and 46% of the genes had hypomethylated regions within 5 Kb upstream of their transcription start sites. In the 5 Kb upstream of the transcription start sites, 4. 4% of the sequence was covered by a CpG islands and 6. 7% of sequences were predicted to be hypomethylated. In this step we identified sites in the human genome where MYC is expected to bind in vivo with high occupancy. Candidate MYC binding sites were identified by scanning the complete human genome sequencing using the TRANSFAC MATCH algorithm [28] and the MYC–MAX position specific weight matrix (PSWM) MA0059 [29] from the JASPAR database [30,31]. Sites that achieved a matrix score of 0. 8 are referred to as motif matches. Four additional sources of data were used to define a subset of these motif matches that are likely to bind MYC in vivo: proximity to transcription start sites, proximity to CpG islands, predicted hypomethylation, and evolutionary conservation. To train our algorithm, we began with the Fernandez et al. [19] data for genomic Myc binding sites in live human cells. This study examined MYC binding to more than 700 E-box sequences in vivo by quantitative ChIP. The authors found that promoter E-boxes were distributed in two groups that bound MYC at distinct frequencies. In the Fernandez et al. dataset, we only used sites where the PCR primers and the associated E-box could be mapped back to the human genome as an exact sequence match. Further, sites were excluded when the dataset contained multiple contradictory assay results. This filtering process resulted in a set of 493 binding sites where ChIP was measured in either U937 or HL60 cell lines (for 40% of the sites, data is available and consistent across both cell lines). A high quality training dataset was defined using sites where ChIP results were consistent across more than one assay. The training set had 43 nonredundant high-occupancy sites and 90 nonredundant low-occupancy sites. Using a Bayesian network classifier, we integrated this high quality training subset of the ChIP dataset with our hypomethylation analysis and genomic sequence conservation data. The resulting classifier predicts which sites in the genome are likely binding MYC with high occupancy. In the training process, we looked for the most important factors determining MYC binding to DNA. The Kolmogorov–Smirnov test showed that the following factors differed significantly (p < 0. 01) between the high- and low-occupancy sites: distance to transcription start site (Figure 1A), distance to nearest CpG island (Figure 1B), distance to nearest hypomethylation region (Figure 1C), phastCons scores [32] (Figure 1D), and distance to nearest chromatin acetylation island (Figure 1E). The CpG islands and hypomethylation regions near or overlapping with the above MYC binding sites were predicted with our Fast Motif Analyzer (FMA), described in the Materials and Methods section. The acetylation island information was derived from the Roh et al. high-resolution genome-wide mapping of Lys 9 and Lys 14 diacetyl histone H3 in resting and activated human T cells [33]. PhastCons scores identify evolutionarily conserved elements using a multiple alignment of genomic sequence weighted using a phylogenetic tree. The PhastCons score is a base-by-base conservation score that can be interpreted as the probability that each base is in a conserved sequence element [32]. This score is chosen as a measurement of sequence conservation over other alternatives because it is available through the widely used University of California Santa Cruz Genome Browser, and could be readily incorporated into our model. Using these datasets and a supervised classification approach [34], we evaluated several algorithms for the prediction of high-occupancy MYC binding sites. The best classification results were obtained using a Bayesian network classifier (see Materials and Methods). This method assigned a probability for every site in the genome. We predicted a site to have high binding if this probability was above 0. 5. The Bayesian network shown in Figure 2 gave the best results on the high-quality training dataset. A 10-fold cross-validation showed a precision of 0. 88 and a recall of 0. 98 on the high-occupancy binding sites prediction. When we applied this classifier to Fernandez' s entire dataset, it correctly identified 130 of the 183 high-occupancy sites and 258 of the 310 low-occupancy sites. To further verify that our algorithm was able to predict high-occupancy MYC binding sites, we evaluated performance of independently obtained test data. We considered binding sites within ±3 Kb of transcription start sites, because almost all of the sites in the training data are within this region. We used the MYC–MAX matrix MA0059 [29] from the JASPAR database [30] to identify putative MYC binding sites, and we applied the rules described above to sort these sites into groups predicted to bind MYC with high and low occupancy. Using TRANSFAC MATCH to analyze the complete human genome sequence, we identified 89,560 MYC sites within ±3 Kb of a transcription start site for 14,387 genes. From these candidate sites, our method classified 14,638 sites in 5,276 genes as likely to bind MYC with high occupancy. We assessed the reliability of these predictions by comparing them with two independently published experimental datasets. One dataset is from Zeller et al. [35] using ChIP–PET to map genomic c-MYC binding sites in human B cells. This study the identified loci as PET sequence tag clusters with varying numbers of tags per cluster. More tags matching a cluster increases the reliability of binding site identification. Zeller' s paper defined 964 PET-2+ clusters with two or more tags per cluster falling within 3 kb of a TSS. PET-2+ may contain a significant number of false positive identifications. Zeller et al. also defined 113 PET-3+ cluster with three or more tags per cluster and within 3 kb of a TSS. These are believed to be highly reliable identifications. For the second evaluation dataset, we used high-density oligonucleotide array ChIP–chip data from Cawley et al. [36]. Looking only at Chromosomes 21 and 22, Cawley et al. defined 181 high-occupancy MYC binding segments within 3 kb of a TSS. In the experimental datasets, MYC binding is localized to a segment of genome, but not necessarily a single E-box. We refer to these segments as “MYC binding loci” and compare these with our predictions at the gene level. Most of the experimentally defined MYC binding loci are associated with a single gene. Table 1 compares the experimentally defined MYC binding loci that are within 3 Kb of a transcription start site with those that were predicted by our methods. First, it is apparent that the different experimental assays yield markedly different results. Only one of the 113 Zeller PET-3+ loci was also scored as a high-occupancy binding site by Fernandez et al. Similarly, only one of six Zeller PET-3+ loci on Chromosomes 21 and 22 was identified as a high-occupancy site by Cawley et al. , and only four of the 181 binding loci identified by Cawley et al. were scored as high occupancy in the Fernandez dataset. Comparing our predictions with the experimental datasets, we see higher levels of agreement than the experimental datasets show among themselves. This is in part a consequence of the fact that our method predicted a larger number of loci (5,276) than are observed in the experimental datasets. Interestingly, the fractional overlap between our predictions and the Zeller et al. dataset increases as we consider more confident clusters. Whereas only about a quarter of the PET-2+ clusters contain a predicted high-occupancy site, half of the PET-3+ cluster and all of the PET-4+ clusters do. One-third of the loci identified by Cawley et al. on Chromosomes 21 and 22 contain a site predicted to bind MYC with high occupancy, and more than a quarter of our predicted high-occupancy sites on these chromosomes are confirmed by Cawley et al. To extend the analysis to more tissues and cell types, we used data from the MYC Target Gene Database. We compared predictions for genes annotated as c-MYC targets with predictions for a collection of genes selected at random from the human genome. The MYC Target Gene Database [17] includes more than 1,000 putative MYC target genes reported to be either regulated or bound by Myc. These MYC target genes tend to have more MYC recognition sites than do randomly selected genes. Table 2 shows the performance of our algorithm in the prediction of high-occupancy binding sites. Each gene in Table 2 contains at least one putative MYC site predicted by a motif match. The MYC target genes are compared with a set of 1,000 randomly selected genes containing 6,259 MYC sites, which served as the control group. Second, we chose 589 genes showing myc binding from the MYC Target Gene Database that were not in Fernandez' s dataset. Third, we chose 417 genes showing Myc regulation for which no binding data was available. The fourth set consisted of 98 genes in MYC Target Gene Database showing both myc binding and myc regulation that were not in Fernandez' s dataset. For each group of genes, we used our algorithm to predict the occupancy of each MYC recognition site. Statistical significance was assessed using Fisher' s exact test on each test group against the control group. We found that each of the putative c-MYC target groups had a significantly higher percentage of high-occupancy binding sites than the genes in the randomly selected group. This demonstrates that our algorithm is able to discriminate biologically functional high-occupancy sites from low-occupancy, presumably nonfunctional binding-site motif matches. We next asked what genes were associated with the high-occupancy binding sites. Many genes have multiple MYC binding sites, but little information is available on how multiple MYC sites affect each other' s binding. Therefore, in our study we treated all the genes having high-occupancy binding sites as potential MYC binding genes. Table 3 shows the result of these predictions. The results show that known c-MYC targets were predicted to bind MYC with a significantly higher frequency than random genes do. Thus, our method has significantly higher accuracy in discriminating the MYC binding genes than does motif match alone. Although our method was built using c-MYC data, the attributes used in the model are general, so we anticipated that the method might also be informative for other transcription factors that are sensitive to DNA methylation. CREB has a CpG dinucleotide in its binding sequence and its binding is sensitive to DNA methylation. We applied our method to the analysis of CREB sites and compared our predictions with a previous ChIP–chip study [26] of 10,209 distinct promoters on the human genome that were predicted to have at least one cAMP-responsive element defined as a match to a simple cAMP-responsive element consensus-site algorithm. Forty of these sites were confirmed by manual ChIP assays. Applying our model to these 40 genes, we correctly classified 21 of the 23 high CREB occupancy genes with seven false positives (Table 4). The ChIP–chip assay identified 2,195 CREB high-occupancy binding sites near promoters. Our model correctly classified 1,713 (78%) of these high-occupancy promoters and gave 3,621 false positives (Table 5). However, many false positives could actually be true positives, because Zhang et al. pointed out in their paper that, although their ChIP–chip method had a high specificity, only 54% of the promoters occupied by CREB in their manual ChIP assay showed positive in their ChIP–chip assay. These predictions on CREB show that our method is useful for the analysis of other transcription factors that are sensitive to epigenetic factors. Myc binding does not always imply c-MYC regulation, and a great deal remains to be learned about how the cell determines which genes are actually regulated by c-MYC under different conditions and in different tissues. To address this issue, we analyzed the co-expression pattern of genomic genes and c-MYC genes in two tissues where c-MYC is reported to play an important biological role: B cell lymphoma and prostate cancer. The c-MYC gene is often deregulated in cancer and induces the expression of many c-MYC target genes. Notable examples include B cell lymphoma and prostate cancers. We selected two datasets for analysis, a human B cell dataset [14] of 336 samples and a human prostate cancer dataset [37] of 102 samples based on number of samples in the dataset, availability of the raw data, and thus use of a standard gene expression analysis platform, the Affymetrix HG-U95Av2 microarray. The raw data for both datasets were reprocessed with Bioconductor [38] and the RMA algorithm [39]. Figure 3 shows the RMA normalized log transformed expression signals of the three c-myc probe sets on the HG-U95Av2 GeneChip for each dataset. In this study we used two c-myc probe sets, “1973_s_at” and “37724_at”, because their expression signals demonstrated a strong and consistent correlation, whereas “1827_s_at” was not strongly correlated with either of the other two (see Figure 3). The Pearson' s correlation coefficients between each c-myc probe set and every other probe set on the chip was calculated, and this value is referred to as the co-expression pattern of c-myc with the other genes. Figure 4 shows the distribution of these co-expression patterns for data derived from B cells and prostate cancers. Overall, 1,217 and 1,418 genes were found to be significantly correlated with c-MYC expression (FDR < 0. 01) in B cells and prostate cells, respectively. Altogether, 2,233 genes were highly correlated with c-MYC in these two datasets; 403 genes were correlated with c-MYC expression in both datasets. Table S1 lists the MYC correlated genes identified in these two tissues. Many known MYC target genes have functions involved in cell cycle progression, apoptosis, and cellular transformation; it is reasonable to hypothesize that unknown MYC targets will share many of these functions. We constructed a set of gene functions that were overrepresented in known MYC targets as represented in GO terms. Using the experimentally verified MYC targets in the MYC target database [17] and considering only the genes on the HG-U95Av2 chip, we found 144 known MYC targets to which 875 terms in the Molecular Function and Biological Process categories were applied. Because some GO terms are too general to be as informative, in subsequent analysis we only used the GO terms with fewer than 500 genes. Applying a hypergeometric test, we found 156 GO terms overrepresented in the MYC target gene sets (p-value < 0. 025; Table S2). Among the top overrepresented GO terms are those related to cell cycle, biosynthesis, nucleic acid binding, and translational regulation. These findings correlate well with expectations based on the biology of c-MYC. Using the predicted MYC high-occupancy binding sites, the c-MYC co-expressed genes in B cells and prostate cells, and genes annotated with GO terms that are overrepresented in known c-MYC targets, we applied a rule-based procedure to define likely c-MYC targets. For a gene to be labeled as a c-MYC target, it must meet four criteria. First, the gene must have at least one MYC high-occupancy binding site. Second, the signal variance of the gene' s probe set must be greater than the mean variance across all probe sets in either the B cell or prostate cancer dataset. Third, the correlation coefficient must be greater than zero and the significance of gene' s co-expression with c-MYC must be less than 0. 01 in either the B cell or prostate cancer dataset. Fourth, the GO annotation for this gene must have at least one overrepresented c-MYC target-related GO term. Applying these rules, we found 440 genes that meet our criteria (see Table S3). Comparing our findings with the Myc Target Gene Database [17], we found that in the literature, 128 of the predicted target genes are reported to bind MYC and 142 are reported to be regulated by MYC. This represents an independent validation of our findings because data from the MYC Target Gene Database was not included in our training data. Sixty-two of the predicted target genes were reported to be both bound and regulated by MYC; we successfully predicted 62 out of the 144 known MYC targets on HG-U95Av2 platform. Among the predicted genes, 264 were correlated with MYC expression in B cells, 277 were correlated with MYC expression in prostate cancers, and 101 were correlated in both datasets. This high level of correlation further validates our predictions. Among the 62 previously identified targets, 42 correlate with MYC expression in the B cell dataset, 39 in the prostate dataset, and 19 in both. These findings show that many MYC targets exhibit some tissue specificity in MYC responsiveness. Figure 5 also shows that many of the known MYC targets in the MYC Target Gene Database [17] have different correlations of gene expression with MYC expression in the two tissues. To further investigate c-MYC tissue specific responses, we examined seven additional microarray gene expression datasets from breast, lung, prostate, and leukemia cancers (Table 6). All these datasets used the Affymetrix HG-U95Av2 platform. Adding correlation of gene expression with c-MYC in these datasets, we were able to identify 20 additional MYC targets. In total, 460 c-MYC target genes were predicted including 215 in the MYC Target Gene Database that had previously been reported to be bound or regulated by MYC (Table S3); 144 were regulated by MYC and 132 were bound by MYC. Further, we found evidence in the literature to validate three additional MYC target gene predictions, ATF3 [40], HSP90A [41], and BAT1 [42]. Overall, 218 of our 460 predictions were validated, including 67 genes that have evidences for both binding and regulation. We believe that this is an underestimate for the true number of MYC targets because we only considered the 8,000 GO annotated genes on Affymetrix HG_U95Av2 platform; this is slightly more than one-third of the genes in the human genome. We compared our predicted 460 genes from nine datasets with the 2,063 genes in the MYC subnetwork predicted by Basso et al. [14] and the 668 high-quality MYC direct responsive genes identified by Zeller et al. [35]. Figure 6 shows the overlaps between these datasets. We see that the overlap number is higher than would be expected purely by chance, but well below complete agreement. Comparing these datasets with the genes in the MYC Target Gene Database (Table 7), we find that our method shows a better specificity than the other two approaches, even without applying GO filtering. Using the GO functional annotation improved the specificity. We investigated the 45 targets identified by Basso et al. that were missed in our 460 predicted targets. Experimentally, these targets exhibit both binding and regulation. Thirteen targets were missed due to their relatively low expression variation or nonpositive correlation with c-MYC across samples, six targets did not have any c-MYC binding motif within 3 Kb of their transcription start sites, 15 targets were false negatives in high-occupancy sites prediction, and the other 11 targets did not have overrepresented GO terms. Using a Bayesian model, we integrate genome sequence data and epigenetic information to identify myc recognition sites in the human genome likely to bind c-MYC with high occupancy. By combining the myc binding probability, gene co-expression data, and functional annotations, we predicted 460 c-MYC targets among the genes presented on Affymetrix HG-U95Av2 platform. The list of predicted c-MYC targets contains many genes found previously in the literature, but also 245 genes not previously identified as c-MYC targets. Our method only predicts upregulated c-MYC targets because downregulated c-MYC targets are not generally mediated by E-box binding. Among the 67 predicted genes that have already been observed to be bound and regulated by MYC, 61 are upregulated by MYC, five genes are reported to be downregulated, and one has evidence for both downregulation and upregulation. Thus, the predicted MYC targets agree well with previous observations. In addition to these 67 independently validated predictions, we also identify 148 genes in the MYC Target Database, 68 of which are bound by MYC in vivo and 80 of which are MYC regulated (Table 6). Among the 250 predicted novel targets, 27 correspond to MYC binding loci reported by Zeller et al. [35] and 11 of these correspond to highly reliable binding loci with PET-3+ clusters. Different from previous studies predicting c-myc targets, our study integrated four sources of data (genomic sequence, gene expression, ChIP, and functional annotation) to improve the specificity of predictions. Tools such as TRANSFAC or MatInspector, which rely on motif matches alone to predict myc binding, have very high false positive rates. The improved specificity obtained by our integrated approach emphasizes the importance of epigenetic factors in modulating c-MYC binding to DNA. Although epigenetic status does vary with tissue type and other factors, recent high throughput studies show that tissue-specific variation in genomic methylation is limited [27]. Adding genomic acetylation islands data obtained on T cells to the model improved the prediction precision from 0. 81 to 0. 88 and the recall from 0. 88 to 0. 98 in cross-validation. Among the attributes considered in the model, we found that the distance to the nearest hypomethylation region is the most informative; this attribute alone could correctly identify 80% of the high- and low-occupancy sites in the cross validation. However, adding the additional attributes does improve performance, and considering all five attributes allows the model to correctly identify 95% of all the cases in cross-validation. MYC is not the only transcriptional factor whose binding is sensitive to the epigenetic factors such as DNA methylation or chromatin acetylation. Because the attributes used in our model are general and the epigenetic factors could influence DNA binding through similar mechanisms, elements of our model may also be useful for other transcription factors. As a validation we applied our MYC binding prediction model to another transcription factor, CREB [26]. Like MYC, the CREB consensus binding sequence has a CpG dinucleotide and CREB binding is sensitive to DNA methylation. Our analysis shows that although our model was specifically built for the study of c-MYC, it is still able to correctly discriminate most of the high-occupancy binding sites and the majority of low-occupancy binding sites. Thus, our approach to modeling chromatin structure effects is transferable to other transcription factors. ChIP–chip technology can partially address the question of where transcription factors bind the genome, but with current technology the resolution of TF binding loci is limited and the data are error-prone. In addition, computational modeling can help to understand the complex transcriptional machinery. For example, adding or removing an attribute to the model and assessing the affects on performance is one way to evaluate the importance of this attribute on the regulation of the DNA binding by a transcription factor. Combining different types of data offsets the shortcoming of each. Obviously, our predictions based on genome sequence alone cannot address tissue specific binding. However, by taking gene expression data of specific tissues into consideration, we restrict our identified targets to those functionally regulated by MYC in the context of certain tissues, which would be of real interest to the biological community. In addition, integration of different types of data also allows us to predict which DNA–protein binding sites are likely to trigger transcriptional regulation. Figure 4 shows that compared with the expression of all genes, a higher portion of genes from our MYC binding and function predictions are highly correlated with c-MYC expression. This demonstrates that our binding prediction and gene function analysis do identify bona fide c-MYC targets and are helpful in improving the specificity of prediction. One of the limitations in our target gene analysis is that we only consider genomic sequences within 3 Kb of transcription start sites. This was done because almost all of the sites in our training data fall within this region, but there are some high-occupancy MYC target sites far from any known transcription start sites. It is possible that there are direct c-MYC targets where the only functional MYC sites are more than 3 Kb from the transcription start site. Increasing the search window for c-MYC recognition sites might improve the sensitivity of prediction, but such a change would also decrease specificity. Another limitation of our method is that we only consider E-box–dependent MYC binding and regulation; MYC targets regulated through other mechanisms will not be identified. Transcriptional inhibition of MYC targets is often mediated by mechanisms unrelated to E-box binding [43–47]. Therefore, we only consider the MYC activation in our predictions and consider only the expression of target genes that is positively correlated with c-MYC expression. We do observe a minority of genes where there is a significant anti-correlation of gene expression with c-MYC expression. The analysis of these cases will be the subject of future work. Distinguishing between direct and indirect targets of MYC is an important issue. In the co-expression analysis we found a large number of genes where expression levels showed positive correlations with c-MYC expression and yet lacked a high-occupancy MYC binding site, either predicted or experimental. We believe that many of the genes are not the direct targets of MYC, but it is difficult to exclude the possibility that they contain a functional MYC binding site not detected by our own or experimental methods. MYC responses vary significantly between different tissues. Less than half of our predicted MYC targets show significant gene expression correlation with c-MYC in both B cells and prostate cancers. We analyzed seven additional independent microarray datasets (Table 6) where the c-MYC gene was deregulated and its probe-set signals showed large variance across samples. In these studies, many predicted MYC targets, including genes that have been experimentally verified as direct MYC targets, failed to show a strong correlation of target gene expression with MYC expression. The differences between these datasets cannot be explained by tissue of origin alone. For example, datasets 2,3, and 4 are all derived from the prostate, but the c-MYC correlated targets from these three datasets do not agree more than those from different tissues (see Figures S3–S5). Some of this variation could be a result of technical variation in gene expression profiles between different laboratories and experiments. This is one of the first studies to systematically analyze c-myc targets in multiple datasets and in multiple tissues; most previous studies focused on a single tissue or cell line. Our analysis confirms the well-known finding that c-MYC is deregulated in many cancers and has a direct influence on the expression of hundreds of other genes. One potential pitfall in using GO as a criterion for predicting MYC targets is the possibility of missing important groups of targets that do not fall into a specific GO category or are not annotated by GO at all. In Table S3 we also provide the list of 1,188 predicted targets without applying GO filtering in the prediction. A second concern is that it is difficult to define a test set that is totally independent of prior knowledge because we cannot exclude the possibility that GO annotators were aware of MYC regulation status in assigning gene annotations. Looking at the target genes we predicted that were not in the Basso et al. prediction, we find variable levels of correlation with MYC expression across the different datasets that we have examined. Even genes that have been identified as c-MYC targets in published literature often have very different co-expression patterns with c-MYC in different microarray datasets. This is consistent with the view that many of the differences among these high throughput studies may result from experimental variation, the noise inherent in these approaches, and the effects of cell density or the number of culture passages [48]. Whether this reflects tissue-specific responses or technical variation in microarray data, it is apparent that a study focusing on any single dataset will be insufficient. As is shown in Table 6, using multiple datasets from different studies improved the power of the prediction. Because the current study only predicted upregulated genes with MYC binding motifs close to the transcription start site, which are on the Affymetrix HG-U95Av2 array and which have GO annotations, we believe that the 460 targets identified here underestimate the number of direct MYC targets in the human genome. Our estimates for the number of MYC targets in the human genome are roughly consistent with the MYC database and the high-confidence Zeller [35] and Cawley [36] studies. They are not inconsistent with the larger numbers of c-MYC targets suggested by some other studies [12–15]. One explanation for this range of findings is that MYC binding in vivo is not a Boolean event and even strong MYC binding sites are unlikely to be occupied with unit stoichiometry. Thus, different studies may be applying different thresholds for defining a MYC target. All the genes in the paper and their Entrez Gene IDs can be found in Tables S1 and S3. The HG17 build of the human genome sequence was downloaded from the University of California Santa Cruz (UCSC) genome database. The transcription start sites were from the annotated transcription starts of RefSeq genes in the UCSC genome database (http: //hgdownload. cse. ucsc. edu/goldenPath/hg17/bigZips/upstream1000. zip). PhastCons scores for multiple alignments of seven assemblies to the human genome hg17 were downloaded from http: //hgdownload. cse. ucsc. edu/goldenPath/hg17/phastCons/mzPt1Mm5Rn3Cf1Gg2Fr1Dr1/. The human gene annotation information was downloaded from ftp: //ftp. ncbi. nlm. nih. gov/gene/DATA/gene2refseq. gz, and the GO information was downloaded from ftp: //ftp. ncbi. nlm. nih. gov/gene/DATA/gene2go. gz. Known MYC target genes in previous literature were obtained from the MYC Target Gene Database [17] (http: //www. myc-cancer-gene. org). The large-scale assay for genomic MYC binding sites in live human cells was obtained from the supplementary data of Fernandez et al. [19]. Chromatin acetylation data were obtained from the supplementary data of Roh et al. [33]. MYC binding loci on Chromosomes 21 and 22 [36] were downloaded from http: //transcriptome. affymetrix. com/publication/tfbs/, and MYC binding data in human B cells using ChIP–PET were obtained from the supplementary data of Zeller et al. [35]. The CREB genomic binding loci [26] was downloaded from http: //natural. salk. edu/CREB. Two microarray gene expression datasets were used in this study, the B cell dataset [14] (GSE2350) and the prostate cancer dataset [37] (http: //www. broad. mit. edu/cgi-bin/cancer/datasets. cgi). The B cell dataset contains 336 samples of normal and transformed human B cells, and the prostate dataset contains 52 tumor and 50 nontumor prostate samples. The sources of other microarray gene expression profile datasets are listed in Table 6. Bayesian network classification was performed using Weka 3. 4. 5 [34], which is available for download at http: //www. cs. waikato. ac. nz/ml/weka/. When learning the Bayesian network, the best performance was obtained when using K2 as the search algorithm for local score metrics and using Simple Estimator with alpha at 0. 5 to find the conditional probability tables. In addition, an empty network is set as the initial structure when learning the network. The training data contained 43 high-occupancy sites and 90 low-occupancy sites from Fernandez' s dataset [19]. These sites and their ChIP primers were mapped back to human genome using the UCSC In-Silico PCR tool. The five attributes for every site were derived from the genome sequence and annotations, including: distance to transcription start site, distance to nearest CpG island, distance to nearest hypomethylation region, nearest hypomethylation region score, and phastCons scores [32]. The CpG islands and hypomethylation status of sequence were predicted as described below. The distances in base pairs and the hypomethylation scores were transformed to natural log scale. CpG islands were predicted by the criteria of Gardiner-Garden [49]. Hypomethylation regions were predicted using a fifth-order Markov model likelihood ratio test. where Lnm is the likelihood ratio score that the hexamers in the interval from n to m are drawn from the frequency distribution for a training set of hypomethylated sequences relative to the likelihood that they are drawn from a frequency distribution describing the genome as a whole. si is the residue occurring at position i in the sequence, f (si+5 | si…si+4) is the frequency of finding residue si+5 given the preceding five residues si…si+4 in the hypomethylated DNA collection, and g (si+5 | si…si+4) is the corresponding frequency in the genome as a whole. The hypomethylation score for a region is the sum of the log likelihood ratio scores for all overlapping hexamers in the region, expressed as log base 2. This score measures how similar the hexamer content of a region is to the hexamer frequencies in the hypomethylated sequences training set described below. The boundaries for a predicted hypomethylation region are chosen to maximize this log likelihood ratio score, with constraints that each predicted region must score at least 23. 1 bits and must not contain any subregion whose score is more negative than −14. 5 bits. An optimized hexamer score cutoff is determined as shown in Figure S1. This algorithm is implemented in the program FMA, a C++ program, which utilizes a hexamer model to predict hypomethylation regions. The output includes the exact locations of CpG islands, the fraction of CpG dinucleotides, and the exact location and score (log2Lnm) of each hypomethylation region. FMA also implements an efficient indexed motif search. In the first phase of the FMA search, each PSWM in the query set is analyzed and the most informative contiguous six-nucleotide core segment is identified. Next, a branch-and-bound strategy is used to enumerate all hexamers matching this core segment and able to be extended over the full PSWM to achieve a specified log likelihood threshold for the full match. Hexamers in this set are stored in a suffix tree. During the search, the target sequence is first scanned using the suffix tree to find core segment matches. These matches are then extended over the full PSWM to see if they achieve the specified log likelihood threshold. Using this approach, we are able to exhaustively search large target sequence libraries for matches to large query sets of PSWM. The training data for hypomethylated human genomic sequences were obtained by aligning the hypomethylated sequence tags collected by Cross et al. [50,51] with the human genome sequence and extending to the nearest MseI site. These hypomethylated sequence tag sequences were obtained by digesting human genomic DNA from peripheral blood leukocytes with MseI, selecting fragments that failed to bind a methyl–CpG binding protein column, methylating these fragments in vitro, and subsequently selecting fragments that bound the methyl–CpG binding protein column. This yields a collection of genomic DNA fragments that were not methylated in vivo, but which contained a CpG dinucleotide that could be methylated in vitro. These fragments were subsequently cloned and subjected to end-sequence analysis. By aligning the end-sequence tag to genomic sequence and extending to the nearest MseI site, we reconstruct the sequence of the full hypomethylated DNA segment. Validation data for hypomethylation predictions were obtained from the Human EpiGenome Project [27]. TRANSFAC was used to scan the human genome sequence for putative c-MYC binding sites with the MYC–MAX position weight matrix [MA0059] [29] from the JASPAR database [30]. The sequence attributes for each putative binding site were inputted into the Bayesian network classifier (see above) to predict the probability of MYC binding for each site. A site assigned a MYC binding probability above 0. 5 was considered to be a high-occupancy site. Because almost all the sites in the training data were within ±3 Kb of transcription start sites, we limited the prediction of high-occupancy binding sites to this region for further studies. The prediction of c-MYC binding genes was based on the high-occupancy c-MYC binding sites prediction. If a gene contained any high-occupancy binding site, it was considered a potential c-MYC binding gene. The HG-U95Av2 platform annotation file was downloaded from Affymetrix. The raw data files for each dataset were firstly normalized with RMA in Bioconductor [38]. For each probe set, the signal variance across the samples was calculated. Only probe sets with a signal variance larger than the mean of all probe sets' variances were used in co-expression analysis. The Pearson' s correlation coefficient r was then calculated for every pair consisting of a c-MYC probe set and another probe set. The significance (probability) of the correlation coefficient is determined using the t-statistic: where r is the correlation coefficient and n is the sample size. The p-values from this multiple testing were adjusted to control the false discovery rate of Benjamini and Hochberg [52]. Only probe set pairs with an adjusted p-value less than 0. 01 were considered to be significantly co-expressed. For a gene to be correlated with c-MYC expression, it must have at least one probe set with p-value less than 0. 01 for both c-MYC probe sets “1973_s_at” and “37724_at. ” We analyzed genes on the HG-U95Av2 GeneChip. From the Myc Target Gene Database [17], we extracted the subset of these genes that were reported to be both bound and regulated by myc. For these genes, we collected the associated GO terms from the Molecular Function and Biological Process trees and tested for significant overrepresentation using a hypergeometric test implemented in the GOHyperG function Bioconductor [38] package GOstats. A GO term is claimed to be significant if the p-value is less than 0. 025.

c-MYC is an important proto-oncogene that controls the expression of many other genes, and MYC regulation is deranged in many cancers. Identifying c-MYC target genes is one of the key steps to understand both the biological role and molecular mechanism of c-MYC action. Defining the complete list of c-MYC target genes and categorizing them as genes that are directly and indirectly modulated remains a challenge. Computational models also help us to understand the mechanisms modulating c-MYC function. We describe a method to predict where MYC will bind in the genome and which c-MYC binding sites will be biologically active. The method integrates multiple sources of data, including both genome sequence and functional annotations, to predict that 460 genes are direct c-MYC targets. These include many genes previously known to be c-MYC targets as well as 245 novel direct c-MYC targets. Using multiple, independent gene-expression datasets improves the sensitivity and specificity of the prediction and demonstrates significant tissue-specific variation in c-MYC action at different genes. Our study suggests that chromatin state plays an important role in modulating both c-MYC binding-site activity and the functional consequences of c-MYC binding.

Abstract Introduction Results Discussion Materials and Methods

primates oncology mammals computational biology homo (human) genetics and genomics

2007

Integration of Genome and Chromatin Structure with Gene Expression Profiles To Predict c-MYC Recognition Site Binding and Function

Scaffolding proteins that direct the assembly of multiple kinases into a spatially localized signaling complex are often essential for the maintenance of an appropriate biological response. Although scaffolds are widely believed to have dramatic effects on the dynamics of signal propagation, the mechanisms that underlie these consequences are not well understood. Here, Monte Carlo simulations of a model kinase cascade are used to investigate how the temporal characteristics of signaling cascades can be influenced by the presence of scaffold proteins. Specifically, we examine the effects of spatially localizing kinase components on a scaffold on signaling dynamics. The simulations indicate that a major effect that scaffolds exert on the dynamics of cell signaling is to control how the activation of protein kinases is distributed over time. Scaffolds can influence the timing of kinase activation by allowing for kinases to become activated over a broad range of times, thus allowing for signaling at both early and late times. Scaffold concentrations that result in optimal signal amplitude also result in the broadest distributions of times over which kinases are activated. These calculations provide insights into one mechanism that describes how the duration of a signal can potentially be regulated in a scaffold mediated protein kinase cascade. Our results illustrate another complexity in the broad array of control properties that emerge from the physical effects of spatially localizing components of kinase cascades on scaffold proteins. In the context of signal transduction, cells integrate signals derived from membrane proximal events and convert them into the appropriate cell decision. Within the complex networks that integrate these signals lies a highly conserved motif involving the sequential activation of multiple protein kinases. Signal propagation through these kinase cascades is often guided by a scaffolding protein that assembles protein kinases into a multi-protein complex. Signaling complexes maintained by scaffolds are intensely studied and have been shown to affect myriad cell decisions [1]–[7]. Despite numerous advances in the understanding of the signaling function of scaffold proteins [8]–[15], many questions remain. For instance, although scaffolds are believed to have profound effects on the dynamics of signal propagation [6], [9], [10], [16], the mechanisms that underlie how scaffolds regulate signaling dynamics are not well understood. One key factor in specifying a cellular decision is the duration of a signal (i. e. the time over which a kinase remains active) [17], [18]. Differences in signal duration have been implicated as the basis of differential decisions in myriad cell processes. For example, it has been suggested that decisions on growth factor induced cell proliferation, positive and negative selection of T cells, apoptotic programs, cell cycle progression, among many others, are regulated by the duration of signaling [19]–[24]. Therefore, the issue of how a signal output, such as the activity of extracellular regulatory kinase (ERK) in a MAPK pathway, is distributed over time, is of considerable interest. There are many ways in which the duration of the output of a kinase cascade can be controlled. Regulation of signaling dynamics can arise from processes upstream of the cascade [25]. For example, degradation of upstream signaling components such as the surface receptors [26] and differential kinetics of GTPase regulators [27], [28] can be essential in regulating MAPK signaling dynamics [25]. Also, multisite phosphorylation is predicted to influence signal duration [29]. It has been also been shown that differential modes of feedback regulation that are manifested under different conditions within the same cascade can regulate signal duration [30]. Scaffold proteins have also been implicated as key determinants in the regulation of signal duration [9], [10], [31]. Because the many factors that control scaffold mediated signaling are difficult to systematically control in a laboratory setting, a precise understanding of how scaffold proteins affect the dynamics of signal transduction has proven elusive. Computational models have been useful in understanding some of the many complex ways in which scaffolds influence signal transduction [16], [32]–[34]. However, it is currently impossible to model theoretically all aspects of any biological signaling process—computational models ultimately require that many gross simplifications be made. Our aim is, therefore, not to attempt to simulate every detail of a specific biochemical pathway but rather investigate the consequences that emerge from a simple scenario of scaffold mediated signaling whereby a model cascade assembles onto a scaffold. In modeling this scenario in itself, we hope to learn more about the functional and mechanistic consequences that these specific physical constraints, imposed by assembling components of a biochemical cascade onto a scaffold, confer to signaling pathways. In parsing these effects from the myriad others that are undoubtedly important, our hope is that our results can serve as a framework for understanding the extent to which these effects are important in specific biological contexts such as the Mitogen Activated Protein Kinase (MAPK) pathway. One theoretical analysis of scaffold mediated cell signaling revealed the presence of non mononotic behavior in signal output as a function of scaffold concentration [34]. If scaffolds are required for signaling, then too few scaffolds will be detrimental to signaling. On the other hand, if scaffolds are present in excess, signaling complexes become incompletely assembled and the signal output is attenuated. As a consequence of this “prozone” effect, scaffolds were shown to also differentially affect the kinetics of signaling. The observation that scaffolds can differentially affect signaling dynamics leads to many questions. How do scaffold proteins control the time scales involved in signal propagation? An important metric of cell signaling is the time it takes for a downstream kinase to become active [35], [36]. As signal transduction is stochastic in nature, the more precise question is: what is the distribution of times characterizing the activation of a downstream kinase? How do scaffolds affect this distribution, and what might be the biological consequences of changes in this distribution as a result of signaling on a scaffold? We compute first passage time distributions [37] using a stochastic computer simulation method to investigate these questions. Specifically, we use a kinetic Monte Carlo algorithm. We have previously used such methods to study a different question concerning the regulation of signal amplitude by scaffold proteins [33]. It is also possible that a differential equation model that considers mean-field kinetics could be used to study the first passage time distribution [37]. However, such an approach would require the imposition of absorbing boundary conditions that can make the numerical analysis difficult. Our simulation results suggest that, depending on physiological conditions, scaffold proteins can allow kinase cascades to operate in different dynamical regimes that allow for large increases and decreases in the speed and characteristic time scale of signal propagation. Furthermore, and perhaps more importantly, scaffolds are shown to influence the statistical properties of the times at which kinases are activated in complex ways. Scaffolding protein kinases cascades can allow for broadly distributed waiting times of kinase activation, whereas in the absence of a scaffold, the time it takes for a kinase to be activated is effectively characterized by a single time scale. These stochastic characteristics of scaffold-mediated kinase cascades are, to our knowledge, elucidated for the first time and may have diverse biological consequences that pertain to how signal duration is regulated. It is also our hope that our results provide a framework for achieving a deeper qualitative understanding of how scaffolding proteins can regulate the dynamics of cell signaling and the statistical properties of signal transduction. For our study, we considered a model three tiered protein kinase cascade such as the MAPK pathway [38]. Since our aim is to study the effects of spatially localizing protein kinases on signaling dynamics, we considered a minimal description of a model kinase cascade. Many factors that are undoubtedly important in regulating signaling dynamics were not considered. These factors include feedback regulation within the cascade, allosteric and or catalytic functions provided by the scaffold, and the effects of multiple phosphorylations of each kinase [11], [25], [26], [30], [39], [40]. In our model, signal propagation occurs in a three step hierarchical fashion: an initial stimulus (S) activates a MAP3K (A) that in turn, activates a MAP2K (B), that subsequently can activate its MAPK (C) substrate—phosphatases can deactivate each activated species and this deactivation occurs regardless of whether or not the active kinase is bound to a scaffold. A schematic is presented in Figure 1A that illustrates the basic processes that are allowed in our model. A steady-state ensemble is considered. That is, simulations are allowed to first reach a dynamic steady-state and once this state is reached, dynamics are studied. We do not consider dynamics from the starting time that requires propagation through a hierarchical cascade. Recent work has studied the statistical dynamics of kinase activation that result from the hierarchical organization of a kinase cascade; in that study, it was shown that the hierarchical structure of the cascade gives rise to broad waiting time distributions of cascade activation. In the regime that we study here, these effects are absent since activation of the cascade requires that an inactive C protein encounter an active B protein; our motivation is thus to investigate how the dynamics of kinase activation can be affected by assembling components of the cascade onto a scaffolding protein that localizes single complexes. Therefore, we do not emphasize how the hierarchical structure of a signaling cascade effects signal propagation and instead focus on how assembly of the cascade onto a scaffold affects signaling dynamics. We also underscore the notion that in our approach, many undoubtedly important effects such as the hierarchical structure of protein kinase cascades, the influence of feedback loops, differential enzymatic mechanisms and allosteric control by scaffolds are neglected. Again, by excising these effects, we restrict our attention to a hypothetical scenario that aims only to investigate the consequences of assembling components of a cascade onto a scaffold protein. The key quantities computed and parameters used are discussed below in Table 1 and Figure 1B. Additional details are provided in the Methods section. To set the context, consider the consequences of signaling in two limiting cases in our model. When the binding affinity of the kinases to the scaffold, E, is low (defined here to be close to the thermal energy, E∼kBT; kB is Boltzman' s constant and T is the temperature) and kinases disassociate rapidly from the scaffold, few proteins on average are bound to a scaffold. Therefore, signaling dynamics corresponds to that of a kinase cascade in solution. For a very strong affinity, E≫kBT all available binding sites to scaffold proteins are occupied by kinases (on average). In this case, signaling dynamics are controlled by the time required for initial stimuli to encounter and interact with each fully assembled complex. Therefore, we consider cases in which kinases can disassociate from their scaffolds and exchange with unbound kinases on time scales pertinent to cell signaling processes. Such time scales correspond to disassociation constants (Kd) on the order of 1–10 µM and off rates, koff∼1s−1. Such Kd values correspond to free energies of binding of roughly 7–9 kcal/mol, an energy scale typical of protein-protein interactions in kinase cascades [41]. We have used 12 kBT as the binding energy in our simulations which corresponds to ∼7. 2kcal/mol. We also discuss the robustness of our results with respect to changes in this value. Scaffold concentration has been identified as a key variable that can regulate the efficiency of signal propagation through a kinase cascade [2], [5], [34]. For the set of parameters used in the simulations (Table 1), signal output (defined as the average steady state value of the final kinase in the cascade) has a non-monotonic (biphasic) dependence on the relative concentration of scaffolds ζ (, where [Scaffold] is the concentration of the scaffold and [MAP3K]0 is the concentration of the first kinase in the cascade) and peaks at an optimal value of ζ = 1 [33], [34]. To quantify signaling dynamics, we consider a survival probability S (t) (methods) that, as mentioned, can be viewed as a type of autocorrelation function. where σ (t) equals 0 or 1 depending upon the activity of the final kinase within the cascade (methods) and the brackets indicate an average over all kinases in the simulation averaged over many simulations. This quantity gives the probability that the final kinase in the cascade remains inactive at time t given that it was inactive at time t = 0. Therefore, signaling dynamics can be monitored by observing the decay of this function with time. In Figure 2A, S (t) is computed for different values of the relative scaffold concentration, ζ. The intrinsic time of signal propagation, τ, is the value at which S (t) decays to e−1 of its original value (S (t = τ) = e−1). Upon increasing scaffold concentration, τ increases. At very high scaffold expression levels, signals propagate so slowly that cell signaling is not observed on experimentally measurable time scales which we take to be in our simulations ≫106 Monte Carlo (MC) steps; 1 MC step∼1 µs assuming a lattice spacing of 10 nm and a diffusion coefficient of 10 µm2/s [42]. The increase in τ spans several orders of magnitude as is observed in Figure 2B. Distinct stages are also observed in the behavior of τ, and are separated by an inflection point occurring shortly past the optimal value of scaffold concentration (ζ∼1). This phenomenon suggests that different physical processes are determining the signaling dynamics at different ranges of scaffold concentration. These results also suggest that the concentration of scaffold proteins can in principle set an intrinsic time scale that determines the speed of signal propagation. Such an intrinsic time scale arises solely from changes in the concentration of scaffold proteins. This time scale can span several orders of magnitude for biologically relevant affinities and diffusion coefficients and increases monotonically with increasing scaffold concentration. Note that these calculations consider only the speed of signaling and do not necessarily imply that signaling is more efficient when τ is small. To observe the total amount of integrated signal flux, the survival probability is conditioned with the probability that a kinase in the pool of signaling molecules is active in the steady state. We compute R (t) defined as S (t) multiplied by the average number of (the final downstream) kinases active at steady state, where fA is the fraction of active kinases at steady state. The time derivative, , can be thought of as a flux of activated kinases being produced. In Figure 2C, R (t) is plotted as a function of time. For low concentrations of scaffolds, the small amount of signal, albeit quickly propagating, is rapidly quenched. As scaffold concentration increases, both the amplitude and duration of the signal increase up to an optimal value. Past the optimal value, higher scaffold concentrations result in signals with small amplitude but the duration of signaling is extended. The behavior of the integrated reactive flux is a direct consequence of the existence of an optimal scaffold concentration and “bell shaped” titration curve since the area under these curves is proportional to the average signal output [33], [34]. Figure 2 emphasizes how the characteristic time for signal propagation is influenced by changes in the relative scaffold concentration. It also appears that the qualitative features of S (t) change as scaffold concentration is varied. The decay of some distributions appears highly concentrated at a particular time while the decay of other distributions appears more broadly distributed. To further investigate this observation, we plotted the survival probability as a function of the dimensionless time, t/τ. If the decay of S (t) is purely exponential, then S (t/τ) will have the form e−t/τ. Figure 3 shows S (t/τ) for different values of scaffold concentration and a decaying exponential function is given as a reference. One notices that S (t/τ) is exponential at negligible scaffold concentrations. As scaffold concentration increases, the behavior of S (t/τ) deviates from a single exponential decay. Near ζ = 1, S (t/τ) shows maximal deviation from purely exponential kinetics. As scaffold expression increases past this point, the shape of S (t/τ) reverts back to an exponential form. A deviation from exponential behavior can be quantified by considering a stretched exponential function, and fitting S (t/τ) to this form for different values of ζ. One desirable feature of the stretched exponential function is the minimal number of parameters, τ and β, that are involved in the least-squares fit; also, the values of these parameters can be physically interpreted. τ gives the characteristic time for one overall timescale of signal propagation, and β is a measure of how much the function, S (t), deviates from a single exponential and thus how broadly distributed are the signaling dynamics. Figure 3B shows how β depends on scaffold concentration. For these simulations, β∼1 for small and large values of scaffold concentrations indicating exponential behavior. For intermediate values, β peaks at a minimum of β∼0. 6, a significant deviation from purely exponential behavior. In the limits of small and large scaffold concentrations, the presence of a single exponential decay, β∼1 indicates that signal propagation, or the relaxation of S (t/τ), occurs at one characteristic time scale. In the intermediate regime, β shows significant deviations from one, thus allowing for a broadly distributed signal. When β is significantly less than one, signals can steadily propagate over several decades. In this regime, the waiting time distribution f (t), has a large tail and the activation of kinases is slowly maintained over many time scales. Why do we observe exponential and non-exponential behavior under different conditions? Signal transduction in our model occurs on a time scale that is much slower than the microscopic time scales associated with diffusion, binding/unbinding, and enzyme catalysis. We might therefore expect that some coarse-graining exists whereby events at these fast, “microscopic” time scales interact with other relevant biophysical parameters (e. g. scaffold concentration) to give rise to emergent properties that evolve on slower times scales. These processes are a manifestation of the collective dynamics of the many processes that occur on faster time scales. Understanding the factors that govern these emergent time scales would then provide insight into the origin of the different temporal characteristics that are revealed by our simulations. In order for a signal to propagate (i. e. for the last kinase in the cascade to become active), a hierarchical sequence of phosphorylation reactions among kinases must occur that leads to the final kinase in the cascade being activated by its upstream kinases. The activation process may occur either in solution or on a scaffold. Also, in the course of signaling, kinases can exchange from a scaffold. Some kinases are bound to a scaffold that contains an incomplete assembly of the necessary signaling molecules, and are not signaling competent. Ultimately, an inactive kinase can exist in one of three states: in solution, bound to a complete complex, or bound to an incomplete complex. Figure 4A contains a diagram of such a minimal picture and arrows denote transitions between the four states. This minimalist description clarifies the behavior in Figures 3A and 3B. For low scaffold concentrations (ζ≪1), kinases predominately exist in solution and signal transduction is dominated by the time it takes for an upstream kinase to encounter its downstream enzyme. Since a steady-state ensemble is used, the rate limiting step for signal propagation is the diffusion limited collision between an active B* molecule with an inactive C molecule. For high scaffold concentrations (ζ≫1), kinases predominately exist in incomplete signaling complexes and signal transduction is limited by a time scale that characterizes the turnover of a signaling incompetent complex to one that is able to signal. For intermediate concentrations, inactive kinases can exist in each of three states and transitions between these states also occur. Thus, the source of the nonexponential relaxation (i. e. β<1) arises from the mixing of many time scales that are relevant for intermediate scaffold concentrations. Figure 4B illustrates this minimal picture of the kinetics of signal propagation derived from these physical considerations. Also note that the sensitivity of our results to changes in model parameters can be understood from this simple picture of scaffold mediated signaling dynamics. For instance, changes in kinase and scaffold concentrations result in changes in the relative amount of kinases existing in the three states in ways that have been previously characterized [33], [34]. Changes to other parameters such as the rates of activation and deactivation and the concentration of phosphatases alter the rates of transitions between these different states. For instance, if phosphatase concentrations are very large, then activation in solution is very slow and occurs predominantly on a scaffold. Also, slower rates of activation (and larger rates of deactivation) result in a larger portion of signaling originating from kinases that are bound to scaffolds. In general, when the activation of kinases originates more (less) predominantly from a particular state in the minimal model, β increases (decreases). When multiple pathways to kinase activation contribute with comparable time scales, β is small, and signaling is broadly distributed over many time scales. We have performed many simulations with varying parameters to test the robustness and parameter sensitivity of our findings and find that that the qualitative behavior of our results follow this simple, qualitative, physical picture. Additional insight can be gleaned from consideration of the power spectrum of S (t). The power spectrum, where, , computed in the frequency domain, resolves the time scale dependence of kinase activation. This approach has proven useful in studying the dynamics of complex biochemical networks in many contexts [43]–[45]. We first note that S (t) obtained from the simulations fits well to the functional form (χ2 values small). Thus, we use the parameters β and τ that were extracted from the fits at low (ζ = 0. 001), optimal (ζ = 1. 0), and high (ζ = 3. 5) scaffold concentrations to compute P (ω) for these three cases. In Figure 5, (τopt) −2 P (ωτopt) is plotted versus ωτopt where the time τopt is the characteristic time scale τ for relaxation at the optimal ζ = 1scaffold concentration. That is, time is rescaled to units of τopt. For each curve, at lowωτopt≪1 frequencies P (ωτopt) is constant (P (ωτ→0) →τ2) signifying that kinase activation has become uncorrelated. At high ωτopt≫1 frequencies, kinase activation is correlated and a power law decay is observed for each curve P (ωτopt) ∼ω−2. As a reference, note that for an exponential decay, S (t) = e−t/τ, the transition between these two regimes occurs at ωτ∼1 and is determined by the Lorentzian: In Figure 5, for high (ζ = 3. 5, blue) and low (ζ = 0. 001, green) scaffold concentrations power spectra closely resemble the Lorentzian with the transition to P (ωτopt) ∼ω−2 behavior occurring at different frequencies. At low ζ = 0. 001 concentrations, the inverse time scale or corner frequency at which kinase activation decorrelates is determined by the diffusion limited rates of activation and deactivation of the final kinase C*. The corner frequency can be estimated fromwherek+ and k− are diffusion limited rates of activation and deactivation and are given by a diffusion limited encounter rate that is on the order of D Ntota where D is the diffusion constant used in the simulation, Ntot is the number of proteins, and a is the size of a protein taken to be the size of a lattice site. Substitution of the numbers used in the simulation (Table 1 and Methods) achieves a value for the relaxation time that is commensurate with the relaxation time for ζ = 0. 001 in Figure 2; i. e. , τc∼105 mcsteps. At high ζ = 3. 5 concentrations, the corner frequency is determined by rates of formation and disassociation of an intact signaling complex. Furthermore, because of these many process that comprise the relaxation rate in this case, a numerical estimate of the corner frequency is difficult. In the case of the optimal (ζ = 1. 0, red) concentration, the transition from constant to P (ωτopt) ∼ω−2 behavior occurs smoothly over many decades from ωτopt∼0. 1 to ωτopt∼10. 0. The plot in Figure 5 also resolves different frequency dependent processes occurring in signal transduction. At high frequencies or short times, ωτopt<10. 0, kinase activation is limited by the diffusive motion of the kinases in the cascade. At intermediate frequencies, 0. 1<ωτopt<10. 0, activation is dominated by transitions between kinases assembled in competent, incompetent, and solution based kinases. For low frequencies ωτopt<0. 1 or long times, kinase activation decorrelates for each scaffold concentration. To illustrate how computed values of S (t) and the distribution of waiting times for kinase activation relate to conventional means of defining signal duration, we consider a differential equation for the time evolution of the activated form of the final kinase within the cascade. In this picture, species become activated at rates derived from the functional form that was fitted to the survival probabilities that were computed from the simulations. The waiting time or first-passage time distribution f (t) is used as a forward rate and the activated final kinase then can be degraded with a kinetics of degradation characterized by a rate constant, kφ. A kinetic equation describing this process is written as: x is the number of active species, τ is the time constant of signal propagation, and β is the stretching parameter that quantifies deviations away from single exponential behavior. In this picture, x (t) represents the average response to a stimulus f (t) that is distributed temporally according to and subject to a first order decay with characteristic time 1/kφ. The equation for x (t) can be solved and using the initial condition, x (0) = 0: x (t) was integrated numerically and is shown for different values of β in Figure 6A. As seen in Figure 5A, decreasing values of β result in the trajectories having longer tails and thus an extended duration of signaling. Also, smaller values of β result in the signal having a larger peak. This property directly follows from the decay of S (t) that was shown in Figure 3A for different values of β. At early times, S (t) decays more quickly when β is smaller; as a consequence, more kinases are activated at these times, thus resulting in a larger peak. This concept of signal duration can be made more precise by considering a threshold amount of signal, T, that is required for the pathway to be considered active. With a chosen value of T, the signal duration, υ, is defined as the time it takes for the signal to decay to some threshold value, T. That is, the equationis satisfied. Figure 6B shows the signal duration, υ, as a function β for values of β ranging from 0. 5 to 1 for different values of T. For smaller values of T, β<1 (i. e. , scaffolds are present) results in a large increase in signal duration compared to the case in which β = 1. Therefore for a fixed value of τ, the most broadly distributed signal leads to the longest signal duration. We first showed that scaffold concentration is a key variable in regulating the speed of signal transduction. Moreover, we showed that the concentration of a scaffold protein can influence signaling dynamics by controlling the distribution of times over which kinases become active. This type of regulation may have many important consequences that are related to the influence of signal duration on cell decisions. Controlling the times over which kinases are activated may also be useful in directing a specific, robust response in a number of ways. Thus, the scaffold concentration itself provides another variable for maintaining signal specificity by controlling signal duration. This is consistent with data from genetic studies involving KSR1 [9], [10], where the authors reported that the concentration of KSR1 can control a cell decision involving commitment to adipogenesis. Our study focused solely on aspects of scaffold mediated regulation of signal transduction and we only considered the times at which kinases are active in the course of signal transduction. Many other factors also control signal duration. For example, our study does not consider the negative feedback loops that are often associated with the upregulation of phosphatases [18], [32]) or the role of receptor downregulation in controlling signal duration. Also we did not explicitly consider the role of positive versus negative feedback loops in shaping signal duration which is undoubtedly important [30]. It was our focus to study how spatially localizing kinases on a scaffold protein influences signal duration. We aimed to untangle this effect of scaffold proteins from other essential features of kinase cascades such as allostery and feedback regulation. Also, other theoretical studies have investigated the first passage time statistics in signal transduction cascades and have found interesting dynamics that result from, in part, the sequential activation of multiple steps in a kinase cascade [35], [36]. Our studies of signaling through scaffold proteins supplement these findings and, to our knowledge, provide the first study that shows how scaffolds affect the statistics of signal transduction. Several predictions from our model of how scaffolds regulate signaling dynamics can be tested. Measurements that monitor the time course of signaling for different scaffold concentrations could potentially resolve the differences in signaling dynamics that are predicted. Also, single molecule or fluorescence correlation based spectroscopic methods [46]–[48] could potentially probe the statistics of signaling dynamics inherent in kinase cascades and study how such statistics are related to reliable cell decisions. Such techniques can monitor the propagation of a signal, at the level of an individual molecule and thus directly measure how kinase activation within a single cell is distributed over time. We simulate a model protein kinase cascade such as the mitogen-activated protein kinase (MAPK) cascade (Figure 1A) in the presence and absence of a scaffold with a kinetic Monte Carlo algorithm [49], [50], which allows us to monitor the relevant stochastic dynamics. Since we are investigating phenomena that occurs on the time scales of signal transduction, we course-grain the system so that proteins are represented as discrete objects, occupying a site on a lattice of dimensions 100×100×100 lattice spacings. Scaffold proteins are modeled as rigid, immobile objects containing three binding sites that are each specific for a particular kinase. When bound to a scaffold, kinases are tethered in nearest neighbor positions that are proximal to their downstream substrates. Allowing the scaffold and scaffold-bound species to move does not affect the qualitative results. Reflecting, no flux (i. e. Neumann) boundary conditions exist at each of the faces of the cubic lattice. The system is not periodically replicated since our simulation box is a size on the order a cell. Proteins can diffuse (i. e. translate on the lattice in random directions), bind and unbind, and undergo state transformations according to the prescribed reaction network involving a three staged cascade of activation and deactivation events (Figures 1A and 1B). Protein motion is subject to excluded volume (steric) constraints in that no two proteins can occupy the same site on the lattice. Chemical (state) transformations and binding events are modeled as thermally activated processes with energy barriers for activation, inactivation, binding and unbinding reactions. Parameters used are given in Table 1. We simulate the dynamics with a fixed time step Monte Carlo algorithm. In a Monte Carlo step, n trials are attempted, where n is the number of proteins in the simulation. For a given trial, a protein is first chosen at random with uniform probability. A displacement move in a uniformly random direction is attempted with probability, where d is the dimensionality of the simulation box, Deff is the probability of attempting a diffusion move and sets an overall time scale to diffuse the length of a lattice site. Excluded volume is accounted for by imposition of an infinite energy barrier, E∞, for hopping to sites containing other proteins; i. e. Upon considering all possible nearest neighbor interactions, reaction moves, as determined by the network topology, are tried with probabililty, where keff is the probability of attempting a reaction move; (keff sets an overall reaction time scale), Ej, j′ is the energy barrier for the j′→j reaction scaled with respect to kbT (Boltzman' s thermal energy). With this Monte Carlo move set, the simulations formally evolve the dynamics of the probability that a chemical species si, j of type i and state j at position at time t according to the Master equationwhere is the transition probability per unit time for a displacement from to of species; ; is the per unit time transition probability at for a species si, j′ to change to state si, j (e. g. binding and unbinding reactions) and is, and imposes the constraint that binding and unbinding occurs only at specified binding sites on the scaffolds at positions (is zero unless a scaffold is located at position; and is the transition probability per unit time for a species at to facilitate (i. e. catalyze) the j″→j transformation at site; and is zero unless the site at is occupied by the appropriate catalyst (i. e. i′ = i″) in which case it is 1; each summation indicates a sum over nearest neighbors. The parameters used in the simulation were first constrained to typical literature values. Energies of disassociation were taken to be 12kbT corresponding of a disassociation constant Kd of roughly 1 µM. 200 stimulatory molecules, S, 200 molecules of kinase A and a 1∶1∶5 ratio of A, B, and C kinases was used. If we assume a lattice spacing of 10nm, a typical diameter of a protein, the concentration of kinases in our simulation box is roughly 1 µM for kinase A and kinase B and ∼5 µM for kinase C. In a physiological context, assuming the radius of the cell is about 10 µm, this approximately corresponds to ∼105 molecules of kinases A and B and a copy number of ∼5×105 for kinase C in our simulation. 600 generic phosphatases are also present. These relative numbers are commensurate with reported kinase concentrations in Yeast and other systems [51], [52]. Chemical kinetics were modeled in the simplest possible way by considering a single elementary reactive collision; i. e. , where the asterisk (*) denotes an active species. For the purposes of our simulations, saturation effects were ignored and the kinetics were taken to be in a linear regime. Such a model is reasonable when reactions are not limited by the availability of the enzyme. However, relaxing this assumption does not affect the qualitative behavior of our results provided that the times scales involved in the formation of an enzyme-substrate complex and subsequent catalysis do not compete with the diffusive processes in solution. If additional processes associated with enzyme catalysis dominate over diffusive motion of the proteins or binding and unbinding to and from the scaffold, then these process would be observed in the autocorrelation function and corresponding power spectrum. Given that catalysis would incorporate additional processes into the mechanism of kinase activation, such effects would serve to broaden the distribution at all scaffold concentrations as we have observed in our simulations (data not shown). We did not explore this scenario in its entirety since our aim was to solely investigate the effects of scaffolding a kinase cascade. As discussed in a previous study [33], an important variable that determines the role of scaffolding a kinase cascade is the amount of time required (τec) for an active kinase to encounter its downstream target. For simple diffusion, in three dimensions, where D is the diffusion constant and C is a typical concentration of kinases. Experiments indicate that τec is on the order of 10−4s–100 s [42]. Our studies focused on these experimentally relevant conditions. Steady-state values are reported. The system is first placed in a random configuration and simulations are allowed to “equilibrate” by letting the dynamics evolve to a time much larger than the time it takes for a kinase to diffuse the length of the simulation box. Kinases that are inactive at time t′ are tagged and waiting times are observed at time t+t′ (i. e. statistics are collected for the times at which the kinases become activated), and t′ is chosen to be a time longer than the time required for equilibration of the Monte Carlo trajectory. Signaling dynamics can be defined microscopically as the distribution of times at which an individual kinase among of pool of available kinases becomes activated. Therefore, we quantify signaling dynamics by first considering the survival probability S (t). S (t) gives the probability that a particular kinase among the pool of signaling molecules has not been activated at time t provided that it was inactive at time t = 0. S (t) is a two time point autocorrelation function: where the brackets denote an ensemble average and σ (t) is a binary variable indicating the state of a kinase; i. e. , The survival probability is related to other dynamical properties; for instance, it can be related to a waiting time probability density function or first passage time distribution, f (t), in the following way: S (t) is the complement of the cumulative probability distribution of the first passage time. S (t) is computed from the simulations by integrating f (t). Such a calculation is analogous to the data obtained from a single molecule experiment that measures the statistics of enzyme dynamics [46]. This distribution of waiting times underlies the intrinsic duration of signal propagation in a protein kinase cascade—the decay of such a quantity is a measure of how fast the signaling cascade responds to stimuli. Important to note is that this quantity gives information only on the timing of the signal and not on its final magnitude. We also consider the product of the survival probability with the probability that a kinase in the pool of signaling molecules is active in the steady state, where fA is the fraction of active kinases at steady state. When normalized, R (t) is a measure of how the activity of the total pool of kinases is distributed over time, and can be thought of as an integrated flux of activated kinases. is seen as a reactive flux in provides a measure of the rate at which downstream kinases are being activated. One can imagine that both quantities could be biologically relevant. If conditions dictate that a biological response requires that a certain number of kinases remain active for extended amounts of time, R (t) may be the more relevant quantity. On the other hand if the cellular decision requires a count of kinases that become active over a specified time window, then S (t) could be the relevant quantity since it provides a measure of how the activation of individual kinases is distributed over time. Both quantities may be used to integrate signals in different contexts but since our study focuses on signaling dynamics we primarily focus on the survival probability and its related quantities. Power spectra were computed numerically. Real and imaginary parts of the Fourier transform were obtained from numerical integration using the trapezoidal rule with a step size Δt = 0. 001. P (ω) is calculated by squaring the real and imaginary parts of X (ω) P (ω) was sampled at N = 100 logarithmically spaced (i. e. , ωmax = ω0 (10δ (n−1) ); n∈[1,100] so that) angular frequencies beginning at: , where T is the total length of the autocorrelation function.

Signal transduction is the science of cellular communication. Cells detect signals from their environment and use them to make decisions such as whether or when to proliferate. Tight regulation of signal transduction is required for all healthy cells, and aberrant signaling leads to countless diseases such as cancer and diabetes. For example, in higher organisms such as mammals, signal transduction that leads to cell proliferation is often guided by a scaffold protein. Scaffolding proteins direct the assembly of multiple proteins involved in cell signaling by providing a platform for these proteins to carry out efficient signal transmission. Although scaffolds are widely believed to have dramatic effects on how signal transduction is carried out, the mechanisms that underlie these consequences are not well understood. Therefore, we used a computational approach that simulates the behavior of a model signal transduction module comprising a set of proteins in the presence of a scaffold. The simulations reveal mechanisms for how scaffolds can dynamically regulate the timing of cell signaling. Scaffolds allow for controlled levels of signal that are delivered inside the cell at appropriate times. Our findings support the possibility that these signaling dynamics regulated by scaffolds affect cell decision-making in many medically important intracellular processes.

Abstract Introduction Results Discussion Methods

physics/interdisciplinary physics cell biology/cell signaling cell biology/cell growth and division biophysics/theory and simulation computational biology/molecular dynamics computer science/numerical analysis and theoretical computing computational biology chemical biology computational biology/signaling networks biophysics biophysics/cell signaling and trafficking structures computational biology/systems biology chemical biology/chemical biology of the cell

2008

Regulation of Signal Duration and the Statistical Dynamics of Kinase Activation by Scaffold Proteins

Ligand virtual screening is a widely used tool to assist in new pharmaceutical discovery. In practice, virtual screening approaches have a number of limitations, and the development of new methodologies is required. Previously, we showed that remotely related proteins identified by threading often share a common binding site occupied by chemically similar ligands. Here, we demonstrate that across an evolutionarily related, but distant family of proteins, the ligands that bind to the common binding site contain a set of strongly conserved anchor functional groups as well as a variable region that accounts for their binding specificity. Furthermore, the sequence and structure conservation of residues contacting the anchor functional groups is significantly higher than those contacting ligand variable regions. Exploiting these insights, we developed FINDSITELHM that employs structural information extracted from weakly related proteins to perform rapid ligand docking by homology modeling. In large scale benchmarking, using the predicted anchor-binding mode and the crystal structure of the receptor, FINDSITELHM outperforms classical docking approaches with an average ligand RMSD from native of ∼2. 5 Å. For weakly homologous receptor protein models, using FINDSITELHM, the fraction of recovered binding residues and specific contacts is 0. 66 (0. 55) and 0. 49 (0. 38) for highly confident (all) targets, respectively. Finally, in virtual screening for HIV-1 protease inhibitors, using similarity to the ligand anchor region yields significantly improved enrichment factors. Thus, the rather accurate, computationally inexpensive FINDSITELHM algorithm should be a useful approach to assist in the discovery of novel biopharmaceuticals. Ligand virtual screen is widely used in rational drug discovery [1], [2]. The first stage of structure-based ligand screening is the prediction of the binding mode adopted by the small molecule complexed to its target receptor protein; a variety of algorithms have been developed to achieve this goal [3], [4]. The next step is to estimate the relative binding affinity of the docked ligands [5], [6]. Of course, it is not sufficient that a given ligand binds favorably to given protein; rather, to minimize side effects, it must also bind selectively. Classical molecular docking has been used to address both goals. However, only is it computationally expensive, but there are significant issues associated with ligand ranking [5], [7]. Thus, fast and accurate methods for both binding pose prediction and ligand ranking need to be developed. With the rapid increase in the number of experimentally solved protein structures, protein homology modeling has become a powerful tool in modern structural biology [8], [9]. Comparative modeling methods identify homologous protein structures and use them as structural templates to model the target protein of unknown tertiary structure. Using a high sequence identity template with a clear evolutionary relationship to the target, the modeled target structure can have a root-mean-square-deviation, RMSD, from the native structure <2 Å [10]. In the “twilight zone” of sequence identity [11], structural information extracted from weakly homologous structure templates identified by threading is sufficient to provide approximately correct 3D models for a significant fraction of protein targets [12], [13]. In contrast to protein structure prediction, information from related 3D structures is rarely used in the large-scale modeling of protein-ligand complexes. One example of an approach that employs such information is CORES, an automated method for building three-dimensional protein-ligand complexes [14]. CORES directly utilizes the conformation and binding pose of key structural elements of the target ligand, termed “molecular frameworks”, found in templates that are closely related to the protein target. Its practical utility was demonstrated on a set of protein kinases in which ligands containing related frameworks were found to bind in the same orientation. A similar approach designed specifically for kinases, kinDOCK, performs ligand comparative docking by using a kinase family profile to align the related kinase-ligand complexes onto the target kinase' s structure and then directly transfers the ligand coordinates [15]. KinDOCK typically docks target ligands into the kinase binding pocket within a 2 Å RMSD from the crystal structure. Moreover, an original clustering procedure based on the binding pose similarity was proposed to highlight the structural similarities and differences within a set of multiple X-ray structures complexed with different ligands [16]. Other examples of ligand docking studies that utilize structural information extracted from closely related protein-ligand complexes include the analysis of cathepsin inhibitor specificity [17], the examination of carbohydrate recognition by the viral VP1 protein [18], screening for selective bacterial sirutin inhibitors [19] and the design of small molecule inhibitors of the macrophage migration inhibitory factor [20]. Typically, modeling templates for the target ligands are extracted from 3D structures of small molecules complexed to closely related proteins. In our previous study, we observed that evolutionarily remotely related proteins identified by threading often share a common ligand-binding site [21]. Both the localization of the binding site and chemical properties of bound ligands are strongly conserved. This forms the basis of the FINDSITE binding site prediction/protein functional inference/ligand screening algorithm [21]. Furthermore, we found that a pocket-specific potential of mean force derived from known protein-ligand complexes identified for a given target sequence by threading is often more specific than generic knowledge-based potentials derived from ligand-protein complexes found in the PDB [22]. This enhanced specificity suggests that the binding mode and protein-ligand interactions in distantly related protein families are conserved during evolution. To confirm this hypothesis, here, we present the results of ligand binding mode analysis of evolutionarily distant proteins identified by state-of-the-art threading methods [23]. The ligands that bind to the common binding site contain a set of strongly conserved anchor functional groups as well as a variable region that imparts specificity to a particular family member. Furthermore, the degree of sequence and structure conservation of residues in contact with the ligand anchor functional groups are higher than those contacting ligand variable regions. Exploiting these observations, we develop FINDSITELHM (LHM stands for Ligand Homology Modeling) that employs structural information extracted from weakly related proteins to perform rapid ligand docking and ranking by homology modeling; we compare its accuracy to classical ligand docking/ranking approaches [4], [22], [24]. The protocol followed in this study is a direct extension of FINDSITE [21], a threading-based method for ligand-binding site prediction and functional annotation that detects the conservation of functional sites and their properties in evolutionarily related proteins. For a given target sequence, FINDSITE identifies ligand-bound template structures from a set of distantly homologous proteins (here, we limit ourselves to target proteins having <35% sequence identity to their closest template, but this arbitrary restriction would be removed in real world predictions) recognized by the PROSPECTOR_3 threading approach [23] and superimposes them onto the target' s (experimental or predicted) structure using the TM-align structure alignment algorithm [25]. Binding pockets are identified by the spatial clustering of the center of mass of template-bound ligands that are subsequently ranked by the number of binding ligands. For each target protein, the template-bound ligands that occupy a top-ranked, predicted binding site are clustered using the SIMCOMP chemical similarity score [26]. The “anchor” substructure is then identified in each cluster as described in Methods. First, we examine the anchor substructure size relative to the average molecule size. Applying the approach to a representative benchmark set of 711 ligand-protein complexes (where the target proteins have pairwise sequence identity to their templates <35%, see Methods), as shown in Figure 1, in most cases, at least 50% of a ligand is comprised of an anchor region whose functional groups are conserved in >90% of the template ligands. Those clusters in which the anchor region is smaller than 50% of the ligand are mostly short oligosaccharides, with a sugar monomer identified as a common substructure. This also explains the high standard deviation in the average ligand molecule size. For some difficult cases, our graph isomorphism analysis didn' t provide a sufficient number of atomic equivalences to recognize a common substructure. In contrast, those targets near the diagonal have an anchor equivalent to the average molecule size and represent strongly conserved ligands with little chemical variability; e. g. hemes. In addition, there are targets with a very small number of templates, all having very similar ligands. Nonetheless, for the majority of targets, a well-defined anchor substructure with a co-occurring variable region is detected. Having identified the anchor substructure, we next investigate the structural conservation of its binding mode. Figure 1 (inset plot) shows the histogram of the average pairwise RMSD among the anchor groups upon global superposition of the template proteins. Note that the properties of the native ligand are not used in any way to identify the anchor region' s properties. Clearly, in most cases, the average pairwise RMSD is <2. 5 Å. We next examine the properties of the protein' s ligand-binding region. Given the chemical conservation of the anchor substructure as well as the strong structural conservation of it' s binding mode, for binding residues, one would expect that residues contacting ligand anchor groups are more conserved than average. The degree of sequence and structure conservation was calculated for consensus binding residues (CBRs), defined as residues contacting a ligand in at least 25% of the threading templates. This criterion was previously found to maximize the overlap between predicted and observed binding residues [21] and provides sufficient statistics to calculate the sequence and structural features of binding residues. We used a probability threshold to define anchor/non-anchor CBRs based on the protein-ligand contacts extracted from the threading templates. The probability of a residue to be an anchor residue simply corresponds to the fraction of contacts formed by all residues in the equivalent position in the template structures with anchor functional groups of bound ligands. Differences in the degree of sequence and structure conservation between anchor and non-anchor CBRs were calculated on increasing the probability threshold from 0. 1 to 0. 9 using Student' s t-test for independent samples. Shannon' s information entropy is used to measure the sequence variability at a particular position in a target protein (see Methods). Analysis of the sequence entropy revealed a significantly higher sequence conservation of residues in contact with the anchor functional groups than those in contact with ligand variable regions (Figure 2A). Next, we analyzed the structural features of CBRs in terms of the experimental B-factors that reflect local mobility [27] and find that the B-factors of residues in contact with the anchor region of the ligands are significantly lower (Figure 2B and 2C which shows the B-factors of the Cαs and side chain heavy atoms, respectively). The conservation of the anchor-binding pose is consistent with the relatively lower B-factors observed for the residues in spatial proximity to the anchor functional groups. These results differ significantly from random (Figure 2D). Given that Figure 1 (inset) strongly suggests that the localization of the anchor substructure and its internal conformation is conserved, we developed FINDSITELHM, a very simple, rapid approach for ligand binding pose prediction. Using the consensus-binding mode of the anchor substructure, we align the ligand of interest to the anchor region and then, optionally, minimize the ligand conformation to remove steric clashes. This procedure can be thought of as “ligand docking by homology modeling”. Here, only weakly related template proteins (<35% sequence identity to the target) selected by threading were used to derive the consensus anchor-binding mode. In Table 1, using the crystal structures as the target receptors for ligand docking for the 711 ligand-protein set, the results are compared to three established ligand docking approaches [4], [22], [24] in terms of the heavy atom RMSD from the crystal structure. Target proteins are divided into three subsets with respect to the coverage of the predicted anchor substructure. For the first subset (full coverage) that consists of proteins for which a portion of their target ligands cover at least 90% of the functional groups in the predicted anchor substructure, simple ligand superposition is quite successful and outperforms regular ligand docking approaches. For these cases, using all-atom minimization with Amber [28], the predicted binding mode can be refined to an average RMSD from the crystal structure of ∼2. 5 Å. An example of successful refinement is presented for the human fibroblast collagenase in Figure 3, where the final ligand heavy atom RMSD is 0. 63 Å. In contrast, the RMSD from AutoDock is 2. 77 Å. The second subset (partial coverage) comprises target ligands that do not fully cover any of the predicted anchor substructure. Here, the average RMSD of the binding mode predicted by FINDSITELHM is higher than AutoDock and is comparable to Q-Dock and LIGIN. However, it is still better than random ligand placement. Finally, if none of the predicted anchor substructures are even partially covered by a target ligand (low coverage), the results of docking using FINDSITELHM are indistinguishable from random. Here, traditional ligand docking approaches, particularly AutoDock, give much better results. In addition to anchor structure coverage, the performance of FINDSITELHM depends on the overall accuracy of binding pocket prediction and the conservation of the anchor-binding mode; this is discussed in further detail below and presented in Figure 4, see below. Here, we note that using the fraction of the anchor region that is aligned for a given ligand (Figure 4A), or the average pairwise RMSD of the anchor ligand functional groups (Figure 4B), we can predict the expected accuracy of binding pose prediction without knowing the experimental result. Not surprisingly (Figure 4C), when the accuracy of the binding pocket prediction as provided by FINDSITE improves, the accuracy of the ligand pose prediction by FINDSITELHM also improves. Weakly homologous protein models frequently have significant structural inaccuracies in side-chain and backbone coordinates and thus, are much more challenging targets for ligand binding pose prediction. The performance of FINDSITELHM, AutoDock, Q-Dock and LIGIN in ligand docking when protein models are used as the target receptors was assessed for the Dolores dataset of 205 proteins [22], [29]; the average Cα RMSD to native of these protein models is 3. 7 Å. Table 2 presents ligand docking results using crystal structures as well as weakly homologous protein models in terms of the fraction of recovered binding residues and specific native contacts. Considering the complete dataset and receptor crystal structures, the accuracy of FINDSITELHM is slightly lower than AutoDock and Q-Dock. This is because the predicted anchor substructure was fully covered (≥90%) by the target ligand only for 62. 4% of the receptors; partial (≥50% and <90%) and low (<50%) coverage of the anchor substructure was found for 25. 4% and 12. 2% of the targets, respectively. This partly reflects the fact that the placement of the ligand variable region has a random component that diminishes the overall accuracy. Consistent with the decrease in ligand RMSD on minimization, the fraction of binding residues and native contacts increases. In contrast, for protein models, FINDSITELHM recovered more binding residues and specific native contacts than both all-atom docking approaches, AutoDock and LIGIN. Considering only the most confident cases for which FINDSITE was likely to predict the binding pocket center with ≤4 Å accuracy (“Easy” targets) and the predicted anchor substructure fully (partially) covered by the target ligand, the fraction of binding residues and specific native contacts recovered by FINDSITELHM is 0. 66 (0. 61) and 0. 49 (0. 43), respectively. However, now the all-atom minimization procedure applied to the binding poses predicted by FINDSITELHM caused a loss of the specific native contacts. This reflects the fact that structure adjustments are required to remove the repulsive ligand-residue interactions that are not accommodated by simple minimization. Nevertheless, these results represent a significant improvement over traditional all-atom docking against modeled receptor structures. We also note the high sensitivity of all-atom docking approaches to the quality of the receptor structures; for weakly homologous protein models, the performance of AutoDock and LIGIN is no better than random ligand placement into the predicted binding sites. The performance of Q-Dock for protein models was notably higher, since it was explicitly designed to deal with structural inaccuracies in predicted receptor models. Finally, in contrast to classical ligand docking approaches, FINDSITELHM is computationally less expensive, and typically requires less than a minute of CPU time (see Table S1). An interesting question that is very important from the practical point of view, is when should we expect a successful binding mode prediction by using ligand docking by homology modeling? In addition to the coverage of an anchor structure that clearly impacts docking accuracy (Figure 4A), we also investigated the relationship between pocket prediction accuracy, expressed as the distance between the predicted pocket center and the geometric center of the native ligand, the conservation of anchor binding mode in terms of the average pairwise RMSD of the anchor functional groups, and the accuracy of FINDSITELHM binding mode prediction assessed by the heavy atom RMSD from the crystal ligand pose. As expected, the average accuracy of the binding mode prediction by FINDSITELHM decreases with decrease in the degree of the conservation of the anchor substructure (Figure 4B). The RMSD of the predicted ligand-binding pose is <2 Å on average for highly conserved anchor substructures whose pairwise RMSD is <2 Å. For moderately conserved anchor substructures with a pairwise RMSD of 2–4 Å, the RMSD of the predicted ligand-binding mode is <3 Å in most cases. Finally, accompanied by weak (>4 Å) structural conservation of an anchor, docking accuracy drops to >3 Å on average. In addition, the drop off in ligand binding pose prediction correlates with the overall accuracy of binding pocket prediction by FINDSITE (Figure 4C). The most accurate ligand binding poses were obtained for precisely detected pockets, where the pocket center was predicted within 2 Å from the geometric center of the native ligand. Considering the structural conservation of the derived anchor substructure, its coverage by a target ligand and the FINDSITE confidence index for pocket detection [21], (all properties which can be calculated without knowledge of the native binding pose), one can roughly estimate the quality of the performance of FINDSITELHM in ligand binding pose prediction. The results of the application of FINDSITELHM to glutathione S-transferase (PDB-ID: 1a0f), MTA phosphorylase (PDB-ID: 1sd2) and lysine aminotransferase (PDB-ID: 2cjd) are presented. Figures 5–10 present the common ligand anchor substructures/variable groups identified from weakly homologous threading templates for these 3 proteins. In Figure 11, the degree of sequence and structure conservation of amino acid residues for these proteins is projected onto the target protein surface and compared to a random distribution. In Figures 12–14, using the target crystal structure, the results of flexible ligand docking by FINDSITELHM (including refinement) are compared to ligand binding poses predicted by classical docking approaches and the consistently better performance of FINDSITELHM is demonstrated. In the case of Figure 14 where the RMSDs of LIGIN and random pose prediction are the same as FINDSITELHM, the pyridoxal-5′-phosphate moiety is clearly better placed by FINDSITELHM. All have the same RMSD mainly due to the incorrect placement of the variable region. Recently, a detailed picture of the evolution and diversification of enzyme function was drawn from the analysis of conservation of substrate substructures in 42 major enzyme superfamilies [30]. Based on graph isomorphism analysis, highly conserved substructures were identified in all substrates of a particular enzyme superfamily. For the remaining substrate substructures, called reacting substructures, substantial variation in chemical properties within the superfamily was found. Systematic analysis of the substrates in 42 major SCOP [31] enzyme superfamilies revealed chemically conserved patterns that typify individual superfamilies [30]. This approach is very similar in spirit to FINDSITELHM; both demonstrate how evolutionary pressure directs the evolution of protein molecular function. The structural and chemical patterns of enzyme substrates, or small ligands in general, have been conserved during evolution due to the strong conservation of the structural and chemical features of the binding site residues. We next analyzed the overlap between the conserved substrate substructures (CSSs) identified at the SCOP superfamily level [30] and the anchor regions in ligands bound to evolutionarily related proteins selected by threading. The results presented in Figure 15 show that the highly conserved substructures of the enzyme substrates identified by Babbitt and colleagues [30] to a large extent overlap with the anchor substructures detected by our threading-based approach; in over 70% of the cases, the anchor substructure covers at least 70% of CSS' s atoms. Detailed results obtained for 4-α-glucanotransferase from T. litoralis (PDB-ID: 1k1w) and D-xylose isomerase from Arthrobacter sp. (PDB-ID: 1die) are presented in Tables S2 and S3, respectively. We find that the highly conserved substructures of the enzyme substrates frequently overlap with the conserved anchor substructures detected by our threading-based approach. The set of ligands that bind to the common binding site in distantly evolutionarily related proteins contain a set of strongly conserved “anchor” functional groups and “variable” regions that account for a specificity toward a particular family member. As a consequence of the ligand clustering procedure that precedes anchor identification, the anchor substructures typically contain more atoms than CSSs and are not confined to enzymes. Both features are important for practical application in ligand docking by homology modeling, as demonstrated by FINDSITELHM simulations, where the consensus anchor-binding mode is used as a reference framework for the superposition of a query ligand. Furthermore, common anchor substructures are observed across ligands bound to weakly related proteins that belong to more than one superfamily. These subtle evolutionary relationships detected by sensitive threading techniques [32], [33] are of paramount importance for novel biopharmaceutical discovery that could be accounted for to identify potential off-site drug targets and reduce side effects. HIV-1 protease plays a crucial role in the life cycle of HIV [34], [35]; thus, it is an important drug target for AIDS treatment with a number HIV-1 protease inhibitors identified [36], [37]. Several (Table 3) are FDA-approved anti-HIV drugs. Here, we selected HIV-1 protease as an example to demonstrate the performance of FINDSITELHM in ligand-based virtual screening using the coverage of anchor substructures as a simple scoring function. The performance of FINDSITELHM alone and in combination with FINDSITE in virtual screening for HIV-1 protease inhibitors is presented in Figure 16. Both FINDSITE and FINDSITELHM perform considerably better than a random ligand selection. The molecular fingerprints constructed by FINDSITE recovered slightly more known active compounds in the top-ranked fraction of the screening library than anchor-based FINDSITELHM; the enrichment factor calculated for the top 1% (10%) is 27. 0 (6. 8) and 23. 3 (5. 9) for FINDSITE and FINDSITELHM, respectively. Clearly, fusion by ranks outperforms the individual scoring functions with the enrichment factor of 38. 1 (7. 3) for the top 1% (10%) of ranked ligands. These results suggest that the anchor-based approach is able to detect active compounds for which the fingerprint-based method assigns relatively low score. Furthermore, using the combined FINDSITE/FINDSITELHM approach, 4 (7) out of 10 FDA-approved HIV-1 protease inhibitors are found in the top 1% (5%) of the screening library (Table 3). Conservation of protein sequence and structural patterns is widely used to study protein molecular function [38]–[40]. Indeed, the structural and chemical characteristics of a binding site are important for understanding ligand selectivity and cross-reactivity [41], [42]. In that regard, our sequence entropy analysis suggests that residues contacting anchor functional groups have been subjected to higher evolutionary conservation pressure than those contacting ligand variable regions. Furthermore, the conservation of the anchor-binding pose is consistent with the relatively low experimental B-factors observed for residues contacting anchor functional groups. The significantly higher structural plasticity of variable region binding residues could reflect the different types/sizes of functional groups found in the ligand variable substructures that might be responsible for ligand specificity for particular protein family members. Binding site analysis also has practical implications. In the simplest case, using the ligand binding modes extracted from closely related structures and incorporated as spatial restraints in protein structure modeling provides better homology models of protein binding sites [43]. In large-scale computational experiments involving ligand docking, using the AnnoLyze approach the transfer of ligands from known structures of closely related protein-ligand complexes is an attractive alternative to CPU-expensive, classical ligand docking approaches [44]. Here, we have shown that this idea is in fact more general and applies to evolutionarily distant proteins. Indeed, evolution provides a type of signal averaging to identify the essential features associated with ligand binding. This insight can be profitably exploited in a variety of contexts. For example, for evolutionary distant proteins, we identify the subset of ligands whose pose is conserved, viz. the anchor region. Then, based on the observation that across a set of weakly related proteins, not only is the chemical identity of anchor functional groups strongly conserved but also the anchor binding mode, with an average pairwise RMSD <2. 5 Å in most cases. FINDSITELHM uses the consensus binding mode of an anchor substructure as the reference coordinates to perform rapid flexible ligand docking by superposition. This results in an average ligand heavy atom RMSD from native of 2. 5 Å for those ligands that contain a significant portion of the anchor region. Moreover, for predicted protein structures, with considerably less CPU time, FINDSITELHM outperforms all-atom ligand docking approaches in terms of the fraction of recovered binding residues and specific native contacts. The accuracy of FINDSITELHM is affected by several factors: First, for a given target, the set of evolutionarily related template structures needs to be identified. Given the improvements in threading approaches [23], [45] and the completeness of the fold library [46], one can expect to obtain a set of templates for the majority of single domain targets. Next, the docking performance of FINDSITELHM is well correlated with the overall accuracy of binding pocket prediction by FINDSITE. Typically, high accuracy in ligand binding pose prediction requires the binding site to be precisely detected within a distance of 2 Å. This level of accuracy in pocket prediction is usually achieved for Easy targets, as classified by FINDSITE [21]. Finally, the average accuracy of the binding mode prediction by FINDSITELHM decreases with the decrease in the coverage of the anchor substructure by the target ligand as well as with the decrease in the degree of the anchor structural conservation. Here, the growing number of protein crystal structures solved in the complexed state with chemically diverse small organic molecules expands the pool of suitable targets for FINDSITELHM. It is noteworthy from the practical point of view that all these properties can be calculated during the modeling procedure, without knowing the native binding pose. Thus the expected accuracy of FINDSITELHM in ligand binding pose prediction can be estimated with fairly high confidence. Also as shown for HIV-1 protease, using just the target protein' s sequence as input, FINDSITE/FINDSITELHM can efficiently and rather accurately rank a large ligand library. Since for the majority of gene products, at least weakly homologous proteins can be identified in structural databases by current threading methods [23] and approximately correct protein models can be generated by protein structure prediction techniques [10], [12], [13], FINDSITELHM offers the possibility of proteome-scale structure-based virtual screening for novel biopharmaceutical discovery. This would have a great advantage over just screening single proteins. It affords the possibility of identifying lead compounds with desired selectivity that could be further exploited at the outset of the drug development process to reduce side effects. We note that similarity in global fold alone is usually insufficient for effective function interference and results in a high false positive rate [47]. For that reason, the most effective function prediction methods, such as ProFunc [48], AnnoLite [44] or Mark-Us [49] typically combine structure- and sequence-based techniques. In that respect, an important component of FINDSITE/FINDSITELHM is the template selection by threading that employs a strong sequence profile term [23]. This allows the detection of evolutionarily distant homologues [21] with clear functional relationships to the protein of interest not only in terms of the localization of the binding site, but also in the detailed chemical and structural aspects of ligand binding, particularly those that impart binding specificity. Thus, threading provides a richness to functional annotation that to date was not fully exploited. High quality protein–ligand complex X-ray structures were taken from the Astex diverse set used to validate the GOLD docking algorithm [50] and from a non-redundant Q-Dock dataset [22]. In the Astex set, we excluded complexes in which the binding site is formed by more than one protein chain. From the Q-Dock set, we exclude proteins with >35% sequence identity to any protein in the Astex set. We only include proteins for which at least 5 ligand-bound weakly homologous threading templates can be identified by protein threading and the binding pocket can be predicted by FINDSITE [21] within 4. 5 Å from the bound ligand; this represents about 67% of protein targets. The final dataset consisting of 711 complexes is found at http: //cssb. biology. gatech. edu/skolnick/files/FINDSITELHM. In addition to the crystal structures used as the target proteins, we evaluated the performance of FINDSITELHM in ligand docking against weakly homologous protein models for the Dolores dataset [22], [29] of 205 protein models generated by our protein structure prediction protocol, TASSER [13]. For a given amino acid sequence, the PROSPECTOR_3 threading algorithm [23] is used to identify weakly homologous structure templates where templates with >35% sequence identity to target protein are excluded. Structures that bind a ligand are identified by FINDSITE [21] and superimposed onto the reference crystal structure by TM-align [25]. FINDSITE employs an average linkage clustering procedure to cluster the centers of mass of template-bound ligands to detect putative binding sites and then ranks them by the number of ligands. Template-bound ligands that occupy top-ranked predicted binding pockets are clustered using a SIMCOMP similarity (SC) cutoff of 0. 7. SIMCOMP is a chemical compound-matching algorithm that provides atom equivalences [26]. Each cluster of ligand molecules is used to detect an anchor substructure. The equivalent atom pairs provided by SIMCOMP are projected onto ligand functional groups. Here, we used the set of 17 functional groups defined in [22]. The anchor substructure is defined as a maximum set of conserved functional groups present in at least 90% of the ligands from a single cluster. The degree of sequence variability was calculated for each consensus binding residue using Shannon' s information entropy [51]: (1) where pk is the probability that the i-th residue position is occupied by an amino acid of class k, with the amino acid classification given in [52]. The sequence entropy was calculated only for ligand-bound threading templates that share a common binding pocket. Residue equivalences were provided by TM-align [25]. Raw experimental B-factors were extracted from the PDB [53] and normalized using the procedure described in [54], with outliers detected and removed using the median-based method [55]. The FINDSITELHM docking procedure superimposes the target ligand onto the consensus binding pose, the anchor conformation averaged over the seed compounds (the largest set of compounds that have their anchor substructures within a 4 Å RMSD from each other), of the identified anchor substructure. We note that no structural information from the crystal structure of the target complex is used. If multiple anchor substructures are detected, we select the one derived from the cluster of template-bound ligands with the highest average chemical similarity to the target ligand, as assessed by its SIMCOMP score [26]. This maximizes the coverage of the selected anchor. If atom equivalences for non-anchor atoms can be established between the target ligand and any template-bound ligand, their positions are also included in the set of the reference coordinates. Often, by including additional coordinates, approximately correct positions of ligand variable groups can provide a good initial conformation for post-docking refinement, e. g. in Figures 3,12, and 13. If none of the identified anchor substructures is covered by the target ligand, it is randomly placed in the predicted pocket. Ligand flexibility is accounted for by the superposition of multiple conformations of the target ligand (for details see classical ligand docking protocols). The conformation that can be superposed onto the reference coordinates with the lowest RMSD to the predicted anchor pose is selected as the final model. Crude models of protein-ligand complexes generated by FINDSITELHM were optionally refined by a simple energy minimization in Amber 8 [28]. We used the Amber force field 03 [56] for proteins and the general Amber force field [57], GAFF, for ligands. The parameterization of ligands was done in a fully automated fashion with the aid of Antechamber 1. 27 [58]. If necessary, the system was neutralized by calculating a Culombic potential on the grid of 1 Å using LEaP (Amber 8) in order to place chloride (sodium) ions at the positions of the highest (lowest) electrostatic potential around the initial protein-ligand complex. Protein atoms were fixed, while the ligand conformation was energy minimized in vacuum by 1000 cycles of a steepest-descent procedure, followed by 1000 cycles of a conjugate gradient procedure. From our dataset of 711 protein-ligand complexes, we selected only enzymes in which the anchor substructure (or multiple anchor substructures) derived for the top-ranked predicted binding pockets consists of ≥50% and ≤90% of the average ligand molecule' s size and matches the native ligand. Subsequently, native ligands were scanned for the presence of CSSs. Here, we used the collection of the CSSs compiled for 42 major enzyme superfamilies by Babbitt and colleagues [30], from which we removed those substructures that consist of less than 5 atoms. A CSS was considered to be present in the native ligand if the native ligand atoms cover at least 90% of its atoms, as reported by SIMCOMP [26]. This procedure resulted in 24 enzymes and 35 ligand clusters. Next, for each cluster and the associated anchor substructure, we examined the fraction of CSS' s atoms covered by the anchor functional groups as well as the fraction covered by the non-anchor groups. The screening library consists of 1089 known HIV-1 protease inhibitors (MDL activity index: 71523) extracted from the MDL Drug Data Report [63] and 123,274 lead-like background compounds from the Asinex Platinum Collection [64]. A weakly homologous model of HIV-1 protease was generated from the amino acid sequence (PDB: 1w5y) using TASSER [13]. Only distantly related (<35% sequence identity to HIV-1 protease) structure templates were used. The predicted model used in this study has a 4. 91 Å (4. 09 Å) RMSD to native calculated for all heavy atoms (Cα atoms). We applied two ligand-based virtual screening techniques to rank the screening library: a fingerprint-based method implemented in FINDSITE and simple scoring by the anchor substructure coverage, where the anchor substructures were identified by FINDSITELHM. In both cases, we used a collection of ligands bound to weakly homologous (<35% sequence identity to the target) threading templates identified by PROSPECTOR_3 with a Z-score ≥4. FINDSITE constructs ligand templates for fingerprint-based virtual screening by clustering the molecules that occupy the top-ranked predicted binding site using the Tanimoto coefficient (TC) [65] cutoff of 0. 7 [21]. Here, we employed the 1,024-bit molecular fingerprints from Daylight Chemical Information Systems [66]. The representative molecules selected from the clusters were used to rank a compound library using a weighted Tanimoto coefficient (mTCave): (2) where n is the number of ligand clusters, wi is the fraction of ligands that belong to cluster i, and is the averaged TC (TCave) calculated for the representative ligand from cluster i and a library compound. The overlap between two fingerprints was measured by TCave [67]–[69]: (3) where TC′ is the TC calculated for bit positions set to zero rather than to one as in the traditional TC [65]. In virtual screening by anchor coverage, we used the anchor substructures detected for HIV-1 protease by FINDSITELHM as described in Methods. For a given library compound, we calculated the coverage of the anchor substructure that was derived from the cluster of template-bound ligands with the highest average chemical similarity, as assessed by SIMCOMP score [26]. The screening library was then ranked by decreasing anchor coverage. Finally, we applied data fusion to combine the results from virtual screening using the fingerprint-based (FINDSITE) and the anchor-based (FINDSITELHM) approaches. Data fusion techniques are commonly used in chemoinformatics to merge screening results generated by different descriptors or scoring functions [70]–[74]. Typically, chemical data fusion employs the combination of rankings from individual screening experiments using one of several different fusion rules, such as MIN, MAX or SUM [75]. Here, we applied the SUM rule that is expected to be less sensitive to noisy input than both extreme rules [70] and is generally preferred when fusion is by rank [71]. For a given library compound k, a combined score (CS) is calculated from: (4) where n is the number of ranked lists (in our case, n = 2: FINDSITE and FINDSITELHM) and ri denotes the rank position of the library compound k in the i-th ranked list. To assess the performance of FINDSITE/FINDSITELHM in virtual screening for HIV-1 protease inhibitors, we calculated the enrichment factor (EF) [76], [77] for the top 1% and 10% of the ranked screening library: (5) where Isampled is the number of known HIV-1 protease inhibitors in the top-ranked fraction of Nsampled compounds, Itotal and Ntotal is the total number of inhibitors and the library compounds, respectively. The maximal enrichment factors for the top 1% and 10% of the ranked library are 100 and 10, respectively. In addition to the enrichment factor, we assessed the results in terms of the enrichment behavior, i. e. the fraction of known inhibitors retrieved in the top-ranked fraction of the ranked screening library.

As an integral part of drug development, high-throughput virtual screening is a widely used tool that could in principle significantly reduce the cost and time to discovery of new pharmaceuticals. In practice, virtual screening algorithms suffer from a number of limitations. The high sensitivity of all-atom ligand docking approaches to the quality of the target receptor structure restricts the selection of drug targets to those for which high-quality X-ray structures are available. Furthermore, the predicted binding affinity is typically strongly correlated with the molecular weight of the ligand, independent of whether or not it really binds. To address these significant problems, we developed FINDSITELHM, a novel threading-based approach that employs structural information extracted from weakly related proteins to perform rapid ligand docking and ranking that is very much in the spirit of homology modeling of protein structures. Particularly for low-quality modeled receptor structures, FINDSITELHM outperforms classical all-atom ligand docking approaches in terms of the accuracy of ligand binding pose prediction and requires considerably less CPU time. As an attractive alternative to classical molecular docking, FINDSITELHM offers the possibility of rapid structure-based virtual screening at the proteome level to improve and speed up the discovery of new biopharmaceuticals.

Abstract Introduction Results Discussion Methods

biophysics/biomacromolecule-ligand interactions biochemistry/biomacromolecule-ligand interactions biochemistry/drug discovery

2009

FINDSITELHM: A Threading-Based Approach to Ligand Homology Modeling

The sparse coding hypothesis has enjoyed much success in predicting response properties of simple cells in primary visual cortex (V1) based solely on the statistics of natural scenes. In typical sparse coding models, model neuron activities and receptive fields are optimized to accurately represent input stimuli using the least amount of neural activity. As these networks develop to represent a given class of stimulus, the receptive fields are refined so that they capture the most important stimulus features. Intuitively, this is expected to result in sparser network activity over time. Recent experiments, however, show that stimulus-evoked activity in ferret V1 becomes less sparse during development, presenting an apparent challenge to the sparse coding hypothesis. Here we demonstrate that some sparse coding models, such as those employing homeostatic mechanisms on neural firing rates, can exhibit decreasing sparseness during learning, while still achieving good agreement with mature V1 receptive field shapes and a reasonably sparse mature network state. We conclude that observed developmental trends do not rule out sparseness as a principle of neural coding per se: a mature network can perform sparse coding even if sparseness decreases somewhat during development. To make comparisons between model and physiological receptive fields, we introduce a new nonparametric method for comparing receptive field shapes using image registration techniques. A central question in systems neuroscience is whether optimization principles can account for the architecture and physiology of the nervous system. One candidate principle is sparse coding (SC), which posits that neurons encode input stimuli efficiently: stimuli should be encoded with maximum fidelity while simultaneously using the smallest possible amount of neural activity [1], [2]. Much evidence suggests that primary visual cortex (V1) forms sparse representations of visual stimuli [1], [3]–[7]. For example, when trained with natural scenes, SC models have been shown to learn the same types of receptive fields (RFs) as are exhibited by simple cells in macaque primary visual cortex (V1) [1], [8]. Throughout this paper, we make reference to the notion of “sparseness”. Intuitively, sparseness is related to there being either a small subset of neurons active at any time (population sparseness), or to each neuron being active only a small fraction of the time (lifetime sparseness) [9]. In the Methods section, we define the precise notions of sparseness that we use in this paper. In further support of the SC hypothesis, measurements of the firing rates of V1 neurons in response to videos of natural scenes show that those rates are low, and that the firing rate distributions are sharply peaked near zero [7]. Similarly, cell-attached recordings in auditory cortex show highly sparse levels of activity [10]. Conversely, other experimenters [11] have observed non-sparse (dense) neuronal activity in visual cortex, although the boundary between “sparse” and “dense” activity is open to interpretation and thus it is unclear how sparse the activity must be in order to confirm the SC hypothesis [12]. Importantly, however, it has been observed that stimulating larger portions of the visual field leads to sparser, and less correlated, V1 neuronal responses [4]–[6]. It has been suggested that this effect arises because of inhibitory recurrent connections between excitatory cells, mediated by the appropriate interneurons [6]. In simulating the development of a sparse coding model, one typically [1], [3] initializes the receptive fields with random white noise — so as to not bias the shapes of the RFs learned by the network — and then presents the network with natural images, in response to which the RFs get modified. As the model (e. g. , [1], [3], [13]) modifies itself in response to the stimuli, neurons gradually learn features that allow for a better encoding of the stimuli, so the sparseness is expected to increase over time. This point was emphasized in recent work [14]. Physiology experiments, however, show something different in the developing visual cortex. Recently, Berkes and colleagues measured multi-unit V1 activity in awake young ferrets viewing natural movies, and found that, as the animals matured, their stimulus-driven V1 activity became less sparse [14], [15] (Fig. 1). The above discussion hints at a major source of confusion in this area of research. In particular, sparseness is discussed as both a relative measure (i. e. : “Is network A sparser than network B? ”), and as an absolute descriptor (i. e. : “Is network A sparse? ”). In this paper, we will first study relative measures of sparseness, and observe how these measures change as a result of development in our recently published SAILnet model [12]; do they increase, or decrease over time? The absolute sparseness values of the final (mature) networks – which vary between 0 (not sparse), and 1 (maximally sparse) – will be used to infer whether the final network is sparse at all. This mirrors the way that Berkes and colleagues discussed sparseness in the developing ferret. The ferret sparseness-over-time data appears to contradict the SC hypothesis. At the same time, that hypothesis has otherwise been quite successful in explaining some key features of peripheral sensory systems. It is therefore natural to ask whether sparse coding models necessarily must exhibit increasing sparseness in order to learn V1-like receptive fields and perform sparse coding in the mature state. In this work, we focus primarily on a recently published variant of sparse coding called SAILnet [12] in which homeostasis regulates the neuronal firing rates while synaptically local plasticity rules modify the network structure, leading to V1-like receptive field formation. We will demonstrate that, depending on the initial conditions of the simulation, SAILnet can exhibit either increasing, or decreasing sparseness, while learning RFs that are in quantitatively good agreement with those observed in V1, and having a reasonably sparse final state. The choices of parameter values in the model determine the equilibrium state to which the network ultimately converges. If the initial conditions are even sparser than this equilibrium point, sparseness will decrease during development, and yet the final state can still be sparse in an absolute sense. We will also see that, for appropriately chosen initial conditions, the same can be true of the canonical SparseNet model of Olshausen and Field [3]. Thus, the apparent contradiction between the ferret developmental sparseness data, and SC models [14] does not necessarily mean that SC is implausible as a theory for sensory computation. Later in this paper, we discuss plausible alternatives for sensory coding other than SC models. Since this paper focuses primarily on our SAILnet model (Fig. 2), we will now provide a brief overview that model, which is described in detail elsewhere [12] and summarized in the Methods section. The model consists of a network of leaky integrate-and-fire (LIF) neurons, which receive feed-forward input from image pixels, in a rough approximation of the thalamic input to V1. The neurons inhibit each other via recurrent inhibitory connections, the strengths of which are learned so as to reduce correlations amongst the units, consistent with recent physiology experiments [4]–[6]. We note that one can modify SAILnet so that interneurons mediate the inhibition between excitatory cells so as to satisfy Dale' s law (E-I Net; [16]). The neurons' firing thresholds are modified over time so as to maintain a target lifetime-average firing rate. For our LIF neurons, this is similar to synaptic rescaling, which has been proposed as a mechanism to stabilize correlation-based learning schemes [17], [18], and has been observed in physiology experiments [17]. Alternatively, the variable firing threshold can be thought of in terms of a modifiable intrinsic neuronal excitability, another well-known homeostatic mechanism [19]. Finally, the feed-forward weights are learned by the network, so that the neuronal activities form an optimal linear generative model of the input stimulus, subject to the constraints imposed by limited firing rates and minimal correlations. The derivation of our learning rules from this objective function is presented in [12]. All information needed for the model' s plasticity rules is available locally at the synapse being modified — updates depend only on the pre- and post-synaptic activity levels. To study the change in sparseness over time, we ran SAILnet simulations, starting with randomized feed-forward weights, recurrent connection strengths, and firing thresholds that were initialized with Gaussian-distributed white noise. At different times during the development process, we recorded the simulated neuronal activity in response to randomly selected batches of natural images. Following a recent experimental study [14], we computed from these network activities three sparseness measures, which are discussed in more detail in the Methods section. Each of these measures varies between (not sparse at all) and (as sparse as possible). The first of these, the “activity sparseness, ”, measures the fraction of units that are inactive in response to a given stimulus, averaged over different stimuli. If this quantity is near 1, then only a small subset of units responds to each stimulus. If every unit is active in response to every stimulus, then. The “population sparseness” [5], [14], [20], , measures the degree to which the population response to a given stimulus is restricted to a small subset of the population, averaged over all stimuli. If only a small number of units have large activities in response to a given stimulus, then will be near, even if many units have small but non-zero activities. By contrast, would be small in that case, because there are not many completely inactive units. If the units all respond equally to every stimulus, then. For the same level of representation error, the formation of efficient representations demands relatively high values of and, such that only a small fraction of the available neural resources are utilized in representing each image. Finally, the “lifetime sparseness” [5], [14], [20] measures how much the responses of individual units tend to be concentrated over a small subset of stimuli, averaged over units. If the units respond very selectively, so that they have strong responses to a small number of stimuli, and weak responses to most stimuli, then will be near 1. Conversely, if each unit responds equally to all stimuli, then. Intuitively, all of these measures are somewhat related (although, see [9] for notable exceptions). At the same time, when it comes to the efficiency of the neural representation, the more relevant quantities are the activity sparseness and the population sparseness, which both have to do with the fraction of neural resources used to represent each image [9]. Furthermore, the values of “lifetime” sparseness one obtains will vary with the time scale over which one performs the measurement. As such, it is a somewhat more ambiguous quantity than are the activity and population sparseness measures. Despite these issues, we include results for the lifetime sparseness for our model in keeping with the experimental study [14] that motivated this theoretical project. In order to further facilitate meaningful comparison with the experiment of Berkes and colleagues, we mimicked a multi-unit activity measurement by randomly grouping together sets of 8 SAILnet neurons, whose activities were then summed to form a multi-unit response. These “multi-unit” activities were used for computing our sparseness measures. This procedure yielded results (Fig. 3) that were qualitatively similar to the single-unit sparseness measures (not shown), but with larger changes in sparseness values. A direct quantitative comparison between our model multi-unit sparseness data and the ferret data is difficult because it is not clear how best to estimate the relevant number of neurons to group together, or even whether all groupings should have the same number of neurons. Furthermore, in the ferret data, the neurons grouped together are physically nearby, which means that, due to retinotopic and orientation maps in V1, they will have similar receptive fields. The SAILnet model has no such notion of spatial organization and our random grouping of cells misses that aspect of the ferret experiment. In Fig. 3, we show a random subset of 196 (out of the 250 total) receptive fields of SAILnet neurons both before and after the network is trained with natural scenes. We also show the evolution of our multi-unit sparseness measures during that training process. Contrary to the idea that sparse coding models must show strictly increasing sparseness during learning [14], our SAILnet model can display decreasing sparseness by all three measures while it is learning localized and oriented receptive fields. We will later show that the popular SparseNet model of Olshausen and Field [3] can also exhibit decreasing sparseness over time. The time course of the sparseness measures depends on the learning rates (parameter modification step sizes), with smaller learning rates leading to slower changes in sparseness measures, as expected (data not shown). The depth of the observed “undershoot” also depends on the initial conditions and the learning rates. The specific activity sparseness values () depend on the chosen threshold: higher thresholds lead to higher sparseness values. The receptive fields learned by the model, while displaying decreasing sparseness, are in good quantitative agreement with a measured corpus of 250 macaque monkey V1 simple cell receptive fields, as we will demonstrate in the section on comparisons of receptive field shapes. The model discussed in this section and shown in Fig. 3 has relatively high firing thresholds and a relatively large amount of lateral inhibition in the initial state. Those properties lead to highly sparse firing. As the network learns, the homeostatic firing rate regulation reduces the thresholds and then inhibitory connections are modified by their own plasticity rules (see Methods section for details). These have the effect of reducing the model' s sparseness over time. For contrast, in the next section we will consider a model that is identical to the one presented here, but with the following modifications to the initial conditions: the firing thresholds are initialized at smaller values and there is initially less lateral inhibition. We find that SAILnet does not require sparseness to decrease over time; rather, it is compatible with decreasing sparseness. To demonstrate this point, we repeated our SAILnet simulations and sparseness measurements with different initial conditions (see Methods section for details). As discussed in the previous section, these initial conditions have lower firing thresholds and less lateral inhibition than in the model shown in Fig. 3. In this case, the relatively low firing thresholds and relatively small amount of lateral inhibition lead to the initial network state being less sparse than the final (equilibrium) state, so sparseness increases over time (Fig. 4). Similar to Fig. 3, in Fig. 4 we show a random subset of 196 neuronal receptive fields both before and after the training procedure and, as we demonstrate below, those RF shapes are in quantitative agreement with those measured in macaque V1. We emphasize that, compared to the model discussed in Fig. 3, which exhibited increasing sparseness, the model discussed here (and in Fig. 4) differs only in the initial conditions; all other parameters were the same for the two models. Consequently, after a long training period, over which the effects of the initial conditions gradually disappear, these two models have very similar final sparseness levels and receptive fields. For the models studied in this paper (Figs. 3 and 4), the final multi-unit sparseness values (after the final training batch) are for the model in which sparseness increases over time (Fig. 4), and for the model in which sparseness decreases over time. While our SAILnet model [12] and a recent extension that obeys Dale' s law [16] are more biophysically realistic than previous sparse coding models, the increasing or decreasing sparseness over time we describe above (Figs. 3 and 4) is not unique to SAILnet. To explore this issue more fully, we return to the canonical SparseNet model of Olshausen and Field [3], [21]. We first note that, as in SAILnet, the “equilibrium” sparseness level in SparseNet is determined by a free parameter in the model (in [3]). Furthermore, in the SparseNet model, there is a homeostatic mechanism that adjusts the magnitudes of the feed-forward weights so as to keep the units' activities near some pre-defined set point. Similar to SAILnet, this process is not instantaneous. In what follows, we use the SparseNet code of Olshausen and Field [3], [21] “out of the box, ” without modifying any parameters except the initialization of the basis functions — these are analogous to the feed-forward weights, or receptive fields, of SAILnet units. We begin by initializing these basis functions with Gaussian white noise of variance, so that the bases have norms of approximately. In this case, sparseness increases over time (Fig. 5a) and the basis amplitudes decrease: the mean norm of these bases is approximately 0. 5 once the model converges, after the training period. These changes can be understood by recalling that, during inference — where the activities of the units are determined in response to a given image — the activities are chosen to minimize the following cost function: . The error, as in SAILnet, is the sum over pixels of the squared error of the difference between the image and a linear generative model formed by multiplying each unit' s activity by its basis function. Thus, if the basis has a small magnitude, then large activations are needed in order to form a decent linear generative model. However, these large activations are punished by the “activities” term in the cost function. As a result, some of the units that add only a modest improvement to the representation are turned off, or nearly so. If the basis has a large magnitude, then only small activations are needed to form the linear generative model, and these small activations are less strongly penalized than are large activations. As a result, the activation can be more distributed over the network when the filters have large magnitudes. To summarize: large basis function magnitudes lead to less sparse network activity and the basis magnitudes are modified with a non-zero timescale. Putting all of this together, if we initialize the bases with Gaussian white noise of variance, so that they initially have norms in the neighborhood of, then the basis norms increase over time during training. After this model converges, the mean norm of the bases is again around 0. 5. Consequently, the sparseness decreases over time (Fig. 5b). As in SAILnet, one way to understand these trends is to recall that the model parameters dictate the final “equilibrium” state of the model, but the initial conditions can be chosen independently of the final state. As such, initial conditions can be chosen to be either more or less sparse than the equilibrium condition, leading to sparseness either decreasing or increasing over time. In many theoretical studies (e. g. , [1], [12], [22], [23]), one learns a sparse coding dictionary for natural stimuli, then compares the shapes of the resultant basis functions to the shapes of the physiologically measured receptive fields. Typically, this comparison is either done by eye (as in [22]), or by fitting both the model RFs and the experimentally measured ones to some parameterized shape functions, and then comparing (again, typically by eye) the distributions of the resultant shape parameters for both the model and the data [1], [12], [23]. More rigorously, one can quantitatively compare the distributions of these shape parameters (Rehn, Warland, and Sommer, CoSyNe 2008 abstract) between the model and the experimental data, although that method fails if one has too few RFs with which to perform the comparison. The by-eye comparisons are not very quantitative, even if they first involve fitting parameterized shape models, and any fitting of parameterized shape models is vulnerable to failures of the shape function: any RFs whose shapes are not well described by the parameterized function will yield nonsense best fit parameter values. To get around these difficulties, we introduce a novel method for directly comparing the shapes of theoretical and experimental receptive fields, using image registration. In this technique, we assume that receptive fields may differ by a translation, rotation, and/or global size rescaling, yet still have the same shape. For example, consider an equilateral triangle within a bounding box. A shifted, rotated, and resized version of that shape is still an equilateral triangle. We apply this intuition to the comparison between our model receptive fields, and a set of 250 macaque V1 receptive fields courtesy of D. Ringach. We do this by taking each experimentally measured V1 receptive field and then for each model RF we find the combination of translation, rotation, and overall rescaling that gives the best match between the experimental and transformed-model RF. We quantify the match by the value (square of the correlation coefficient) between the pixel values of the model and the experimental RF. Choosing the -maximizing RF is similar to seeking the model RF that can account for the largest fraction of the variance of the experimental RF, but allows for an overall multiplicative constant in front of the model RF. Specifically, the value tells us the fraction of the experimental RF variance that could be explained by a linear function of the best model RF (i. e. , possibly including an additive constant to all pixel values and an overall amplitude change). Once we have done this for all model RFs, we take the one whose best transform yields the largest and take that as the best-fit model RF for the given experimental RF. In this way, we answer the question: Once we account for possible translations, rotations, and size rescalings, how much of the variance in the experimental RF pixel values can be accounted for by the library of model RFs? values near indicate that the experimental RF is reproduced perfectly by the theoretical model, while values near indicate that it is not. We then repeat this procedure for all of the RFs in the experimental dataset, yielding one value per experimental RF. Below, we discuss the averages over these values. Looking at the macaque RFs in Fig. 6A, it is clear that the region of support of the RF is often smaller than the size of the image window over which the RF is measured, and any noise outside of the region of support (or within it, for that matter) will be unaccounted for by the image registration process, thus lowering the apparent goodness-of-fit. To attempt to quantify this level of noise and to give a solid benchmark with which to assess our experiment-vs. -model comparison, we repeat the image registration fitting described above, but instead of using model RFs as comparators, we compare each experimental RF against the corpus of other experimental RFs. In other words, we do a leave-one-out analysis where we try to fit each macaque RF and, for that fit, we take the 249 other macaque RFs and pretend that they are “model” RFs. We then find the largest value possible for the best transformation of each of the 249 comparison RFs and use that number to quantify how well we could realistically expect to fit the macaque RF. We term this maximal value for the macaque-to-macaque comparison the fraction of variance in the data that is “explainable” by our image-registration technique (although not necessarily captured by the dictionary of shapes learned by our sparse coding model). In using this technique to estimate the noise level, we essentially assume that the RF shapes are repeated in the experimental data (which one can see in Fig. 6), and use that intuition to ask, “How much of the data variance is due to noise rather than the RF properties themselves? ” In Fig. 6, we show the experimental RFs, the best-transformed (translation, rotation, and overall size rescaling) model RFs learned with either increasing or decreasing sparseness values, and the quantitative comparisons between the RF shapes (average values for how well the macaque RFs can be explained by the model RFs). The macaque-to-macaque comparisons (Fig. 6E) show that, on average, of the variance in the RF pixel values can be explained using other RFs from the macaque V1 dataset. We term this the “explainable” variance and it sets an upper bound on how well we could expect our model RFs to match the macaque RFs. For comparison, the model RFs account for, on average, of the variance in the RF pixel values, regardless of whether the learning of those RFs was accompanied with either increasing or decreasing sparseness. The difference between the average explained variance for the two different models (those of Figs. 3 and 4) is not statistically significant (, paired t test;), while the differences between each of the model' s mean values and that of the macaque-vs-macaque comparison are statistically significant (, paired t test;). Because the model RFs can explain an average of roughly () of the explainable variance in the RF data, regardless of whether the model experienced increasing or decreasing sparseness during training, we conclude that the model RFs are in quantitatively good agreement with the experimental RFs, independent of whether the learning of shapes was accompanied with increasing or decreasing sparseness. Generally, larger networks have a greater diversity of receptive field shapes (this can be easily seen by comparing the RFs shown in this paper to those in [12]), and will thus tend to perform better in our image-registration comparisons. The networks studied in this paper were relatively small, in order to allow us to study the evolution of networks with many different initial conditions. Thus, we expect that even better model-to-experiment RF matches are possible if one were to study larger networks. On the other hand, the macaque V1 database to which we compared our model contains only 250 receptive fields, so a “fair” comparison to a larger simulated network would require more experimental data, or the selection of a random subset of the simulated RFs. Our quantitative non-parametric RF comparison method could be used to compare many different theories to experimental data, and thus to ascertain which ones provide the best fit. That comparison is beyond the scope of this paper. We have demonstrated that a computational model (SAILnet [12]) can learn V1-like receptive fields while simultaneously exhibiting either a decrease, or an increase, in the sparseness of neuronal activities. In both cases, the sparseness of the final (mature) network state is high enough to be reasonably considered “sparse. ” We further showed that these same trends in sparseness over time can be achieved with the less biophysically realistic SparseNet model, despite the fact that it does not incorporate the same form of homeostasis as SAILnet. In order to quantify the similarity between experimentally measured V1 receptive fields and the receptive fields learned by our SAILnet model, we have further introduced a novel non-parametric RF comparison tool based on image registration techniques. Since sparseness can decrease during development, with the mature network state still performing sparse coding, the type of active sparseness maximization disproven by recent experiments [14] is not necessary to produce observed V1 receptive field shapes, nor is it required to learn a sparse representation of natural scenes. The trends in sparseness over time can be so strongly affected by the initial conditions of the network that those trends are not very informative about the objective function being optimized. Thus, developmental data, such as those shown in Fig. 1, cannot strictly rule out the general notion that V1 simple cell receptive fields develop so as to form sparse, efficient, representations of natural scenes. One possibility that requires consideration is that the sparseness data of Berkes and colleagues [14] should not be compared at all with theoretical models when assessing the hypothesis that V1 performs sparse coding. Recall that the sparse coding models criticized by Berkes and colleagues showed increasing sparseness during receptive field formation, and that their claim was that, since the ferret data instead showed decreasing sparseness during development, the sparse coding models do not provide a good description of V1. In order for this comparison to be “fair, ” one must ensure that receptive fields undergo significant change during the developmental period over which Berkes and colleagues measured sparseness. Indeed, other experimenters have observed that the orientation and direction selectivity of the neurons in ferret V1 increase [24], [25] during the same developmental period over which Berkes and colleagues observed decreasing sparseness levels, and that the mature state of the visual cortex for both ferrets and cats is sensitive to visual experiences in this period [25]–[28]. Combining these observations, we note that sparseness in ferret V1 seems to decrease while the visual cortical maps and receptive fields are being refined by experience, in apparent contradiction to the SC hypothesis. The largest contribution of this paper is to resolve that apparent contradiction. Of course, there could always be other reasons — beyond the scope of this paper — why the developmental data fail to be relevant to the sparse coding hypothesis. We leave that question for future work. For the sake of completeness, we note that Rochefort and colleagues have observed that spontaneous slow-wave activity in the anesthetized mouse visual cortex becomes sparser immediately after eye-opening [29]. At first glance, this might appear to contradict the ferret data of Berkes and colleagues [14]. However, since the ferret data is stimulus-evoked activity and the mouse data is spontaneous activity, it is not clear that a comparison between these datasets is meaningful. Because the spontaneous activity is not very easily related to the sparse coding hypothesis, which does not have much to say about activity in the absence of sensory input, we have focused on the ferret data (stimulus-evoked activity) in this paper. We are not the first to propose that homeostasis might underlay experience-dependent modification of the nervous system. Indeed, Marder and others have strongly and persuasively argued that neural systems might have a desired operating point such that when perturbed they use homeostatic mechanisms to return to that desired functional state [18], [19]. Moreover, Miller and others have shown that homeostatic activity regulation can facilitate learning in model-neuronal systems [30]–[32]. Finally, recent work by Perrinet [30] also used homeostatic activity regulation in learning sparse codes, so that model may also show either increasing or decreasing sparseness over time, for appropriately chosen initial conditions. We have demonstrated that the mature network state can perform sparse coding regardless of whether learning is accompanied by an increase or decrease in sparseness. At the same time, sparse coding is not the only principle that has been proposed in order to understand V1 function. Of particular interest in this regard is a recent study by Berkes and colleagues [15]. In that work, the authors showed that both stimulus-evoked and spontaneous (in the absence of any stimulus) activity in V1 became more similar as the animals aged, suggesting that V1 might be learning a Bayesian prior on the statistics of the environment. They further observed that part of this change in activity distributions came about due to increases in the correlations between V1 neurons with age. This observation is contrary to the redundancy reduction arguments often used to support efficient coding models, such as traditional sparse coding models and SAILnet. Thus, there is some evidence that sparse coding may not be the best theory of V1 function, and that others, such as that advanced in [15], may be better in some regards. Moreover, it is not clear how the sparse coding hypothesis could account for many of the properties of V1 complex cells despite its success for simple cells, and differences in the level of activity between awake and anesthetized V1 may pose additional challenges for sparse coding models. However, sparse coding models have successfully accounted for the shapes of V1 simple cell receptive fields, while the Bayesian-type optimality models have yet to do so. There is thus much room for further advances in our understanding of sensory coding, and measurements that can rule out theoretical models are key to that advancement. Our results show that the decrease in sparseness during development, which has been argued to rule out sparse coding in V1 [14], is not, on its own, sufficient to make that claim. There are many different ways to measure sparseness. In this work, we follow the experimental study of Berkes and colleagues [14] and use the following three measures. First, the “activity sparseness” (), which is the fraction of units inactive in response to any given stimulus: (1) where is the number of units whose activities rose above some threshold number of spikes in response to the stimulus, and N is the total number of units for which data were recorded. We set the threshold to 4 spikes for the multi-unit sparseness data shown herein. We performed this measurement by averaging over different input stimuli. The activity sparseness is very similar to the norm – which is actually a pseudo-norm – of the unit activities; low norms correspond to high values. In addition, we recorded two other sparseness measures, originally due to Treves Rolls [20] (TR), and subsequently modified by Vinje Gallant [5]. First, let us consider what we will call the “TR population sparseness” measure [15], , (2) where is the activity of unit. Note that is assessed in response to a single image, although for our purposes, we will average this measure over different image stimuli, to infer the average TR population sparseness. Similarly, we will define the “TR lifetime” sparseness of a given unit, , the same way (Eq. 2), but with the replacement that represents the unit' s activity in response to a given image, and will be the number of different image stimuli () for which activities are recorded. Similar to the TR population sparseness, we will average these values over the entire population for our measurement. The SAILnet model [12] consists of a network of leaky integrate-and-fire neurons that receive feed-forward input from image pixels (a rough approximation of the thalamic input to V1), and inhibit each other through recurrent connections. The feed-forward weight from pixel (with value) to neuron () and the inhibitory recurrent connections between neurons and () are learned by the network. In response to a given image, neuron emits some number of spikes, which can be zero. Homeostasis, enforced via modifiable firing thresholds forces the neurons to all have the same lifetime-average firing rate of spikes per image. After the image presentation, the network parameters are updated via (3) where the (positive) constants, and define the rates at which the parameters are learned. In order for them to remain inhibitory, after each update, the recurrent connections are rectified so that. For all simulations shown herein, the feed-forward weights were initialized with Gaussian white noise, and the learning rates were set to. We used neurons which viewed pixel image patches, hence the neuronal representation was approximately critically sampled (overcomplete) with respect to the number of pixels. The target firing rate was set to spikes per neuron per image in all cases. In all cases, the norm of the feed-forwards weights to each neuron was in the initial condition. For the data shown in Fig. 4, the initial recurrent connection strengths were drawn randomly from a normal (Gaussian) distribution with zero mean and unit variance, the firing thresholds were initialized to, and the model neuronal activity became more sparse during training. For the results shown in Fig. 3, the initial recurrent connection strengths were drawn randomly from a lognormal distribution (the negative of their logarithms were drawn from a normal distribution with zero mean and unit variance), the firing thresholds were drawn uniformly over, and the model neuronal activity became less sparse during training. We note that in all cases, the specific shape of the sparseness vs. time plot depends on the choice of initial conditions. However, for a large class of initial conditions, the sparseness will decrease over time, and for another large class of initial conditions, it will increase (data not shown). Thus, our qualitative conclusion is not particularly sensitive to the exact numerical values described above. The SparseNet results were generated using code publicly distributed by Bruno Olshausen (http: //redwood. berkeley. edu/bruno/sparsenet/). The code was used “out of the box”, without modifying the parameter values. For the data shown in Fig. 5a, then bases (matrix A in the code) was initialized with Gaussian white noise of variance 1. For the data shown in Fig. 5b, matrix A was initialized with Gaussian white noise of variance 0. 01. In both cases, 256 units were used, and the model was trained on pixel image patches: the model is overcomplete with respect to the number of input pixels. To perform the quantitative comparison of receptive field shapes, we used the image registration tool in MatLab [33]. This package allowed us to quickly and easily compute the optimum combination of translation, rotation, and stretch – called a “similarity” transform in MatLab – to match each physiology RF with an RF from the appropriate comparison class (either model data or other experimentally measured RFs). For the optimizer, we used the “monomodal” option. Our data consists of 250 Macaque V1 receptive fields, measured using reverse-correlation methods in the lab of Dario Ringach [34]. These data are either, , or pixel images, showing the extent to which the neuron responds to each pixel. In order to standardize these data for the comparison, and because many of the RFs occupy only a tiny fraction of the image, we pre-process those RFs that are larger than pixels, as follows. First, we find the peak absolute pixel value in the image – nominally, this is somewhere in the region of support of the “real” RF –, and we cut out a pixel region surrounding that peak. We then perform our image registration fitting on these standard-sized RFs. For each macaque RF, we performed an exhaustive search over all model RFs, wherein we found the best similarity transform to match each model RF to the macaque RF, then took the best-matching model RF (with the appropriate best similarity transform), as the fit. The value between this best-fit transformed-model RF and the data RF was used to quantify the goodness of fit. To generate a benchmark to assess how good a “good” value is for this problem, we repeated our fitting process, but instead of fitting the data to SAILnet model RFs, we fit each macaque RF with the corpus of other macaque RFs. In so doing, we could estimate the fraction of data variance that could be explained in a best-case scenario. The ratio between these numbers – the data-vs. -model value and the data-vs. -data value – gives us an estimate of the fraction of explainable variance that is captured by the model.

The popular sparse coding theory posits that the receptive fields of visual cortical neurons maximize the efficiency of the neural representation of natural images. Models implementing this idea typically minimize a combination of the error in reconstructing natural images from neural activities, and the average level of activity in the model neurons. In simulations, these models are presented with natural images and the RFs then develop so as to increase representation efficiency. After a long developmental period, the model RFs typically agree well with those observed experimentally in visual cortex. Since the models seek to minimize (for a given level of reconstruction error) the neural activity levels, the average levels of neural activity might be expected to decrease as the models develop. In the developing mammalian cortex, visual RFs are also modified during development, so the sparse coding hypothesis might appear to suggest that activity levels should decrease during development. Recent experiments with young ferrets show the opposite trend: mature animals tend to have more active visual cortices. Herein, we demonstrate that, depending on the models' initial conditions, some sparse coding models can exhibit increasing activity levels while learning the same types of RFs that are observed in visual cortex: the developmental data do not preclude sparse coding.

Abstract Introduction Results Discussion Methods

systems biology developmental neuroscience theoretical biology neural homeostasis computational neuroscience sensory systems biology neuroscience

2013

Sparse Coding Models Can Exhibit Decreasing Sparseness while Learning Sparse Codes for Natural Images

It is well established that RNA viruses exhibit higher rates of spontaneous mutation than DNA viruses and microorganisms. However, their mutation rates vary amply, from 10−6 to 10−4 substitutions per nucleotide per round of copying (s/n/r) and the causes of this variability remain poorly understood. In addition to differences in intrinsic fidelity or error correction capability, viral mutation rates may be dependent on host factors. Here, we assessed the effect of the cellular environment on the rate of spontaneous mutation of the vesicular stomatitis virus (VSV), which has a broad host range and cell tropism. Luria-Delbrück fluctuation tests and sequencing showed that VSV mutated similarly in baby hamster kidney, murine embryonic fibroblasts, colon cancer, and neuroblastoma cells (approx. 10−5 s/n/r). Cell immortalization through p53 inactivation and oxygen levels (1–21%) did not have a significant impact on viral replication fidelity. This shows that previously published mutation rates can be considered reliable despite being based on a narrow and artificial set of laboratory conditions. Interestingly, we also found that VSV mutated approximately four times more slowly in various insect cells compared with mammalian cells. This may contribute to explaining the relatively slow evolution of VSV and other arthropod-borne viruses in nature. RNA viruses show extremely high genetic variability and rapid evolution, ultimately due to their elevated rates of spontaneous mutation, which range from 10−6 to 10−4 substitutions per nucleotide per round of copying (s/n/r). However, mutation rate estimates vary considerably, even for the same virus [1], [2]. Since viral mutation rates have implications for pathogenesis [3], [4], vaccine development [5], [6] antiviral therapy [7], [8], and epidemiological disease management [9], [10], it is important to have accurate data and a clear understanding of the factors determining these rates. As a case in point, the risk of cross-species transmission is determined, in addition to the ecology of virus-host interactions, by the input of new adaptive mutations in the viral population [11], and a recent phylogenetic analysis of rabies virus isolates suggested that the waiting time required for host jumps depends on the number of positively selected mutations involved in cross-species transmission [12]. In RNA viruses, mutation rates are determined by the intrinsic base selection specificity of the viral polymerase [13]–[16], the presence/absence of proofreading mechanisms such as 3′exonuclease activity [17]–[19], or the mode of replication [20], [21]. However, in addition to these virus-encoded factors, viral mutation rates can be host-dependent. For instance, it has been suggested that the replicase of cucumber mosaic virus exhibits different fidelity in pepper and tobacco plants [22], [23]. In retroviruses, replication fidelity may be affected by intra-cellular dNTP imbalance and total concentration, which vary among cell types [24]–[26], although a recent study revealed no differences in the HIV-1 mutation rate in various cell types including T lymphoblast, glioblastoma and human embryonic kidney cells [27]. Also, the expression of host genes may influence the viral mutation rate as is the case of APOBEC3 cytidine deaminases, which can edit the HIV-1 cDNA and produce G-to-A hypermutations [28]–[30]. A similar role was postulated for the cellular RNA-dependent adenosine deaminase (ADAR) which could lead to A-to-G hypermutation in several RNA viruses, including rhabdoviruses [31], paramyxoviruses [32], and retroviruses [33]–[35]. Finally, cell metabolism may also have an impact in viral mutation rates, since it has been shown that ethanol-derived reactive oxygen species (ROS) can damage the RNA of hepatitis C virus, whereas other compounds such as glutathione and iron chelators were found to have the opposite effect [36]. Vesicular stomatitis virus (VSV) is a non-segmented negative-stranded RNA virus belonging to the family Rhabdoviridae with an extremely wide host tropism. The virion attaches to phosphadtidyl serine or other ubiquitous cell surface receptors and can productively infect most mammalian cells [37]. In nature, VSV infects a very large number of mammal species including livestock (cattle, horse, swine, goats, etc.) and wild animals (rodents, bear, lynx, bats, etc.), and also infects insects (sandflies, blackflies, mosquitoes, etc.) [38], [39], which act as transmission vectors [40]–[42]. Therefore, VSV replicates in widely different cellular environments, but the impact of this heterogeneity on the viral mutation rate is unknown. Actually, nearly all mutation rate estimates for animal viruses have been obtained in standard laboratory cell lines, which are usually immortalized or cancerous and thus show aberrant metabolic/mitotic rates and gene expression patterns. For VSV, most studies are conducted using hamster kidney cells, despite the fact that the brain is the main target organ of rhabdoviruses. Furthermore, all viral mutation rate studies have been conducted under atmospheric oxygen levels but these are substantially higher than those found in most tissues [43], and the impact of this type of environmental stress in the estimates is unknown. Here, we measured the mutation rate of VSV in primary and tumoral cell types including murine fibroblasts of various origins and neural cells, and under different oxygen levels, as well as insect cells. We found that the VSV mutation rate was relatively constant in all mammalian cells tested. However, VSV mutated four times more slowly in insect cells than in mammalian cells, a finding that may have implications for our understanding of arboviral evolution. We measured the mutation rate of VSV by the Luria-Delbrück fluctuation test, a standard estimation method [44] that has been used previously in several viruses including poliovirus [45], vesicular stomatitis virus [46], influenza A virus [47], measles virus [48], turnip mosaic virus [49], and bacteriophages φ6 [20] and Qβ [50]. To score mutants, we used a monoclonal antibody against the envelope glycoprotein G and determined the probability of appearance of monoclonal antibody resistance (MAR) mutants in independent cultures (null-class method). First, we performed six independent tests in baby hamster kidney cells (BHK-21), for which we had previous results [46]. This gave an average mutation rate to the MAR phenotype of m = (1. 64±0. 27) ×10−5 per round of copying (Table 1). This rate can be converted to per-nucleotide units as, where T is the set of observable mutations leading to the phenotype (mutation target) and three stands for the number of possible nucleotide substitutions per site [2]. Sequencing of the glycoprotein G gene from 15 MAR plaques allowed us to identify four different nucleotide substitutions, which led to amino acid changes D257N, D259A, D259N, and S273T, whereas previous work reported the same substitutions at position 259 of the G glycoprotein in addition D257G, D257V, D257Y and A263E [51]. Taking T = 8, the estimated mutation rate is μ = 6. 15×10−6 substitutions per nucleotide per round of copying (s/n/r). To verify the reliability of the above estimate, we used a molecular clone sequencing approach. This allowed us to score mutations more directly than in fluctuation tests and to analyze a wider genome region, although the interpretation of the data is complicated by the fact that the observed mutation frequency is dependent on selection, the number of generations elapsed, etc. BHK-21 cells were infected with a single infectious particle (i. e. plaque forming unit, pfu) by limiting dilution, and the resulting viral bursts (1. 55×107 final pfu on average) were used for RNA purification, RT-PCR, molecular cloning, and sequencing of three genome regions mapping to genes P, G, and L. We observed four single-nucleotide substitutions in 77500 bases in total, giving a mutation frequency of f = 5. 16×10−5 (Table 2). For a per-cell burst size of B = 1250 [46], the number of infection cycles (i. e. viral generations) elapsed should be. Therefore, the per-generation increase in mutation frequency was. To account for the effect of selection, we used the previously characterized distribution of mutational fitness effects (see Methods). Based on this, the expected fraction of observable mutations after 2. 3 generations was 53% and, thus, the estimated per-cell mutation rate is. The exact number of round of copying per cell is unknown but a previous work suggested rC = 5. 8 rounds/cell, implying that μ = 7. 30×10−6 s/n/r. This estimate is fully consistent with the results provided by the Luria-Delbrück fluctuation test. Subsequent experiments were done using fluctuation tests only because they provided a faster and simpler approach. Previous mutation rate studies with VSV have been conducted in BHK-21 cells only [2], [46], [52]. However, these are immortalized/tumoral cells, as opposed to those typically encountered bythe virus in vivo. Furthermore, VSV has a tropism for neural cells, and kidney fibroblasts are not a natural target of the virus. Toaddress the potential effect of immortalization on the viral mutation rate, we performed fluctuation tests in primary mouse embryonic fibroblasts (MEFs) and isogenic, p53 knock-out, MEFs. The average rate was similar in normal (m = 1. 27×10−5) and p53knock-out MEFs (m = 0. 82×10−5), revealing no significant effectof cellular immortalization (Figure 1; t-test: P = 0. 232, n = 6). However, many cell lines are tumoral and show other genetic andmetabolic alterations in addition to p53 inactivation. To check the potential effects of these changes, we performed fluctuation tests in CT26 cells from an undifferentiated grade IV colon adenocarcinoma of a BALB/c mouse [53], but we found no significant differences with primary MEFs (m = 1. 18×10−5; t-test: P = 0. 885, n = 6). Of note, BHK-21 are also tumor-forming cells, and the mutation rate was similar to the rate observed in MEFs or CT26 cells (one-way ANOVA: P = 0. 293, n = 12). This homogeneity in the VSV mutation rate was not an obvious a priori, because metabolic and mitotic activity should alter the availability of NTPs [54] and hence could impact RNA replication fidelity, although VSV replicates in the cytoplasm and may not be strongly affected by these alterations. This result has implications for the field of oncolytic virotherapy [55], since it is critical to assess the genetic stability of these therapeutic viruses during large-scale manufacturing and clinical use. In particular, CT26 cells have been used in mice as a model for testing the oncolytic activity of VSV [56]. Also, the above results suggest that VSV replicates with similar fidelity in different cell types, but we sought to test whether this would also hold for neural cells. We therefore performed fluctuation tests in Neuro-2a cells from a mouse neuroblastoma [57]. Again, we found that the average mutation rate did not significantly differ from the rate obtained in BHK-21 cells (m = 1. 06×10−5; t-test: P = 0. 461, n = 9). Finally, to test for other potential effects of cell physiology, we also varied oxygen levels. The VSV mutation rate in BHK-21 cells cultured under hypoxic conditions (1% oxygen) was slightly higher but not significantly different to the rate obtained under standard conditions (m = 2. 71×10−5; t-test: P = 0. 122, n = 9). Oxidative stress should lead to the release of ROS, which have been previously shown to be mutagenic for hepatitis C virus [36]. However, VSV does not appear to be sensitive to oxidation levels. This might be related to the fact that the nucleocapsid of mononegavirales forms a tunnel-like structure which wraps the viral genomic RNA and remains assembled during the entire infection cycle [58], [59], effectively isolating the viral RNA [60]. Since VSV alternates between mammalian and insect hosts in nature, we sought to measure the viral mutation rate in insect cells (Figure 1). In S2 cells from D. melanogaster embryos, the average estimate from three independent fluctuation tests was m = 4. 08×10−6, representing a fourfold decrease compared with BHK-21 (t-test: P = 0. 009, n = 9). To further investigate this, we selected two additional insect cells lines: Sf21 ovarian cells from the moth Spodoptera frugiperda, and C6/36 from Aedes albopictus mosquito larvae. Also, since insect cells were infected at 28°C and mammalian cells at 37°C, we performed four additional tests in BHK-21 at 28°C. We used estimates obtained in mammalian (BHK-21, BHK-21 at 28°C, MEF, MEF p53−/−, CT26, and Neuro-2a) and insect cells (S2, Sf-21, and C6/36) to jointly test for the effects of host type and temperature (fixed factors) in a two-way ANOVA in which the specific cell line was treated as a random factor nested within host type. This confirmed that VSV shows lower mutation rate in insect cells than in mammalian cells (ANOVA: P<0. 001), and also that temperature cannot account for this result because the estimates in BHK-21 were actually higher at 28°C than at 37°C (P = 0. 001). Using log10-transformed data, the estimated effect size of the host type in the above model was 0. 590±0. 205, which implies a 3. 9 fold mutation rate decrease in insect cells. One possible explanation for this difference is that our sensitivity to detect MAR mutants varied between assays performed in mammalian and insect cells. To address this, we first verified that MAR plating efficiency was similar in BHK-21, S2, Sf21, and C6/36 cells using a genetically engineered MAR mutant (D259A). Second, we tested for differences in the mutation target size (T). To do this, we sampled 15 individual MAR plaques from fluctuation tests performed in S2 cells and sequenced the region of the G protein controlling this phenotype. We found the same amino acid replacements as in fluctuation tests performed in BHK-21 cells (D257N, D259N, S273T, see above) except for D259A. However, because the D259 mutant is viable in insect cells [61], failure to detect it was probably due to insufficient sampling depth. We also found substitution A263E, which was reported previously in BHK-21 cells [51]. Therefore, insect S2 and BHK-21 cells shared a similar mutational repertoire and plating efficiency, supporting the consistency of the observed mutation rate difference. Interestingly, VSV [62] and arboviruses in general [63], [64] tend to evolve more slowly than directly transmitted viruses. Our own meta-analysis using 170 previously published evolutionary rates confirmed that, after accounting for phylogenetic relatedness and the timespan of sequence sampling, arboviruses showed a significantly lower evolution rate than directly transmitted viruses (Figure 2; two-way ANOVA: P = 0. 006), the geometric mean rates being 5. 7×10−4 substitutions per site per year (s/s/y) and 1. 3×10−3 s/s/y, respectively. This has been often interpreted in terms of fitness tradeoffs, whereby neutral or beneficial mutations in mammals can be deleterious in insects, and vice versa, thus restricting viral evolution. However, whether arboviruses show similar mutation rates in mammalian and insect cells has not been addressed before, and our results offer a new possible explanation for the relatively slow arboviral evolution. Future experiments with other arboviruses could help elucidate the generality of these findings and, if so, to delineate the mechanisms behind the observed differences in replication fidelity. Viruses were obtained from an infectious cDNA clone by transfecting baby hamster kidney (BHK-21) cells [65], [66], purified by filtration (0. 22 µm), and stored at 70°C in aliquots until use. BHK-21 cells (American Type Culture Collection, ATCC) were cultured in DMEM supplemented with 10% fetal bovine serum (FBS), 0. 02 mM L-Glutamine, a mix of non-essential amino-acids, 100 µg/mL streptomycin, 60 µg/mL penicillin, and 2 µg/mL fungizone. MEFs and their p53−/− derivatives were obtained from Dr. Carmen Rivas (Centro Nacional de Biotecnología, Madrid) and cultured in the same medium but with 12% FBS. Neuro-2a cells were obtained from Prof. José M. García-Verdugo (Department of Cell Biology, University of Valencia) and cultured in MEM supplemented with 2 mM L-Glutamine, 1 mM sodium pyruvate, 10% FBS, non-essential amino acids and the above antibiotics. CT26 cells (ATCC) were cultured in DMEM with 10% FBS, 2 mM L-glutamine, 10 mM HEPES and antibiotics. All the above cells were incubated at 37°C with 5% C02 and passaged upon confluence. D. melanogaster Schneider (S2) cells were obtained from Dr. Rubén Artero (Department of Genetics, University of Valencia) and cultured in Schneider' s medium supplemented with 10% FBS and antibiotics at 25°C in the absence of C02, and infected at 28°C. Sf21 cells were obtained from Dr. Salvador Herrero (Department of Genetics, University of Valencia) and were cultured in Grace' s insect medium supplemented with 10% FBS and antibiotics at 28°C in the absence of C02. C6/36 cells (ATCC) were cultured in DMEM supplemented with 10% FBS, 2 mM L-glutamine, non-essential amino acids, 1 mM sodium pyruvate and antibiotics at 28°C under 5% C02. Hypoxia was achieved by displacing oxygen with nitrogen, using a Galaxy 170R incubator (Eppendorf). We inoculated 32 identical cultures each containing 104 confluent cells with approximately 300 pfu/well (Ni) and incubated them until approximately 3×104 pfu/well were produced (Nf). After a round of freeze-thawing to release intracellular particles, we used eight cultures for titration and 24 for plating the entire undiluted volume (100 µL) in the presence of a monoclonal antibody against the surface glycoprotein G at a concentration that neutralizes completely the wild-type virus and selects for MAR mutants. The antibody, in the form of a hybridoma supernatant, was added to the plating medium (25% v∶v) to avoid phenotypic masking [52]. Plating assays were done in DMEM gelled with 0. 4% agarose containing 2% FBS. After 24 h, monolayers were fixed with 10% formaldehyde and stained with 2% crystal violet to visualize plaques. Since mutation is a rare event, the number of mutations per culture is expected to follow a Poisson distribution of parameter and therefore the probability of observing no mutants in a culture is, where m is the mutation rate from the wild-type to the MAR phenotype (null-class method). However, if there is incomplete plating, some cultures may contain undetected MAR mutants. If we define z as the plating efficiency (relative to BHK-21 cells), the probability of observing no mutants can be expressed as, where is the probability of k actual mutants in a culture. Using a Poisson distribution of parameter for k, we numerically solved Q0 given P0, Ni, Nf, and z and calculated the mutation rate as. For each cell type tested, the plaque assay for scoring MAR mutants was done in BHK-21 cells for technical feasibility and to control for differences in plating efficiency among cells. However, since plaque assays to score MAR mutants were done without dilution, antiviral cytokines or other compounds released from the cells in which the virus was grown could modify plating efficiency (plaque assays for determining Nf were done at a roughly 1/100 dilution and thus were much less affected by this problem). For instance, BHK-21 cells are at least partially responsive to interferon [67], potentially inhibiting growth of MAR mutants and biasing mutant counts down. To calibrate this effect, we titrated a MAR clone obtained by site-directed mutagenesis (substitution D259A in the surface glycoprotein G) in the presence of undiluted supernatants harvested from cells previously infected with the wild-type virus (Ni≈300 pfu and Nf>≈104 pfu, similar to fluctuation tests), adding monoclonal antibody to the plates to observe MAR plaques only. The wild-type infections were performed under each of the experimental conditions (BHK-21, MEF, MEF p53−/−, CT26, Neuro-2a, BHK-21 with 1% O2, BHK-21 at 28°C, S2, Sf21 and C6/36 cells). Addition of supernatants from BHK-21 cells infected under standard conditions did not alter the titer of the D259 MAR clone, hence the relative plating efficiency was z = 1. The relative plating efficiency for each of the other conditions is shown in the Supporting Information “Text S1” and was based on at least six independent plating assays. To ascertain the number of possible mutations conferring the MAR phenotype, we plated approximately 105 pfu in the presence of antibody, incubated them for 24 h, and pipetted individual plaques. Viral RNA was purified, reverse-transcribed using AccuScript High Fidelity Reverse Transcripatse (Agilent Technologies), and the cDNA was PCR-amplified using Phusion High Fidelity DNA polymerase (New England Biolabs). We used specific primers to amplify and sequence a region of the G protein (genome sites 3361 to 4501 in GenBank accession EF197793) which controls the MAR phenotype [51]. PCR products were sequenced by the Sanger method and analyzed using Staden software. A 96-well plate containing 104 cells per well was inoculated with a limiting dilution of the viral stock such that approximately 10% of wells were infected. Plates were incubated at 37°C for 24 h, inspected under the microscope for cytopathic effects, and freeze-thawed to allow release of intracellular viruses. Viral RNA was purified from the supernatant of each of five positive wells and reverse-transcribed using AccuScript High Fidelity Reverse Transcripatse, and the cDNA was PCR-amplified using Phusion High Fidelity DNA polymerase and specific primers located in the P, G and L genes, as indicated. PCR products were cloned and used for E. coli transformation, and 10 colonies were picked and amplified by colony PCR using Phusion High Fidelity DNA polymerase. PCR products were sequenced by the Sanger method and analyzed using Staden software. To obtain the mutation frequency, the number of observed mutations was divided by the total number of bases sequenced. We used the empirically characterized distribution of mutational fitness effects of random single-nucleotide substitutions in VSV to correct for the effect of selection on mutation frequency and obtain the mutation rate per cell infection. We did so numerically by simulating the combined effects of mutation and selection. The statistical distribution of fitness effects (s) for viable substitutions can be roughly captured using an exponential distribution truncated at (lethality) plus a class of lethals occurring with probability pL: if 0<s<1, if s = 1, and otherwise. In a previous work using the same VSV strain as here, it was estimated that and that [65], [68]. Fitness effects were measured as growth rate ratios, , where r is the exponential growth rate and subscripts i and 0 refer to the mutant and wild-type, respectively [65]. These s-values were transformed to per cell infection units as, where B is the burst size. After simulating fitness effects using the truncated exponential plus lethal distribution and applying the per cell infection transformation, selection was applied by picking individuals for the next cell infection cycle with weighted probability 1−s′, and the process was iterated. This provided an expected mutation frequency f and therefore a relationship between μ and f. Genetic drift was ignored since it should not modify the expected value of f. Also, for simplicity, mutations were assumed to have independent fitness effects (no epistasis) and back mutations were ignored, which seems reasonable in the short-term, because single forward mutations will greatly outnumber secondary and back mutations. Simulations were performed using Wolfram Mathematica and Excel. A graphical representation of this correction can be found in a previous work [2]. In a previous meta-analysis, we collected evolutionary rate estimates that were originally inferred from field isolates using Bayesian analysis of dated sequences after validation of the molecular clock [69]. Here, we used 170 of these estimates, which corresponded to 62 different riboviruses. We sought to compare viruses transmitted directly through respiratory secretions, blood, sexual contact, feces, or animal bites (n = 113) against arboviruses (n = 57). We used a two-way ANOVA in which the following factors were included: transmission mode (fixed), viral family (random) to account for phylogenetic relatedness, and sampling timespan (covariate) to account for the known time-dependency of evolution rate estimates. Since rates ranged several orders of magnitude log-transformed data were used.

RNA viruses show high rates of spontaneous mutation, a feature that profoundly influences viral evolution, disease emergence, the appearance of drug resistances, and vaccine efficacy. However, RNA virus mutation rates vary substantially and the factors determining this variability remain poorly understood. Here, we investigated the effects of host factors on viral replication fidelity by measuring the viral mutation rate in different cell types and under various culturing conditions. To carry out these experiments we chose the vesicular stomatitis virus (VSV), an insect-transmitted mammalian RNA virus with an extremely wide cellular and host tropism. We found that the VSV replication machinery was robust to changes in cellular physiology driven by cell immortalization or shifts in temperature and oxygen levels. In contrast, VSV mutated significantly more slowly in insect cells than in mammalian cells, a finding may help us to understand why arthropod-borne viruses tend to evolve more slowly than directly transmitted viruses in nature.

Abstract Introduction Results/Discussion Materials and Methods

microbial evolution virology population genetics biology microbiology evolutionary biology

2014

Variation in RNA Virus Mutation Rates across Host Cells

The mean age of dengue has been increasing in some but not all countries. We sought to determine the incidence of dengue virus (DENV) infection in adults and children in a prospective cohort study in the Philippines where dengue is hyperendemic. A prospective cohort of subjects ≥6 months old in Cebu City, Philippines, underwent active community-based surveillance for acute febrile illnesses by weekly contact. Fever history within the prior seven days was evaluated with an acute illness visit followed by 2,5, and 8-day, and 3-week convalescent visits. Blood was collected at the acute and 3-week visits. Scheduled visits took place at enrolment and 12 months that included blood collections. Acute samples were tested by DENV PCR and acute/convalescent samples by DENV IgM/IgG ELISA to identify symptomatic infections. Enrolment and 12-month samples were tested by DENV hemagglutination inhibition (HAI) assay to identify subclinical infections. Of 1,008 enrolled subjects, 854 completed all study activities at 12 months per-protocol undergoing 868 person-years of surveillance. The incidence of symptomatic and subclinical infections was 1. 62 and 7. 03 per 100 person-years, respectively. However, in subjects >15 years old, only one symptomatic infection occurred whereas 27 subclinical infections were identified. DENV HAI seroprevalence increased sharply with age with baseline multitypic HAIs associated with fewer symptomatic infections. Using a catalytic model, the historical infection rate among dengue naïve individuals was estimated to be high at 11–22%/year. In this hyperendemic area with high seroprevalence of multitypic DENV HAIs in adults, symptomatic dengue rarely occurred in individuals older than 15 years. Our findings demonstrate that dengue is primarily a pediatric disease in areas with high force of infection. However, the average age of dengue could increase if force of infection decreases over time, as is occurring in some hyperendemic countries such as Thailand. Dengue virus (DENV) is the leading cause of vector-borne viral disease globally with an estimated 390 million infections and 96 million symptomatic cases occurring annually [1]. Ecologic and demographic changes are thought to be major contributing factors to the emergence of dengue over the past few decades [2]. DENV infection can present as asymptomatic infection, subclinical infection, undifferentiated fever, dengue fever, dengue hemorrhagic fever (DHF) with or without dengue shock syndrome (DSS), and other severe forms of dengue [3]. Estimates of the proportion of subclinical infections among all DENV infections have ranged widely depending on the year, location, population, and surveillance method [4]. Prospective longitudinal cohort studies can characterize the burden and clinical spectrum of DENV infections including subclinical infections [5]. Multiple exposures to DENV are generally presumed to result in protective immunity leading to lower incidence of clinically overt disease in adults in areas with hyperendemic transmission [6,7]. However, when dengue does occur in adults, the clinical manifestations may be more apparent than in children, perhaps due to differing physiologies and more frequent co-morbid conditions in adults [8–11]. At the same time, some dengue hyperendemic countries have reported an increase in the mean age of dengue [12–15]. An important contributing factor to this age increase may be demographic transition in some countries where decreasing birth and death rates may lead to decreasing force of infection (FOI) due to fewer susceptible individuals entering the population [16,17]. Yet, very few prospective longitudinal cohort studies undergoing active surveillance have been conducted in adults to assess overall incidence and disease burden [18,19]; and even fewer have evaluated dengue incidence and relative proportion of subclinical infections in adults and children within the same cohort. In the Philippines, a dengue outbreak was first reported as early as 1906 [20], and the first epidemic of severe dengue was documented in Manila in 1953 [21]. Since then, dengue has been hyperendemic in most areas of the country with a general increase in the number of dengue cases over time [22]. We conducted a prospective longitudinal cohort study in Cebu City, Philippines, among subjects of all ages ≥6 months, with the objective of determining the age-stratified incidence of DENV infection and the ratio of symptomatic to subclinical infection in adults and children. We found that in a setting with high force of infection with consequently high DENV seroprevalence in adults, symptomatic infections were rare in individuals older than 15 years with almost all infections in this age group being subclinical. The study was approved by the Institutional Review Boards of Vicente Sotto Memorial Medical Center (VSMMC) in Cebu City, Philippines, and the Walter Reed Army Institute of Research (WRAIR). Written informed consent was obtained from subjects ≥18 years old and from parents of subjects <18 years old. Written assent was obtained from children ≥12 years old. The study was conducted in the barangay (i. e. , village equivalent) of Punta Princesa, an urban community in Cebu City, capital of Cebu province located within the Visayas region of the Philippines. Punta Princesa encompasses 0. 96 sq km with a total population of about 27,000 people. Public outpatient care is available to local residents at the Punta Princesa Health Center with patients who require higher level medical care or hospitalization referred to Cebu City Medical Center, the tertiary public hospital of Cebu City, or to VSMMC, the tertiary public hospital of Cebu province. Study enrolment and surveillance procedures have been previously described [23]. Subjects were enrolled from March to May 2012 in a prospective longitudinal seroepidemiological cohort study using convenience sampling within the community. Inclusion criteria included age ≥6 months and residence within Punta Princesa. Exclusion criteria included known active pulmonary tuberculosis, in order to decrease risk to study staff. Enrolment was limited to one subject per household. Approximately 1,000 subjects were targeted for enrolment with roughly equal distribution among five different age groups: 6 months-5 years, 6–15 years, 16–30 years, 31–50 years, and >50 years old. At enrolment and at 12 months, subjects were administered demographic and health questionnaires, and underwent blood collections. Blood was processed into serum aliquots within 24 hours of phlebotomy, frozen at approximately -70°C, and eventually shipped on dry ice to the Armed Forces Research Institute of Medical Sciences (AFRIMS) in Bangkok, Thailand for testing. The serum was tested by hemagglutination inhibition (HAI) assay for all four DENV serotypes and Japanese encephalitis virus (JEV). Enrolled subjects underwent active community-based surveillance to detect acute febrile illness which was defined as reported fever or measured temperature >38. 0°C. Subjects were instructed to report any fevers, and were contacted by study staff once a week through telephone, short message service, and/or home visits. A history of fever within the prior seven days triggered an acute fever investigation consisting of an acute illness visit followed by subsequent visits at 2,5 and 8 days, and a convalescent visit at 3 weeks. Clinical assessments were performed at all visits, and blood was collected at the acute and 3-week convalescent visits. Blood was processed into serum aliquots within 24 hours of phlebotomy, frozen at approximately -70°C and shipped on dry ice to AFRIMS in Bangkok, Thailand. Acute serum was tested by hemi-nested reverse transcriptase polymerase chain reaction (PCR) to detect DENV RNA. Paired acute/convalescent sera were tested by an in-house DENV IgM/IgG capture enzyme-linked immunosorbent assay (ELISA) and by DENV/JEV HAI. Acute symptomatic DENV infection was defined as an acute febrile illness with positive DENV PCR in the acute sample, and/or positive DENV IgM/IgG ELISA in the paired acute/convalescent samples. Subclinical DENV infection was defined as a ≥four-fold rise in DENV HAI titer for any serotype in paired enrolment/12-month samples with no acute symptomatic DENV infection detected during the intervening surveillance period. Subjects with an HAI titer <10 to all four serotypes were considered to be dengue naive. Those with an HAI titer ≥10 to only one DENV serotype were considered to have monotypic dengue immune status. Those with HAI titers ≥10 to two or more serotypes were considered to have multitypic immune status. Descriptive statistics for infection rates, symptoms and other characteristics were performed. To inform the underlying infection dynamics in this population, we employed the chi-square or Fisher’s exact test as appropriate to evaluate the association between DENV infection and both age >15 years and the nominal categorical baseline immune status. To estimate FOI, age-stratified DENV HAI seroprevalence data at enrolment was analyzed such that HAI titer <10 for all four DENV serotypes indicated a dengue naive subject while HAI titer ≥10 for at least one DENV serotype indicated a non-naive subject. A catalytic model [16,27] was used to estimate FOI (λ) for each DENV HAI serotype assuming independence between serotypes (details available in supplementary information). In this catalytic model, (1—exp [-λ]) was the proportion of the dengue naive population infected each year by each serotype. FOI was estimated in a Markov Chain Monte Carlo (MCMC) framework. Means and 95% credible intervals were reported. When using seroprevalence at a single time point, age and time could have been confounded so that changes in FOI during the last five years, for example, could have been due either to differential exposure in individuals <5 years old or to different FOI in the last five years. Two models with different assumptions were assessed; one in which FOI was assumed to be constant over all years (model 1), and one in which FOI could be different during the previous five years (or equivalently, among individuals <5 years old) from the years before then (model 2). For each model, an average FOI was estimated across all serotypes. Ages were rounded down so those <1 year old were shown as 0. All analyses were performed using R version 3. 0. 2 (R Foundation for Statistical Computing, Vienna, Austria). A total of 1,008 subjects were enrolled from March to May 2012. A description of the enrolled cohort is presented in Table 1. During 985 person-years of active surveillance from enrolment to 12 months, 274 acute febrile illnesses were detected with 268 acute and 261 convalescent blood samples collected (Fig 1). Of 50,125 weekly contacts for the 1008 enrolled subjects, only 1. 2% were unsuccessful due to the subject being “unavailable. ” These weekly contacts were completed by household visit (22. 4%), SMS (37. 6%) or telephone call (38. 8%). Sixteen acute symptomatic DENV infections were identified during this surveillance period. Thirteen of the 16 symptomatic infections were PCR-positive with serotype data available: 10 DENV-1, two DENV-2 and one DENV-3. A total of 854 subjects completed all study activities by the 12-month visit constituting the “per-protocol” subjects. Of the 154 subjects who did not complete all study activities, 66 relocated out of the study area, 64 withdrew consent, 16 were lost to follow up, and eight developed other physical conditions or problems. Subjects in the youngest age group had the largest number who did not complete all study activities (Table 1). Subclinical infections were only possible for per-protocol subjects since only this group had DENV HAI testing performed at 12 months. During 868 person-years of active surveillance in just the per-protocol subjects, 13 acute symptomatic DENV infections and 61 subclinical infections occurred. The incidence of symptomatic DENV infection per 100 person-years in the overall cohort (including non-per-protocol subjects) was 1. 62 (95%CI: 0. 97,2. 58). Only one symptomatic infection occurred in subjects >15 years of age. The incidence of subclinical infection per 100 person-years in per-protocol subjects was 7. 03 (95%CI: 5. 42,8. 96), ranging from 4. 16 in the 31–50 year group, to 10. 04 in the 6–15 year group (Table 2). Three subjects were hospitalized for DHF, and five subjects received outpatient medical care (S1 Table), and no fatalities occurred. DENV HAI seroprevalence increased sharply with age (Fig 2). The proportion of each age group that had a negative HAI profile at enrolment decreased from 56. 1% in 6 month-5 year olds to 0. 6% in 31. 50 year olds, while the proportion having multitypic HAIs increased from 39. 9% in 6 month-5 year olds to 98. 8% in 31–50 year olds. The seroprevalence of multitypic HAIs was >98. 3% for all ages >15 years. The proportion of subjects who developed symptomatic DENV infection differed depending on baseline DENV HAI profile (Fig 3). Symptomatic infections occurred more often with monotypic (2 of 19) and negative (4 of 125) HAIs at enrolment, and less often with multitypic (7 of 710) HAIs (chi-square, p<0. 001). In addition, the proportion of symptomatic infections among all DENV infections (i. e. , symptomatic and subclinical) tended to differ depending on the HAI profile at enrolment (chi-square, p = 0. 086) (Fig 4). FOI was estimated by analyzing age-stratified DENV HAI seroprevalence at enrolment. Model 1, which assumed a constant FOI, yielded an average FOI of 0. 044/year (95% CI: 0. 039,0. 049) for each HAI serotype over the previous 20 years, with a log likelihood of -272 (nearly universal HAI seropositivity in adults older than 20 years did not allow FOI estimates beyond the previous 20 years). The model fit is shown in dark blue in Fig 5 with uncertainty shown in light blue. Model 2, which assumed non-constant FOI, yielded an average FOI of 0. 061/year per serotype (95% CI: 0. 051,0. 072) during the previous five years and 0. 028/year per serotype (95% CI: 0. 020,0. 036) for the 15 years before then, with a log likelihood of -265. The model fit is shown in red in Fig 5, with uncertainty shown in pink. Model 2 fit the data slightly better, particularly in the younger age groups, suggesting either that there was a higher FOI in the previous five years, or that subjects ≤5 years of age experienced a higher FOI than older subjects. To assess whether different age groups may have had differing FOIs, DENV HAI seroconversions at 12 months in subjects who were dengue naive at enrolment were analyzed by age group (Fig 6). The age given in Fig 6 is the age at the start of the year in which seroconversions were considered. Because the number of dengue naive subjects was small (n = 125), there were large confidence intervals for these proportions. Nevertheless, there was no evidence of differential seroconversion in subjects ≤5 years old than those >5 years, suggesting that the observed results from model 2 were due to FOI being higher during the previous five years than before then, rather than to differential exposure with age. In our study in a hyperendemic area of the Philippines, symptomatic DENV infections rarely occurred in subjects older than 15 years. Infections that did occur in this older age group were more likely to be subclinical than in those ≤15 years. This age pattern of DENV infection occurred in the setting of high seroprevalence of multitypic DENV HAIs in adults likely due to high force of infection. This is one of the first prospective longitudinal cohort studies using active surveillance in which dengue incidence in adults and children have been determined within the same cohort. The historical force of infection based on age-stratified seroprevalence at enrolment was found to be high in our study. Presuming a constant FOI, the infection rate among dengue naïve individuals was approximately 16% per year over the previous 20 years (based on FOI of 0. 044/year per serotype). Presuming a non-constant FOI, the infection rate was about 22% per year during the previous five years (FOI of 0. 061/year per serotype) and 11% per year before then (FOI of 0. 028/year per serotype). Whichever of these two models was used, the results indicate that sustained DENV transmission has been ongoing in Cebu for many years. By the time subjects reached 16 years of age, over 96% had multitypic DENV HAIs. The high population turnover in the Philippines leading to large numbers of dengue naïve individuals entering the population may be a contributing factor to the high FOI. In 2014, the birth and death rates in the Philippines were 25. 34 and 5. 02 per 1,000 population, respectively, as compared to 12. 95 and 7. 29 in Thailand (World Health Statistics 2014, World Health Organization). Nevertheless, other factors may also be playing a role in the Philippines, such as permissive environmental/ecological factors and/or incomplete vector control. The predominant viral serotype and strain during different years may have also affected the epidemic potential and, thus, FOI. The current epidemiology of dengue in the Philippines may mirror past epidemiology in other countries such as Thailand before demographic transition occurred. We were unable to perform a sub-analysis of adult symptomatic dengue because so few symptomatic infections in adults occurred. This lack of cases existed in the setting of high seroprevalence of multitypic HAIs in adults. So while the independent effects of age and HAI profile on symptomatic dengue could not be evaluated, the lack of symptomatic infections in adults was likely related to the high seroprevalence of multitypic HAIs, which in turn was probably due to high historical FOI. It is notable that so few symptomatic infections occurred in subjects >15 years old despite a vigorous active surveillance system. The incidence of symptomatic dengue in children in our study was comparable to previous prospective cohort studies conducted in children in which the average rate has ranged from 0. 6% to 3. 9% per year [28], supporting the validity of our surveillance procedures for detecting dengue cases. Our findings demonstrate the primarily pediatric focus of dengue when FOI is high, and why dengue has been considered a childhood disease in many hyperendemic countries. Our study does not, however, clarify whether the clinical manifestations of dengue are more or less severe in adults since so few symptomatic adult cases were observed. Symptomatic DENV infections occurred more frequently with baseline negative or monotypic HAIs than with multitypic HAIs. A multitypic HAI profile can result from two or more past DENV infections by different serotypes, although a single infection is still possible. If a monotypic HAI profile was presumed to have resulted from a single past infection, then our findings suggest that a second infection with a different serotype may be more likely to cause symptomatic infection. There were several limitations to our study. First, only one year of surveillance data was available. Given the known cyclical nature of DENV transmission, other years may have different epidemiological and clinical patterns. For example, different predominating DENV serotypes or strains in the context of different population immunity in different years may lead to different ratios of symptomatic and subclinical disease [29]. Second, relatively few symptomatic infections occurred in the overall cohort limiting the statistical significance of our analyses. Third, the use of DENV HAI to determine both seroprevalence and seroconversion may have underestimated the actual rates given the unclear longevity of HAI titers. Measurement of more persistent markers of past infection would have been preferable, but were not performed due to resource limitations. It is, therefore, possible that the incidence of total DENV infections was actually higher than indicated by our study results. Nevertheless, the lack of symptomatic infections observed in adults would not have been affected. In summary, our study demonstrates that symptomatic dengue is primarily a pediatric disease in hyperendemic areas with high force of infection. However, the average age of dengue could increase if force of infection decreases over time, which may be occurring in some dengue hyperendemic countries such as Thailand.

The average age of dengue has been increasing in some but not all dengue endemic countries. To investigate the age pattern of dengue in people of all ages ≥6 months old, a prospective community-based cohort study was undertaken in Cebu City, Philippines where dengue virus has been circulating for many decades. Active surveillance for acute fevers was performed, and acute/convalescent blood samples were tested for evidence of symptomatic dengue. Blood was also collected at enrolment and one year later, and tested serologically to identify subclinical infections. Overall, 1. 62 symptomatic and 7. 03 subclinical infections per 100 person-years of surveillance were detected. Among people older than 15 years, only one symptomatic dengue case occurred while 27 subclinical infections were identified. By analyzing age-specific dengue serology data, the historical infection rate among people with no prior dengue virus infection was found to be high at around 11–22% per year. Our results show that dengue is primarily a childhood disease in endemic settings where the historical infection rate has been high. However, the average age of dengue could increase if the infection rate decreases over time as is happening in some endemic countries like Thailand.

Abstract Introduction Methods Results Discussion

dengue virus children medicine and health sciences body fluids pathology and laboratory medicine pathogens microbiology pediatrics viruses age groups research design adults rna viruses signs and symptoms cohort studies pediatric infections families research and analysis methods medical microbiology microbial pathogens hematology people and places blood anatomy flaviviruses fevers viral pathogens physiology population groupings biology and life sciences organisms

2016

Incidence of Dengue Virus Infection in Adults and Children in a Prospective Longitudinal Cohort in the Philippines

Hemorrhagic fever with renal syndrome (HFRS) and hantavirus pulmonary syndrome (HPS) are diseases caused by hantavirus infections and are characterized by vascular leakage due to alterations of the endothelial barrier. Hantavirus-infected endothelial cells (EC) display no overt cytopathology; consequently, pathogenesis models have focused either on the influx of immune cells and release of cytokines or on increased degradation of the adherens junction protein, vascular endothelial (VE) -cadherin, due to hantavirus-mediated hypersensitization of EC to vascular endothelial growth factor (VEGF). To examine endothelial leakage in a relevant in vitro system, we co-cultured endothelial and vascular smooth muscle cells (vSMC) to generate capillary blood vessel-like structures. In contrast to results obtained in monolayers of cultured EC, we found that despite viral replication in both cell types as well as the presence of VEGF, infected in vitro vessels neither lost integrity nor displayed evidence of VE-cadherin degradation. Here, we present evidence for a novel mechanism of hantavirus-induced vascular leakage involving activation of the plasma kallikrein-kinin system (KKS). We show that incubation of factor XII (FXII), prekallikrein (PK), and high molecular weight kininogen (HK) plasma proteins with hantavirus-infected EC results in increased cleavage of HK, higher enzymatic activities of FXIIa/kallikrein (KAL) and increased liberation of bradykinin (BK). Measuring cell permeability in real-time using electric cell-substrate impedance sensing (ECIS), we identified dramatic increases in endothelial cell permeability after KKS activation and liberation of BK. Furthermore, the alterations in permeability could be prevented using inhibitors that directly block BK binding, the activity of FXIIa, or the activity of KAL. Lastly, FXII binding and autoactivation is increased on the surface of hantavirus-infected EC. These data are the first to demonstrate KKS activation during hantavirus infection and could have profound implications for treatment of hantavirus infections. The Bunyaviridae family encompasses viruses that cause numerous hemorrhagic fever diseases in humans. The genus Hantavirus includes Old World and New World viral lineages. Old World hantaviruses are widespread throughout Asia and Europe and are associated with the clinical syndrome, hemorrhagic fever with renal syndrome (HFRS). The prototype hantavirus, Hantaan virus (HTNV), can cause severe HFRS with a case fatality rate as high as 15% [1], [2]. The New World hantaviruses are the causative agents of hantavirus pulmonary syndrome (HPS) and are found in the Americas [1], [2]. The case fatality rate for HPS is greater than that of HFRS and has been reported to be as high as 50% for Andes virus (ANDV) [1]. While HFRS and HPS differ in the organs exhibiting pathogenic consequences; i. e. , kidneys for HFRS and lungs for HPS, both diseases primarily affect blood vessels and cause systemic vascular leakage that can lead to hypotension and shock [1]–[3]. In vivo and in vitro studies have identified EC as a primary site of viral replication, although hantaviruses can infect epithelial and vascular smooth muscle cells (vSMC) as well [4], [5]. Importantly, despite high levels of viral antigens, the capillary endothelium displays no obvious cytopathology [5], [6]. The mechanism by which hantaviruses cause pronounced vascular leakage when the lining of the endothelium is still intact has remained elusive. It has been assumed that during viral infection, since EC are not damaged, there must be some alteration to the infected cells directly or indirectly through immune mediated processes that result in vascular leakage. One hypothesis implicates the indirect effects of cytokines released from immune cells such as monocytes or T cells. Support for this hypothesis stems from several clinical studies showing that hantavirus-infected patients develop high levels of T cells and cytokine-producing cells, which have been correlated with disease severity (reviewed in [7]). However, in laboratory studies, depletion of T cells was found to have no effect on the outcome of disease in a hamster model that accurately reflects HPS [8]. A second hypothesis centers around observations that vascular endothelial growth factor (VEGF) levels are elevated in HPS and HFRS patients, and that VEGF could affect vascular permeability by increasing the degradation of VE-cadherin [9]–[12]. VE-cadherin is an endothelial specific adhesion molecule that is critical for maintaining cell contact integrity and for regulating vascular permeability [13]. However, in the lethal hamster model of HPS, an upregulation of VEGF was not observed in plasma of ANDV-infected hamsters [14]. In the studies we report here, we developed and tested an alternative hypothesis; i. e. , that permeability changes in hantavirus-infected EC are due to increased kallikrein-kinin system (KKS) activation and the liberation of bradykinin (BK). The KKS consists of serine proteinases; factor XII (FXII) and prekallikrein (PK), and the cofactor, high molecular weight kininogen (HK) [15]. Initiation of this pathway can occur via two mechanisms. The first is dependent upon autoactivation of FXII. FXII can bind to artificial and biological surfaces to induce a conformational change and activation to FXIIa. FXIIa mediates the proteolytic cleavage and activation of PK to kallikrein (KAL) and subsequent cleavage of HK. Alternatively, activation of this cascade can occur independently of FXII. Polycarboxylpeptidase (PRCP) and heat shock protein 90 (HSP90) have been shown to be activators of PK and mediate conversion of PK to KAL and cleavage of HK [16], [17]. However, if FXII is present after initial KAL formation, both enzymes can reciprocally activate causing an amplification of the cascade. In both circumstances the outcome is cleavage of HK resulting in release of a 9 amino acid peptide, BK [18]. BK is an extremely potent inflammatory peptide that exerts numerous effects on the vasculature including vasodilation and increased vascular permeability [19]. BK binds to the bradykinin B2 receptor (BKB2R) and triggers an increase in intracellular Ca2+, endothelium-derived hyperpolarizing factor (EDHF), prostacyclin (PGI2), and nitric oxide (NO) production, eliciting relaxation of vSMC and a drop in blood pressure [20]. Vascular permeability is also increased allowing protein and fluid extravasation across the endothelium via endothelial gaps or trans-endothelial transport [21]. BK has been implicated in the pathogenesis of numerous diseases and disorders including sepsis, pancreatitis, asthma, and most notably the genetic disorder, hereditary angioedema (HAE) [22]–[25]. HAE patients are deficient in C1 esterase inhibitor (C1-INH), an important regulator of activated FXII, which leads to excessive BK generation resulting in edema. Several therapeutics that block BK formation and BK binding have been developed and approved for human use to treat this disorder. The similar disease parameters seen in HAE, HFRS, and HPS led us to examine the importance of BK in the pathogenesis of hantavirus infections. Taken together, our in vitro findings of increased KKS activation, clinical data demonstrating activation of PK, and the reported successful treatment of patient with severe HFRS due to Puumala virus infection with a BK antagonist [26] implicates BK in the vascular leakage associated with pathogenic HFRS hantavirus infections. In vitro organotypic blood vessel systems are becoming more common for studying the processes associated with angiogenesis. Work by Evenson et al. resulted in such a system in which co-culture of human umbilical vein EC (HUVEC) with human mesenchymal stem cells (hMSC) or human pulmonary artery SMC (PaSMC) produced spontaneous formation of capillary-like blood vessels [27]. The capillary-like blood vessels underwent phenotypic changes that are commonly observed in vivo such as stabilization of endothelial capillary tubes by PaSMC or hMSC-differentiated cells, formation of adherens junctions, and the presence of basement membrane like structures [27]. We adapted this system as a platform for studying hantavirus pathogenesis. To generate and then verify the formation of blood vessel-like structures in vitro, we co-cultured HUVEC with PaSMC or hMSC on silicon chips and processed them for scanning electron microscopy. Upon ultrastructural examination, long and branched EC tubes were visualized throughout the samples (Figure 1A and B). Tubes varied in size with the largest tubes measuring approximately 12 µm in diameter (Figure 1A and B). The tubes were presumably stabilized and wrapped by additional cells that consisted of PaSMC or hMSC-differentiated cells. A matrix morphologically consistent with collagen was observed supporting the capillaries. To more specifically identify cell types, we fixed and immunostained HUVEC and PaSMC or hMSC co-cultures with antibodies specific for von Willebrand' s factor (vWF) or smooth muscle actin (SMA) to identify EC or SMC and collagen XVIII, a vascular basement membrane protein that is normally localized to the peri-endothelial cell region and is indicative of vessel maturity (Figure 1C and D). Numerous EC capillary tubes could be observed in both co-culture samples (Figure 1C and D). Taken together these data demonstrate that we were able to generate an in vitro capillary system that could be used as a novel model for studying multicellular involvement during hantavirus pathogenesis. Most in vitro studies of hantavirus pathogenesis have been performed by infecting EC alone, and although hantaviruses do predominantly infect this cell type, there are also reports of viral antigen within the perivascular SMC of infected patients [4], [5]. The in vitro capillary blood vessel model contains EC and SMC and recapitulates the natural site of infection during hantavirus diseases, so we sought to determine if pathogenic hantaviruses infect both cell types in this model. We found that after infection with HTNV or ANDV, we could detect viral nucleocapsid proteins (N) in the EC capillary tubes (Figure 2A, arrows). Surprisingly, we discovered that HTNV and ANDV antigens were detected in cells not expressing vWF, suggesting that they were also able to infect PaSMC in the capillary blood vessel system (Figure 2A, arrowheads). Similarly, hantavirus N staining patterns were also observed in HUVEC and hMSC co-cultures (data not shown). For subsequent experiments, which compared individual cell types and co-cultures, we used only capillaries formed with HUVEC and PaSMC because PaSMC do not require differentiation. To determine if vSMC infection was unique to the capillary blood vessel system or if the vSMC alone could be infected, PaSMC were infected with HTNV or ANDV, fixed, and stained with antibodies specific for SMA and hantavirus N. Both HTNV and ANDV N were detected within PaSMC (Figure 2B) indicating that a co-culture system is not required for infection of vSMC. To determine if viral replication was occurring, we also performed a time course infection experiment over 4 days. PaSMC were mock-infected or infected with HTNV or ANDV and whole cell lysates were harvested on 1,2, 3, and 4 days post-infection and western blots performed to measure the viral N proteins. Increases in N were apparent for both hantaviruses over the 4 days, demonstrating that viral infection and replication in PaSMC were occurring (Figure 2C). VEGF potently mediates vascular permeability through a complex signaling process in which binding to the receptor VEGF-R2 results in loss of the endothelial barrier due to internalization and degradation of VE-cadherin [13]. VEGF has been proposed to mediate hantavirus pathogenesis, as inferred from studies in which adding VEGF to hantavirus-infected HUVEC resulted in hyperphosphorylation of VEGF-R2 and internalization and degradation of VE-cadherin [10]. Another study, however, showed that infection of primary human pulmonary microvascular EC (HMVEC-L) with ANDV alone can direct degradation of VE-cadherin [12]. To further explore this possible mechanism of hantavirus pathogenesis, we tested the effects of HTNV and ANDV infections on VE-cadherin degradation in the in vitro capillary blood vessel model. This model should more faithfully represent in vivo conditions than EC alone because it contains both cell types normally present in human blood vessels and secretes VEGF [27]. HUVEC and PaSMC were co-cultured to form in vitro capillary blood vessels and mock-infected or infected with HTNV or ANDV for 3 days without VEGF-containing media. Whole cell lysates of HTNV- or ANDV-infected samples were examined for the presence of VE-cadherin by western blotting. There was no noticeable degradation of VE-cadherin in the virus-infected samples (Figure 3A), despite the presence of HTNV and ANDV antigens in both HUVEC and PaSMC as determined by IFA. VEGF was detected in all supernatants at 3 days post-infection with no statistically significant levels in hantavirus samples when compared to mock (Figure 3B). In addition, we did not observe abnormalities in vessel structures, which we would expect if VE-cadherin levels were reduced as VE-cadherin is also recognized to be essential for maintaining the structural integrity of newly formed blood vessels [28]. These results demonstrate that infection with pathogenic hantaviruses in a blood vessel model that secretes VEGF does not cause loss of vascular integrity. Hantavirus-infected EC would likely be exposed to blood plasma proteins in vivo, providing an opportunity for their participation in dysregulating endothelial cell barrier functions through the release of BK, which in turn would immediately affect vascular permeability. To test whether the KKS can result in altered levels of BK, HUVEC and PaSMC were plated and co-cultured to form in vitro capillaries and then were mock-infected or infected with HTNV or ANDV. The plasma factors FXII, PK, and HK were added to cultures and BK in the supernatants was measured using an enzyme immunoassay. Significantly higher levels of BK (p≤0. 0001) were identified in supernatants from both HTNV- and ANDV-infected capillaries as compared to the mock-infected samples (Figure 4). To determine if BK elevations occur in individual cell types, PaSMC and HUVEC were also cultured separately and mock-infected or infected with HTNV or ANDV. Plasma components FXII, PK, and HK were added to the cultures, and BK levels in cell supernatants were measured. We detected significantly higher levels of BK in the supernatants of HTNV- (3. 5-fold) and ANDV-infected (3-fold) HUVEC as compared to mock-infected samples, and a modest increase in PaSMC (p≤0. 001) (Figure 4). While KKS activation can occur on the surface of normal uninfected HUVEC, the increased levels of BK that we observed in the supernatants of hantavirus-infected capillaries or individual cultures of PaSMC or HUVEC suggested an increase in cleavage of HK was occurring. FXII can amplify the KKS through reciprocal activation with PK and cleavage of HK yielding BK; however, in the absence of FXII, other proteins such as PRCP and HSP90 can mediate the activation of PK and subsequent cleavage of HK. To determine if FXII was required for increased HK cleavage in hantavirus-infected EC, we mock-infected or infected HUVEC with HTNV or ANDV and incubated with FXII, PK, and HK or only PK and HK. When FXII was present, we observed a significant decrease of full length HK in HTNV- and ANDV-infected samples as compared to the mock and evident by densitometry (Figure 5A). We also observed the appearance of cleavage products, a heavy chain (62 KDa) and two light chain products (56 KDa and 46 KDa) [29] (Figure 5A). Moreover, treatment of the mock- or hantavirus-infected HUVEC with the FXIIa inhibitor, corn trypsin inhibitor (CTI), resulted in increased full length HK in all samples, indicating that FXII was important for this process (Figure 5A). In contrast, when exogenous FXII was not present, similar amounts of full length HK and cleavage products were detected in mock-, HTNV-, and ANDV-infected samples (Figure 5B). Treatment with the KAL inhibitor, PKSI-527, but not CTI was able to inhibit cleavage of full length kininogen indicating that KAL was present in our samples (data not shown). Additionally, when only HK was added, we did not observe a significant difference in the levels of HK on the surface of mock- and hantavirus-infected HUVEC by western blotting, indicating that hantavirus infection by itself does not increase HK binding (data not shown). Although it has not been clearly established how PRCP or HSP90 activates PK, we considered whether viral infection might increase the levels of these proteins and lead to increased PK activation. To eliminate a role for these proteins in increased PK activation, we tested levels of PRCP and HSP90 in mock-, HTNV-, and ANDV- infected HUVEC and detected no apparent differences by western blot (Figure 5C). Increased levels of BK and increased cleavage of HK suggest that increased activation of the enzymes FXII and PK is occurring on the surfaces of HTNV- and ANDV-infected HUVEC. To assess enzyme activities, we incubated mock- or hantavirus-infected HUVEC with FXII, PK, and HK and used a chromogenic substrate, S2302, with specificity for the active enzymes FXIIa and KAL. Consistent with our data showing increased cleavage of HK and subsequent release of BK, we also detected significantly higher FXIIa/KAL hydrolytic activity in the presence of HTNV and ANDV when compared to mock-infected samples (Figure 6). In the absence of FXII, no significant difference in KAL activity was detected between mock and hantavirus-infected HUVEC, further supporting the importance of FXII for the increase in KKS activation of hantavirus-infected EC (data not shown). Our data showing that HTNV and ANDV can alter enzyme activation of FXII and PK, cleavage of HK, and liberation of BK, led us to hypothesize that BK would induce profound changes to EC and increase vascular permeability. Generally, changes in permeability are identified using transwell systems, in which leakage of a tracer molecule is measured by passage from the upper chamber to the lower chamber over a course of 1–2 h. However, the biological half-life of BK is extremely short; lasting approximately 15 sec and its effects are rapid [30]. To capture this rapid process, we implemented electric cell-substrate impedance sensing (ECIS), which allows for real-time measurements of resistance and alterations in permeability of cultured cells. Mock- or hantavirus-infected HUVEC were cultured on ECIS arrays and the effects of added FXII, PK, and HK on resistance measurements were continuously recorded throughout the experiment. Time 0 h indicates that point at which plasma factors with or without inhibitors was added to cells already mock-, HTNV-, or ANDV-infected. An initial drop in barrier function was observed, even in mock-infected cells, upon addition of plasma factors FXII, PK, and HK. This is expected as addition of any type of solution to the wells will cause a negligible and transient drop in resistance due to temperature fluctuations, shear stress that is induced when media is added to wells, and effects of the factors themselves. The maximum decrease in resistance that we observed in the mock-infected cells was 10% (Figure 7A). These findings are consistent with previously reported changes on the ECIS after treatment of cells with BK alone [31]. In stark contrast, wells that contained HTNV- or ANDV-infected HUVECs exhibited a profound decrease in barrier function of approximately 50% as compared to the initial resistance measured at 4,000 Hz. The effect occurred at a more rapid rate than observed in the mock-infected wells and was seen immediately upon addition of the factors (Figure 7A). To further explore the role of FXII, PK, and BK on the reduction in barrier function, we also treated cells with either the FXIIa and KAL inhibitors CTI or PKSI-527, respectively or the BK receptor antagonist, HOE 140 (Figure 7B, C, & D). Addition of these compounds had a marked positive impact on the ability of the endothelium to circumvent virus-directed permeability changes. Indeed, CTI was able to completely protect the endothelial monolayer from decreases in barrier function (Figure 7C & D). These data support and demonstrate functionally the critical role of FXII in the cleavage events leading to permeability changes associated with hantavirus infections. Cells treated with HOE 140 also exhibited a strong protective effect, with cells maintaining a greater percentage of initial barrier function over untreated controls and again exhibiting a more rapid recovery (Figure 7C & D). Finally, addition of PKSI-527 is partially protective as cells treated with this inhibitor exhibited a reduced change in barrier function as compared to untreated cells and they recovered faster as well (Figure 7C & D). Together these data suggest that blocking the activities of FXIIa and the BK receptor itself have the greatest protective effect for maintaining endothelial integrity of hantavirus-infected HUVECs. Collectively, our data suggest that increased bradykinin liberation during hantavirus infection is dependent upon FXII. We hypothesized that FXII binding was increased on the surface of hantavirus-infected EC allowing for more activated enzyme in the cascade loop to amplify the reaction. We tested this by mock-infecting or infecting HUVEC with HTNV or ANDV. Cells were then trypsinized, washed, and left in suspension. FXII alone was added to samples in the presence and absence of zinc. Western blot analysis indicated that FXII levels were markedly increased on HTNV- and ANDV-infected HUVEC in the presence of zinc concentrations optimal for binding (Figure 8A). However, even in the absence of zinc, we were able to detect small amounts of FXII bound to HTNV- and ANDV-infected HUVEC (Figure 8A). Since FXII alone is increased on the surface of hantavirus-infected EC, we also considered whether FXII could still autoactivate and if autoactivation was increased over levels of normal uninfected EC. Autoactivation of FXII to FXIIa requires 60–120 min. incubations on the surface of HUVEC [32]. HUVEC that were mock-infected or infected with HTNV or ANDV were incubated with FXII for 2 h. At 1 h, we detected no discernable difference in the generation of FXIIa as measured by the chromogenic substrate S2302 (data not shown). However, by 2 h enzyme activity was detectable in mock-, HTNV-, and ANDV-infected HUVEC suggesting that autoactivation was occurring (Figure 8B). Furthermore, FXIIa activity was significantly increased in HTNV- and ANDV-infected HUVEC when compared to mock (Figure 8B). This activity was further increased upon addition of PK, suggesting that FXIIa was present to convert PK to KAL (Figure 8B). In the absence of FXII, PK alone was not sufficient to hydrolyze S2302. Samples incubated in buffer alone revealed no increase in hydrolytic activity of S2302 between time 0 and 2 h, demonstrating that contaminating KAL was not present in our assay to account for the FXIIa formation (Figure 8B). Here we show that KKS activation and BK liberation directly increases EC permeability in vitro, and we hypothesize that a similar mechanism contributes to the vascular leakage manifestations of hantavirus diseases in vivo. Previous studies suggested that hantavirus pathogenesis was related to degradation of VE-cadherin due to a hypersensitization of hantavirus-infected EC to VEGF [10], [11]. In another study, ANDV infection of cultured EC was found to result in increased levels of VEGF and VE-cadherin degradation, leading investigators to postulate that the stimulation of VEGF release by EC coupled with their sensitization to the effects of VEGF could result in the loss of EC barrier function [12]. Contrary to this hypothesis, our results obtained using an in vitro capillary blood vessel model comprised of both EC and vSMC and able to secrete VEGF, did not show increased VE-cadherin degradation or VEGF levels after infection with either the HFRS-causing virus HTNV, or the HPS-causing virus, ANDV. Our findings do not preclude the importance of VEGF and VE-cadherin in hantavirus permeability, but suggest that additional studies are needed to resolve these discrepancies in the different types of cell culture systems used. As we showed, HTNV and ANDV can infect vSMC when cultured alone or in the in vitro capillary blood vessel model. Endothelial cells regulate barrier function and vascular tone in close collaboration with vSMC. Endothelial dysfunction as a result of injury or stress can trigger the production of EDHF, PGI2, and NO production [33]. These EC-derived factors act directly on vSMC to induce vasodilation. As a response to EC-derived factors, vSMC can upregulate the transcriptional activity of certain factors such as VEGF [34]. Testing has not been performed to determine if hantavirus-infected vSMC respond appropriately to EC-derived mediators. Our findings warrant additional studies on the role that hantavirus-infected vSMC play during pathogenesis. This will help to further dissect pathogenic mechanisms occurring locally in infected blood vessels. The manifestations of HFRS and HPS; e. g. , edema, hypotension, shock, and coagulation abnormalities, resemble clinical symptoms of bacterial diseases for which systemic KKS activation is thought to play an important role. Most notably, excessive KKS activation has been observed during meningococcal septic shock and streptococcal toxic shock syndrome [25], [35]. Both gram positive and gram negative bacteria have the ability to bind to one or more of the plasma components and directly cleave HK or proteolytically activate FXII and/or PK to indirectly release BK (reviewed in [36]). These types of enzymatic activities are not known to be associated with hantavirus proteins making it unlikely that the virus itself is directly activating FXII or PK through cleavage mechanisms. Endogenous activators of FXII have also been identified such as nucleic acids, misfolded proteins, and aggregates [37]–[39]. Since hantaviruses have no lytic effect on EC, it seems improbable that viral RNA or misfolded/aggregated proteins would be released from cells initiating FXII and PK activation. Tissue factor and anionic phospholipid is found incorporated into the virion of Herpes Simplex Virus-1 which allows for binding of plasma factors and initiation of coagulation cascades [40]. We have not observed any increase in activation of the KKS with HTNV alone. The more likely explanation is a hantavirus-directed effect that increases KKS activation. In addition to bacterial diseases, a rare genetic disorder, HAE, has been associated with plasma factor activation dysfunctions. HAE is characterized by recurrent attacks of edema that can occur in the airway, extremities, face, and abdomen [41]. While the trigger for attacks is not well understood, the mechanism that produces edema has been very well characterized. Patients are deficient in C1 esterase inhibitor (C1-INH) which is important in suppressing FXIIa [41]. With reduced levels of C1-INH, patients experience uncontrolled FXII activation and subsequently elevated levels of BK resulting in edema formation [41]. Fortunately, there are three types of FDA approved drugs that target individual components of this pathway to treat disease. C1-INH concentrate (Berinart, Cinryze, Rhucin) can be administered to suppress FXIIa and KAL activity, Ecallantide (Kalbitor) is a specific inhibitor of KAL, and HOE 140 (Icatibant, Firazyr) targets the BKB2R to prevent binding of BK [42]. Although most of these drugs have not been tested in HFRS or HPS patients, recently, the BKB2R antagonist, icatibant, was given to a single patient with a severe case of Puumala virus-associated HFRS [26]. The patient improved and recovered after receiving one dose of icatibant, supporting the theory that BK might play a role in HFRS pathogenesis and suggesting that further controlled clinical studies of BK inhibitors are warranted. Our in vitro studies demonstrate that increased cleavage of HK and permeability can be significantly inhibited in the presence of CTI, a specific inhibitor of FXIIa. Our data also shows that FXII binding is increased on the surface of hantavirus-infected EC. FXII activation can occur by two distinct mechanisms; proteolytic activation on the surface of EC by formed KAL or autoactivation through binding to biological surfaces. Our data suggests that both mechanisms could be important for increased KKS activation on the surface of hantavirus-infected EC. Increased binding of FXII to hantavirus-infected EC would facilitate increased formation of PK and subsequent liberation of BK, while binding of FXII alone could autoactivate to FXIIa and initiate the KKS. FXII binding is receptor specific on the surface of EC with a predominant complex utilized for preferential binding of FXII that includes urokinase plasminogen activator receptor (u-PAR) and cytokeratin 1 (CK1) [43]. FXII also autoactivates upon binding to transiently expressed globular heads of complement C1q receptor (gC1qR) [44]. Under physiologic conditions, binding of FXII is highly restricted by the concentration of zinc and HK present in the plasma [45]. In vitro studies have demonstrated that activated platelets can serve as a sufficient source of zinc for FXII activation [46]. In vitro, HTNV and ANDV are known to direct the adherence of platelets to the surface of EC, but there is yet no available data on whether platelets are activated. Future studies aim to identify changes occurring in hantavirus-infected EC that facilitates increased binding of FXII and whether platelets can contribute to binding and activation of FXII. There are currently no data available on FXII activation during hantavirus infections; however, clinical studies have already been performed specifically examining KAL and vascular permeability in HFRS patients, providing evidence that KKS activation is also altered in vivo. In one clinical study, plasmas from 86 patients were collected over the course of disease and KAL levels were measured during each phase [47]. Additionally, patients were categorized based on the severity of disease; mild, moderately grave, and grave. The findings were dramatic with elevated levels of KAL detected in almost all of the patients. Even in mild cases, these levels were 8 times greater than the control values [47]. Moderately grave and grave cases demonstrated 10 and 12 fold increases in KAL activity providing a correlation to severity of disease [47]. In a separate study, a direct link between KAL and the permeability index was identified in 56 HFRS patients [48]. Rates of fluid and protein loss were directly associated with elevated levels of KAL. In this study, PK levels were also measured and were determined to be decreased except in patients with mild forms of disease [48]. Similar findings are commonly observed in other diseases, such as septic shock, where excessive activation of FXII and PK results in consumption and thus decreased total levels of factor [49]. There are currently no documented clinical or animal studies that have extensively examined FXII, PK, KAL, or levels of BK during the course HPS. However, examination of studies published from ANDV-infected hamsters and patients reveal indirect evidence of intrinsic pathway activity during disease including prolonged bleeding times, decreased fibrinogen, and fibrin deposition found in pathology [50]–[52]. The hamster model accurately reflects many aspects of HPS caused by ANDV infection in humans; it will be extremely interesting to examine these pathways in detail in the hamsters and the potential of drugs that impact them during the course of diseases. To date, there are no therapeutics available to counteract acute symptoms of HFRS or HPS. If, as we conjecture, KKS activation and BK liberation are responsible for edema formation and hypotension during HFRS and HPS, then inhibitors targeting the effects of BK would presumably be effective as therapeutics for patients that are already presenting with symptoms. Figure 9 depicts three specific pathway junctions that we were able to target in vitro to block hantavirus-induced permeability. Our findings suggest the possibility of identifying multiple points of intervention using therapeutics. Some of these points can already be targeted by existing drugs used for other human diseases. Furthermore, identification of effective therapeutics that are already approved for use in humans could shorten time and expense to bring them to market. Lastly, although many hemorrhagic fever-causing viruses induce coagulation abnormalities, the underlying mechanisms of pathogenesis remain to be determined. It is possible that the pathways we have identified as contributing to hantavirus pathogenesis will also have broader implications for other viral hemorrhagic illnesses. Human umbilical vein endothelial cells (HUVEC) were maintained in endothelial cell basal growth media (EBM-2) that was supplemented with 2% fetal bovine serum (FBS), 0. 2 ml hydrocortisone, 2 ml hFGF-B, 0. 5 ml VEGF, 0. 5 ml R3-IGF-1,0. 5 ml ascorbic acid, 0. 5 ml hEGF, 0. 5 ml GA-1000,0. 5 ml heparin (Lonza). Human mesenchymal stem cells (hMSC) were maintained in mesenchymal stem cell basal growth media (MSCGM) supplemented with 10% mesenchymal cell growth serum (MCGS), 0. 5 µg/ml of A/B, and 100 IU of Pen-Strep (Lonza). Human pulmonary artery smooth muscle cells (PaSMC) were maintained in smooth muscle basal media (SmBM-2) supplemented with 5% FBS, 0. 5 ml heparin, 1 ml hFGF-B, 0. 5 ml GA-1000, and 0. 5 ml hEGF (Lonza). All primary cells were cultured in wells coated with 0. 1% gelatin. Cercopithecus aethiops kidney cells, Vero C1008 clone E6 cells (Vero E6) were maintained in Eagle' s Minimum Essential Medium (EMEM) supplemented with 10% FBS. All cells were incubated at 37°C, 5% CO2 for the indicated period of time. HTNV strain 76–118 [53] and Andes virus strain Chile-9717869 (from Center for Disease Control, 4th pass of 808034, July 24th, 1998) were propagated in Vero E6 cells. Supernatants were harvested and clarified by centrifugation and titers of aliquots were determined by plaque assay. Single chain HK, PK (specific activity of 28 PEU/mg), and FXII (specific activity of 29 U/mg) were purchased from Enzyme Research Laboratories. The substrate H-D-Pro-Phe-Arg-p-nitroanilide (S2302) was purchased from Diapharma (Franklin, OH). The antagonist, HOE 140 (icatibant) was used to block the BKB2R and Trandolapril to prevent cleavage of BK (Sigma Aldrich). PKSI-527 is a specific inhibitor of KAL activity (Enzo Life Sciences) and corn trypsin inhibitor (CTI) inhibits FXIIa (Haematologic Technologies Inc.). For immunohistochemistry and immunofluorescence of in vitro capillaries, HUVEC and PaSMC or hMSC were seeded, cultured and maintained for 5 days as previously described [27]. Samples were fixed with 10% formalin and permeabilized with 0. 2% Triton X-100 in PBS. After blocking with 10% goat serum, samples were stained for markers to specifically detect HUVEC [vWF (Santa Cruz Biotechnology) ], smooth muscle cells [SMA (Santa Cruz Biotechnology) ], or Collagen 18A1 (Santa Cruz Biotechnology) overnight at 4°C. In HUVEC and PaSMC co-cultures, horseradish peroxidase (HRP) and alkaline phosphatase (AP) conjugated antibodies were used to visualize SMC and EC capillary tubes, respectively. DAB and vector blue (Vecta Labs) were used as chromogenic substrates. In HUVEC and hMSC co-cultures, Alexa Fluor 568 and 488 secondary antibodies were used to detect vWF and Collagen 18A1, respectively. For hantavirus infection, in vitro capillaries or PaSMC alone were seeded on chamber slides and mock-infected or infected with HTNV or ANDV at a multiplicity of infection (MOI) of 2. For samples containing in vitro capillaries, the MOI was based on the initial number of HUVEC and PaSMC seeded in wells. After infection for 3 days, samples were fixed, permeabilized, and blocked as described for immunohistochemistry. Cells were stained with antibodies specific for vWF (Santa Cruz Biotechnologies), SMA (Santa Cruz Biotechnologies), or hantavirus nucleocapsid (N) (Abcam) overnight at 4°C. Alexa Fluor 568 and 488 secondary antibodies were used to detect vWF or SMA and hantavirus N, respectively. All coverslips were mounted on slides using Prolong Gold antifade reagent with DAPI (Life Technologies) and examined with a Leica DMI6000 B inverted microscope. A co-culture of HUVEC and PaSMC or hMSC were grown (as described above to induce formation of capillary tubes) on 5×7 mm silicon chips (TedPella). Specimens were then fixed with 2. 5% glutaraldehyde in 0. 1 M sodium cacodylate buffer, post fixed with 1% osmium tetroxide in 0. 1 M sodium cacodylate, and dehydrated with a graded ethanol series. The samples were critical point dried under carbon dioxide in a Bal-Tec model cpd 030 Drier (Balzers, Liechtenstein), mounted on aluminum studs, and sputter coated with 100 A of iridium in a model IBSe/SBT ion beam sputterer (South Bay Technologies) prior to viewing in a Quanta Field Emission Gun Scanning Electron Microscope (FEI) at 5. 0 kV and a working distance of 10 mm. Mock-infected, HTNV- or ANDV-infected cell lysates were prepared with NP-40 lysis buffer (150 mM NaCl, 50 mM Tris-HCl pH 8. 0,1% NP-40, and protease inhibitor cocktail), NuPAGE LDS sample buffer, and NuPAGE reducing agent (Life Technologies). Proteins were separated on 4 to 12% gradient polyacrylamide gels and transferred to polyvinylidene difluoride (PVDF) membranes. Blots were blocked with 5% non-fat milk in Tris-buffered saline (TBS) and subsequently probed overnight at 4°C with antibodies directed against HK (Abnova), FXII (Santa Cruz Biotechnology), PRCP (Santa Cruz Biotechnology), HSP90 (Santa Cruz Biotechnology), hantavirus N (ANDV NR-9673 BEI Resources), VE-cadherin (Santa Cruz Biotechnology), or GAPDH (Santa Cruz Biotechnology). Blots were then washed with TBS containing Tween 20 and probed with the appropriate species-specific HRP-conjugated secondary antibody. Blots were developed using a femto substrate detection system (Pierce Biotechnology) and signal captured using the G: Box (Stratagene). Capillaries were seeded and mock-infected or infected with HTNV or ANDV at an MOI of 2 for 3 days. Supernatants were subsequently removed and measured using a VEGF ELISA kit according to the manufacturer' s instructions (Life Technologies). Capillaries, HUVEC, or PaSMC were seeded and mock-infected or infected with HTNV or ANDV at an MOI of 2 for 3 days. To accurately measure BK levels in our supernatants, cells were pre-treated with 1 µM HOE 140 and 5 µM Trandolapril for 30 min. and throughout incubation with the plasma factors (FXII, PK, and HK) to prevent receptor binding and degradation of BK. FXII, PK, and HK were diluted to 50 nM each in HEPES (14. 7 mM) -Tyrodes' s buffer (Sigma Aldrich) containing 8 µM Zn2+, added to cells, and incubated for 1 h at 37°C. Supernatants were subsequently removed and measured using a BK EIA kit according to the manufacturer' s instructions (Bachem). HUVEC were seeded and mock-infected or infected with HTNV or ANDV at an MOI of 4 for 4 days. Media was removed and cells were washed with HEPES-Tyrode' s buffer. HK was diluted to 50 nM in the presence of 1 µM Zn2+ HEPES-Tyrode' s buffer, added to cells, and incubated at 37°C for 1 h. Cells were washed and incubated with FXII (50 nM) and PK (50 nM) diluted in 8 µM Zn2+ HEPES-Tyrode' s buffer. In some experiments, CTI (1 µM) was added to inhibit FXIIa. At the end of 1 h incubation at 37°C, cells were washed and lysed in NP-40 lysis buffer and examined by western blotting. In certain experiments, only PK (50 nM) and HK (50 nM) were present in order to measure HK cleavage that is independent of FXII. These experiments were performed as described above except 1 µM Zn2+ concentrations were used throughout the incubation steps. HUVEC were seeded and mock-infected or infected with HTNV or ANDV at an MOI of 4 for 4 days. Media was removed and cells were washed with HEPES-Tyrode' s buffer. HK was diluted to 20 nM in the presence of 1 µM Zn2+ HEPES-Tyrode' s buffer, added to cells, and incubated at 37°C for 1 h. Cells were washed and incubated with FXII (20 nM), PK (20 nM), and S2302 (0. 8 mM) in the presence of 8 µM Zn2+ HEPES-Tyrode' s buffer. Hydrolysis of the chromogenic substrate S2302 was measured by taking absorbance readings at 405 nm after 1 h incubation at 37°C. ECIS 8W10E+ arrays (Applied Biophysics) were coated with 10 µg/ml of poly-D-lysine (PDL) and washed with sterile water. EmbryoMax Ultrapure water with 0. 1% gelatin (Millipore) was then added to each well and allowed to incubate for 1 h followed by washing with sterile water. After the final wash, 400 µl of EBM-2 media was added to each well. Arrays were then loaded into the ECIS 16-well array station and stabilized using ECIS software v1. 2. 92. 0 (Applied Biophysics). HUVEC cells were mock-infected or infected with HTNV or ANDV and cultured for 4 days in flasks. After 4 days, cells were trypsinized and approximately 1×105 cells were plated (per well) in the coated arrays. The array station was placed in a standard 37°C incubator with 5% CO2. ECIS software was used to run a multi-frequency time scan at 400,4, 000,32,000, and 64,000 Hz to allow for collection of data related to resistance, impedance and capacitance. After approximately 18 h, data were visualized at 64,000 Hz and capacitance levels were uniformly below 10 nF (indicative of confluent monolayers). The instrument was paused and the array station moved to a Class II BSC. Media was removed from wells and replaced with 200 µl of phenol red free EBM-2 containing Zn2+ and inhibitors [CTI (1 µM), PKSI-527 (5 µM), or HOE 140 (1 µM) ]. Cells were allowed to re-equilibrate at 37°C after which time an additional 200 µl of phenol red free EBM-2 containing Zn2+, FXII, PK, and HK (100 nM each) were added to wells in a Class II BSC. Data were collected real-time throughout the experiment and analyzed using ECIS Software. HUVEC were seeded and mock-infected or infected with HTNV or ANDV at an MOI of 4 for 4 days. Cells were trypsinized and washed with HEPES-Tyrode' s buffer. FXII was added to cells in suspension at 100 nM in the presence or absence of 10 µM Zn2+ HEPES-Tyrode' s buffer and incubated at 37°C for 1 h. Cells were washed and prepared as described for western blotting. To exclude the possibility that FXII was non-specifically binding to extracellular matrix proteins produced during culturing, cells were trypsinized and binding performed in suspension. HUVEC were seeded and mock-infected or infected with HTNV or ANDV at an MOI of 4 for 4 days. Prior to the addition of plasma factors, HUVEC were placed in serum free media overnight. Cells were washed with HEPES-Tyrode' s buffer and incubated with FXII (25 nM), PK (25 nM), or FXII and PK, and S2302 (0. 8 mM) in the presence of 8 µM Zn2+ HEPES-Tyrode' s buffer. Hydrolysis of the chromogenic substrate S2302 was measured by taking absorbance readings at 405 nm after 2 h incubation at 37°C. ELISA and enzyme activity data were analyzed using a Student' s t test to determine significant differences between mock and HTNV or mock and ANDV samples. P values<0. 05 were considered significant. The following GenBank accession numbers refer to the proteins mentioned in the text: FXII, AAM97932. 1; PK, AAF79940. 1; HK, AAB59551. 1; CTI, CAA37998. 1.

Primary manifestations of disease due to hantavirus infections include systemic vascular leakage and hypotension for which the underlying mechanism is not known. A particularly perplexing finding is that the vascular endothelium remains intact during hantavirus infection and with no apparent cytopathic effects to explain leakage and edema. Our studies show for the first time that hantavirus-infected EC have increased KKS activation resulting in liberation of the inflammatory peptide, BK. BK is a potent inducer of vascular permeability, edema formation, and hypotension; thus, our results provide a novel mechanism for hantavirus-induced vascular abnormalities. Additionally, we describe the use of an in vitro capillary blood vessel model to examine responses occurring locally in blood vessels during infection. This model could be used in future studies by others for assessing further aspects of hantavirus pathogenesis or that of other vascular tropic viruses.

Abstract Introduction Results Discussion Materials and Methods

medicine biochemistry infectious diseases zoonoses hantavirus coagulation factors proteins virology biology microbiology viral diseases

2013

Endothelial Cell Permeability during Hantavirus Infection Involves Factor XII-Dependent Increased Activation of the Kallikrein-Kinin System

Due to worldwide increased human mobility, air-transportation data and mathematical models have been widely used to measure risks of global dispersal of pathogens. However, the seasonal and interannual risks of pathogens importation and onward transmission from endemic countries have rarely been quantified and validated. We constructed a modelling framework, integrating air travel, epidemiological, demographical, entomological and meteorological data, to measure the seasonal probability of dengue introduction from endemic countries. This framework has been applied retrospectively to elucidate spatiotemporal patterns and increasing seasonal risk of dengue importation from South-East Asia into China via air travel in multiple populations, Chinese travelers and local residents, over a decade of 2005–15. We found that the volume of airline travelers from South-East Asia into China has quadrupled from 2005 to 2015 with Chinese travelers increased rapidly. Following the growth of air traffic, the probability of dengue importation from South-East Asia into China has increased dramatically from 2005 to 2015. This study also revealed seasonal asymmetries of transmission routes: Sri Lanka and Maldives have emerged as origins; neglected cities at central and coastal China have been increasingly vulnerable to dengue importation and onward transmission. Compared to the monthly occurrence of dengue reported in China, our model performed robustly for importation and onward transmission risk estimates. The approach and evidence could facilitate to understand and mitigate the changing seasonal threat of arbovirus from endemic regions. The substantial growth and reach of human travel in recent decades has contributed to the global spread of infectious diseases [1–4]. In particular, air travel has allowed human hosts or carriers of pathogens to move long distances within the incubation period of infections [5], such as the viruses that cause severe acute respiratory syndrome (SARS), H1N1, Ebola, Zika, and yellow fever [6–11], or the parasites that cause malaria [12–14]. Regarding to the continual growth of international tourist arrivals, from 25 million in 1950 to 1. 2 billion in 2015 [15], understanding the global dynamics of infectious disease has become a major 21st-century challenge, and mechanistic or mathematical models built with air-transportation data been widely used to measure risks of arriving infected humans, growth rate of an introduced epidemic and the impact of specific surveillance and control strategies [2,6, 16,17]. Some relevant factors for assessing the risk of disease importation from endemic regions into a country are: 1) the risk of a person acquiring the disease in the origin country; 2) the risk of a person traveling to the destination country of interest while infectious; and 3) the likelihood of subsequent local transmission in the destination country [18]. However, most previous modelling studies have only focused on some of these components, and the seasonal and inter-annual risks of international spread of infectious diseases have rarely been quantified [6,16,19–22]. Moreover, the relative exposure risk and importation probability in travelers are likely to differ between local residents in endemic regions and residents of non-endemic areas traveling to endemic countries [21,23,24]. Given the global expanding distribution of Aedes mosquitoes [1], dengue has established itself throughout the world’s tropical and subtropical regions in both endemic and epidemic transmission cycles, causing significant morbidity and mortality, particularly highly endemic in South-East Asia (SEA) [24–26]. However, dengue remains a seasonal disease in China, with epidemics occasionally triggered by imported dengue viruses (DENV) [27]. More than 90% of imported cases between 2005 and 2014 originated from SEA [27–31]. Following China’s economic boom in the last two decades, the number of Chinese citizens travelling abroad has increased from 5 million in 1996 to 128 million in 2015 [32]. Recent government led initiatives to further foster international trade may contribute to increased flows between SEA and China [33], which could also increase the number of importations of pathogens including DENV. While the risk of dengue in China is apparent and growing [27], the seasonal pattern and changing risk of importation and subsequent transmission are unclear, a challenge amplified by a dearth of models for assessing seasonal risk for pathogen spread globally [18,34]. As international travel between SEA and China by airplane is fast and common, based on the assumption that human mobility via commercial air travel is an important conduit for the spread of infectious diseases internationally, we constructed a branching process model by focusing on the seasonal and multiannual movement of DENV from the endemic countries in SEA into China via air travelers of Chinese and SEA residents between 2005 and 2015. We then retrospectively quantified and validated the seasonal risks, ranging from zero to certain (1), of DENV importation from nine SEA countries and leading to autochthonous transmission (introduced transmission) in China, identified geographic and seasonal patterns of emerging origin-destination routes, and estimated the number of imported infections in Chinese travelers and SEA residents into China. With rising concerns about global pathogen dispersal, this study provides approaches and evidence that can inform efforts to mitigate the spread of DENV and other arboviral pathogens including Zika, chikungunya, and yellow fever viruses from endemic regions. Ethical clearance for collecting and using secondary data in this study was granted by the institutional review board of the University of Southampton, England (No. 18152). All data were supplied and analyzed in an anonymous format, without access to personal identifying information. The volume of airline travelers from 17 SEA countries into China nearly quadrupled from 3. 6 million in 2005 to 13. 8 million in 2015, with the most (69. 3% of all 73. 9 million passengers) departing from Thailand, Singapore, and Malaysia (Fig A in S1 Appendix). Nine SEA countries with available monthly dengue incidence data for risk analysis had a total of 63. 4 million airline travelers (85. 8% passengers from 17 SEA countries) into 165 cities in China between 2005 and 2015, including 38. 7 million (61. 1%) Chinese travelers and 24. 7 million residents (38. 9%) from nine SEA countries with Chinese increased rapidly from 1. 4 million (44. 8%) in 2005 to 9. 5 million (79. 0%) in 2015 (Figs B and C in S1 Appendix). Fig 1 shows the volume of travelers from SEA and the number of corresponding imported dengue cases into China have positive correlations by year and by origin (Spearman’s rank correlation, both p<0. 001). Seasonal patterns of dengue transmission in nine countries were also seen with annual amplitude positively correlated to the latitude of each country. Furthermore, there was a significant synchrony between dengue incidence in SEA and importation to China, and the seasonal epidemics in China were also highly coherent with dengue transmission in SEA and importation into China (Fig 2 and Figs D-F in S1 Appendix). The monthly DENV importation risk from nine countries of SEA into a province of China have increased from a median of 0. 18 (IQR 0. 03–0. 57) in 2005 to 0. 98 (0. 72–1. 0) in 2015 (Fig 3, panel A). Both Chinese travelers and SEA residents contributed to increasing risk over that decade, but Chinese travelers (median 0. 26, IQR 0. 03–0. 88) were more likely to introduce dengue into China than SEA residents (0. 14,0. 02–0. 56), particularly since 2011 (Figs G-I in S1 Appendix). Across all years, the lowest risk (median 0. 22, IQR 0. 03–0. 82) was in March, and the highest (0. 65,0. 12–1. 0) was in August when 23 cities (13. 9% of 165 cities) had an average risk greater than 0. 5 between 2005 and 2015 (Fig 3, panel B and Figs J and K in S1 Appendix). The percentage of cities with a median importation risk higher than 0. 5 increased from 4. 8% (8/165) in 2005 to 21. 8% (36/165) in 2015 with most emerging destinations in central and western China (Fig 4), and cities with a median probability of risk greater than 0. 5 due to Chinese travelers increased from 7 to 35, versus 5 to 18 for SEA residents (Fig L in S1 Appendix). Thailand, Malaysia and Singapore were consistently amongst the locations with the highest risk for DENV importation into China; while Sri Lanka and Maldives were emerging as important origins due to the increasing travel, particularly in Chinese (Fig 5 and Figs M and N in S1 Appendix). Meanwhile, among the 1485 routes from nine SEA countries to 165 cities of China, those with a median risk higher than 0. 5 rose from 15 (1. 0%) in 2005 to 84 (5. 7%) in 2015. A total of 11,901 infections (95% UI 6923–16,917) via air travel was estimated to import from nine SEA countries into China between 2005 and 2015, which was 13. 5 times (7. 8–19. 2) of the 879 imported cases reported in dengue surveillance system of China. The estimates had positive correlations with the reported numbers by month and country, and by nationality (Fig 6 and Figs O and P in S1 Appendix). Furthermore, the estimated time series with a one-month lag could significantly predict the numbers of cases reported in surveillance (F = 203. 7, p<0. 001) (Fig Q in S1 Appendix). The probability of DENV introduced local transmission from nine SEA countries into China also rose, with a median risk increasing from 0. 10 (IQR 0. 01–0. 30) in August 2005 to 0. 56 (IQR 0. 21–0. 91) in August 2015 at provincial level (Fig 3, panel C). Significant seasonal variation was evident, with high risk during the warm months between May-October, but very low risk in other months (Fig 3, panel D). Compared to cities with intensive importation in cold regions of northern China, e. g. Beijing and Shenyang, the introduced risks in the lower latitude cities, e. g. Guangzhou, Shenzhen and Haikou, were much higher and extended over longer time periods (Fig 4 and Fig K in S1 Appendix). The countrywide change in the probability of introduced transmission between 2005 and 2015 led to a much larger population being at risk: Guangzhou, Shanghai and Xiamen with 32 million people were the only three cities with a risk greater than 0. 5 in August 2005, while there are 102 million people in 10 cities (Guangzhou, Shanghai, Fuzhou, Xiamen, Shenzhen, Hangzhou, Haikou, Nanning, Wuhan, Changsha and Chongqing) with the same risk in August 2015 (Fig R in S1 Appendix). The dynamics of dengue in SEA, the volume, demography and immunity level of airline travelers, and the environmental suitability of DENV local transmission in China have been changing the high-risk routes for importation and introduced transmission (Figs S and T in S1 Appendix). For instance, the Maldives-Guangzhou, Philippines-Fuzhou and Malaysia-Hangzhou routes have had increasing risks since 2005, and cities in central China and middle coastal regions, e. g. Hangzhou, Chengdu and Wuhan, are emerging as destinations with an increasing risk of introduced transmission. Additionally, compared to the monthly occurrence of cases reported at the provincial level in China, the ROC curves showed our model performed robustly with an AUC of 0. 86 for importation risk estimates and 0. 92 for introduced transmission estimates (Fig U in S1 Appendix). Moreover, the importation risk estimates for SEA residents had a slight better performance than for Chinese travelers (AUC 0. 91 vs 0. 86). Being able to identify areas at risk for introduction and spread of pathogens in a timely manner is critical for situational awareness and for tailoring strategies for preparedness and response, e. g. allocation of finite health and human resources [27]. In this study, we constructed a branching process modelling framework to elucidate seasonal probability of international spread of mosquito-borne viral disease from endemic countries in SEA via air travel. We have identified the emerging origins in SEA and locations in China that are most susceptible to dengue importation and onward transmission, and we also revealed the seasonal patterns and increasing risks in routes of DENV spread by air travel over a decade. The spatiotemporal heterogeneities of DENV importation risk have also been seen in the travelers of Chinese and SEA residents. The risk of introduced transmission from particular routes identified can be used to inform efforts to prevent and control the spread of DENV, and are particularly important for currently neglected, high-risk locations, i. e. Chengdu, Wuhan and Hangzhou. Moreover, with the increasing risk of dengue importation and transmission from SEA, China can also be the source of exportation, and this was amply shown in the introduction of dengue from Guangzhou into Japan resulting in a small outbreak in Tokyo [43]. The geographic, historical, and cultural ties between SEA countries and China, as well as increasing economic and tourism links, has contributed to the growing travel volume. We demonstrate here the epidemiological significance of this travel in the context of DENV importation from these countries into China over a decade. Compared to SEA residents travelling into China, the accelerated growth in the volume of international Chinese travelers over time has also facilitated increased DENV importation from SEA. For instance, the growth in risk of dengue importation from Sri Lanka since 2010 can likely be attributed to the increasing investment and workers from China [44], while the rising risk from the Maldives is probably related to increasing numbers of Chinese tourists [32]. The megacities in China, e. g. Beijing, Shanghai and Guangzhou, each regional aviation hubs, have consistently received large volumes of international air passengers, leading to high risks of dengue importation from SEA. However, the rapid growth of travel abroad for tourism, business and migrant workers from cities in central and southwest China is also sufficient to cause substantial risk of importation. A similar pattern has been described for malaria importation from Africa and SEA into these areas [45]. The increase in imported DENV from SEA has also increased subsequent transmission risk in China, with Guangdong, Yunnan and Fujian provinces frequently reporting outbreaks following dengue importation throughout the last decade. Meanwhile, other provinces (e. g. Henan, Shandong and Shanghai) have reported autochthonous cases of dengue for the first time [27,46]. The increasing importation risk, together with increasing temperatures and the spatial spread of Ae. aegypti [47], are all contribute to increased risk of introduced transmission and the potential for year-round autochthonous transmission of DENV and other flaviviruses in several tropical and subtropical regions of China (e. g. Hainan, Guangdong and Yunnan). The variation in DENV serotypes introduced from different origins over time is especially relevant considering the potential for adverse effects from dengue haemorrhagic fever after infection with a different serotype of DENV [27]. The number of imported cases reported in surveillance systems could be predicted by the estimated time series with one-month lag, which might be due to the longer period of travel, and the delay identification and reporting of imported infections by the routine surveillance. The gaps between the estimates and reported numbers found in this study also highlight the needs to improve the capacity of surveillance systems and formulate strategies to mitigate these imported contagions, and public health authorities and partners in areas with huge volume of imported infections and high risk of autochthonous transmission should consider implementing appropriate actions at an early stage of potential seasonal transmission. These could include health education in Chinese travelers and early identifying the infections in entry points, and improve the capacity of surveillance, vector control, laboratory diagnosis, and clinical management. However, the risk of introduction is a more complex function that reflects more than travel volumes [48,49], e. g. the incidence of the disease in the country of disembarkation, the probability of being infected/viremic at the time of travel and arrival in the destination country, the duration of viremia, the presence of favorable conditions (vectors and seasonality) in the destination. Therefore, our findings must be considered in the context of several assumptions and data limitations. First, the quality of incidence data on dengue incidence in SEA and China likely varies due to differences in surveillance systems including case definitions, reporting methods, availability of healthcare and laboratory diagnosis, under reporting, and the completeness and accuracy of data reported. Second, the risk of dengue infection in SEA was assumed to be identical across each country, without considering the immunity of different serotypes in Chinese travelers and SEA residents. Third, we only estimated the risk of the Chinese and SEA residents, but the actual nationalities in travelers might be much complex, not only the Chinese and SEA residents, but also the residents from other countries passing through SEA on their way to China. Fourth, we regarded Ae. albopictus as an equally competent vector to Ae. aegypti for DENV, with similar temperature dependency and extrinsic incubation period. Fifth, our estimates did not address variability in the public health and health-care capacity and resources for different years and locations in China and SEA countries in response to dengue. Furthermore, due to the availability of monthly disease incidence in SEA countries and the absence of monthly travel data by land and water from SEA into China and within China, we only estimate the seasonal risk of introduction through air travel for nine SEA countries. Therefore, the total risk of dengue introductions from all SEA countries into China must be underestimated in this study. To solve these problems in future studies, the monthly dengue incidence for all countries could be estimated by mathematical models based on epidemiological and entomological parameters and climate data, and the seasonal and multiannual cross-border population movements could be further estimated by gravity-type spatial interaction models or using novel sources of data, e. g. mobile phone data or social media data [50–52]. Nonetheless, the models and findings presented here leverage previous work suggesting that a probabilistic model of pathogen spread over a heterogeneous network by multiple populations could capture most of the information in much complex stochastic simulation models [18,34]. Moreover, our retrospective validation showed that the predicted seasonal risk of DENV into China coincided with a surge in the number of imported cases and volume of airline travelers arriving into China from SEA countries with reported dengue virus activity. Our model framework is sufficiently flexible to incorporate new forms of data and adapt to different vector-borne diseases. Moreover, it may be used to project into the future given different scenarios and to quantify the effects of different control methods. It also highlights the need for high-quality, accessible travel and surveillance data at national, regional, and global levels. As shown here, travel dynamics have a direct and drastic impact on regional and global infectious disease dynamics and having accessible data to assess those risks in real time can support appropriate risk assessment and prevention, and control activities.

Given the global expanding distribution of Aedes mosquitoes, dengue has established itself throughout the world’s tropical and subtropical regions in both endemic and epidemic transmission cycles, causing significant morbidity and mortality. Moreover, the rise of air travel over the past century has resulted in a highly inter-connected world, where geographical distance is becoming less of a barrier to pathogen dispersal. However, few studies have quantified and validated changes in seasonal and long-term risks of international spread for infectious diseases. In China, dengue remains a seasonal disease occasionally triggered by imported dengue viruses in travellers, with more than 90% of imported cases between 2005 and 2014 originated from South-East Asia. Therefore, taking dengue importation from South-East Asia into China as an example, we constructed a branching process modelling framework to integrate three components for assessing the risk of dengue introduction into a country: 1) the risk of a person acquiring the disease in the origin country; 2) the probability of a person traveling to the destination country of interest while infectious; and 3) the likelihood of subsequent local transmission in the destination country. This model has revealed the seasonal patterns and increasing risks in routes of dengue spread by air travel over a decade. The spatiotemporal heterogeneities of dengue importation risk have also been seen in the travelers of Chinese and SEA residents. The risk of introduced transmission from particular routes highlighted could be used to inform efforts to dengue prevention and control, particularly in currently neglected, high-risk locations.

Abstract Introduction Methods Results Discussion

dengue virus invertebrates medicine and health sciences air travel pathology and laboratory medicine china pathogens geographical locations microbiology social sciences human mobility animals viruses rna viruses insect vectors human geography infectious diseases geography medical microbiology sri lanka microbial pathogens disease vectors insects arthropoda people and places mosquitoes eukaryota asia flaviviruses viral pathogens earth sciences biology and life sciences species interactions organisms maldives

2018

Seasonal and interannual risks of dengue introduction from South-East Asia into China, 2005-2015

In rhesus macaques (RMs), experimental depletion of CD4+ T-cells prior to SIV infection results in higher viremia and emergence of CD4-independent SIV-envelopes. In this study we used the rhesus recombinant anti-CD4 antibody CD4R1 to deplete RM CD4+ T-cells prior to SIVmac251 infection and investigate the sources of the increased viral burden and the lifespan of productively infected cells. CD4-depleted animals showed (i) set-point viral load two-logs higher than controls; (ii) macrophages constituting 80% of all SIV vRNA+ cells in lymph node and mucosal tissues; (iii) substantial expansion of pro-inflammatory monocytes; (iv) aberrant activation and infection of microglial cells; and (v) lifespan of productively infected cells significantly longer in comparison to controls, but markedly shorter than previously estimated for macrophages. The net effect of CD4+ T-cell depletion is an inability to control SIV replication and a shift in the tropism of infected cells to macrophages, microglia, and, potentially, other CD4-low cells which all appear to have a shortened in vivo lifespan. We believe these findings have important implications for HIV eradication studies. The interaction between HIV and CD4+ T-cells is complex, and may result in contrasting effects with respect to virus replication. On the one hand, CD4+ T-cells have a beneficial role as mediators of antiviral immune responses, both directly and by providing help for HIV-specific CD8+ T-cells and B cells [1]–[4]. On the other hand, CD4+ T-cells are key targets for infection and sustain virus replication [5], [6]. To better understand the relationship between CD4+ T-cell availability and HIV replication, we recently conducted a CD4+ T-cell depletion study in rhesus macaques (RMs) prior to SIV infection [7]. This previous study showed that antibody-mediated depletion of CD4+ T-cells was associated with increased virus replication and rapid disease progression [7]. Furthermore, using in vitro systems we demonstrated the emergence of CD4-independent SIV envelopes capable of mediating entry into cells expressing CCR5 without CD4. The absence of antibodies targeting conserved CD4-inducible epitopes has been proposed as one of the mechanisms allowing CD4-independent SIV to emerge in CD4-depleted RMs [8]. Of note, in that study one RM with the least effective CD4+ T-cell depletion showed the lowest viremia and survived throughout the entire study, suggesting that intermediate levels of CD4+ T-cells may be the ideal balance between the beneficial and harmful contribution of CD4+ T-cells to disease progression. This previous study raised some critical questions, including: (i) is partial depletion of CD4+ T-cells beneficial? (ii) What cells are the main sources of virus replication in the absence of CD4+ T-cells, and where are they located? (iii) What is the in vivo lifespan of these productively infected cells? And finally, (iv) can we identify correlates of the high viremia associated with CD4 depletion? To answer these questions, we designed a new study where we used a concentration of the CD4 depleting antibody CD4R1 which generated variable levels of CD4+ T-cell depletion allowing us to test how CD4+ T-cell availability impacts SIV infection. Furthermore, we performed an extensive combined immunohistochemistry (IHC) and in situ hybridization (ISH) analysis on colon, jejunal and brain tissues collected at necropsy from the eight SIV-infected RMs - four CD4-depleted and four controls - included in Ortiz A. M et al [7]. In contrast to our current study, these animals were not ART-treated when euthanized. As such, they represent the ideal samples to investigate the presence and the phenotype of productively infected cells. Of note, both the present and the former studies were performed at the same facility, using the same animal species, virus, route and dose of infection. We found that depletion of RM CD4+ T-cells prior to SIV infection is associated with dramatic changes in the course of disease, including post-peak viral load two-logs higher than undepleted controls, expansion of pro-inflammatory monocytes, and massive activation and infection of macrophages and microglia that appear to be the predominant population of productively infected cells. Finally, our analysis of the slope of viremia decline after initiation of antiretroviral therapy (ART) suggests that, in the absence of CD4+ T-cells and in the presence of high levels of activation, the lifespan of these virus targets is significantly longer than controls, but markedly shorter in comparison to those previously estimated for macrophages. CD4+ T-cells were depleted in eight RMs using a single administration of CD4R1 antibody at 50 mg/kg, as recommended by the “NIH Nonhuman Primate Reagent Resource” protocol (Figure 1a). Of note, this regimen generated variable levels of CD4+ T-cell depletion in our pilot studies. Four untreated animals were included as controls. All 12 RMs were infected with SIVmac251 (i. v. 3,000 TCID50) six weeks post CD4R1 treatment, in order to avoid the possible confounding effect of direct antiviral activity of the CD4R1 antibody. At day-52 post-infection (p. i.) all animals were treated with a three-drugs antiretroviral regimen (PMPA, FTC, raltegravir). Blood, bone marrow aspirate (BM), lymph nodes (LN) and rectal biopsies (RB) were collected longitudinally and at necropsy (Figure 1a). The efficacy of CD4+ T-cell depletion was determined in blood and tissues by flow cytometry. Figure 1 shows the longitudinal levels (mean±S. E) of the circulating CD4+ T-cells expressed as the absolute number (b) or fraction of CD3+ T-cells (c). Treatment with 50 mg/kg CD4R1 resulted in a severe depletion in three RMs (orange), but an intermediate depletion in the remaining five treated animals (blue). Of note, there was no association between the extent of CD4+ T-cell depletion and the age, sex, and weight of the animals. Representative flow plots of the percentage of CD4+ T-cells pre- and post-depletion are shown in Figure 1d for a severely (top) and an intermediately (bottom) depleted RM. Severe depletion was characterized by a nadir CD4+ T-cell frequency of 3. 18±0. 356% (93. 5% depletion from baseline), and CD4+ T-cell counts of 25. 64±6. 877 cells/µL (Figure 1b, c). Intermediately-depleted animals had a nadir CD4+ T-cell frequency of 18. 6±3. 021% (57% depletion from baseline) and CD4+ T-cell counts of 181. 2±58. 98 cells/µL. In both severely- and intermediately-depleted RMs, the percentage and absolute count of CD4+ T-cells were significantly lower (P<0. 01) than controls at all experimental points post-depletion (D-35, D-11, D0; Figure 1b, c). The same gradient of CD4+ T-cell depletion was confirmed in BM and LN tissues (Figure 1e). Since we used an anti-CD4 antibody clone for immunophenotyping that is not cross-blocked by CD4R1, and additionally found a significant reduction in the fraction of CD3+ T-cells and no selective increase in the fraction of CD3+CD4−CD8− T-cells, we are confident that the loss of CD4+ T-cells observed following CD4R1 administration is indeed a true depletion of these cells, rather than a masking. As expected based on previous studies [7], depletion of CD4+ T-cells was minimal at mucosal sites. In severely-depleted RMs, depletion of CD4+ T-cells induced high levels of proliferation, with in average 72% of the remaining CD4+ T-cells expressing Ki-67 at D0 (Figure 1f). Although we cannot exclude the contribution of other mechanisms, including reactivation of latent infections, we interpreted this high proliferation as an attempt of the immune system to reconstitute the depleted CD4+ T-cell compartment. In summary, the variable efficacy of CD4+ T-cell depletion achieved in this study allowed us to study SIV viral dynamics and disease progression in RMs with normal, low or extremely low levels of circulating and tissue resident CD4+ T-cells, therefore representing an ideal model to address how CD4+ T-cells impact SIV infection. Six weeks post CD4R1 administration all twelve RMs were infected with SIVmac251 (i. v. 3,000 TCID50). Peak viral load (day 10 p. i.) was similar between CD4-depleted and control RMs (Figure 2a, b). However, whereas control animals showed a rapid drop in viral load after the peak, CD4-depleted animals maintained viremia close to peak levels and ∼2-logs higher than controls in later infection. This difference in viral load was statistically significant (P<0. 001) at each experimental point between day 21 and day 52 p. i. (Figure 2a, b). Intermediately-depleted animals exhibited viral kinetics remarkably similar to those of severely-depleted animals (Figure 2b). Furthermore, following SIV infection the absolute numbers of CD4+ (a), CD4+Ki-67+ (b) or CD4+CCR5+ (c) T-cells were comparable in intermediately-depleted RMs when compared to severely-depleted animals, and significantly lower than those found in controls at all the experimental points (Figure S1). Since severely- and intermediately-depleted RMs had comparable viremia and numbers of activated/proliferating T-cells, for the remaining of the study all eight treated animals were grouped together and defined as CD4-depleted RMs. These results show that a partial depletion of CD4+ T-cells is sufficient for the establishment of an infection phenotype of persistently high viremia. Furthermore, the presence of a 2-log higher viremia together with a dramatic loss of CD4+ T-cells suggests a different source of viral burden in CD4-depleted animals. Increased activation and turnover of monocytes predict progression to AIDS in SIV-infected RMs even better than CD4+ T-cell number [9]–[11]. Hence, we quantified the levels and Ki-67 expression of the monocyte subsets in CD4-depleted and control RMs. Monocytes were defined as classical (CD14+CD16−), pro-inflammatory (CD14+CD16+), and non-classical (CD14−CD16+) based on the expression of CD14 and/or CD16 (Figure 3a). In the control animals, stable levels of all monocyte subsets at day 52 p. i. followed their initial expansion. However, classical and pro-inflammatory monocytes in CD4-depleted RMs continued to increase, with numbers of CD14+CD16+ monocytes significantly higher than those of controls at days 28 (P = 0. 0283) and 52 (P = 0. 0283) p. i. (Figure 3b). Furthermore, and consistent with an activated/pro-inflammatory status and an increased output from bone marrow, the number of CD14+CD16+ monocytes expressing Ki-67+ in CD4-depleted RMs was significantly higher as compared to controls at days 28 (P = 0. 0162) and 52 (P = 0. 0485) p. i. (Figure 3b). At day 52 p. i. CD4-depleted RMs also have a higher number of proliferating CD14+CD16− and CD14−CD16+ monocytes than controls, although the difference was not statistically significant (Figure 3b). In CD4-depleted RMs plasma levels of soluble CD163 (sCD163), a marker of monocyte activation associated with rapid disease progression [9], [12], [13], were significantly higher than in controls at day 28 (p = 0. 016) and day 52 (p = 0. 048) p. i. (Figure 3c). Of note, at day 52 p. i. , levels of sCD163 strongly correlated with the numbers of total (r = 0. 7483, P = 0. 0070) and Ki-67+ (r = 0. 6923, P = 0. 0155) CD14+CD16+ monocytes (Figure 3d), as well as with viremia (r = 0. 8322; P = 0. 0013) (Figure S2). Thus, depletion of CD4+ T-cells prior to SIV infection results in increased number, activation and turnover of monocytes during early SIV infection. As indicated in Figure 1a, all 12 SIV-infected RMs were started on ART at day 52 p. i. The numbers of pro-inflammatory monocytes (CD14+CD16+; P = 0. 0145) and their levels of Ki-67 expression (P = 0. 0189) after 12 days of ART (day 64 p. i) were significantly decreased as compared to pre-ART (day 52 p. i.) levels, and become comparable to those found in controls (Figure 3b). Classical (CD14+CD16−) and non-classical (CD14−CD16+) monocyte levels were not significantly affected by ART (Figure 3b). Levels of sCD163 remained significantly higher in CD4 depleted animals than controls after 12 days of ART (Figure 3c). Of note, the interpretation of the effects of ART in the aforementioned parameters, in particular for sCD163, is complicated by the fact that we were able to only treat the animals for a short period. Indeed, and perhaps as a consequence of the fact that CD4-depletion resulted in very high virus replication, seven of the eight depleted RMs had to be euthanized for AIDS-related reasons briefly after ART initiation (Table S1). We next investigated the levels of SIV infection of macrophages in peripheral lymph node and intestine from CD4-depleted and control RMs by immunofluorescence staining for cell lineage markers combined with fluorescence in situ hybridization (F-ISH) for SIV vRNA. Since monocyte/macrophage express CD4, we used CD3 to determine the infection frequency of CD4+ T-cells. At day 42 p. i. SIV vRNA+ cells were more frequent in the LN of CD4-depleted animals as compared to controls (Figure 4a), consistent with the ∼2-log higher plasma viremia found at the same experimental point. The vast majority of SIV vRNA+ cells expressed the T cell marker CD3 in undepleted controls, but macrophage markers (CD68 and/or CD163) in CD4-depleted RMs. We then performed the same F-ISH staining on day 42 rectal biopsies from the RMs included in the current study as well as on colon and jejunal tissues at necropsy from the four CD4-depleted animals included in Ortiz A. M et al [7]. In contrast to our current study, those latter animals were not ART-treated and thus showed high viremia when euthanized. The same phenomena of increased levels of total SIV vRNA+ cells and expression of macrophage markers by infected cells were observed in the intestine of CD4-depleted (Figure 4b), but not of control RMs. Quantitative image analysis of LN and intestinal tissues showed that in undepleted controls more than 80% of SIV vRNA+ cells were CD3+ T-cells, while in CD4-depleted RMs ∼80% of SIV vRNA+ cells were CD68+ and/or CD163+ macrophages (Figure 4c). Although the majority of infected cells in CD4-depleted animals were macrophages, it is possible that infection of these cells was a result of high viral loads, and they did not contribute substantially to the total viral load because they produced little virus per infected cell. In order to determine whether viral production from macrophages was likely to contribute to the total viral load, we sought to compare the relative abundance of SIV vRNA in productively infected CD4+ T cells to that of macrophages in four CD4-depleted RMs. We measured the volumetric sum of the SIV vRNA intensity integration values in defined productively SIV infected CD3+ and CD68/CD163+ cells in situ from confocal images collected under identical laser settings. Using this approach we were able to demonstrate that, within a particular host, macrophages have similar or even higher (on average two-fold more) SIV vRNA content per cell than productively infected CD4+ T-cells (Figure 4d). Altogether, these data indicate that depletion of CD4+ T-cells prior to SIV infection results in activation of monocyte and massive infection of tissue-resident macrophages. These infected macrophages have higher levels of SIV vRNA than infected CD4+ T cells, and, since they constitute 80% of all infected cells in CD4-depleted animals, they are most likely the major source of systemic viremia in these animals. We then investigated if the observed high viremia, monocyte activation, and infection of macrophages in peripheral tissues were associated with central nervous system (CNS) virus dissemination and pathology. First, we investigated the presence of infected cells in brain sections collected at necropsy from the CD4-depleted and non-depleted (SIV+ controls) SIV-infected RMs included in our previous study (in contrast to the current study, these animals were not treated with antiretroviral drugs at the time of necropsy [7]). For the SIV negative controls, we used brain sections from historically SIV-uninfected RMs present at the Tulane National Primate Research Center (SIV− controls). In CD4-depleted RMs SIV vRNA+ cells were found throughout the parenchyma (Figure 5a) and had a stellate morphology typical of microglia. The high number of infected cells led us to evaluate, in the same brain sections, the levels of activation by IHC/IF staining for CD163, HLA-DR, and proliferating cell nuclear antigen (PCNA, a marker of perivascular macrophages that has been used as a marker of DNA repair in macrophages [14]). As shown in the representative images (b) and in the quantitative analysis (c) of Figure 5, the expression of CD163 (P = 0. 0013; P = 0. 0047), HLA-DR (P = 0. 0571; P = 0. 0571), and PCNA (P = 0. 0026; P = 0. 0121) in cells with location and stellate morphology typical of microglia were significantly higher in CD4-depleted than SIV-uninfected or SIV-infected controls. The surprising finding that a significant number of microglia are activated and productively infected was confirmed by triple fluorescent labeling for SIV vRNA (red), CD163 (blue) and the microglia-specific marker IBA-1 [15] (green) (Figure 5d). Quantification of these stainings showed numbers of SIV vRNA+ IBA-1+ microglia markedly higher in CD4-depleted compared to non-depleted SIV-infected animals (11. 3±8. 4 vs. 0. 2±0. 2; P = 0. 0286). Of note, we found a direct correlation between the number (cells/mm2) of SIV vRNA+ cells and PCNA+ cells in the eight CD4-depleted animals (r = 0. 8289, P = 0. 0302). Collectively, these results indicate that depletion of CD4+ T-cells in RMs results in massive activation and infection of microglia. One way of testing the longevity of infected cells is to block new rounds of infection with ART and measure the rate of viral decay (which reflects the death rate of productively infected cells present at the time of treatment). Thus, if infected macrophages have a longer survival than infected CD4 T cells, we should expect to see a much slower decay of virus under therapy in CD4-depleted RMs, where infected macrophages are the major source of systemic viremia. To directly estimate the in vivo lifespan of the productively infected cells, all 12 SIV-infected RMs were started on ART at day 52 p. i. We then measured and modeled the slope of viral load decay during ART as described in previous studies [16]–[19]. As described above, seven of the eight depleted RMs had to be euthanized for AIDS-related reasons briefly after ART initiation (Table S1). As a result of this, we were able to perform only three sample collections, one at ART initiation (day 0) and the other two at day 7 and day 12 on-ART (Figure 6a). Since at day 12 on-ART we already lost three CD4-depleted animals, the lifespan of productively infected cells was estimated based on viral load changes between day 0 and day 7, although the same conclusions were reached if we used the smaller subset of animals alive at day 12. As expected, the viral decay in control animals was rapid, with an estimated half-life of infected cells of 0. 775±0. 01 days, which was consistent with previous estimates. The lifespan of productively infected cells was significantly longer in the CD4-depleted animals (1. 33±0. 108 days, P = 0. 0238; Figure 6b), but was considerably shorter than expected for HIV infected macrophages [20]–[22]. Thus, in CD4-depleted RMs, macrophages constitute the main productively infected cell type, produce the majority of virus, and show a relatively short in vivo lifespan. Further analysis of cell death in a subset of animals showed a statistically significant higher number of TUNEL positive cells (the vast majority of which express CD163) in the brain parenchyma in CD4-depleted RMs than in controls (p = 0. 0439) and CD4-depleted animals on-ART (p = 0. 0002) (Figure S3a). The increased cell death was not Caspase-3 mediated since the number of cells positive for the active form of Caspase-3 was comparable between SIV-infected CD4-depleted RMs and controls (Figure S3b). We previously showed that a severe depletion of CD4+ T-cells is associated with increased viremia, emergence of CD4-independent viruses and rapid progression to AIDS in SIV-infected RMs [7]. In this new study we used CD4R1, a rhesus recombinant anti-CD4 depleting antibody, to achieve a gradient of CD4+ T-cells in vivo. This strategy allowed us to experimentally infect with SIVmac251 RMs with normal, low or very low level of CD4+ T-cells in blood and lymphoid tissues. The critical questions we aimed to answer in this study were the following: (i) is partial depletion of CD4+ T-cells beneficial? (ii) What cells are the main sources of virus replication in the absence of CD4+ T-cells, and where are they located? (iii) What is the in vivo lifespan of these productively infected cells? And (iv) can we identify correlates of the high viremia associated with CD4 depletion? Answering these questions may have important implications for our understanding of HIV pathogenesis and latency. Viral loads were indistinguishable between CD4-depleted and control RMs for the first 14 days p. i. However, severely- and intermediately-depleted RMs maintained viremia ∼2-logs higher than undepleted controls at all subsequent experimental points. Thus, partial depletion of CD4+ T-cells prior to infection did not result in a more benign course of infection and was sufficient to generate the higher viremia phenotype. CD4R1 was efficient in depleting CD4+ T-cells from blood, BM, and LN but not from mucosal sites. Thus, mucosal CD4+ T-cells may be responsible for maintaining peak viral loads in depleted RMs similar to that in controls, before being severely depleted in the first weeks of infection due to the high levels of SIV replication. Of note, our data on the level of proliferating cells does not support the possibility that CD4-depleted RMs have higher numbers of T-cells that can support viral replication as a result of the high level of activation/proliferation in the remaining CD4+ T-cells. Thus, in RMs the association of viremia that is ∼2-logs higher with a profound loss of CD4+ T-cells is consistent with an alternative critical source of virus replication. The observed high viremia in the context of very low levels of CD4+ T-cells prompted us to investigate the susceptibility of macrophages/myeloid cells to SIV infection. Remarkably, we showed directly in situ that CD4-depletion induces a significant shift in the nature of productively infected cells, with 80% of SIV vRNA+ cells in LN and mucosal tissues being CD3+ T-cells in undepleted controls but CD68/CD163+ macrophages in CD4-depleted RMs. It has recently been described that increased activation and rapid turnover of monocytes, in particular those with a CD14+CD16+ phenotype, associates with macrophage destruction in tissues and predicts the tempo of progression to AIDS in SIV-infected RMs [9]–[11]. In response to activation, CD163 is cleaved from the cellular surface of monocytes/macrophages and is shed as sCD163 [23]. Recently, sCD163 has been indicated as a strong correlate of monocyte activation and turnover as well as disease progression in SIV-infected RMs [9]. Consistent with these studies, CD4-depleted RMs, but not undepleted controls, experienced a substantial increase in the number of circulating CD14+CD16+ monocytes. Furthermore, plasma levels of sCD163 were significantly higher in CD4-depleted than in undepleted control RMs, and directly correlated with viral load as well as with the numbers of total and proliferating CD14+CD16+ monocytes. Although in vivo BrdU labeling studies were not performed, the findings of higher frequency of Ki-67+ monocytes and sCD163 plasma levels are consistent with a model in which CD4-depletion prior to SIV infection results in increased activation/turnover of monocytes. The key findings we described in CD4-depleted animals - high viral load, expansion of pro-inflammatory monocyte, increased level of sCD163, and massive infection of tissue-resident macrophages/myeloid cells – are all potential markers of CNS infection [7], [9], [13], [24]–[27]. In this study, CD4-depleted animals were euthanized during ART, thus the analyzed brain tissues did not represent the ideal samples to investigate the presence of productively infected cells. To overcome this limit, we stained brain sections collected in our previous study of CD4+ T-cell depletion that did not include ART [7]. It is worth mentioning that both the present and the former studies were performed at the same facility, using the same species, virus, route and dose of infection. Furthermore, the high viremia and rapid disease progression phenotype was remarkably similar between the two studies. Our combined IHC/ISH approach provided three surprising results. First, a large fraction of microglial cells from CD4-depleted animals expressed high levels of activation/proliferation/DNA repair markers, such as CD163, HLA-DR and PCNA, that were absent or exclusively present in perivascular macrophages in brain tissue of uninfected RMs and undepleted SIV-infected RMs. Second, the amount of SIV vRNA+ cells was remarkably higher in the brain of CD4-depleted RMs when compared to undepleted SIV-infected controls, thus consistent with what we found in LN and RB tissues. Third, and uniquely for the CD4-depleted RMs, SIV vRNA+ cells include not only perivascular macrophages but also cells with anatomic location, morphology, and phenotype typical of microglia. Demonstrable in vivo infection of microglia in SIV-infected RMs, particularly at the extent found here, is an exceedingly rare event even in the well-developed models of accelerate SIVE induced by depletion of CD8+ T-cells or by joint infection with an immune-suppressive and a neurovirulent SIV-variants [9], [15], [28]. Based on the paucity of CD4+ T-cells and the infection of macrophages/myeloid cells, which are typically thought to be long-lived cells, we hypothesized that in CD4-depleted RMs viral load decay during ART would be slower when compared to undepleted controls. In CD4-depleted RMs the lifespan of the main productively infected cells was 1. 3 days, considerably shorter than what is generally expected for macrophages [20], [21], [29]. Two caveats must be considered when discussing this result. The first is that the decay model assumes that viral load is at steady-state and ART is equally effective at inhibiting viral replication in depleted and control animals. The second is related to a specific limitation of our study. Due to the poor clinical conditions of CD4-depleted RMs at ART initiation and their rapid disease progression, we could not sample these animals as frequently and as long as originally planned. Thus, the estimates of viral decay rates are based on a few relatively early experimental time points on-ART. However, given the rapid rate of decay in both groups, we were able to measure a substantial drop in viral load over seven days in both cases, and observed a significant difference in the decay rate of infected cells between control and CD4-depleted animals. These caveats notwithstanding, we concluded that the bulk of the available data are consistent with the idea that– in the context of CD4+ T-cell depletion and high levels of activation/inflammation—macrophages can be highly infectable, exhibit rapid turnover, and short in vivo lifespan when infected. In this regard, our data are consistent with recent findings indicating that rapid turnover of monocyte/macrophages determined by in vivo Brdu labeling predicted AIDS progression better than viral load or lymphocyte activation [9], [10], [30]. An alternative interpretation of the short lifespan of productively infected cells in CD4-depleted RMs is that the few remaining productively infected CD4+ T-cells produce much more virus than productively infected macrophages on a per cell basis. Thus in this scenario, T-cells could still constitute the major source of virus production measured in the plasma. However, this seems extremely unlikely because the differences in virus production would need to be dramatically skewed in favor of productively infected CD4+ T-cells, considering that they represent only 20% of all productively infected cells. In fact, when we measured in situ the relative volumetric abundance of SIV vRNA within CD3+ T cells and CD68/CD163+ macrophages, we found that macrophages on average have even higher per cell SIV vRNA content (∼2-fold greater) compared to productively infected CD4+ T-cells within the same host. With macrophages representing approximately 80% of all productively infected cells in all examined tissues, and with higher levels of SIV vRNA per cell in infected macrophages, it is safe to conclude that macrophages are the main productively infected cells in CD4-depleted animals. It is important to note that, due to the short length of ART-treatment, our data do not exclude the existence of a subpopulation of SIV-infected, long-lived macrophages. However, this would constitute a very small percentage of total viral production at most, because of the observed rapid viral load seen in CD4-depleted animals. In conclusion, our study demonstrates that, in SIV-infected RMs, the net effect of CD4+ T-cell depletion is the inability to control SIV replication and a shift in the pattern of infected cells from CD4+ T-cells to macrophages, microglia, and, potentially, other CD4-low cells. These findings have important implications for functional cure and eradication studies as they indicate that macrophages and microglia can be critical target for virus infection in the context of a severely compromised immune system. Furthermore, our finding that HIV-infected macrophages can be short-lived if highly activated raises a suggestive hypothesis that eradication of HIV from this reservoir could be enhanced by therapeutics able to modulate monocyte/macrophage turnover. All animal experimentations were conducted following guidelines established by the Animal Welfare Act and the NIH for housing and care of laboratory animals and performed in accordance with Institutional regulations after review and approval by the Institutional Animal Care and Usage Committees at the Yerkes National Primate Research Center (YNPRC). All efforts were made to minimize suffering. All the blood and tissue collections were obtained from RMs housed at the Yerkes National Primate Research Center, which is accredited by American Association of Accreditation of Laboratory Animal Care. RMs are fed standard monkey chow (Jumbo Monkey Diet 5037, Purina Mills, St Louis, MO) twice daily. Consumption is monitored and adjustments are made as necessary depending on sex, age, and weight so that animals get enough food with minimum waste. SIV-infected RMs are singly caged but have visual, auditory, and olfactory contact with at least one social partner, permitting the expression of non-contact social behavior. The YNPRC enrichment plan employs several general categories of enrichment. Animals have access to more than one category of enrichment. IACUC proposals include a written scientific justification for any exclusions from some or all parts of the plan. Research-related exemptions are reviewed no less than annually. Clinically justified exemptions are reviewed more frequently by the attending veterinarian. This study was performed in strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health, a national set of guidelines in the U. S. and also to international recommendations detailed in the Weatherall Report (2006). This work received prior approval by the Institutional Animal Care and Use Committees (IACUC) of Emory University (IACUC protocol #2000353, entitled “Homeostasis of CD4+ T cells in non-human primates”). Appropriate procedures were performed to ensure that potential distress, pain, discomfort and/or injury was limited to that unavoidable in the conduct of the research plan. The sedative Ketamine (10 mg/kg) and/or Telazol (4 mg/kg) were applied as necessary for blood and tissue collections and analgesics were used when determined appropriate by veterinary medical staff. Twelve female RMs were included in this study. Eight of them were treated with a single administration of the anti-CD4 antibody CD4R1 at 50 mg/kg (intravenous), as recommended by the “NIH Nonhuman Primate Reagent Resource” protocol. CD4R1 has rhesus constant regions and rhesus variable framework sequences. Only the CDRs (and a few amino acids critical for Ig conformation) are derived from the original mouse antibody. Four untreated animals were included as controls. All 12 RMs were infected with SIVmac251 (i. v. 3,000 TCID50) six weeks post CD4R1 treatment. At D52 post-infection all animals were treated once a day with Tenofovir (30 mg/kg; s. c.), Emtricitabine (30 mg/kg; s. c.) and Raltegravir (100 mg; oral). Furthermore, we used paraffin embedded tissues, including brain, collected at necropsy from the eight SIV-infected RMs (four CD4-depleted and four controls) included in Ortiz A. M et al [7]. All animal experimentations were conducted following guidelines established by the Animal Welfare Act and the NIH for housing and care of laboratory animals and performed in accordance with Institutional regulations after review and approval by the Institutional Animal Care and Usage Committees at the Yerkes National Primate Research Center (YNPRC). All efforts were made to minimize suffering. Collections and processing of blood, bone marrow aspirate (BM), lymph nodes (LN) and rectal (RB) biopsies were performed longitudinally and at necropsy as previously described [31]–[34]. Briefly, blood samples have been used for a complete blood count, and plasma separated by centrifugation within 1 h of phlebotomy. Peripheral blood mononuclear cells were prepared by density gradient centrifugation. BM aspirates were performed using an aspiration kit to remove approximately 1 mL. For rectal biopsies, an anoscope has been placed a short distance into the rectum and up to 20 pinch biopsies obtained with biopsy forceps. RB-derived lymphocytes have been isolated by digestion with 1 mg/ml collagenase for 2 h at 37°C, and then passed through a 70-µm cell strainer to remove residual tissue fragments. For lymph node biopsies, the skin over the axillary or inguinal region have been clipped and surgically prepared. An incision has been made over the LN, which has been exposed by blunt dissection and excised over clamps. Biopsies have been homogenized and passed through a 70-µm cell strainer to mechanically isolate lymphocytes. All samples were processed, fixed (1% paraformaldehyde), and analyzed within 24 hours of collection. Twelve-parameter flow cytometric analysis was performed on WB, LN, RB and BM derived cells according to standard procedures using a panel of monoclonal antibodies that we and others have shown to be cross-reactive with RM [35]–[37]. Predetermined optimal concentrations were used of the following antibodies: anti-CD3-Alexa700 (clone SP34-2), anti-CD3-APC-Cy7 (clone SP34-2), anti-CD4-Pacific Blue (clone OKT4), anti-CD8-APC-Cy7 (clone SK1), anti-CD95-PE-Cy5 (clone DX2), anti-CCR5-APC (clone 3A9), anti-Ki-67-Alexa700 (clone B56), anti-Ki-67-FITC (clone B56), anti-CD14-Pe-Cy7 (clone M5E2), anti-CD16-BV421 (clone 3G8), anti-CD62L-PE (clone SK11), anti-CCR7-PE-Cy7 (Clone 3D12) (all from BD Pharmingen); anti-CD28-ECD (clone CD28. 2) (Beckman Coulter); anti-CD8-Qdot705 (clone 3B5) and Aqua Live/Dead amine dye-AmCyan (Invitrogen). Intracellular staining for Ki-67 was performed at room temperature for 30 minutes following permeabilization with cytofix/cytoperm (BD Bioscience). Flow cytometric acquisition was performed on an LSRII cytometer driven by the FACS DiVa software. Analysis of the acquired data was performed using FlowJo software (TreeStar) and graphs were prepared using Prism version 6. 0 (GraphPad). Quantitative real-time RT-PCR assay to determine SIV viral load was performed as previously described [38]. To detect viral RNA in the tissues, in situ hybridization for SIV was performed using riboprobes as described previously [39]. Briefly, 7 µm-thick formalin-fixed, paraffin-embedded tissue sections were de-paraffinazed through a graded series of xylenes and ethanol and rehydrated in water threated with diethylpyrocarbonate (DEPC; Sigma; Aldrich, St. Louis, MO) before antigen retrieval by boiling in a microwave for 20 minutes in citrate buffer pH 6. The tissue sections were blocked with hybridization buffer containing 50% formamide with denatured herring sperm DNA and yeast tRNA at 10 mg/ml each in a humidified chamber at 45°C for 1 hr. Hybridization was performed with SIV-digoxigenin-labeled anti-sense riboprobes (Lofstrand Labs Ltd; Gaithersburg, MD) that were applied to the tissue sections at 10 ng/slide in hybridization buffer and incubated overnight at 45°C. After hybridization, slides were washed sequentially with 2× SSC, 1× SSC, and 0. 1× SSC. The slides were incubated in TBS (tris buffer saline, 10 mM, pH 7. 4) and followed with blocking solution for 1 h. Alkaline phosphatase-conjugated sheep anti-digoxigenin antibody diluted at 1∶200 (Roche; Penzberg, Germany) was used to detect hybridized digoxigenin-labeled probes. Either Dako Liquid Permanent Red or NBT/BICP substrate-chromogens system (Dako, Inc. Carpinteria, CA) were prepared according to manufacturer' s instructions and added to the tissue for 20 min at room temp to develop the reaction. Controls included matched positive and negative tissues hybridized with digoxigenin-labeled sense RNA-labeled probes and processed in parallel. Formalin-fixed, paraffin-embedded tissue sections (5–6 µm thick) were examined by immunohistochemical staining using the following primary antibodies: Anti CD163 (clone EDHu-1, AbD Serotec), HLA-DR (clone LN3, eBioscience), and PCNA (clone PC10, DAKO). Reactivity to primary antibodies was detected using the MACH3 alkaline phosphatase polymer detection kit from Biocare Medical (Concord, CA) with either NBT/BCIP substrate system for light microscopy or Permanent red for fluorescent microscopy both from Dako (Carpinteria, CA) after which sections were counterstained using YO-PRO nuclear stain (Life Technologies, Grand Island, NY). As controls, duplicate sections were processed in the absence of primary antibodies. Triple label confocal microscopy was performed to co-localize SIV-RNA with cell type specific markers to determine the immunophenotype of infected cells as described previously [39]. Immunofluorescent labeling for microglia (rabbit polyclonal against IBA-1, Wako) and macrophages (Mouse IgG1 monoclonal to CD163, novocastra) was performed after ISH as previously described [15]. After incubation with the primary antibodies and subsequent washes, the appropriate species-specific secondary antibodies were applied; AlexaFluor 488 (green) conjugated goat-anti-rabbit (Invitrogen, Carlsbad, California) and AlexaFluor 647 (far-red, shown in blue) conjugated goat-anti-mouse IgG1 (Invitrogen, Carlsbad, California), respectively. To image the sections a Leica TCS SP2 confocal microscope equipped with 3 lasers (Leica Microsystems, Exton, PA) with a resolution of 512×512 pixels was used. The confocal imaging was performed using sequential mode to separately capture the fluorescence from the different fluorochromes. Volocity Software (v 5. 5, Perkin-Elmer) was used to render the images from the Leica z stacks. Adobe Photoshop software (Version CS6; Adobe Systems) was used to assemble the images. Quantitation of positive cells was performed manually for the SIV-RNA detection with NBT/BICP (purple-black color) by counting positive cells per field in 10 randomly selected fields encompassing a minimum of 120 mm2. The numbers of infected cells are expressed as cells per mm2. Quantitation of cells labeled for the cell type specific antigens was performed on immunolabeled slides using Inform v2 1. 2 software after capturing 10 randomly selected fields encompassing a minimum of 30 mm2 with a CRI-multispectral camera using Nuance software v 3. 0. 2 (Perking-Elmer). The positive cell numbers are express as cells in mm2. Prism v5 software (GraphPad software) was used to present the data in a graphic form. Fluorescent in situ hybridization (F-ISH) was performed using our newly designed SIV riboprobes as previously published [7]. In brief, 5-µm tissue sections were mounted on Superfrost Plus Microscope Slides (Fisher Scientific), heated at 60°C for 1 h, dewaxed in xylenes and graded ethanols, and rehydrated in double-distilled H2O. Heat-induced epitope retrieval was performed by placing slides in in 10 mM Citrate (pH 6. 0) containing 0. 05% Tween-20 or Pretreatment-2 buffer (Advanced Cell Diagnostics, Inc.) at 100°C for 5 min followed by immediate immersion into HyPure molecular biology–grade H2O (Hyclone; Thermo Scientific) at room temperature. The slides were then incubated for 5 minutes at 37°C in a Tris-buffered solution containing 2 mM CaCI2 and proteinase K (1. 25 µg/ml) then washed in HyPure molecular biology–grade H2O (Hyclone; Thermo Scientific), acetylated (0. 25% acetic anhydride) for 20 minutes, and placed in 0. 1 M triethanolamine (pH 8. 0) until hybridization. Sections were then covered with hybridization solution (50% deionized formamide, 10% dextran sulfate, 0. 6 M NaCI, 0. 4 mg/ml yeast RNA (Ambion Inc.), and 1× Denhardt medium in 20 mM HEPES buffer (pH 7. 2) with 1 mM EDTA) containing 100–400 ng/ml pooled SIV riboprobes and hybridized for 18 hours at 48°C. After hybridization, slides were washed in 5× SSC (1× SSC = 0. 15 M NaCL+0. 015 M Sodium Citrate) at 42°C for 20 minutes, 2× SSC in 50% formamide at 50°C for 20 minutes, and 1× RNA wash buffer (RWB: 0. 1M TRIS-HCL PH 7. 5,0. 4M NaCL, 0. 05M EDTA PH 8. 6) at 37°C with ribonuclease A (25 µg/ml) and T1 (25 U/ml) for 30 minutes. After washing in RWB buffer, 2× SSC, and 0. 1× SSC at 37°C for 15 minutes each, sections were transferred to 1× Tris-buffered saline (TBS; Boston BioProducts) containing 0. 05% Tween-20 (TBS-Tw). Tissues were blocked in TBS-Tw containing 2% donkey serum for 1 hour at room temperature, then incubated with goat anti-digoxigenin-DyLight-594 (Vector Labs; 1∶1,1000), mouse anti-CD68 (1∶400; clone KP1, Dako), mouse anti-CD163 (1∶400; clone 10D6; Novocastra/Leica), and rabbit monoclonal anti-CD3 (clone SP7; Labvision; 1∶200) in TBS-Tw containing 2% donkey serum overnight at 4°C. Slides were washed in TBS-Tw, incubated with donkey anti-goat–Alexa594, donkey anti-mouse–Alexa488, and donkey anti-rabbit–Alexa647 (all Invitrogen; 1∶400) for 1 hour at room temperature in the dark and washed in TBS-Tw. Slides were incubated with 0. 1% Sudan Black B in 70% ethanol (Cat. No. 4410; ENG Scientific) and 1× TBS for 30 minutes at room temperature to quench autofluorescence then incubated with 300 nM DAPI for 10 min. Slides were washed, mounted with Prolong® Gold (Invitrogen) and imaged on an Olympus FV10i confocal microscope using a 60× phase contrast oil-immersion objective (NA 1. 35) imaging using sequential mode to separately capture the fluorescence from the different fluorochromes at an image resolution of 1024×1024 pixels. Volumetric confocal images (Z-stack) were taken from lymph node and jejumal tissues following SIV F-ISH using the Nyquist sampling method using an Olympus FV10i confocal microscope with a 60× phase contrast oil-immersion objective (NA 1. 35) with imaging using sequential mode to separately capture the fluorescence from the different fluorochromes at an image resolution of 1024×1024 pixels. Laser settings (gains and power) were set to ensure no pixel saturation in the SIV vRNA (Alexa594) channel and kept constant for all sample collections for each animal imaged. Images were opened in the Olympus FV10-ASW software (v3. 1) and well-defined cellular regions were manually drawn around SIV vRNA+ macrophages and T cells from the compressed Z-stack image. The integrated intensity within these defined cell regions was measured in each Z-plane. The integrated intensity (or intensity integration) is the sum of the pixel intensities for a channel of interest within a specified object, and thus, the sum of the integrated intensities from each Z-stack is the total of the pixel intensities within each object volume. Thus, the “relative abundance” of SIV vRNA within the volume of each cell of interest was determined by summing the integrated intensities from each Z-plane. Soluble CD163 (sCD163) plasma levels were quantified by ELISA according to manufacturer' s protocol (Trillium Diagnostics) using a 1∶50 dilution of plasma samples. Repeated-measures analyses for each outcome (CD4+ T cells; CD4+Ki-67+ T cells; CD14+CD16+; CD14+CD16+Ki-67+; and viral load) were performed with a means model with SAS Proc Mixed (version 9) providing separate estimates of the means by weeks post-depletion and infection between groups. A compound-symmetric variance-covariance form in repeated measurements was assumed for each outcome and robust estimates of the standard errors of parameters were used to perform statistical tests and construct 95% confidence intervals [40]. T-test or Mann Whitney test were used to compare the differences between the model-based treatment means (least-squares means) at each time point within the framework of the mixed effects linear model. Statistical tests were 2-sided. Pearson product-moment correlation coefficients were utilized to estimate linear associations for normally distributed data and Spearman rank correlation coefficients were used for skewed and other non-normal distributions. A P value ≤0. 05 was considered statistically significant for the correlation analyses. The mean ± SEM were used as descriptive statistics for each continuous outcome. The rate of decay of virus was estimated by taking the linear slope of the natural log-transformed data of viral load from day 0 to 7 of treatment. The half-life is calculated as ln (2) /decay rate.

CD4+ T-cells are both mediators of antiviral immune response and critical targets for HIV replication. We have previously shown that experimental depletion of CD4+ T-cells prior to SIV infection in rhesus macaques results in higher viremia and the emergence of CD4-independent SIV-envelopes. The findings reported in this new study of CD4 depletion address key unanswered questions about the phenotype, location, and lifespan of the sources of the increased viral replication in the absence of CD4+ T-cells. Altogether, our new data indicate that depletion of CD4+ T-cells prior to SIV infection results in activation of monocyte and massive infection of tissue-resident macrophages, which appear to be the predominant population of productively infected cells. Furthermore, our analysis of the slope of viremia decline after initiation of antiretroviral therapy suggests that the lifespan of these virus targets is markedly shorter than those previously estimated for macrophages. In summary, in the context of CD4+ T-cell depletion macrophages can be highly infectable, exhibit rapid turnover, and short in vivo lifespan. These finding raises a suggestive hypothesis that eradication of HIV from this reservoir could be enhanced by therapeutics able to modulate monocyte/macrophage turnover.

Abstract Introduction Results Discussion Materials and Methods

infectious diseases hiv infections medicine and health sciences clinical immunology aids immune deficiency biology and life sciences immunology viral diseases

2014

CD4 Depletion in SIV-Infected Macaques Results in Macrophage and Microglia Infection with Rapid Turnover of Infected Cells

A growing body of evidence highlights the importance of the cellular microenvironment as a regulator of phenotypic and functional cellular responses to perturbations. We have previously developed cell patterning techniques to control population context parameters, and here we demonstrate context-explorer (CE), a software tool to improve investigation cell fate acquisitions through community level analyses. We demonstrate the capabilities of CE in the analysis of human and mouse pluripotent stem cells (hPSCs, mPSCs) patterned in colonies of defined geometries in multi-well plates. CE employs a density-based clustering algorithm to identify cell colonies. Using this automatic colony classification methodology, we reach accuracies comparable to manual colony counts in a fraction of the time, both in micropatterned and unpatterned wells. Classifying cells according to their relative position within a colony enables statistical analysis of spatial organization in protein expression within colonies. When applied to colonies of hPSCs, our analysis reveals a radial gradient in the expression of the transcription factors SOX2 and OCT4. We extend these analyses to colonies of different sizes and shapes and demonstrate how the metrics derived by CE can be used to asses the patterning fidelity of micropatterned plates. We have incorporated a number of features to enhance the usability and utility of CE. To appeal to a broad scientific community, all of the software’s functionality is accessible from a graphical user interface, and convenience functions for several common data operations are included. CE is compatible with existing image analysis programs such as CellProfiler and extends the analytical capabilities already provided by these tools. Taken together, CE facilitates investigation of spatially heterogeneous cell populations for fundamental research and drug development validation programs. Emerging pieces of evidence stress the importance of a cell’s local microenvironment as a regulator of cellular phenotype and gene expression heterogeneity within cell populations. Microenvironmental parameters such as mechanical forces, cell to cell contact and endogenous signaling, all vary between cells at different positions in a well [1,2]. The spatial heterogeneity of these factors leads to variability in efficiency of endocytosis and the vulnerability to viral infection [3], influences epithelial tissue growth [4], impacts the expression of angiogenic factors in tumor cells [5] and influences the differentiation potential of mouse and human pluripotent stem cells (mPSCs, hPSCs) [2,6]. Microenvironmental heterogeneity is also a potential confounding factor behind contradictory findings in the response of different cell types to key signaling pathway activity [7–9] and could limit the interpretation and reproducibility of experiments. A comparative analysis [10] of two large scale pharmacogenomic studies, the Cancer Genome Project [11] and Cancer Cell line Encyclopedia [12], revealed a surprisingly poor correlation between cell line drug response phenotypes between laboratories, which prevented meaningful extraction of drug-gene relationships. Correlation remained low even when using matched protocols and cell lines with highly correlated gene expression profiles. Although the exact source of variation in this study is unknown, a separate analysis of single cell data from 45 high-throughput (HTP) screens revealed that population context is indeed a ubiquitous source of variation between screens, and accounting for population context can improve experimental reproducibility between cell lines and laboratories [9]. Although it is acknowledged that understanding population heterogeneity is critical in biomedical research [13,14], the scientific community has been slow to adopt approaches to reduce heterogeneity, such as controlling microenvironmental variables. The increasing affordability of high content screening instruments, emergence of core screening facilities and technological advancements such as micropatterning in multi-well plates [15], enable investigation of population context dependent variables with unprecedented throughput and veracity [16]. By patterning extracellular matrix (ECM) proteins on a tissue culture surface, cells can be restricted to adhere to an array of spots of predefined shapes and sizes [17,18]. An advantage of such patterning is enhanced control over microenvironmental variation within each well and improved assay robustness [19,20]. Growing cells in colonies of defined size and shape facilitates analyses of inter- and intra-colony variation in protein expression. For example, we observe that hPSCs growing in such patterned colonies express varying levels of pluripotency markers, including SOX2 and OCT4, depending on colony size [2,19]. Elucidating the impact of the population context dependent variables on cellular phenotype will not only add to our understanding of fundamental cell biology, but will also allow us to optimize culture conditions and cell assays, provide possible explanations for current seemingly conflicting research and inform in silico models. These aspects are critical to next-generation drug development strategies and systems biology approaches. We have previously developed an HTP platform for micropatterning of cells on ECM spots of defined shapes in multi-well plates [19,21]. To augment this platform, we here present a computational tool, context-explorer (CE), which facilitates colony level analyses and cell patterning quality control. The CE software is meant to extend the functionality of currently available software solutions, both open source and commercial, for analyzing features of imaged cells [22–24]. While some implementations already exist that can be used to identify arrays of cells on glass slides [25] or to study differential gene expression of cells in different spatial locations within the same colony [26], our software aims to improve the HTP workflow for analysing cells in micropatterned multi-well plates by facilitating evaluation of patterning fidelity, enabling identification of colonies within a well, and improving spatial analyses of heterogeneous protein expression within colonies. As HTP technologies become more widespread, it is increasingly important to provide user friendly data analysis software targeted towards these platforms (Fig 1A). Here, we demonstrate the utility of CE by investigating the impact of intra-colony location on hPSC pluripotency marker expression. Designed to complement existing imaging software, CE fits into the analysis pipeline following the extraction of cellular features from microscope images (Fig 1B). The input to CE is a CSV-file, which contains single cell xy-coordinates, well label, and at least one other measurement of interest, such as protein fluorescent intensity values. These single cell coordinates can be clustered into colonies within which spatial trends for the measurements of interest can be visualized (Fig 1C). By leveraging existing image extraction software and processing the resulting text files, CE has low system requirements and runs smoothly on modern laptop computers. CE is implemented in Python, and utilizes the scientific open source ecosystem SciPy [27]. Specifically, NumPy [28] and Pandas [29] are used for array manipulations, while Matplotlib [30] and Seaborn [31] generate the graphical visualizations. To make CE easily accessible to a broad scientific community of various technical backgrounds, all functionality is available via a graphical user interface designed in Qt. To interrogate organized cell behavior within colonies, the concept of cellular colonies must first be introduced by classifying closely positioned cells as belonging to the same colony. Manually labeling individual cells is infeasible in HTP assays that often include millions of cells. There are many existing algorithms for automatically identifying dense clusters of data points [32] and CE employs the Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm [33], as implemented in the scikit-learn Python package [34], to identify sets of points at high two dimensional density. Clustering cells into colonies based on local cell density is similar to how these communities are defined biologically since cellular communication is restricted by the distance between cells. The DBSCAN algorithm is capable of identifying colonies of any geometrical shape and performs well on any cell constellation where the distance between neighboring cells within a colony is shorter than the distance between neighboring colonies. DBSCAN scan also has the advantage that it has a notion of outliers, cells far away from any colony, and can classify such cells as noise rather than trying to force all cells to belong to a colony as many other cluster algorithms would do. DBSCAN performs unsupervised clustering and does not require prior knowledge of the number of colonies within each well, only specification of the neighborhood search radius (Eps) and the minimum number of points (MinPts) within the neighborhood to start propagating a cluster. For each point found within the Eps neighbourhood of the starting point, a search for additional points will be performed. If the number of points found within a point’s Eps neighborhood is greater or equal to MinPts, that point is considered a core point of the cluster. Points that fail to meet this criteria, but that are density reachable from a core point, are considered border points and are classified as part of the colony. Points that fail to meet either of these two criteria are labelled as noise and not part of any colony. While there are implementations of DBSCAN that automate parameter optimization, these increase time complexity [35,36]. As an alternative to automatic parameter estimation, CE allows for the Eps and MinPts parameters to be adjusted via the graphical user interface while viewing the resulting colony identification accuracy. The immediate visual feedback enables intuitive and accurate colony classification and decreases the time it takes to optimize Eps and MinPts. DBSCAN clustering is deterministic for the core cells of each cluster and only border points which are density reachable from more than one cluster core can be assigned to different clusters between runs. Colonies in ECM patterned wells rarely grow close enough for border cells to be density reachable from more than one colony, so for this application cells are routinely clustered deterministically. To further increase colony identification accuracy, CE includes filters for colony size, density and roundness, which refine the colony identification procedure and are particularly useful to deal with imaging artefacts and overgrown colonies. These filters are controlled via sliders in the GUI and the resulting colony identification is immediately visualized for the selected well. Each cluster of points returned by DBSCAN corresponds to cells growing together in a colony. To define colony attributes, CE utilizes the geometric analyses package Shapely [37]. Generally, the polygonal boundary area of a colony can be defined as either the convex or concave hull of its cells. CE uses the convex hull algorithm as it is less expensive to compute than the concave hull and performs well with commonly used micropatterned spot shapes. By finding the colony bounding area, additional geometric attributes such as colony diameter, circumference, area and cell density can be calculated for each colony. Additionally, each cell can be assigned Cartesian coordinates relative the colony’s centroid or the closest edge of the colony’s boundary area. These relative cell coordinates can be used to group cells at similar positions from multiple colonies into concentric bins. The process for grouping cells is initiated by deriving an aggregated value of all the cells in a colony that are within the same location bin. These colony values are then aggregated for multiple colonies and the error estimation reported describes the variation between colonies rather than between cells within one colony. The visualizations built into CE further facilitates the analyses of spatial trends within colonies. Cells can also be grouped according to a hexagonal grid, which aggregates cells from all colonies in the same bin within the grid. To demonstrate the colony classification process, we patterned mouse PSCs on circular ECM spots 500 μm in diameter in a 96-well plate using an in-house HTP UV-lithography method [21]. To enable extraction of cellular coordinates within the well, cell nuclei were labelled with DAPI and analysed by a primary image analysis program, such as CellProfiler. After imaging and extraction of single cell features, the resulting CSV-file was processed by CE to classify cells into colonies. Cells robustly adhered to the patterned regions, and were grown for 48 h in pluripotent conditions (LIF and Serum, see Supplementary Methods for detailed culture conditions). These micropatterned, well-separated clusters of cells were easily identified by CE as separate colonies. However, cells occasionally bridge adjacent confluent colonies, effectively merging two or more colonies together. Such colonies would still be classified as valid clusters by DBSCAN, since all the cells are density reachable from each other. Another important caveat is that the imaging hardware may not allow the entire well to be captured, resulting in partial images of many of the colonies. Including either of these merged or partial colonies in the downstream analyses could confound the interpretation of the underlying biology. When limited to only the default DBSCAN parameters MinPts and Eps, partial and merged colonies are difficult to discriminate from colonies of desired shape and size. We found that an efficient way to eliminate these undesired colonies, was to apply a set of filtering criteria to the colonies detected by DBSCAN. Filtering on colony roundness and size were the most effective criteria to exclude merged and partial colonies from the DBSCAN output (Fig 2A). The effect of applying size and roundness filters was striking when comparing the number of cells per colony before and after filtering. Prior to filtering, several clusters of colony sizes were detected, including bigger merged colonies and smaller partial colonies cut off by the imaging limitations (Fig 2B). These colonies were omitted from the final analyses as they would skew the calculations of both the mean number of cells per colony and spatial trends within the colonies. After excluding colonies of undesirable size and shape, we observed only one cluster of colony sizes and the mean number of cells per colonies was notably consistent between wells, indicating reproducible patterning of cells (Fig 2C). The result of the filtering in CE was almost identical to manually identifying merged and partial colonies (406 of 412 colonies were correctly classified, S1A Fig and S1B Fig). Another useful metric for assessing patterning fidelity is the number of colonies per well, which is also computed by CE. Comparing the number of filtered and unfiltered colonies identifies wells containing many fused and partial colonies that were detected by DBSCAN, but then excluded by the filters. The filtered count was nearly identical to the counts obtained from visually inspecting the images from each well (Fig 2D), and were computed in a fraction of the time of manual counting. CE can also accurately identify clusters of cells growing in micro-patterned ECM spots of different sizes within the same well, cells patterned in non-circular colony shapes, and colonies in unpatterned wells (Fig 2E). The flexibility of the parameter and filter adjustments makes it possible to identify colonies within unpatterned wells with high accuracy (the mean difference from a manual colony count was 2. 6% (sd 1. 3%) for ten wells with around 100 colonies in each, S1C Fig and S1C and S1D Fig). These results demonstrate the capacity for CE to semi-automatically identify colonies of a wide array of geometries both in patterned and unpatterned wells with an accuracy similar to that of visual image inspection. To apply CE to hPSCs analysis, we first patterned hPSCs in 200 μm diameter colonies in 96-well plates using microcontact-printing. While most cells adhere to the ECM spots in the patterned plates, there is also limited non-specific cell adhesion in-between patterned ECM spots. Compared to UV-lithography, microcontact printing has a higher proportion of cells growing in tiny colonies and as single cells outside patterned ECM spots, which makes this technology suitable for comparing the behavior of cells outside and inside micropatterned colonies. To test whether cells that adhere non-specifically display differences in protein expression compared to cells within colonies, we assessed cellular response 42 h after treatment with serum free medium containing BMP4 (SF+BMP4, induces trophectoderm and primitive endoderm [38,39]), or MEF conditioned medium (CM, maintains pluripotency [40]). Expression of the pluripotency-associated transcription factors SOX2 and OCT4 was analyzed to quantify cellular differences. As expected, in SF+BMP4 medium, pluripotency signals were repressed and no difference was observed in SOX2 and OCT4 expression between cells inside and outside colonies (Fig 3A & S1E Fig). In contrast, CM induces the expression of SOX2 and OCT4 in cells within colonies, while cells outside colonies express the marker to a lesser extent (Figs 3A & S1E). This is visible both as a change of shapes and a shift in means of the protein expression distributions. These differences suggest that cells inside and outside colonies do not respond similarly to added factors in the medium, further highlighting the importance of controlling for microenvironmental parameters such as population context when assessing cellular responses to experimental conditions. In addition to facilitating inter-colony analyses, CE allows for investigation of intra-colony variation in protein expression. Radial gradients of protein expression have previously been reported in non-patterned and patterned colonies of hPSCs [2,6, 21,41]. Analysing these trends through visual inspection and manual data analyses is feasible in a low throughput platform, but becomes error-prone and time-consuming in HTP systems with hundreds or thousands of colonies. To visualize spatially biased protein expression, CE can automatically aggregate colonies within replicate wells and display a heatmap of spatial protein expression variation within these colonies (Fig 3B). This analysis can be applied to colonies of various sizes and shapes. When investigating colonies of hPSCs grown in either CM or SF+BMP4, we observed distinct radial gradients of SOX2 and OCT4 expression between the two conditions (Figs 3C & S1F). To distinguish differences attributed to local spatial factors from those attributed to exogenous factors, all intensities were normalized relative to the expression at the centroid of the colony. In hPSCs grown in CM, SOX2-expression decreased in a linear fashion towards the edge of the colony, with cells at the colony border only displaying half the fluorescence intensity of cells near the colony centroid. Meanwhile, cells grown in SF+BMP4, exhibited low SOX2 expression throughout the entire colony, which can be attributed to BMP4 inducing differentiation and overriding any local pluripotency supporting signals. OCT4 expression followed similar patterns in both CM and SF+BMP4. Quantitative evaluation of protein expression levels relative to the location of a cell within a colony was performed by aggregating cells in radial bins according to their distance from the colony centroid rather than their relative xy-coordinates. The mean or median expression values could then be compared between ring-shaped bins across multiple colonies. This analysis technique highlights the spatial trends of SOX2-expression for hPSCs grown in CM (Fig 3C), where expression of SOX2 was highest in cells at the center of the colony and linearly decreased toward the colony edge. Statistical significance at p = 0. 01 can be roughly inferred from non-overlapping pairs of 95% confidence intervals [42]. However, it should be noted that statistical significance is easily achieved with sample sizes this large, even at small effect sizes [43], so it is important to assess the magnitude of the differences. In hPSCs cultured in SF+BMP4, a weak radial gradient of SOX2 expression emerged exhibiting no more than a 10% difference in expression level between cells at the centroid and the edge of the colony. In contrast, hPSCs cultured in CM exhibit more than double the level of SOX2 expression at the center of the colony compared to the edge. For cells grown on ECM patterns of non-circular geometries, we evaluated how grouping cells into bins based either on the distance from the colony border or the colony centroid affected our interpretation of spatial trends in protein expression. To illustrate the different biological interpretations that could arise from these two metrics, we investigated the SOX2 expression of hPSC colonies grown on triangular ECM patterns in SF+BMP4. In addition to aggregating cells in hexagonal bins as previously described, colonies were segmented into concentric annular or colony-shaped (in this case triangular) bins (Fig 3D). The annular and triangular segmentations were created based on the distance from the colony centroid or the closest colony border, respectively. When the intensity was visualized based on the hexagonal binning strategy a clear spatial bias in SOX2 expression was reveled (Fig 3E). To further quantify this organized expression, the average expression levels of cells grouped according to the annular and triangular bins was compared. Importantly, the choice of binning metric could influence the interpretation of the resulting protein expression trends. In this example, SOX2 expression decreased more rapidly as a function of the distance from the colony centroid compared to from the colony edge (Fig 3E). Depending on the binning strategy, cells were grouped to different bins and the number of cells in each bins differed greatly (Fig 3E). There is overwhelming evidence that increased control and monitoring of population context parameters is needed to improve assay reproducibility and to understand heterogeneous responses between cells in the same population. However, addressing this challenge has proven difficult in the broader biomedical community. To augment the power of HTP analysis of population context parameters in the cellular microenvironment, we previously developed cell patterning techniques to control population context parameters, and here we demonstrate a software tool for improved monitoring of microenvironmental variables and interrogation of community driven cell fate acquisitions in HTP assays. In this study, CE was utilized to explore and quantify radial spatial trends in SOX2 and OCT4 expression within micropatterned hPSC colonies of various shapes and sizes. We observed that the protein expression levels vary as a function of cells’ location within a colony, further highlighting the importance of understanding variation in population context dependent factors. CE is compatible with existing HTP imaging software and standard fluorescent microscopy based assays. By developing a GUI-driven workflow and releasing it under an open source license, we provide a solution to facilitate colony-level analysis for a wide scientific community. Members of our group regularly use CE for colony level analyses as evidenced in previously published and ongoing studies [19–21,44]. To further broaden the utility and applications of the software, there are built-in visualizations to assist with fluorescent intensity thresholding and pattern fidelity assessment, and there are interface components for assigning wells to treatment groups. To lower the threshold for wide adoption, CE is distributed as a Python package through the conda and pip package managers, and runs under Linux, OS X and Windows. The setup process does not require any use of the command line and can be done entirely from the Anaconda Navigator GUI. The source code is distributed under the open source BSD 3-clause license, which enables incorporation of its features into existing image analysis pipelines. Source code and installation instructions are available online at https: //gitlab. com/stemcellbioengineering/context-explorer, and the documentation can be found at https: //contextexplorer. readthedocs. io.

Cell behavior is influenced by cues that cells receive from their surrounding environment such as signals secreted from other cells and cell-to-cell contact. These factors are spatially heterogeneous and cells at different positions within a colony will experience varying degrees of influence from such environmental cues. In vitro assays often do not allow control over environmental variables and there is a lack of easy to use software to investigate the effect of spatial variation in these factors. We have developed a software package to address this gap and facilitate the quantification of spatially heterogeneous cell responses. Our software accurately identifies colonies of cells within a well and individual cells can be grouped according to their position within these colonies, which enables quantification of cell response as a function of cellular location. To support broad scientific accessibility, the full functionality of the software is available through a graphical user interface. Using this software to analyze data from a screening-optimized micropatterning platform, we show that human pluripotent stem cell-derived colonies grown either under pluripotency maintenance or differentiation-inducing conditions exhibit cell responses that are dependent on spatial organization. This technology should enable more accurate and predictive context-dependent drug screening and cell-fate investigation.

Abstract Introduction Design and implementation Results Availability and future directions

graphical user interfaces engineering and technology protein expression human factors engineering stem cells molecular biology techniques research and analysis methods computer and information sciences cell potency imaging techniques animal cells man-computer interface gene expression high throughput screening pluripotency molecular biology molecular biology assays and analysis techniques gene expression and vector techniques cell biology computer architecture computer software software engineering genetics biology and life sciences cellular types software tools image analysis user interfaces

2019

Context-explorer: Analysis of spatially organized protein expression in high-throughput screens

Arboviruses are viruses transmitted to humans and other animals by the bite of hematophagous arthropods. Infections caused by chikungunya virus (CHIKV), dengue virus (DENV), Zika virus (ZIKV), and the deadlier yellow fever virus (YFV) are current public health problems in several countries, mainly those located in tropical and subtropical regions. One of the main prevention strategies continues to be vector control, with the elimination of breeding sites and surveillance of infested areas. The use of ovitraps for Aedes mosquitos monitoring has already demonstrated promising results, and maybe be also useful for arboviral surveillance. This work aimed to detect natural vertical transmission of arboviruses in Aedes aegypti and Aedes albopictus. Mosquito egg collection was carried out using ovitraps in Itacoatiara, a mid-size city in Amazonas state, Brazil. Collected eggs were allowed to hatch and larvae were tested for CHIKV, DENV, and ZIKV RNA by RT-qPCR. A total of 2,057 specimens (1,793 Ae. aegypti and 264 Ae. albopictus), in 154 larvae pools were processed. Results showed one positive pool for CHIKV and one positive pool for ZIKV. The active ZIKV infection was further confirmed by the detection of the negative-strand viral RNA and nucleotide sequencing which confirmed the Asian genotype. The Infection Rate per 1,000 mosquitoes tested was assessed by Maximum Likelihood Estimation (MLE) with 0. 45 and 0. 44 for CHIKV and ZIKV, respectively, and by Minimum Infection Rate (MIR) with 0. 45 for both viruses. To our knowledge, this is the first detection of ZIKV in natural vertical transmission in the Ae. aegypti, a fact that may contribute to ZIKV maintenance in nature during epidemics periods. Furthermore, our results highlight that the use of ovitraps and the molecular detection of arbovirus may contribute to health surveillance, directing the efforts to more efficient transmission blockade. The arboviruses transmitted by mosquitoes of the genus Aedes, like chikungunya virus (CHIKV), dengue virus (DENV), Zika virus (ZIKV), and yellow fever virus (YFV) have reached threatening numbers in the last years, with a huge impact on public health systems in several countries throughout the world [1–7]. Nevertheless, the detection and identification of circulating arboviruses are most often taken from human cases, mainly when an outbreak is already in place. With the fast-worldwide expansion of new emerging or reemerging arboviruses such as CHIKV, DENV, and ZIKV, the need to establish the role of each mosquito species in the spread of these pathogens is clear. This knowledge is fundamental to the implementation of effective surveillance and control measures against these vectors in order to avoid the early establishment of an arboviral epidemic [8]. The first report of natural infection of mosquitoes with ZIKV in Brazil was from Aedes aegypti adults collected in Rio de Janeiro [9] and this species was also considered as the primary vector during the epidemic. Indeed, the detection of ZIKV in Ae. aegypti mosquitos occurred soon after the emergence of ZIKV in this city. A recent study showed the first description of Ae. aegypti infected with CHIKV ECSA genotype in Brazil [10]. These authors reinforce the role of this species as an important vector of CHIKV in urban areas of northeastern Brazil and emphasize the benefits of entomological surveillance programs for public health, regarding the immediate implementation of diseases prevention. A study conducted in the Amazonas State, Brazil, demonstrated that the entomological surveillance using ovitraps could be successfully used to monitor the different DENV serotypes circulating in the municipalities of the interior of the state. Moreover, this study suggests that the use of arboviral monitoring strategies in routine surveillance helps for early detection of virus circulation before outbreaks, contributing to more efficient and effective control measures [11]. The continuous monitoring of Ae. aegypti infestation, associated with the early detection of arbovirus circulation, may contribute to the development of epidemic prediction models for diseases transmitted by this vector. Besides, vertical transmission, even at a low rate, contributes to the preservation of an arbovirus in nature, without a well-known cycle involving invertebrates and vertebrates, and may be of paramount importance in endemic areas as an alternative arboviral maintenance mechanism [12,13]. In the present study, we monitored and detected the natural vertical transmission of arboviruses in a mid-size city of the Amazonas State, Brazil, during the emergence of ZIKV in 2015–2016. The collection of Aedes eggs doesn' t require special permission in Brazil. All the house owners agreed with and allowed the installation of ovitraps in their properties. The study area was the city of Itacoatiara in the Amazonas State, located in Northern Brazil. Amazonas is the largest state of the Brazilian federation with 1,559,161 km2 and an estimated population of 4,063,614 inhabitants (2. 23 inhabitants/km2). Amazonas has international borders with Venezuela, Colombia, and Peru, and domestic borders with the states of Roraima, Pará, Mato Grosso, Rondônia, and Acre. Itacoatiara (latitude 03° 08' 35" S; longitude 58° 26' 39" W) is located 189km from the capital Manaus and is the third most populous municipality in the Amazonas with an area of 8,891. 9 km2 and a population estimated at 100,000 inhabitants (9. 77 inhabitants/km2). The predominant climate is equatorial (tropical monsoon), characterized by high temperatures and a significant rainfall over ≥ 2500 mm per year, but with a dry season also known as Amazonian winter. An important fluvial harbor is located in Itacoatiara for the transport of inhabitants and agricultural products [14]. Our group has been using ovitraps to monitor Aedes spp. infestation. The ovitraps consist of a dark plastic container with a capacity of 700 mL containing 300 mL of a 0. 04% brewer' s yeast solution as an attractant for mosquito females, and an oviposition substrate on which the eggs are laid (Eucatex, Brazil) with the rough part facing the inner area of the trap for oviposition. On each monitoring cycle, a total of 100 ovitraps were installed around selected houses in the city, with a weekly exchange of the pallets. Ovitraps distribution was equidistant on every 200 meters, covering the entire inhabited area of the city. All of the properties where the ovitraps were installed were georeferenced with the use of a GPS device (Garmin Ltd, USA#MP62SC) with UTM—SIRGAS 2000 projection. The coordinates were inserted into a geographic information system (GIS) in the software QGIS 2. 16. 2, where each property had its spatial location identified as a point (attribute) in a layer (shape) of the city. Initially, collected pallets were sent to the Itacoatiara entomology laboratory, placed to dry and analyzed under a stereoscopic magnifying glass for Aedes spp. eggs count. Once every 2 months, the positive pallets were sent to the entomology laboratory of the Fundação de Vigilância em Saúde do Amazonas (FVS-AM), where the pallets were individually immersed in glass flasks containing 200 mL of dechlorinated water for egg hatching. The larvae were raised until the third stage when species were identified and separated into pools of a maximum of 30 specimens, placed in cryotubes, and sent frozen to Instituto Leônidas e Maria Deane (ILMD) –Fiocruz Amazônia, where they were kept in a -80°C freezer, until molecular analysis for arboviral RNA detection. Firstly, each larvae pool was spiked with the Escherichia coli bacteriophage MS2 (ATCC 15597-B1) to be used as an internal positive control (IPC) using the same conditions previously described [15]. All pools were individually disrupted in 2 mL microtubes containing a 5 mm stainless-steel bead and 250 μL of TRIzol Reagent (Invitrogen, USA, #15596026) with the aid of the TissueLyser LT bead mill (Qiagen, Germany, #85600), 50Hz for 5 minutes. The reservoir containing the microtubes was frozen and kept in an ice bath during the process. Posteriorly, the homogenized pools were clarified by centrifugation, and the supernatant was added with 750 μL of Trizol. Therefore, RNA extraction followed the manufacturer’s recommendations. The RNA pellet was resuspended in 40 μL of nuclease-free water and evaluated for quantity and quality with the BioDrop DUO UV spectrophotometer (BioDrop Ltd, United Kingdom, #80-3006-61). A total of 2. 5 μL microliters of the extracted RNA was used for the RT-qPCR assays. The remaining volume was stored at -80°C for further analyses. All samples were evaluated for the detection of three arboviral RNAs by RT-qPCR in a StepOnePlus Real-Time PCR System (Applied Biosystems, USA, #4376598) located at the Real-Time PCR Platform of ILMD. The protocols used for the detection of each virus were previously published: DENV [16]; CHIKV [17]; and ZIKV [18], but we conducted the assays with some modifications. All probes were used at a final concentration of 0. 1μM, whereas all primers were used at a final concentration of 0. 3μM. All reactions were performed with TaqMan Fast Virus 1-Step Master Mix (Applied Biosystems, USA, #4444432), following manufacturer’s recommendations, except for the number of cycles that was increased to 45. For each lot of analyzed samples, three blank reactions (nuclease-free water as the template) and external positive controls (RNA extracted from viral culture) were included. The number of viral copies in each positive sample was estimated by RT-qPCR using absolute quantification by the standard curve method and reported as viral RNA copies/μL (of the eluted RNA). With the objective to evaluate the ZIKV replication in naturally infected larvae, we conducted the same RT-qPCR assay used for detection of ZIKV RNA, but in a two-step protocol targeting the negative-strand RNA. Firstly, two different cDNA assays were made: I) with only the reverse primer (which hybridizes to positive-strand RNA) and II) with only the forward primer (which hybridizes to negative-strand RNA). All cDNAs were made with SuperScript IV Reverse Transcriptase (Invitrogen, Carlsbad, CA, #18090050), according to the manufacturer’s recommendations. Subsequently, a qPCR assay was conducted with TaqMan Fast Advanced Master Mix (Applied Biosystems, USA, #4444558), following manufacturer’s recommendations, except for the number of cycles that was increased to 45. Three different master mixes we used: I) with only the reverse primer; II) with only the forward primer and III) with both reverse and forward primers. The positive pools were submitted to a conventional RT-PCR amplification protocol using the primers ZIKA_ASIAN_FNF1 (5’–CCGCGCCATCTGGTATATGT– 3’) and ZIKA_ASIAN_FNR (5’–CTCCACTGACTGCCATTCGT– 3’) designed to target the NS5 coding region of Asian ZIKV lineages. For the CHIKV sample, we used a protocol already described for alphaviruses amplification [19]. Thereafter, amplicons were precipitated with PEG-8000 and submitted to the nucleotide sequencing reaction with BigDye Terminator v3. 1 Cycle Sequencing Kit. Capillary electrophoresis was conducted in an ABI3130 sequencer located at the genomics platform of ILMD, Fiocruz Amazônia. The final FASTA sequences were initially submitted to BLAST analysis [20] and further evaluated by a web-based Dengue, Zika and Chikungunya Subtyping Tool (Version 1. 0), freely available at http: //bioafrica. mrc. ac. za/rega-genotype/html/index. html. The virus infection rate (IR) was calculated with PooledInfRate, version 4. 0 by Biggerstaff, a Microsoft Excel add-in that computes the IR using data from pooled samples (even with different pool sizes) by both Minimum Infection Rate (MIR) and Maximum Likelihood Estimation (MLE) methods. Freely available at https: //www. cdc. gov/westnile/resourcepages/mosqsurvsoft. html. Between January and April 2016,154 larvae pools containing 2,057 specimens (1,793 Ae. aegypti and 264 Ae. albopictus) were analyzed for CHIKV, DENV, and ZIKV RNA. By using the RNA extraction method described under Material and Methods we were able to obtain high quality RNA for most samples (260/280 median value: 1. 96; IQR: 1. 87–2. 01) and all IPC reactions were positive (Ct median value: 30. 9; IQR: 30. 0–31. 4). One pool containing only one larvae, obtained from the trap P056/ITA, collected at 15-Feb-2016, was positive to ZIKV (47 viral RNA copies/μL, approximately 1. 8 x 103 ZIKV RNA copies in the infected larvae). Another pool containing three larvae, obtained from the trap P026/ITA, collected at 25-Feb-2016, was positive to CHIKV (2 viral RNA copies/μL) Fig 1. The positive ZIKV pool was further evaluated for productive infection by the specific detection of the negative-strand RNA. We detected amplification for both cDNAs, derived from the positive or negative-strand RNA, with a positive-to-negative strand RNA ratio of approximately 2: 1, represented by the difference of one Ct (Table 1). Subsequently, conventional RT-PCR amplification was performed on both positive pools as described in Material and Methods, but unfortunately, only the Zika sample could be amplified. The final ZIKV sequence corresponds to a fragment of 450bp that was deposited in GenBank under the accession number MG279550. The analysis of the viral genotyping confirmed the Asian ZIKV lineage. The virus IR per 1,000 mosquitoes tested was calculated for CHIKV and ZIKV based in the total of pools and individuals tested. The CHIKV IR was 0. 45 for both MIR and MLE methods (MIR: lower Limit = 0. 00; upper limit = 1. 32 and MLE: lower Limit = 0. 03; upper limit = 2. 15). The ZIKV IR was 0. 45 for MIR (MIR: lower Limit = 0. 00; upper limit = 1. 32) and 0. 44 for MLE (MLE: lower Limit = 0. 03; upper limit = 2. 15). The first report about the detection of Ae. aegypti in Manaus, the capital of Amazonas State, was in November 1996 and for Ae. albopictus in September 1997. Since then, Aedes spp. mosquitos began to be found in other municipalities of the interior of the Amazonas State, and successive Aedes-related arboviral infections have been reported [21–25]. Between 2015 and 2016, a total of 878 suspected cases of CHIKV and 4,485 cases of ZIKV were reported in the Amazonas State [26,27]. Preventive and reactive measures regarding the vector control were carried out by the health authorities under the coordination of FVS-AM, to decrease Aedes infestation in different municipalities. In addition to the routine actions, ovitraps were installed with the purpose of directing the field efforts more efficiently. The primary aim of our study was to evaluate natural arboviruses vertical transmission in the field. Since this study was conducted under circumstances where the cold-chain could not be guaranteed, we decided to use a study design that favored the molecular detection, protecting viral RNA as soon as possible. Therefore, we decided to disrupt the Aedes larvae directly in Trizol reagent, which consists of a solution of phenol and guanidinium isothiocyanate that concurrently solubilizes biological material and immediately inactivates RNases [28], but also inactivates viral particles, preventing the possibility of viral isolation. The present study confirms other studies demonstrating natural vertical transmission of CHIKV [29–31], in a municipality with no human case previously confirmed. It is important to emphasize that most acute febrile cases in Brazil are diagnosed by clinical examination, without specific laboratory confirmation, especially during an ongoing outbreak. At the time we collected our samples, a Zika outbreak was already established, which could compromise the diagnosis of chikungunya cases. Although the CHIKV positive sample was amplified in duplicate in the probe-based RT-qPCR assays, we were unable to amplify this sample using conventional RT-PCR, preventing its sequencing. RT-qPCR would have been more sensitive compared to conventional RT-PCR, particularly for samples with low viral load. Another important point is that the sequence variation at the primers sites that may also decrease the efficiency of nucleic acid amplification methods. In an experimental study of vertical ZIKV infection, a total of 69 pools of Ae. aegypti adult mosquitos (F1) were tested, and six were positive in an indirect immunofluorescent antibody assay [32]. While other studies have demonstrated the natural vertical transmission of DENV [11,33–37], the detection of ZIKV in naturally infected larvae had not yet been described. We report the first detection of Zika virus vertical transmission in an Ae. aegypti larvae under the natural conditions found in the field. Therefore, we evaluated if there was a productive infection in the Ae. aegypti larvae by strand-specific amplification of viral RNA. ZIKV is a positive-sense, single-stranded RNA virus that belongs to the Flaviviridae family. During flaviviruses replication, a complementary negative-strand RNA is produced, which is used as a template for synthesizing new positive-strand RNA copies. Viral replication progresses asymmetrically, producing more positive-strand than negative-strand RNA. The positive-strand RNA molecules are packaged into the virions, acts as templates for viral protein production and promotes evasion of innate cell response [38,39]. Therefore, the specific detection of negative-strand flavivirus RNAs is an indicator of active viral replication, and different studies have been using this approach [40–43]. In this study, ZIKV cDNA was produced with the reverse primer, which hybridizes with the positive-strand RNA, e. g. , the genomic RNA found into the virion particle and also detected during viral replication, or with the forward primer, which hybridizes with the negative-strand RNA, only found during replication. The qPCR results clearly showed that both cDNAs were amplified in a positive-to-negative strand RNA ratio of approximately 2: 1. To the best of our knowledge, there is only one plausible explanation that could explain the detection of ZIKV RNA in the larvae, besides a canonical viral infection. Recently, different studies showed that naked viral RNA from hepatitis C virus (HCV), as well as from human pegivirus, two members of the Flaviviridae family (genus Hepacivirus and Pegivirus, respectively), may spread infection in exosomes vesicles [44,45]. Importantly, these studies showed the phenomenon in vitro, and further investigations are required to prove if similar events occur during ZIKV infection in vivo. Furthermore, the same studies also showed that, regardless of the way of viral RNA release in the cytoplasm of newly “infected” cells, productive viral RNA replication was observed, leading to the release of infectious particles. Given this, although our study design does not allow us to assert if ZIKV RNA reaches larvae cells by a classic route of viral infection, the specific detection of the negative strand RNA provides substantial evidence that active viral replication has occurred in Ae. aegypti naturally infected larvae. We detected arboviral RNA in larvae, which does not necessary means that larvae would become infected adults due to possible transstadial loss of infection during development to adulthood. On the other hand, if some of the infected larvae achieve maturity still infected, they will be readily able to transmit ZIKV to other mosquitos by venereal transmission [46] or, in the case of females, to human hosts. Thus, this phenomenon may contribute to the epidemic potential of this arbovirus because mosquitoes that emerge as virus-infected adults will have more opportunity to transmit virus than mosquitos that become infected after blood meal in an infected vertebrate. Further studies are necessary to evaluate all variables contributing to maintaining a virus circulating in a specific area until the number of new susceptible human subjects raise, by immigration or births, sufficiently to support a new epidemic cycle. According to previous work, the rapid detection of arbovirus in specimens collected in the field may contribute to the effectiveness of vector control measures, decreasing the viral transmission among the human population [47]. Altogether, the results showed in the present manuscript strengthen the importance of continuous monitoring of arboviral infections in both mosquitoes, as well as in human hosts, before the establishment of a new outbreak.

The control of the vast majority of arbovirus infections relies on entomological measures to reduce mosquito infestation. Therefore, this study analyzed the use of ovitraps for arboviral surveillance in a mid-size city of the Amazonas state, Brazil. We found one larva pool infected with chikungunya virus, before the first human case confirmed in this municipality. Another pool was infected with Zika virus, demonstrating the first evidence that vertical transmission occurs in naturally infected Aedes aegypti mosquito populations.

Abstract Introduction Materials and methods Results Discussion

invertebrates medicine and health sciences pathology and laboratory medicine togaviruses chikungunya infection pathogens tropical diseases microbiology geographical locations animals alphaviruses viruses developmental biology chikungunya virus rna viruses neglected tropical diseases insect vectors infectious diseases south america aedes aegypti medical microbiology microbial pathogens life cycles arboviral infections disease vectors insects brazil arthropoda people and places mosquitoes eukaryota arboviruses flaviviruses viral pathogens biology and life sciences viral diseases species interactions larvae organisms zika virus

2018

Evidence of vertical transmission of Zika virus in field-collected eggs of Aedes aegypti in the Brazilian Amazon

Despite several leprosy control measures in Nigeria, child proportion and disability grade 2 cases remain high while new cases have not significantly reduced, suggesting continuous spread of the disease. Hence, there is the need to review detection methods to enhance identification of early cases for effective control and prevention of permanent disability. This study evaluated the cost-effectiveness of three leprosy case detection methods in Northern Nigeria to identify the most cost-effective approach for detection of leprosy. A cross-sectional study was carried out to evaluate the additional benefits of using several case detection methods in addition to routine practice in two north-eastern states of Nigeria. Primary and secondary data were collected from routine practice records and the Nigerian Tuberculosis and Leprosy Control Programme of 2009. The methods evaluated were Rapid Village Survey (RVS), Household Contact Examination (HCE) and Traditional Healers incentive method (TH). Effectiveness was measured as number of new leprosy cases detected and cost-effectiveness was expressed as cost per case detected. Costs were measured from both providers' and patients' perspectives. Additional costs and effects of each method were estimated by comparing each method against routine practise and expressed as incremental cost-effectiveness ratio (ICER). All costs were converted to the U. S. dollar at the 2010 exchange rate. Univariate sensitivity analysis was used to evaluate uncertainties around the ICER. The ICER for HCE was $142 per additional case detected at all contact levels and it was the most cost-effective method. At ICER of $194 per additional case detected, THs method detected more cases at a lower cost than the RVS, which was not cost-effective at $313 per additional case detected. Sensitivity analysis showed that varying the proportion of shared costs and subsistent wage for valuing unpaid time did not significantly change the results. Complementing routine practice with household contact examination is the most cost-effective approach to identify new leprosy cases and we recommend that, depending on acceptability and feasibility, this intervention is introduced for improved case detection in Northern Nigeria. Leprosy is a communicable disease caused by a bacillus, Mycobacterium Leprae, which can lead to permanent disability among sufferers with significant psychosocial and economic burden. The disease causes skin lesions and nerve damages which progress to deformities of the eyes, hands and feet [1]–[2]. These physical disabilities are the prominent features of the disease which impose stigmatisation and discrimination on the sufferers [3]. The negative impact of leprosy due to stigmatisation is more than most other infectious diseases, making sufferers with physical disability dread stigmatisation and discrimination by the society [4]. In recognition of the burden of leprosy the World Health Organisation (WHO) in 1991 set a goal to eliminate the disease by 2000, defined as reducing the prevalence to less than one case per 10,000 populations [5]. Key targets of strategy for the reduction of burden of leprosy are the timely detection of new cases and prompt treatment with Multi Drug Therapy (MDT), which is the standard treatment for leprosy [6]. Since the progression of leprosy prognosis is insidious, taking an average of 2–5 years to manifest due to slow growth and multiplication of M. Leprae, early identification of the disease is very critical for effective control. This reduces both transmission of M. leprae and prevents disability. Nigeria achieved WHO global leprosy elimination target of less than 1 case per 10,000 in 1998, but the country remains among those that still report relatively high number of registered leprosy cases [7]. According to the NTBLCP Annual Report 2009, implementation of MDT in the country as the strategic intervention to eliminate leprosy resulted to a rapid decline of registered cases from approximately 200,000 in 1989 to 6,906 in 2008 [7]. Disability Grade 2 (DG2) among the new cases were 14% while child proportion was 12%. The National Strategic Plan for Tuberculosis and Leprosy Control in Nigeria [8] is in line with the WHO' s Enhancing Global Strategy for Further Reducing Disease Burden Due to Leprosy 2011–2015 [9]. The plan set the goal for reducing DG2 cases by 35% from 2010 figures by the end of 2015. Objectives include timely case finding and treatment, monitoring and limiting the progress of DG2 to not more than 5% annually. The plan recommends cost-effective approach to leprosy control interventions to ensure achievement of set goals. Several measures have been in place to control the disease in Nigeria, with support from international agencies such as the Netherlands Leprosy Relief (NLR) agency and other International Federation against Leprosy (ILEP) members. However the fact that indicators such as child cases remain high at 10% while DG2 stay at 14%, higher than target [7], [8], raises concern about hidden cases and continuous spread of the disease as well as the effectiveness of current detection methods in identifying early cases. This calls for the review of various detection methods to ensure more effective control of the disease in the region. Existing strategies include the routine/passive case detection (PCD), active case finding through the mini Leprosy Elimination Campaign (Mini- LEC) and contact examination. Published studies suggest that effectiveness of different detection methods vary between settings and contact levels. For instance the effectiveness of HCE varies between countries and this tends to depend on leprosy endemicity and definition of ‘contact’. In high endemic communities in Indonesia, it was found that almost 80% of new cases could be defined as contacts [10]. In Orissa 72 additional cases could be detected through HCE of 400 index cases [11] in higher endemic areas, although in Bangladesh spatial clustering of new patients at household levels could not be clearly established [12], while in low endemic areas relatively higher proportions of new cases could be found among household contacts of index cases [13]. The RVS strategy has played important roles in India, Indonesia, China and Thailand [14], [15], [16], [17]. As part of the mini-LEC in Indonesia, it detected twice higher leprosy prevalence than routine programme activities [15]. It is known to detect cases early which mean shorter treatment for PB cases and less number of disabilities and leprosy reactions. Evidence of Traditional healers' contribution to leprosy case detection has also been documented [18]. They have gained the confidence of the community given their level of successes and affordable costs of treatment. Due to stigmatisation status of leprosy many suspected carriers would rather visit traditional healers for privacy than public health facilities or programmes. Hence, the need to collaborate with recognised practitioners to improve identification of leprosy cases in the community. The objective of this study was to establish the cost-effectiveness of the three alternative leprosy case detection strategies in comparison to current practice with the aim to identify more efficient method for achieving the goals of the national programme in limiting child proportion and disability grade 2 cases. The study was conducted in Adamawa and Gombe states, two neighbouring north-eastern states of Nigeria with a combined population of over 5. 8 million people in 2009, based on the 2006 population projection. The people are subsistent farmers and nearly 50% of the population are below 15 years old. There are over 1905 health facilities which include 2 Federal Medical Centers, 1 dermatology hospital, (which serves as a Leprosy referral center in the area), 1 specialist hospital, 15 general hospitals, and over I100 primary health centers (PHCs). Only about 11% (195) of the health facilities are private while the rest are public. 127 facilities provide leprosy services in Adamawa state while the dermatology hospital serves as the referral center. In line with national guideline Leprosy services are combined with Tuberculosis as National Tuberculosis and Leprosy Control Programme (NTBLCP) established in 1988. Both services are supported by the Netherland Leprosy Relief agency. Although leprosy prevalence rate in the area in 2009 was similar to the national rate of less than 0. 5 case per 10,000 population there exists the presence of high and low endemic communities. Eight of the 32 Local Government Areas (LGA) making up the study states registered prevalence rates of more than 1 case per 10,000 population in 2009 [19], similar to new case detection rate (registered incidence). The two states have DG2 cases and child proportion higher than the national targets of 5% among new cases [8]. In Gombe state alone child proportion was 13. 4% in the same period [19] indicating possible active transmission of leprosy in these communities. A cross-sectional study was designed across the high- and low-endemic communities of the states to evaluate a one year operational cost-effectiveness of the different case detection methods based on retrospective data available from 2005. The communities represented by the 32 LGAs were first categorised as either high-endemic or low-endemic using prevalence data from the NTBLCP 2009. Communities with registered prevalence rate of more than 1 case per10,000 population were categorised as high-endemic while those with less than 1 case per 10,000 population were categorised as low-endemic. Five communities were randomly selected from each endemic area. The ten sampled communities have a combined population of about 2,036,400. Selection was carried out to ensure geographical balance across the study area. Other criteria for selection were access to clinics providing MDT services for confirmed leprosy cases, data availability and comparable socio-economic status. Data was collected between 2005 and 2011. Effectiveness was measured as the number of new cases detected over a one year period and outcome expressed as cost per case detected. Case detection or finding methods are categorised as passive and active methods. The passive case detection method, which constitutes routine practice in this study generally involves voluntary or self reporting by patients with suspected cases, and the active case finding methods at which health personnel visit patients at homes (contact tracing) or vicinities. Each of the detection strategies was implemented as a complement to routine method. The following methods were evaluated and compared. This represents the routine leprosy case detection method which integrates leprosy detection and control into the general health care system. It involves self-referral or referral of suspected cases by local health workers to leprosy units or peripheral health centre for examination by specialised health workers. The method involves the engagement of healthcare workers for the provision of leprosy services, regular staff training and supervision by state and local government Tuberculosis and Leprosy Staff (LGTBLS and STBLS), visits to health facilities (voluntary reporting) by patients, social mobilisation and health education. For our study contacts of index cases who share the same house and a kitchen are examined, and does not include neighbours and social contacts. It is characterised by the following activities; Key activities in RVS include; Practitioners were identified through the Traditional Healers Association. whicu is recognised by government. They were trained to identify and refer suspected patients (based on presentation of skin conditions) to recognised centers for leprosy diagnosis. Two types of financial incentives were provided as motivation for the services. N50 ($0. 33) was paid for each referred suspected case in the first instance while N500 ($3. 30) was paid when leprosy was confirmed. The initial lower amount was paid as an incentive to encourage them to refer suspected skin conditions. It was made minimal to discourage abuse by excessive referrals. Activity summary include The study was carried out as part of routine leprosy control programme with no major contact with patients, requiring no ethical clearance. However the rights of the clients to supply information on family contacts to be examined were respected based on signed consent notes by the index patients. Agreement to participate in the questionnaire survey to estimate patients cost of seeking care was considered as consent. Cost data were collected from expenditure records and reports. The costs of implementing services in all three methods were identified and measured using the ingredient approach, complemented by activity-based data. Bottom-up approach was used to estimate the economic costs, (where information on resource use and costs were available [20]) which involved identification and valuation of all resources required in the detection of new leprosy cases from the provider' s perspective, as well as from the patient and family perspectives. Where detailed information was not available, top-down calculations was performed [21]. Table 1 presents sources of cost data and the method of estimation. Resources were first classified as capital and recurrent costs. Capital costs, which are used up beyond one year were obtained by annualization of the capital items over their expected life-span. All costs were converted to the US Dollar ($) at the 2010 exchange rate, ($1 = N152). Major areas of resource use include personnel, training/workshops, social mobilisation, transport, incentives. Personnel costs (staff salary and allowances) were based on the proportion of health worker/staff time devoted to leprosy case detection and allowances paid in the process, and these include doctors, nurses, healthcare workers etc. Salary data was collected from standard Nigerian payroll scale. Training and workshops comprised of short-term (recurrent) and long-term (capital) trainings and workshop costs. Costs of social mobilisation included such items as advocacy visits, Information, Education and Communication (IEC) materials, radio/TV adverts and promotion. Capital items included vehicles, motorcycles, long-term training and start-up costs. They were annualised over their useful time periods and discounted at 3% rate based on World Bank recommendation, capturing their depreciated costs as opportunity costs of time. The routine programme costs consisted mainly of personnel, training and workshop (short and long-term), social mobilisation, vehicles and patient/family costs. The main cost elements for the RVS were allowances paid to field staff, training and workshop, social mobilisation and incentives provided in the form of de-worming agents given out to encourage attendance. No patient/family cost was recorded for RVS and HCE methods except as part of routine practice. HCE generated costs mainly from transportation and long-term training of field workers. Major cost elements for the THs approach were start-up expenditures (such as advocacy visits and mobilisation), long-term training for the traditional healers, allowances paid to field staff and the incentives paid out for referred cases by the traditional healers. However accuracy of these allocations was subject to availability of reliable data because resource use documentation was not detailed and properly defined (non-specific). Greater efforts were made at identifying and separating cost items for allocation to appropriate categories for analysis. From patient and family perspective, costs incurred by the patients at the household were collected from a structured interview which targeted 50 outpatients from hospital/health facilities in the area of study. These costs included direct out-of-pocket expenses incurred in transportation and hospital/diagnostic fees (including multiple visits to obtain diagnosis). These costs include indirect costs of travel time and hospital (waiting) time. Hospital fees were not charged for leprosy at the time (data not available). The cost of time (time loss) was based on the prevailing minimum subsistence wage rate in Nigeria during the period of analysis. Newly approved rate was however used in sensitivity analysis to assess the impact on the study results. Most of the case finding activities particularly for HCE and RVS methods involved field activities which included the simultaneous provision of TB services. Costs incurred in the process were treated as shared costs. Consequently cost data for many services were adjusted for leprosy programme at proportions that reflected the level of resource use for leprosy case detection. For instance 5% of the cost of training general health care workers (GHCW) was assumed for leprosy case detection based on interviews with the workers. Similarly 30% and 40% personnel costs were estimated for state and local government controllers and supervisors respectively for case detection based on shared activities with TB and other leprosy services. Step-down approach was used to measure and allocate shared costs. Major proportions were however varied in a sensitivity analysis to explore the impact on the result. Some cost items are generated at the beginning of some activities which are one-off and therefore expected to last for longer than one year over the life of the project. Such expenditures include trainings and the purchase of new equipments. Effectiveness data was measured in terms of number of newly confirmed leprosy cases detected within the period. These comprise of all patients diagnosed and confirmed with either PB or MB leprosy including child cases and disability grades 1 and 2 cases. All the data were tabulated and analysed using the Excel spreadsheet (Version 2007). Costing, worksheets were first created to collect relevant items for each method. The sheets contain the lists of likely resources used by each method. Data from the worksheets were then entered into the spreadsheet already programmed to calculate the required programme costs based on standard methods. Comparison of costs and effects was based on incremental cost-effectiveness ratio (ICER). ICER seeks to identify an alternative that replaces an existing practice in the form of mutually exclusive option. It measures the additional cost that would be required to achieve more superior benefits (case detection) than the baseline [22]. This is defined as additional costs divided by additional benefits. This means that the differences in programme costs between each of the methods and routine case detection will be calculated and compared with differences in the new cases detected between each of the methods and routine PCD, to determine the ICER. Thus: Where ‘A’ stands for alternative method (in this case UC+RVS, UC+HCE or UC+THs) and UC stands for Usual Care (PCD, the routine practice). The method that yields the lowest ICER value is considered the most cost-effective alternative. Hence the study used the ICER criteria to identify the most cost-effective method to replace existing practice for leprosy case detection. However using average measure (average cost-effectiveness ratio, ACER) would mean that the method that produces the highest number of outcomes at a given/constant cost generates the lowest ACER, making it the most cost-effective and preferred option [22]. This is defined as total implementation costs divided by the total benefits. Using the average measure becomes relevant only when there is no existing practice and comparison between alternatives will be based on the ACER. The ACER values are however presented for this study for complementary and comparative analysis. Some parameters that showed certain level of variability with potentials of affecting the study results were subjected to one-way sensitivity analysis to assess their impact and determine the robustness of the results. They include the discount rate, personnel costs, allocation factors for shared costs, subsistent (minimum) wage. Personnel costs were included because they constituted a major cost component. The cost-effectiveness values were recalculated using different values of these parameters in the sensitivity analysis. Table 2 summarises new leprosy cases detected by each method. From routine practice, an annual average of 164 cases was detected between 2005 and 2009. This was made up of 10. 4% child cases and about 4. 9% DG2. There were 138 cases in endemic areas and 26 cases in non-endemic areas. The RVS within the same period generated an annual average of 182 new cases with about 9. 3% child cases and 6. 0% DG2. For the HCE and THs method which were assessed between 2010 and 2011,181 and 187 cases were detected respectively. As a major indicator for evaluating case detection and monitoring disease control progress [7], [9], the active detection methods identified more DG2 cases among the new cases than the routine passive method. The THs strategy detected the highest proportion of DG2 at about 8. 6%. On the other hand the proportion of child cases, a indicator of leprosy disease transmission [23] was most detected by the routine detection method at 12. 3%. MB cases accounted for over 82% of all the new leprosy cases. The summary of the annual programme costs estimated for each method is shown in Table 3 by category, presented for both the provider and patients/family perspectives. It shows for each method, the relative composition of resource input for detection of one leprosy case in the study area for a period of one year. The RVS strategy was shown to produce the highest cost input per patient, at about $210 per case. Personnel cost was the largest single item in all methods accounting for 57–65% of total costs, apparently due to high personnel requirements, especially for routine practice. It accounted for 65% of routine practice cost which recorded $198 per case. Training and workshop as well as social mobilisation were the other major cost items generating 10–21% of total resource inputs among the strategies. The high cost of social mobilisation was also due to the number of activities involved in mobilising communities for leprosy detection such as advocacies, education, communications etc. At $193 per case detected the HCE produced the lowest unit cost among the strategies with personnel responsible for 61%. Transport made the most significant impact at 9. 1% of HCE programme cost. THs produced the second lowest unit cost per case at about $197. Interestingly incentive was not a significant item in the THs method. Patient/family costs contributed only about 1% of the total programme cost per patient, generated only by routine practice and the THs strategies. Average transportation costs to and from health care facility was $1. 32 per patient while on the average a patient visits healthcare facilities 1. 5 times before being diagnosed for leprosy, ranging from 1 to 4 times. Average travel and waiting times totalled 92 minutes (61 min and 31 min respectively), which translates to approximately $0. 41 per patient based on the minimum subsistent wage rate during the period of study. New minimum wage rate was used to assess the impact in sensitivity analysis. The respective ICERs for the different options are presented in Table 4. It is presented for providers/patients perspectives at all contact levels, to guide appropriate implementation decisions. The ICER indicates the cost (in US dollars) for each additional leprosy case. For the HCE plus usual care strategy, additional $2,416 was expended to detect 17 additional cases during the period, resulting to $142 per additional leprosy case detected. It was most cost-effective in high- endemic areas at $81 per additional case detected than in low-endemic areas where it produced a total of $604 per additional case detected. The HCE was closely followed by the THs plus usual case strategy which produced additional 23 leprosy cases at a total cost of $4,447 to yield $194 per additional case detected. Results show that RVS was not cost-effective given by the dominance of the THs method which generated lower cost and higher effectiveness than the RVS. Findings were similar from both provider' s and patient' s' perspectives and at different contact levels. At both high and low-endemic areas, the HCE generated the lowest ICERs, making it the most cost-effective method. The RVS method was similarly not cost-effective in those areas as it was again dominated by the THs method. Average values produced similar results which show the HCE having the lowest ACER (Table 4). Results of varying major parameters on the CERs are presented in Table 5 which showed that personnel costs, shared costs proportions, subsistence wage, discount rate and diagnostic accuracy did not significantly alter study results although impact on programme costs were significant for changes in the shared costs proportions and personnel costs. When the shared costs proportions for leprosy detection were increased by 50%, programme costs for each strategy were increased by an average of 40% but the ICERs did not change. When the proportion was reduced by 50%, programme costs reduced by an average of 27% across the methods, with little or no change in the ICER results. However the cost-effectiveness of the RVS significantly improved by 27%, from being dominated to $221 per additional case detected. When the subsistent wage was increased from N7,000 to N18,900 in view of the current wage structure, the impact on the study results was insignificant. When the personnel cost was increased by 30% only the programme costs increased by about 18%, with no impact on the ICER results. Interpretations of the result need to be carried out in relation to some limitations of the study which is important for generalising the findings. Based on retrospective data from routine practice, it is subject to recall biases. Resource use data were not very detailed and specific, resulting in greater efforts at separating and allocating resources which, in many cases were arbitrary. The gap in data availability may have resulted in some costs not adequately captured. Hence cost allocations may not be very appropriate and accurate requiring the need for field implementation to obtain more accurate results, by correcting identified challenges. This was general for all methods. Diagnostic accuracy was also not very specific as separation of cases between the PCD and RVS methods was in some cases challenging due to poor data documentation, which again would be addressed in a field study. Analysis of effectiveness data did not consider future impact beyond a one year period as the mix of leprosy cases detected have varying benefits from preventing the progression to permanent disability. The higher the number of early disease cases the greater the potential benefits of being prevented from progression to permanent disability, with MDT treatment. The use of DALY as a measure of outcome will involve the measure of utility for each leprosy grade detected such that the higher the number of early cases the higher the utility and hence the benefits. Lack of data on utility values and inadequate information on the proportion or distribution of leprosy categories limited this approach. However this may not have changed the study findings since the methods yielded similar proportions of child and DG cases. Definition of HCE in the study was the narrowest, limiting contact of index cases to household members, excluding neighbours and neighbours of neighbours which have been shown to increase case detection [23], [24]. Hence broader definition of contact tracing would increase the cost-effectiveness of HCE. Lastly given the dynamic nature of leprosy transmission as a communicable disease, the study did not capture the benefits of prevention beyond one year, but was only based on analysis of primary prevention. A dynamic model would have captured the benefits of further prevention beyond the index cases but this was not possible as information on the long term impact of prevention is not available. This would also have increased the cost-effectiveness of the various methods, especially for HCE. In addition to the study limitations it would also be necessary to consider other factors such as economic differences between countries when generalising the study results. Programme costs will likely differ due to varying price levels of items such as personnel salaries and allowances, etc, leading to possible variation in cost-effectiveness results. In view of the findings from the study we make the following recommendations to enhance effective detection and control of leprosy in Nigeria. The study has shown that the Household Contact Examination is the most cost-effective approach for identifying new leprosy cases when implemented to complement routine practice in Nigeria. It offers the best option for scale-up for improved coverage having the lowest average cost per case detected. Broadening the definition of contact tracing will enhance detection of more leprosy cases for improved control. The combination of HCE with chemoprophylaxis would further enhance the cost-effectiveness of the intervention. Integration of the strategy into routine practice is therefore recommended for improved case detection of leprosy in Nigeria, depending on its acceptability and feasibility in the area. The finding also justifies the WHO recommendation for inclusion of contact tracing into routine case detection. However due to challenges in data collection (and generally in cost-effectiveness studies), large differences in results make more sense than the little differences recorded in this study. Hence the cost-effectiveness of the three methods can be considered similar and implementation of the methods in a budget neutral way may be necessary to maximise the detection of leprosy in the area, targeting specific areas where each method is considered more effective.

Reported increases in child proportions and disability grade 2 cases in Nigeria suggest that leprosy disease is still spreading in the country. This indicates the need to review case finding strategies to improve case detection for effective control of the disease. It was necessary that available methods be assessed for their value for money in view of limited resources. We evaluated the cost-effectiveness of three available leprosy case detection methods using data available from 2005 to 2011. We explored data to determine which strategy when implemented in addition to routine practice would detect most additional leprosy cases at a given cost (measured in U. S. dollars). Hence, cost-effectiveness was expressed as incremental cost-effectiveness ratio (ICER). Our findings show that at the rate of $142 per additional case detected, the household contact examination was the most costeffective strategy for detecting additional leprosy cases when implemented to complement routine practice.

Abstract Introduction Methods Results Discussion

medicine clinical research design

2012

Cost-Effectiveness Analysis of Three Leprosy Case Detection Methods in Northern Nigeria

Precise temporal and spatial control of cell division is essential for progeny survival. The current general view is that precise positioning of the division site at midcell in rod-shaped bacteria is a result of the combined action of the Min system and nucleoid (chromosome) occlusion. Both systems prevent assembly of the cytokinetic Z ring at inappropriate places in the cell, restricting Z rings to the correct site at midcell. Here we show that in the bacterium Bacillus subtilis Z rings are positioned precisely at midcell in the complete absence of both these systems, revealing the existence of a mechanism independent of Min and nucleoid occlusion that identifies midcell in this organism. We further show that Z ring assembly at midcell is delayed in the absence of Min and Noc proteins, while at the same time FtsZ accumulates at other potential division sites. This suggests that a major role for Min and Noc is to ensure efficient utilization of the midcell division site by preventing Z ring assembly at potential division sites, including the cell poles. Our data lead us to propose a model in which spatial regulation of division in B. subtilis involves identification of the division site at midcell that requires Min and nucleoid occlusion to ensure efficient Z ring assembly there and only there, at the right time in the cell cycle. Mechanisms that regulate cell division in time and space are of fundamental importance to biology because they ensure equal partitioning of DNA into newborn cells. Failure to do so can lead to cell death. The earliest observable event in cell division in rod-shaped bacteria such as Escherichia coli and Bacillus subtilis is the polymerization of the highly conserved tubulin-like protein, FtsZ, to form a contractile structure called the Z ring, at midcell [1]–[6]. The Z ring then recruits several division proteins to form the division complex, known as the divisome, to enable cytokinesis. In this way FtsZ acts as a so-called founder protein that recognizes a sub-cellular location, and instructs other proteins to assemble there through a series of protein interactions [7]. The key question concerning the regulation of cell division is therefore, what dictates the sub-cellular recruitment of this founder protein, FtsZ, precisely to midcell? For many years the paradigm for division site positioning in rod-shaped bacteria such as E. coli and B. subtilis has been that it is determined through the combined action of the Min system and nucleoid occlusion. Both systems negatively regulate Z ring formation by preventing Z rings forming anywhere in the cell except midcell. The Min system prevents Z rings assembling at the poles where there is little or no DNA, whereas nucleoid occlusion prevents Z rings assembling over the nucleoid or chromosome [1]–[3], [8]. It is generally believed that when chromosomes segregate, the DNA-free space created at midcell relieves nucleoid occlusion, allowing a Z ring to form precisely at this site [9]–[14]. In E. coli, the Min system consists of three different proteins: MinC, MinD and MinE [15], [16]. The B. subtilis Min system comprises four proteins: MinC, MinD, MinJ and DivIVA, with MinC and MinD being homologs of the corresponding E. coli counterparts [17]–[21]. In both organisms, MinC is the primary inhibitor of the system; and it appears to inhibit Z ring formation by interacting and destabilizing FtsZ polymers directly [16], [22], [23]. Recent evidence suggests that MinC inhibits FtsZ polymer assembly by preventing lateral interactions between FtsZ protofilaments in B. subtilis and E. coli [24], [25]. Spatial regulation of Z ring assembly by the MinC protein in both organisms relies on its dynamic localization throughout the cell cycle [26]–[28]. Nucleoid occlusion in B. subtilis and E. coli involves the Noc and SlmA proteins, respectively. These proteins appear to perform similar roles, but are not similar in protein sequence [29]–[31]. Both of them bind to DNA and inhibit Z ring assembly over the chromosome [29], [30], [32]–[34]. Noc and SlmA bind to specific DNA sequences that are particularly sparse or absent near the terminus region of the chromosome [32]–[34]. It is proposed that as the round of replication nears completion and the terminus region occupies a midcell location, Noc and SlmA move away from midcell as the bulk of the chromosomes segregate, allowing a Z ring to form there [32]–[34]. The critical role of both of these proteins appears to be in preventing guillotining of the DNA by the septum when chromosome replication or segregation is perturbed [29], [30]. Interestingly, neither the Min system nor the Noc/SlmA proteins are essential in E. coli or B. subtilis. Furthermore, it has been shown previously that MinC and MinD are not required for the precise positioning of Z rings at midcell in B. subtilis [35]. However in B. subtilis and E. coli cells deprived of both Noc or SlmA proteins and the Min system, Z rings form very infrequently and division is defective, resulting in long cells [29], [30]. It has been proposed that this is due to the ability of FtsZ to form polymers at multiple locations along the length of the cell, including over nucleoids, none of which can accumulate enough FtsZ to form a ring [29], [30]. Thus the prevailing belief is that Noc/SlmA and the Min proteins together are primarily responsible for identifying the division site in bacteria by allowing Z rings to form precisely at midcell [1]–[3], [8]. However, while it is clear that these proteins play an important role in influencing placement of the Z ring at the correct site, whether they are solely responsible for the precise positioning of the Z ring at midcell remains to be tested. Cumulative evidence suggests that mechanisms in addition to Noc/SlmA and the Min system are involved in positioning the Z ring at midcell. Firstly, in Noc− Min− B. subtilis and SlmA− Min− E. coli double-mutant cells overproducing FtsZ, division is partially restored and Z ring positioning appears biased towards inter-nucleoid spaces [29], [30]. The significance of this bias however has never been investigated and thus it is still not known to what extent the double-mutant defect actually affects Z ring positioning. Given that neither noc or minCD single mutants of B. subtilis affect Z ring positioning at midcell, it brings into question the exact degree to which the combined action of Noc and the Min system contribute to division site positioning. Indeed, previous studies point to a Noc-independent link between the early stages of DNA replication and division site positioning in B. subtilis [36]–[38]. In addition many bacteria do not have the inhibitory Noc (or SlmA) or Min homologues [1], [8]. The observations that Noc is not essential in Staphylococcus aureus, which does not contain a Min system [39]; and that two proteins in Streptomyces coelicolor, SsgA and SsgB, positively regulate FtsZ positioning between nucleoids during sporulation [40], highlights the distinct possibility that in other bacteria, including B. subtilis, other mechanisms are involved in establishing the position of the division site at midcell. Here we make use of the outgrown spore system to investigate the extent to which Min and nucleoid occlusion are responsible for positioning the Z ring in B. subtilis. We show that, remarkably, in the absence of both Min and Noc, Z rings are positioned precisely at midcell; although their assembly is delayed and less efficient. We also show that in the absence of both Min and any type of nucleoid occlusion, there is a substantial preference for Z rings to form precisely at midcell between two nucleoids rather than between the nucleoid and the pole. Our data strongly support a model in which correct positioning of the Z ring at the division site in B. subtilis is not solely determined by the combined effect of Min and nucleoid occlusion. Furthermore the data raises the possibility that the primary role of Min and nucleoid occlusion is to ensure the efficient utilization of the division site at midcell in B. subtilis by ensuring Z ring placement there, while other mechanisms are responsible for the actual identification of this site. In the absence of Noc and the Min system, Z rings do form in vegetatively-growing cells, albeit very infrequently [29]. However it remains unclear whether, in these long cells, Z rings form precisely at the division site under these conditions. The outgrown spore system of B. subtilis allows an accurate measurement of Z ring position during the first one or two cell cycles following germination. We therefore examined the precision of Z ring positioning in the absence of MinCD and Noc in outgrown B. subtilis spores. We created a minCD noc double-mutant strain (SU681), by introducing a minCD: : cat cassette into a strain containing a noc: : tet cassette (SU656; for all strains refer to Table 1). At 30°C, vegetatively growing cells of the minCD noc double-mutant were on average three-fold longer (11. 8±0. 54 µm, mean cell length ± SEM, Figure S1A and S1E) than wild-type cells grown under the same conditions (3. 8±0. 06 µm; Figure S1A and S1C). At 37°C, the double-mutant cells were even longer as previously reported [29] but in similar proportion relative to wild-type cells at the same temperature (4. 7±0. 08 µm and 18. 1±0. 9 µm, respectively; Figure S1B, S1D and S1F). Similar cell lengths were obtained when MinD was depleted in a Noc− background (data not shown). Spores of the minCD noc double-mutant strain (SU681), from now on referred to as the double-mutant strain, were obtained despite the lower sporulation efficiency compared to the wild-type strain, SU5 (data not shown). Using immunofluorescence microscopy we examined Z ring positioning in outgrown spores of the double-mutant strain at 34°C (the optimum temperature for spore germination), and compared it to wild-type cells (SU5) outgrown under the same conditions. If Min and Noc are solely responsible for determining the division site in this organism, then we would expect that no or very few cells will form Z rings precisely at midcell. At 120 min after germination, wild-type cells had an average cell length of 3. 20±0. 06 µm (Figure 1A). The vast majority of these cells (81%) contained a Z ring exclusively at midcell, with 91% of these positioned between 0. 45 and 0. 5 (Figure 1D and 1F). Z ring position is defined by the distance from the Z ring to nearest pole, divided by the cell length, with 0. 5 being exactly midcell. At the same time point, the double-mutant strain (SU681) had an average cell length of 5. 3±0. 12 µm (Figure 1A) and the frequency of outgrown cells with Z rings was extremely low (3. 5%; 12 out of 340 cells counted). Most remarkably however, 8 of these cells had a bright Z ring located precisely at midcell (see delineated cell in Figure 1G), with the other 4 cells having polar Z rings. Furthermore, several other cells had a distinct accumulation of FtsZ in their central region (see carets in Figure 1G). At 150 min of spore outgrowth, cells of the double-mutant strain reached an average length of 10. 9±0. 45 µm (Figure 1A) and the proportion of cells with Z rings increased significantly to 30% allowing us to measure Z ring positioning in a larger number of cells. Remarkably, at this time point, of the cells containing a Z ring, 60% had one exclusively at midcell (see triangles in Figure 1H). The remaining 40% had Z rings that were positioned either at the poles (5%), at positions other than the pole or midcell (27%) or had more than one Z ring (8%). The remainder of the cell population contained diffuse patterns of FtsZ that appeared as faint bands, helical-like structures, or accumulations of FtsZ as previously reported [29]. We measured the precise position of the medially-located Z rings in the double-mutant strain relative to wild-type Z rings at 120 min. We used 120 min for wild-type cells (SU5) for this comparison as this is when we first observe Z rings after germination. In the double-mutant strain, essentially all these Z rings were located between 0. 45–0. 5 (99%; Figure 1E; compare with Figure 1D). In other words, Z rings that assemble at midcell in the absence of both Min and Noc are at least as precisely positioned as those observed in wild-type cells. To determine the position of all the non-polar Z rings in the absence of both Noc and Min at 150 min, we plotted their position on a graph showing the theoretical division sites along the length of the bacillus rod; with 1/4 representing the second division sites and 1/8 and 3/8 representing the third division sites as described previously ([41]; Figure 1C). As expected, in wild-type (SU5) cells all Z rings were positioned at the 1/2 position (Figure 1B). Interestingly, in the double mutant the non-polar Z rings were either positioned at midcell or at the second division sites (Figure 1C). In other words, in the absence of both Min and Noc, there is a preference for Z rings to form at the first division site at midcell, as well as at the second division site. At 180 min the average cell length of the double-mutant increased to 13. 9±0. 74 µm (Figure 1A), and the majority (70%) of cells contained at least one Z ring (Figure 1I; see white triangle). These cells resemble the vegetative phenotype of the minCD noc double-mutant. Importantly, septation was often observed along these longer cells, showing that at least some of these Z rings are functional (data not shown). Moreover, these longer cells often contained a number of accumulations of FtsZ regularly spaced along their length (Figure 1I). These non-ring localizations of FtsZ are likely precursors to the Z rings at the division site [42], [43]. In summary, the above data establishes that in the absence of both Noc and Min, in addition to cells forming polar Z rings as expected, a significant number of Z rings can form at midcell as well as at future division sites. Furthermore, the precision with which the Z ring is positioned at midcell is unchanged in the absence of both Noc and Min. So, while these factors play an important role in determining where a Z ring will form in the cells, other mechanisms must exist to ensure precise placement of the Z ring at midcell in B. subtilis. While Z rings assembled precisely at midcell in the absence of both Noc and the Min system, their assembly was significantly delayed and resulted in much fewer Z rings at midcell compared to wild-type cells at specific times following spore germination. This delay was accompanied by the ability of FtsZ to accumulate at the second division site. To determine whether Z rings could form at midcell earlier in the cell cycle in the double-mutant strain, and maintain their precision in the absence of both Min and Noc, we increased the cellular level of wild-type FtsZ in the double-mutant during spore outgrowth. This was achieved by introducing a second copy of ftsZ as an IPTG-inducible Pspachy-ftsZ construct, integrated at the amyE locus of the chromosome of the double-mutant to give SU685, and in wild-type cells to give SU558. In a wild-type background, high cellular levels of FtsZ can inhibit division [44]. We therefore performed these experiments using the highest concentration of IPTG (0. 05 mM) that allowed essentially normal division of SU558 cells during vegetative growth (see Figure S2), and spore outgrowth in the wild-type background. FtsZ overproduction with 0. 05 mM IPTG significantly rescued the division phenotype of vegetatively-growing double-mutant (SU685) cells both at 30°C and 37°C (see Figure S2 and Figure S3). Western blot analysis confirmed that FtsZ was overproduced 1. 3-fold in both the double-mutant and wild-type strains containing Pspachy-ftsZ in the presence of 0. 05 mM IPTG compared to the equivalent strains that did not contain Pspachy-ftsZ (Figure S2E). FtsZ overproduction in the wild-type background (SU558) during spore outgrowth resulted in 64% of the cells in the population containing Z rings at 120 min. This is lower than when FtsZ is not overproduced in wild type cells (88%; Figure 2E). The reason for this decrease in Z ring frequency is not known, however it could be due to the observed assembly of midcell Z rings slightly earlier in the cell cycle (at 90 min rather than 120 min) under these conditions, leading to early division and subsequently an increase in the number of newborn cells in the population at 120 min that do not yet have a Z ring (data not shown). Importantly, of the wild-type FtsZ-overproducing cells containing Z rings (64%), the vast majority (89%) contained a Z ring exclusively at midcell (Figure 2A and 2E), with a small number of cells containing polar Z rings (1. 6%). Cells with Z rings located at the quarter positions, and cells with more than one Z ring represented 9. 4% of the total population of cells with Z rings (Figure 2E). In the double-mutant even at 120 min following spore germination FtsZ overproduction resulted in a dramatic increase in the frequency of cells containing a Z ring; from 6% with no overproduction, to 53% in the presence of IPTG (Figure 2E; compare Figure 2B to 2C and 2D). Of the cells containing Z rings (318 cells counted), 42% had a Z ring exclusively at midcell (Figure 2C and 2E), 28% a single polar Z ring (Figure 2E and 2C, see closed triangle), and 6% contained a Z ring at the quarter positions (Figure 2E; PDS). The remainder of these cells (24%) contained more than one Z ring - predominantly a Z ring at midcell and another Z ring at the pole (Figure 2D and 2E). Moreover, the precision of Z ring positioning at midcell in these FtsZ-overproducing mutant cells was the same as for wild type cells (only cells with a single medially-located Z rings were measured; compare Figure 2F and 2G). These results demonstrate that overproduction of FtsZ in the absence of both Min and Noc significantly reduces the delay in midcell Z ring assembly, and maintains the precision of Z ring placement at this site. In the above experiments Z rings were visualized using immunofluorescence, which does not preserve the nucleoid well. To confirm that Z rings actually formed between nucleoids in the double mutant when FtsZ is overproduced, we co-visualized the Z ring and the nucleoid in live cells of both the double-mutant and wild-type strains. We constructed both wild-type (SU492) and double-mutant strains (SU663) that contain a xylose-inducible copy of ftsZ-yfp located at amyE, allowing visualization of the Z ring in live cells in the presence of xylose. Western blot analysis confirmed that in the presence of 0. 5% xylose, total cellular FtsZ levels (FtsZ-YFP plus native FtsZ) were overproduced to approximately 1. 25-fold in the double-mutant (SU663) and wild-type (SU492) vegetatively-growing cells, relative to FtsZ levels in the equivalent strains not containing Pxyl-ftsZ-yfp (Figure S4D). As expected, during vegetative growth, while the addition of 0. 5% xylose did not cause any significant changes to the wild-type cell length, it did result in a significant rescue of cell division in the double-mutant (Figure S4A). Examination of double-mutant and wild-type cells overproducing FtsZ (as FtsZ and FtsZ-YFP) confirmed that all Z rings do indeed form between replicating nucleoids during spore outgrowth (Figure 3A and 3B; see triangles) and vegetative growth (Figure S4B and S4C). Moreover, non-ring localizations of FtsZ, that are likely precursors to Z rings at the division site [42], [43], were also readily observed between nucleoids in live cells of the double mutant (see carets, Figure 3C). As with overproduction of native FtsZ, overproduction of FtsZ using FtsZ-YFP rescued the timing of Z ring assembly, although to a slightly less extent than just FtsZ (data not shown). At 120 min, in the double mutant strain overproducing FtsZ-YFP (SU663), 34% of the outgrown spores had Z rings, 60% of which had a Z ring exclusively at the medial division site, 30% had a single polar Z ring and the remaining 10% had more than one Z ring. An analysis of Z ring positioning in live outgrown spores germinated in the presence of 0. 5% xylose at 120 min of outgrowth confirmed that in double-mutant cells that had a Z ring exclusively at midcell, it was positioned with wild-type precision (Figure 3D and 3E). In summary, these results show that overproduction of FtsZ in the complete absence of Noc and the Min system, can significantly rescue the observed delay in Z ring assembly at midcell in these cells, enabling Z rings to form at midcell earlier in the cell cycle and division to occur at midcell more efficiently. This suggests that while Min and Noc are not required for the precise positioning of a Z ring at midcell, they are needed for the efficient assembly of Z rings at this site, as well as preventing Z rings forming at other division sites. Our data show that in the absence of Noc and the Min system, midcell Z rings still form precisely at midcell. We considered the possibility that, in the absence of Noc and Min, another nucleoid occlusion factor ensures precise positioning of the Z ring at midcell [29], [45]. To test this, we developed an approach that allowed replicated nucleoids to separate substantially after a single round of chromosome replication, such that there is no DNA (no nucleoid occlusion) in the central region of the cell. We then induced FtsZ production to determine whether a Z ring would assemble there. We performed these experiments in the absence of the Min system, the Noc protein, or both. Assembly of Z rings precisely at midcell under these conditions would support the idea of a nucleoid occlusion-independent mechanism for Z ring positioning. Furthermore, preferential positioning of the Z ring precisely at midcell as opposed to the cell poles under these conditions would argue strongly for the identification of a specific site at midcell for Z ring assembly in B. subtilis that does not require Min or nucleoid occlusion. Our experimental approach is shown in Figure 4A (refer to legend). We used the temperature sensitive dnaB mutant, dna-1, to allow only one round of DNA replication during spore outgrowth. This mutation provides a complete block to initiation of DNA replication at the non-permissive temperature (48°C; [46]). Spores were germinated at the permissive temperature for long enough to allow initiation of one round of replication, and then shifted to the non-permissive temperature to prevent initiation of a second round of replication. This allows two replicated nucleoids to separate substantially prior to induction of FtsZ production. We also used an alternative approach following a similar rationale. This involved spore outgrowth at 34°C, but instead of inhibiting DNA replication using the temperature-sensitive DNA replication mutant dna-1, we added HPUra, a potent inhibitor of B. subtilis DNA polymerase III [47], at pre-determined time points to create a significant population of cells with two separated chromosomes, prior to inducing FtsZ production. This alternative approach was performed in RecA− cells to avoid the SOS response-mediated inhibition of Z ring formation and division [48]–[50]. The results we obtained were essentially the same as those obtained using the dna-1 approach described below (data not shown). Using the dna-1 approach shown in Figure 4A, for all strains containing the dnaB (ts) mutation, at least 40% of cells contained two replicated nucleoids that had separated to produce a clear DNA-free gap in the central region of the cell at 135 min, prior to induction of ftsZ expression (Figure 5A and Figure 4C–4F). The remaining cell populations contained either a single nucleoid (no replication) or more than two nucleoids (more than one initiation; data not shown). ftsZ expression was then induced from Pspac-ftsZ (controlling ftsZ expression at the native locus) with 1 mM IPTG to allow Z ring assembly. These cells also contained Pxyl-ftsZ-yfp (located at amyE) induced minimally (with 0. 01% xylose) to enable visualization of the Z ring in live cells. To ensure that nucleoid occlusion was completely relieved at midcell before Z ring assembly had begun, we induced ftsZ expression when the DNA-free gap in cells with two nucleoids was on average almost twice as large as it is in wild-type outgrown cells (SU492, DnaB+) that were grown out for 105 min at 34°C then shifted to 48°C for 30 min (0. 45 µm, see Figure 4B; Figure 5A; compare the DNA-free space in Figure 4B to Figure 4C–4F). As expected, in the absence of xylose and IPTG we very rarely observed Z rings (using immunofluorescence; Figure S5). For all four strains containing the dnaB (ts) mutation (Min+ Noc+, Min− Noc−, Min− and Noc−) we determined the number of cells that contained Z rings and the frequency of cells with two separated nucleoids containing a single Z ring (at 165 min; Figure 4A; Figure 5B). In Min+ Noc+ dnaB (ts) cells (SU671), the percentage of cells in the population with Z rings was 66%. More than half of these cells contained a single Z ring and two nucleoids (Figure 5B), with 94% being positioned between the nucleoids (Figure 5B and Figure 6B). The remaining 6% were positioned between the nucleoid and the pole (Figure 5B and Figure 6C; for more examples refer to Figure S6A–S6D). The proportion of cells with Z rings is reduced to 24% when both Min and Noc are absent (SU680; Figure 5B). This is consistent with the data above showing that the combined absence of the Min system and Noc results in a significant reduction in the frequency of cells with Z rings. Cells with two nucleoids and a single Z ring represented more than half (>50%) of the number of cells containing Z rings (Figure 5B). Remarkably, when Noc and the Min system are both absent (SU680) the vast majority (87%) of Z rings are still positioned between nucleoids (Figure 5B and Figure 6D and 6E); with the remaining 13% positioned between the nucleoid and the pole (Figure 5B and Figure 6F). More examples of cells with two nucleoids and a single Z ring in these cells are shown in Figure S6E–S6M). This surprising result demonstrates that even when there is no DNA (nucleoid occlusion) in the central region of the cell, no Noc and no Min system, almost all the Z rings that form are positioned between nucleoids rather than between the nucleoid and the pole. We then determined whether the Z rings that formed in the double-mutant, dnaB (ts) cells are positioned at midcell with wild-type precision. As the control strain, we used wild-type outgrown spores expressing ftsZ-yfp (DnaB+; SU492) and outgrown for 105 min at 34°C and then shifted to the non-permissive temperature (48°C) for 30 min. Under these conditions the majority of cells (80%) had a single medially-located Z ring (Figure 6A), and 76% of these Z rings were positioned in the 0. 45 to 0. 5 region of the cell (Figure 7A). In Min+ Noc+ cells containing the dna-1 mutation (SU671) 75% of the Z rings were positioned within the 0. 45–0. 5 region (Figure 7B). In Min− Noc− cells (SU683) 74% of Z rings were positioned within this wild-type range (Figure 7E). A statistical analysis was performed (Kolmogorov-Smirnov test) to compare the positioning of Z rings (including those outside the 0. 45–0. 50 range) that formed between two separated nucleoids in Min+Noc+ dnaB (ts) (SU671) and Min− Noc− dnaB (ts) (SU680) cells relative to that of the wild-type cells replicating their DNA normally (SU492, DnaB+). There was no statistical difference in the distribution of Z rings (p>0. 05; data not shown) for both conditions relative to the wild-type control (SU492, DnaB+). These data strongly suggest that neither the Min system, Noc, or in fact any nucleoid occlusion is actually required for the placement of a Z ring precisely at midcell in B. subtilis. We also examined the effect of single deletions of noc and minCD under the conditions of our experimental approach described above (that is, in a dnaB (ts) mutant background). The results were very similar to the double-mutant (Figure 5B; Figure 6G–6J). In both single mutants however, the proportion of cells with Z rings in the population was slightly higher than in the double-mutant [34% and 31% in Noc− (SU683) and Min− (SU678) cells, respectively compared to 24% for the double-mutant (SU680); Figure 5B]; as was the percentage of Z rings positioned between two nucleoids (92% and 94% for single mutants versus 87% for the double-mutant strain; see Figure 5B). More examples of cells with two nucleoids and a single Z ring are shown in Figure S7. In these single mutant dnaB (ts) cells, the remainder of the Z rings were located between the nucleoid and the pole in Min− dnaB (ts) cells (Figure 6H) and over a nucleoid in Noc− dnaB (ts) cells (see Figure 6J). Again, there was no statistical difference (p>0. 05, Kolmogorov-Smirnov test; data not shown) between the position of Z rings (including those outside the 0. 45–0. 50 range) that formed between two separated nucleoids in Min− (SU678) and Noc− (SU683) dnaB (ts) strains relative to that of the wild-type replicating cells (SU492, DnaB+) (Figure 7C and 7D). Under these conditions immunoblotting analysis showed essentially the same cellular level of FtsZ for all four dna-1 strains, including the Min+ Noc+ strain, SU671 (data not shown). Thus the reduction in Z ring formation observed when Min and Noc or both are deleted from the dna-1 strain (SU678, SU683 and SU680), is not likely to be due to decreased cellular levels of FtsZ. Interestingly however, we did observe approximately 30%–40% less FtsZ in all four dna-1 strains at 49°C compared to wild-type (DnaB+) cells (SU5; data not shown). The reason for this is unclear. As with the double-mutant cells, in most single-mutant cells with two separated nucleoids that did not contain Z rings, very faint FtsZ helical structures, dots and accumulations of FtsZ were observed (see Figure S8). Collectively these results demonstrate that neither Min nor nucleoid occlusion is required for positioning the Z ring precisely at midcell in B. subtilis. Most remarkably, our data indicate that there is a substantial preference for Z rings to form precisely at midcell, as opposed to other DNA-free regions, in the absence of these systems. This strongly suggests the existence of an additional mechanism for identifying the cell centre in this organism. Thus the role of Min and nucleoid occlusion in Z ring positioning appears to be in ensuring efficient Z ring assembly specifically at the midcell site by preventing Z rings forming at inappropriate sites, as well as ensuring the overall efficiency of Z ring assembly. Our data show that relieving nucleoid occlusion at midcell beyond normal wild-type limits and prior to Z ring assembly has no significant effect on the precise positioning of Z rings at the division site at midcell. In wild-type cells replicating DNA normally (DnaB+), Z rings were observed at midcell within a small DNA-free gap between segregating nucleoids (0. 45±0. 01 µm) - this corresponds to relief of nucleoid occlusion at midcell (see Figure 4B and Figure 5A). If nucleoid occlusion has a role in determining precisely where the Z ring will be placed between two segregating nucleoids, one might expect that as the internucleoid (DNA-free) gap at midcell increases, the precision of Z ring positioning at midcell would decrease. To test this idea, we determined whether there was any correlation between the amount of relief of nucleoid occlusion (i. e. size of the internucleoid space) and position of Z rings between two separated nucleoids. We used the same cells in Figure 6 - this is shown Figure 8. The scatter plots reveal no obvious relationship between the size of the internucleoid distance and precision with which a Z ring is positioned at the division site. Essentially, for all conditions tested [Min+Noc+ (SU671), Min− (SU678), Noc− (SU683), and Min− Noc− (SU680) ] the majority of Z rings are clustered at positions 0. 45–0. 5, regardless of the internucleoid distance (Figure 8B–8E). A statistical analysis also showed that, while there was a very small negative correlative relationship between Z ring positioning and internucleoid distance (as internucleoid distance increases, midcell Z ring precision decreases) for all strains including the wild-type replicating cells (DnaB+, SU492; Figure 8A), this was not statistically significant (p>0. 05; Kendall' s tau coefficient and Spearman' s rho coefficient; data not shown). Thus, Z ring positioning does not appear to correlate with the amount of relief of nucleoid occlusion (i. e. size of the internucleoid gap) at midcell. This convincing result establishes that nucleoid occlusion (or its relief) is a weak determinant, if at all, in defining the position of the division site in B. subtilis. Thus, while nucleoid occlusion plays a dominant role in determining whether the Z ring forms at midcell or not, it does not identify the division site, nor is it responsible for the precision of Z ring positioning at this site. It has already been shown convincingly that both the Min system and nucleoid occlusion play important roles in positioning the Z ring in both B. subtilis and E. coli [1]–[3], [31], [53]. Our data are indeed consistent with this role and show that these two factors are absolutely required for the efficient formation of Z rings in B. subtilis, as well as for ensuring that Z rings form at midcell and not at other positions. However, several observations made here lead us to propose a model in which nucleoid occlusion and Min do not identify the correct division site at midcell per se, but ensure that the Z ring forms there and only there, at the right time in the cell cycle. This is illustrated in Figure 9. The distinct preference for Z rings to assemble at midcell, instead of the poles, in the absence of Min and Noc is entirely consistent with a previous study showing that in minimal medium almost all Z rings are positioned at midcell in the absence of minCD [54], and with a more recent study by Pogliano and co-workers that supports a primary role for the Min system in preventing polar (or secondary) Z ring assembly at new cell poles [28]. In doing so, it prevents FtsZ from being titrated to polar sites, thus contributing to the correct timing of cell division at midcell [28]. Thus, rather than acting to provide positional information regarding the actual site of Z ring assembly, we propose that the Min system in B. subtilis acts to ensure efficient and timely usage of this site. The significant decrease in Z ring formation when minCD was deleted in cells in which two replicated nucleoids were allowed to separate is consistent with this idea. Rudner and co-workers have recently shown that even in the absence of Noc, other nucleoid occlusion factors appear to prevent Z ring assembly over DNA [45]. Our findings support the existence of additional nucleoid occlusion factors since we rarely observed Z rings over nucleoids in the absence of Noc in all experiments. Our data however strongly argue against the involvement of any of these other nucleoid occlusion factors in defining the site for Z ring assembly at midcell in B. subtilis (see Figure 9). How then does a bacterial cell find its middle? One possibility is that a signal or structure marks the midcell position for Z ring assembly independently of nucleoid occlusion and the Min system [1], [37], [55]–[58]. This could be a protein or a physical change in the conformation or type of membrane lipids at midcell [59]–[61]. The “Ready-Set-Go” model linking identification (potentiation) of the midcell division site for Z ring formation with the progress of the initiation phase of DNA replication proposes such a signal [38]. The recent discovery of two proteins SsgA and SsgB that are required to target FtsZ to division sites during sporulation in Streptomyces coelicolor [40] highlights the presence in bacteria of factors independent of the Min system and nucleoid occlusion that identify the division site. Our data strongly support this possibility in B. subtilis, and the idea of a midcell marker for Z ring assembly definitely merits further investigation. The dramatic delay in midcell Z ring assembly observed in outgrown cells in the absence of both noc and minCD suggests an inability of these cells to restrict FtsZ accumulation to a single division site, resulting in the accumulation of FtsZ at second division sites, as well as at the cell poles. Consistent with this idea, we show that FtsZ overproduction partially restores both the efficiency (see also [29]) and the timing of midcell Z ring formation. This is a similar conclusion to that described recently for min mutants by Pogliano and co-workers [28]. To test whether Min and Noc function to maximise Z ring formation at midcell solely by preventing titration of FtsZ to second division sites, we attempted to determine the effect of different growth rates on the efficiency of Z ring assembly specifically at midcell in Noc− Min− cells. While it was not possible to interpret the results fully, the data did suggest that at slower growth rates when there are less division sites, the efficiency of midcell Z rings increases in the double-mutant cell population. However, since the average length of noc minCD double-mutant cells was still significantly longer than those of the wild type (data not shown), this does not fully explain the phenotype of the double-mutant cells. We therefore propose that even in cells that are growing slowly enough such that they contain only one division site competent for Z ring assembly, these two factors have an additional role in concentrating FtsZ to the central region of the cell, to allow a Z ring to form there. The possibility that Min proteins act away from the pole towards midcell in B. subtilis to prevent accumulation of FtsZ in non-central regions of the cell has been raised previously [29]. Two very recent studies [28], [62] have suggested that the cell poles are potential division sites in B. subtilis in the sense that Min appears to prevent the re-formation of the divisome at the nascent pole during septation. In this respect, potential division sites refer to midcell, future division sites at the quarter positions, and to cell poles. In other words, we propose that Min and Noc act together to prevent Z rings forming at all these potential division sites. During spore outgrowth in the absence of Noc and Min we observed very early accumulation of FtsZ at several division sites. Since Noc can prevent a significant number of Z rings forming precisely at midcell even without completion of the initiation stage of DNA replication [36]–[38] it seems likely that when Noc is absent, FtsZ can start to accumulate at these division sites very early in the round of replication, even before the elongation phase, and Noc plays a role in hindering midcell Z ring assembly over unreplicated DNA [38]. This role for Noc is consistent with recent data showing that it also appears to act as a subtle timing device for cell division in cells replicating DNA normally [34]. Recent evidence indicates the existence of other Noc-independent nucleoid occlusion factors that prevent premature Z ring assembly over DNA, at least during the elongation phase of replication [45]. Interestingly, we observed that, although delayed, the first Z rings that form in double-mutant outgrown cells replicating DNA normally were predominantly located at the medial division site (older division site; over 60%). Given that these longer-than-normal double-mutant cells already contained multiple chromosomes at the time of Z ring formation (and thus more than one division site) it appears that some factor is attracting Z ring assembly to the oldest (first) division site at the cell centre. This could occur through a putative positive signal for Z ring assembly (mentioned above) that is present at the first division site earlier than subsequent division sites, causing accumulation of FtsZ, and finally a functional Z ring at midcell. However in the absence of Min and Noc, very soon after this (at or after initiation of DNA replication) the second division sites become available and FtsZ starts to accumulate also at these sites (Figure 9). Since FtsZ accumulates at midcell earlier than subsequent division sites, there is always more FtsZ here than at other sites. The perhaps slower but constant accumulation of FtsZ at the midcell site in the absence of Noc and MinCD enables the concentration of FtsZ to reach the required level to form a Z ring first. We cannot exclude the possibility that the Min system and Noc function to ensure the efficient and rapid transitioning between the different helical-like states adopted by FtsZ in normal wild-type cells during the cell cycle, as described by recent models for Z ring assembly [42], [43]. This is an attractive idea and warrants further investigation. B. subtilis strains are listed in Table 1. The transformations to obtain strains produced as part of this work are also described in Table 1. All strains containing a minCD: : cat knockout cassette originate from SU561. In SU561 only the first 22 amino acids of minC and the final 23 amino acids of minD remain intact. The remaining portion of these genes was replaced by the cat resistance gene. All of these strains showed a minCD-deletion phenotype with cells having a 30–40% longer average cell length relative to wild-type and the production of minicells. All strains containing a noc: : tet knock-out cassette strain originate from the noc: : tet-containing strain 1284 [29]. Strain SU661 was obtained by congression of SU5 with SU46 chromosomal DNA. The presence of the dna-1 mutation was confirmed by sequencing the dna-1 open reading frame. Transformations involving the dna-1 mutation were also confirmed to be temperature sensitive for DNA replication. Bacteria were grown vegetatively in liquid PAB (antibiotic medium 3, Difco, USA) or semi-solid TBAB (tryptose blood agar base, Difco, USA) at 30°C or 37°C. Thymine where required was added to a final concentration of 20 µg ml−1 unless otherwise indicated. Spectinomycin (40–60 µg ml−1), tetracycline (10 µg ml−1), chloramphenicol (5 µg ml−1), erythromycin (0. 5 µg ml−1), neomycin (1–2 µg ml−1) and phleomycin (2 µg ml−1) were added as appropriate to TBAB plates or overnight cultures. To induce expression of amyE: : Pxyl-ftsZ-yfp, 0. 01% (w/v) xylose was included in the medium. To induce expression of ftsZ: : Pspac-ftsZ, 1 mM IPTG was included in the medium and for amyE: : Pspachy-ftsZ, different concentrations of IPTG were added as applicable. Spores were prepared and harvested as described previously [63]. Spore germination and outgrowth was performed with 2×108 spores ml−1 in PAB at 34°C [36], [37]. Antibiotic selection was not applied during spore outgrowth. For outgrowth in the experimental approach outlined in Figure 4A, spores of the dnaB (ts) strains SU671 (Min+, Noc+), SU678 (Min−), SU680 (Min−, Noc−), and SU683 (Noc−) were germinated in 3 ml of GMD [37] for 105 min at 34°C and then transferred to 48°C to prevent re-initiation of DNA replication and allow separation of replicated nucleoids. At 135 min (still at 48°C) 1 ml of culture was removed for visualization and measurement of the extent of DNA separation. To the remaining 2 ml of culture specific volumes of IPTG and xylose were added to a final concentration of 1 mM and 0. 01%, respectively, to induce ftsZ expression at the level of Pspac ftsZ and Pxyl ftsZ-yfp, respectively. This 2 ml culture continued incubation at 48°C for another 30 min (165 min in total) after which cells were collected for visualization of Z rings and nucleoids. Wild-type control strain SU492 (DnaB+) was outgrown at 34°C in GMD with 0. 01% xylose for 105 min and then transferred to 48°C for 30 min, to allow visualization of the first Z rings formed during spore outgrowth. For all other outgrowth experiments, spores were incubated at 34°C with shaking for the time indicated. Cells (1 mL) were collected and pelleted in a microcentrifuge. The pellet was resuspended in 300 µL PBS and cells were fixed with the addition of 700 µL cold ethanol (95% (v/v), 4°C). Cells were fixed at 4°C (24–48 h). The suspension was pelleted in a microcentrifuge and washed twice by resuspension and centrifugation in 200 µL PBS. The pellet was resuspended in 200 µL PBS. The fixed, washed cells (10 µL) were transferred to poly-L-lysine treated multi-well slides (ICN Biochemicals) and incubated at room temperature for 5 min. The excess liquid was aspirated off and the wells were washed once with PBS. The slide was mounted in PBS containing glycerol (50%) and the edges of the coverslip were sealed with nail polish. Immunofluorescence microscopy was performed using affinity-purified rabbit anti-FtsZ antibodies (raised against B. subtilis FtsZ) and Alexa 488-conjugated secondary antibodies (Molecular Probes, Invitrogen, USA) as described previously [42]. Live cells were visualized on 2% (w/v) agarose pads using Gene Frame (AB Genes) to create a flat surface on the glass slide. DAPI was added at a final concentration of 0. 4 µg ml−1 before placing the cells on the agarose pad. The same procedure was performed for co-visualization of the nucleoid and Z ring in live cells. All images were taken on a Zeiss Axioplan 2 fluorescence microscope equipped with a Plan ApoChromat (100×, NA 1. 4; Zeiss) objective lens or a UPlan Fluorite phase-contrast objective (oil immersion objective; 100×; NA 1. 3; Olympus) and a Zeiss AxioCam MRm cooled CCD camera. The light source was a 100 W high pressure mercury lamp passed through the following filters: for visualising Alexa 488 (Filter set 09, Zeiss), for visualising DAPI (Filter set 02, Zeiss; 365), and for visualising YFP (Filter set 41029, Chroma Technology). Image processing was performed using AxioVision 4. 5 or 4. 6 software (Zeiss) and scoring of these images was performed using AxioVision software. Image quality was improved by applying the cubic spline interpolation algorithm (to reduce pixelation) available on AxioVision software. Images were exported as TIFF files. Pixel to µm scaling for images collected by an AxioCam MRm camera were used to calculate cell lengths and other cell measurements from digital images. Numerical values for each measurement were exported from the AxioVision software as a text file, and imported into Excel (Microsoft) for final analysis. Excel was used to calculate mean, standard deviation, standard error of the mean (SEM) and the number of cells counted (n) for each data set, as well as for the production of scatter plots and histograms. We scored FtsZ localization as a ring if it was a transverse band oriented perpendicularly to the long-axis of the cell. FtsZ localization patterns that showed a band oriented diagonally to the long-axis of the cell were not scored as Z rings, and were not included in the Z position analysis. Z ring position was determined by measuring the distance from the Z ring to the closest cell pole divided by the total cell length, with 0. 5 being precisely midcell. A Z ring was considered to be at midcell if it was positioned within the range of 0. 45–0. 50, where essentially all Z rings formed in the wild-type background under the same conditions. The internucleoid distance in cells with two nucleoids was determined by measuring the distance between two regions of DAPI-stained nucleoids, from edge-to-edge. For all measurements, cells were considered for analysis and measurement only if completely contained within the field of view. For this reason, the average cell length of the noc minCD double-mutant cells should be considered an estimate only. Statistical analyses were carried out using SPSS software version 17 (IBM). Statistical analyses included the non-parametric “Kolmogorov-Smirnov” test, the “Kendall' s tau coefficient” and “Spearman' s rho coefficient”. The “Kolmogorov -Smirnov” test was used to compare the precision of midcell Z ring positioning in wild-type and mutant strains. The “Kendall' s tau coefficient” and “Spearman' s rho coefficient” were used to identify correlations between the size of internucleoid distances and Z ring positioning. All statistics were performed using a 95% confidence interval, where p<0. 05 indicates a statistically significant difference between the comparisons made. B. subtilis cells (10 mL of exponentially-growing vegetative cells; OD600 0. 4–0. 6) cell lysates were prepared as described previously (Harry et al. 1993). Sample volumes representing an equal loading of protein were loaded for SDS-PAGE into pre-cast NuPAGE 4–12% Bis-Tris gels (Invitrogen). The iBlot (Invitrogen) semi-dry transfer system was used, according to the manufacturer' s instructions. After Western transfer, the blot was incubated in blocking solution [PBS+5% (w/v) skim milk powder] for 2 h at RT with rocking. The blot was then incubated with a primary FtsZ antibody (raised in rabbit), diluted 1 in 10,000 in blocking solution, for 2 h. Next, the blot was washed three times in blocking solution containing 0. 05% (v/v) Tween 20 and incubated with HRP-conjugated secondary antibody (Promega) diluted 1 in 2,500 in PBS for 1 h. Band detection was carried out using enhanced chemiluminescence reagents (GE Healthcare) according to the manufacturer' s instructions, and the ChemiDoc XRS+ imaging system (Bio-Rad). Band intensities were analysed by densitometry using Quantity One software (version 4. 6. 1, Bio-Rad). To ensure quantification of protein band intensities within a linear range, a standard curve of was obtained using 2-fold serial dilutions of protein samples. Only bands within the linear range of the standard curve were quantified.

How organisms regulate biological processes so that they occur at the correct place within the cell is a fundamental question in research. Spatial regulation of cell division is vital to ensure equal partitioning of DNA into newborn cells. Correct positioning of the division site at the cell centre in rod-shaped bacteria is generally believed to occur via the combined action of two factors: (i) nucleoid (chromosome) occlusion and (ii) a set of proteins known collectively as the Min system. The earliest stage in bacterial cell division is the assembly of a ring, called the Z ring, at the division site. Nucleoid occlusion and Min work by preventing Z ring assembly at all sites along the cell other than the cell centre. Here we make the surprising discovery that, in the absence of both these factors, Z rings are positioned correctly at the division site, but there is a delay in this process and it is less efficient. We propose that a separate mechanism identifies the division site at midcell in rod-shaped bacteria, and nucleoid occlusion and Min ensure that the Z ring forms there and only there, at the right time and every time.

Abstract Introduction Results Discussion Materials and Methods

genetics biology microbiology molecular cell biology genetics and genomics

2012

The Min System and Nucleoid Occlusion Are Not Required for Identifying the Division Site in Bacillus subtilis but Ensure Its Efficient Utilization

In all sensory modalities, the data acquired by the nervous system is shaped by the biomechanics, material properties, and the morphology of the peripheral sensory organs. The rat vibrissal (whisker) system is one of the premier models in neuroscience to study the relationship between physical embodiment of the sensor array and the neural circuits underlying perception. To date, however, the three-dimensional morphology of the vibrissal array has not been characterized. Quantifying array morphology is important because it directly constrains the mechanosensory inputs that will be generated during behavior. These inputs in turn shape all subsequent neural processing in the vibrissal-trigeminal system, from the trigeminal ganglion to primary somatosensory (“barrel”) cortex. Here we develop a set of equations for the morphology of the vibrissal array that accurately describes the location of every point on every whisker to within ±5% of the whisker length. Given only a whisker' s identity (row and column location within the array), the equations establish the whisker' s two-dimensional (2D) shape as well as three-dimensional (3D) position and orientation. The equations were developed via parameterization of 2D and 3D scans of six rat vibrissal arrays, and the parameters were specifically chosen to be consistent with those commonly measured in behavioral studies. The final morphological model was used to simulate the contact patterns that would be generated as a rat uses its whiskers to tactually explore objects with varying curvatures. The simulations demonstrate that altering the morphology of the array changes the relationship between the sensory signals acquired and the curvature of the object. The morphology of the vibrissal array thus directly constrains the nature of the neural computations that can be associated with extraction of a particular object feature. These results illustrate the key role that the physical embodiment of the sensor array plays in the sensing process. Animals use movements to acquire and refine incoming sensory data as they explore and navigate the environment. This means that – except under rare conditions most often found in the laboratory – sensing is an active process, constrained and shaped by the biomechanics of the muscles and by the material properties and morphology of the sensing organs. It is impossible to meaningfully characterize sensory input to the nervous system during active behavior without considering the physical embodiment of the sensor array. The rat vibrissal (whisker) system is one of the oldest models in neuroscience for studying sensorimotor integration and active sensing [1], [2]. Approximately 30 macrovibrissae are arranged in a regular array on each side of the rat' s face [3]. Rats move their whiskers at frequencies between 5–25 Hz to acquire tactile information about objects in the environment, including size, shape, orientation, and texture [4], [5], [6], [7], [8]. The morphology of the vibrissal array directly constrains the spatiotemporal patterns of mechanosensory inputs that will be generated as the rat actively explores an object. These patterns of whisker-object contact in turn shape all subsequent patterns of neural activation along the vibrissal-trigeminal pathway, from the brainstem to primary somatosensory (“barrel”) cortex. To date, however, the shape and structure of the rat vibrissal array has not been quantified, and there is thus no rigorous way to predict the input patterns that will occur during a given exploratory sequence. The present study was undertaken to quantify the morphology of the rat vibrissal array and demonstrate its influence on the whisker-object contact patterns associated with tactile exploration of an object. A set of equations is developed that describes every point on every whisker in the entire vibrissal array. Given only a whisker' s identity (that is, its row and column within the array), the equations establish the whisker' s two dimensional (2D) shape as well as three-dimensional (3D) position and orientation. The final result is a model that accurately describes the 3D location of every point along a whisker to within 5% of the point' s distance from the whisker base. Simulations demonstrate that alterations in array morphology dramatically alter the mechanosensory signals that the rat would obtain as it whisks against an object. Specifically, for a particular head orientation, the average angle at which the whiskers contact a cylindrical object is uniquely related to the radius of the cylinder. The nervous system could potentially learn this relationship to allow the rat to determine object radius within the time span of a single whisk. If the morphology of the array is altered, however, the same head orientation no longer produces a unique relationship between average angle of contact and object radius. The morphology of the vibrissal array thus directly modulates the information available to the nervous system and constrains the nature of the computations that can be associated with extraction of a particular object feature. These results underscore the critical importance of physical embodiment in the sensing process. The heads and vibrissal arrays of three rats were scanned in a 3D volumetric scanner (see Materials and Methods). In post-processing, the 3D point cloud of the rat' s head (Figure 1A) was placed in a standard position and orientation, defined using three criteria. First, the rat' s nose (the centroid of the two nostrils) was defined to lie at the origin (0,0, 0). Second, the “rostrocaudal midline” of the head was required to lie in the yz-plane, with the caudal-to-rostral vector pointing in the positive y-direction. The rostrocaudal midline was defined as the line between the mean coordinate of all macrovibrissal base-points and the origin. Finally, the “whisker row planes” were forced to lie parallel to the xy-plane (Figure 1B). For each row, the “whisker row plane” was defined as the best-fit plane to all whisker base-points within that row on both left and right sides of the face. The normal to each whisker row plane was averaged across all rows to obtain a single direction vector. Greek whiskers were omitted when computing best-fit planes. The maximum sum-squared-error across all planes was 1. 59 mm2 (maximum residual was 0. 86 mm). For each rat, the whisker row planes were all closely parallel (mean angle between normal vectors = 6. 86°, maximum difference = 17. 92°). Averaging the normal vectors of all best fit planes for a given rat yielded the head orientation that aligned the average row-plane normal vector with the positive z-axis. This ensured that all five planes were oriented as close as possible to parallel to the xy-plane. To evaluate the accuracy and precision of the model, an uncertainty analysis was performed. We used error propagation techniques to compute the variance of any point on the array given the variances of each of the model parameters [15]. To perform this propagation of error analysis, we determined the variability in each parameter across animals, and then examined how that variability would propagate through the final equation (eq. 11) used to generate the model. This analysis gives the variance of the model. Uncertainty in the model was determined by quantifying how variance in each of the parameters (θmp, φmp, ψmp, θbp φbp, s, and a) affected the location of every point along every whisker in the model. Note that orientation angles are not appropriate parameters to include in the error propagation analysis because these angles are actively controlled by the rat during whisking behavior and may be considered “inputs” to the model (see Discussion). Results of the error analysis for the locations of the whisker tip and base-points are shown in Figure 10. The tips of the whiskers are particularly useful locations to quantify error for two reasons: first, the tips were not explicitly modeled, and second, the tips are the most distal part of the array, and consequently will have the largest error of any point along the whisker. The whisker tip positions thus serve as the most conservative upper bound on the error in the model. Across all whiskers, the average standard error in tip position was 3. 2±1. 7 mm (standard deviation: 7. 8±4. 2 mm), and is depicted graphically in Figure 10A. The error in tip position was computed as the Euclidean distance between the location of the whisker tip and the tip point plus the standard deviation of the tip' s (x, y, z) coordinates. Across all whiskers, the average standard error at the whisker base was 1. 3±0. 26 mm (standard deviation 3. 7±0. 65 mm), and is depicted graphically in Figure 10B. The error in base position was computed as for tip positions. The 95% confidence interval for each point along the whisker arc length is shown graphically in Figure 10C, 10D, and 10E using colored surfaces. A conservative estimate of maximum error in our model is that tip positions are accurate to within a standard error of ±5 mm (3. 2+1. 7 mm). To generalize across whiskers, error was also quantified as a function of arc length. The standard error of any point along any whisker divided by the arc length between that point and the whisker base was typically 5%±1% of the arc length. This means, for example, that a point at an arc length of 10 mm will have a standard error of 0. 5 mm. Simulations using the model of the vibrissal array can provide insight into the patterns of mechanosensory input that might be generated as the rat actively explores an object. The morphology of the array directly affects how the whiskers contact the object, and thus the mechanical information available to the nervous system. This information in turn constrains the set of neural computations that can be used to extract particular object features. Here we consider – under a particular set of assumptions about head and whisker movement – how the morphology of the vibrissal array constrains the type of neural computations that could enable the rat to determine the curvature of an object. Embodiment of the vibrissal array can be conceptualized as spanning at least four levels: the material properties of a single whisker (e. g. elasticity and damping), the morphology of a single whisker, the musculature and tissue surrounding each whisker, and the morphology of the entire whisker array. The present study examines this fourth level. Previous studies of vibrissal array morphology qualitatively characterized the locations of whisker base-points and identified systematic variations in whisker length [11], [12]. The grid-like arrangement of the vibrissae is one of the most easily observable features of mystical pad anatomy [1] [3]. Brecht et al. demonstrated that a set of governing geometric principles, conserved across species, could qualitatively explain this grid-like arrangement [11]. The present study now quantifies the three-dimensional vibrissal array architecture. Specifically, we provide a single equation that describes every point of every whisker. This work adds to the understanding of vibrissal array morphology in several important ways. First, the locations of the whisker base-points are quantified in three-dimensions, capturing the strong row- and column-based structure of the array, while also incorporating the underlying curve of the rat' s mystacial pad. Second, equations define the 2D shape of each whisker in terms of its intrinsic curvature as well as its length, and the angle ζ describes the orientation of each whisker' s intrinsic curvature. The orientation of a whisker during protraction will affect both the angle and time at which it will make contact with an object, and a recent study has shown that the rat can change ζ during a protraction [10]. Thus the intrinsic curvature as well as ζ are essential parameters for accurate models of whisker movement. Finally, quantifying vibrissal array features in analytical form allows for systematic, cross-species comparisons of structure-function relationships in the context of behavior and ethological niche, as described below. The present study uses basic techniques from geometric morphometrics to analyze the morphology of the rat vibrissal array. In general terms, morphometrics refers to the quantification of variations in the shapes of objects. When applied to biology, morphometrics can be used to quantify and compare the shapes of organisms within and across species [20]. In some “landmark-based” analyses, for example, deviations of individual specimens from an average morphology have revealed subtle morphological differences between taxa, sexes, ages, and geographic locations [21]. Importantly, morphometric analysis has also been used to distinguish morphological differences attributable to phylogeny from those arising from behaviors that a species or sub-species has adopted in response to the local environment. Once morphological features have been quantified, statistical techniques can provide estimates of which morphological variations are best explained by phylogenetic differences and which by environmental factors (for an example, see [22]). The model developed in the present study lays the groundwork to investigate the origin of cross-species differences in the morphology of the vibrissal array. For example, murid rodents closely related to the rat (e. g. , the mouse) may have a vibrissal morphology that simply scales with body size (a phylogenetic effect). More distant species may show changes in morphology that reflect adaptive traits enabling behaviors essential for species survival. The orientation angles of the whiskers (θ, φ, ψ, ζ) are the angles at which the whiskers emerge from the rat' s face. The facial muscles of the rat control the orientation angles, and these muscles have many degrees of freedom. They can move the whisker through a large range of orientation angles during natural whisking behavior. Because the 3D whisker scans were done post-mortem, rigor mortis of the facial muscles could have pulled the whiskers to orientation angles anywhere within their natural range of movement. Unsurprisingly, these angles were the parameters that exhibited the largest variability across animals, although smooth trends are clearly visible (Figure 8). The variability associated with the static, post-rigor measurement of orientation angle is not particularly meaningful, as the muscles could have pulled the whiskers into any number of orientation angles. Accordingly, this variability was deliberately excluded from the error analysis of the model. The orientation angles should not be thought of as fixed parameters of the model, but should instead be used as inputs to the model to simulate whisking behavior. Numerous studies have quantified changes in the horizontal angle θ during natural whisking [5], [7], [9], [10], [13], [14], [24], [25], [26], [27], [28], [29]. At least two studies have produced data from which it is possible to determine how the angles φ and ψ change over the course of a whisk [10], [25]. Finally, Knutsen et al. provided the first evidence that the whisker rolls about its own axis during active whisking. The roll can be described by the value of the angle ζ over the whisking cycle. Taken together, these studies determine the equations for changes in orientation angles throughout the whisking trajectory. In addition to the equations that describe whisker movements in terms of changes in orientation angles, a complete model of whisking kinematics will require an equation that relates the protraction angle θ to mystacial pad parameters. The mystacial pad translates [9] and also changes shape during the whisk cycle (unpublished observation), causing the whisker base-points to translate. In the model, positions of the whisker base-points translate with the underlying mystacial pad [9], and therefore can be determined based solely on ellipsoid shape. There are several potential uses of an accurate kinematic model of the vibrissal array; most importantly, it could be used to simulate the expected spatiotemporal patterns of whisker-object contact during exploratory behaviors. As shown in the present study, even without information about whisker velocity, the model can be used to simulate relationships between the sensory signals acquired and an object feature. However, these simulations only predicted the angles of whisker-object contact, while an accurate kinematic model could also predict the temporal sequence of contacts. Inter-whisker contact intervals are important in determining whether neural responses in barrel cortex will be suppressed or facilitated [23], [24], [25], [26]. A predictive model of kinematics could thus be used to probe expected neural responses given the head' s position and orientation relative to an object. A second possible use of the model is as a predictive tool to describe where whiskers should be. This could be especially helpful when tracking whiskers in behavioral datasets. In high speed videography, whiskers are often observed to cross, go out of focus, and blur. Predicting the 2D projection of a whisker could reduce the search space for a whisker-tracking algorithm, and provide an estimate of whisker identity at the same time. A growing theme in sensory neuroscience is the need to link movement and sensing behaviors [27]. The present study reflects the increasing need for simulations of dynamics in sensory neuroscience, as the model represents a first step towards the creation of an entirely “digital rat, ” that is, a simulation platform to test theories of whisker movement, mechanical modeling of whisker-object collisions, mechanotransduction, and neural coding. The morphology of the vibrissal array directly constrains the mechanosensory inputs that will be generated during behavior. The morphological model can be used in conjunction with models of whisker kinematics or dynamics to develop increasingly accurate predictions about exploratory patterns of freely behaving animals, and thus the neural computations that can be associated with extraction of particular object features. Responses of sensory neurons are likely to be conditioned on, or tuned to, the particular movements used to extract the data. An improved understanding of the dynamics of motor systems that acquire sensory data will improve our understanding of the complex interactions between sensorimotor structures, the nervous system, and the specific environments in which they function. Animal protocols for this study were written in strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health, and were approved in advance by the Animal Care and Use Committee of Northwestern University. A total of six adult, female Sprague-Dawley rats (>3 months old, approximately 300 grams) were used. The head and vibrissal arrays of three of the six rats were scanned in a three dimensional (3D) volumetric scanner. Rats were euthanized with a lethal dose of pentobarbital. Tiny incisions (∼1 mm) were made in the scalp to accommodate the implantation of four skull screws. Rigid positioning rods were attached to the screws with dental acrylic. The rat was decapitated and the positioning rods were used to mount the head stably within the 3D scanner. Three-dimensional scans occurred within two hours of euthanasia (after rigor mortis had set in). Immediately after the 3D scan, each macrovibrissa was grasped firmly at its base with tweezers and plucked in a single swift motion from the mystacial pad. Each isolated macrovibrissa was then scanned in two dimensions (2D) on a flatbed scanner within 4–12 hours after euthanasia. The remaining three rats were euthanized in unrelated experiments, and did not undergo 3D scanning. Their macrovibrissae were scanned in 2D, again within 4–12 hours after euthanasia. Neither the animals nor the whiskers were preserved prior to scanning in any way. Scans of whiskers in the A, B, C, D and E rows, along with the Greek column α, β, γ and δ, were obtained from both right and left whisker arrays. The first six whiskers of each row were analyzed because these whiskers are associated with sling muscles [3] and are therefore considered macrovibrissae. The arrangement of the microvibrissae was not quantified. In our specimens, the A row contained only four whiskers. The B row consistently contained four whiskers on the left side and five whiskers on the right side; this peculiarity is perhaps due to the specific strain of rats that used in the study (Sprague-Dawley). The C, D and E rows all contained six or more whiskers. This arrangement of whiskers is consistent with that described in previous reports [3], [11]. A Surveyor DS-3040 3D laser scanner (Laser Design Incorporated) was used for the 3D scans. The final scanner output was a finely digitized 3D point cloud (20 micron volumetric accuracy) as shown in Figure 1A of Results. The complete point cloud – including both head and whiskers – was imported into the software package RAPIDFORM XOR. Within this software package, data points corresponding to each macrovibrissae were extracted from the point cloud. An example of an extracted whisker is shown in Figure 13A. The extraction process involved manual rotation of the 3D scanned image to visually determine the set of points that clearly belonged to each whisker through all angles of rotation. The whisker base-point was then identified as the centroid of a small number of points (typically 8–20) on the mystacial pad that rotated the least relative to that identified macrovibrissa. Whisker identity was assigned using the well-known topographic arrangement of whiskers on the mystacial pad [3]. In the case that two or more whiskers emerged from the same follicle (as occurred for approximately 10% of follicles), only the largest whisker was used. After manual extraction, the 3D points belonging to each macrovibrissa were exported to Matlab and a moving average (21-sample window) was used to smooth the shape. The macrovibrissae from all six rats were plucked and scanned in 2D using a flatbed scanner. Isolated vibrissae were scanned at spatial resolutions ranging from ∼10. 6 microns/pixel to ∼3 microns/pixel using either an Epson Perfection 4180 or a UMAX Powerlook 2100 XL scanner. Two scanners were used to quickly and efficiently image the large number of whiskers. The 2D whisker shape was extracted using custom image processing algorithms. Each image was converted to black and white either in Adobe Photoshop v. 7 or in Matlab, and the whisker outline was extracted in Matlab (Figure 13B). When magnified, the upper and the lower edges of the whisker became apparent (heavy black lines in the inset of Figure 13B). The midpoint between the upper and lower whisker edges was determined at small increments along the extracted whisker length. The midpoints were connected to obtain the centerline of the whisker, which closely matched the overall whisker shape (grey line in the inset of Figure 13B). Throughout Results, 2D whisker shape was quantified using the centerline of the whisker. The vibrissal array was parameterized in terms of variables that are relatively easy to measure in behavioral studies. This parameterization relies on seven parameters specific to the mystacial pad and eight parameters specific to each whisker. These 15 parameters are listed in Table 1 of Results and described in more detail below: Resolution limits of the 3D scanner meant that the scan often did not capture the whisker' s most distal region. The fraction of each whisker that was captured in the 3D scan was calculated as the ratio of the 3D scan whisker length to the 2D scan whisker length. The 2D scan was sufficiently high resolution (between 3 and 10 pixels/micron) that it captured all of the whisker tips and can be considered “ground truth” for whisker length. In general, the 3D scan captured 50%±30% (STD) of the 2D whisker length. This is sufficient to estimate the whisker' s base-point location and angles of emergence, which require only the whisker' s most proximal portion. The parameter most greatly affected by the limited 3D scanner resolution is the ζ angle, which requires estimation of the whisker plane. For example, if only 10% of the whisker is scanned in 3D, the data will most likely be linear, and a plane would be poorly conditioned. The ζ angle was therefore only computed when a well-conditioned plane could be found (84 out of 158 whiskers scanned in 3D). To combine the mystacial pad parameters and the whisker specific parameters across all rats into a single model, we found the underlying relationship between each of the eight whisker specific parameters (θBP, φBP, s, a, θ, φ, ψ, ζ) and two independent variables: row and column identity. A two-way ANOVA was performed for each parameter with the row and column as possible factors to find identity-based relationships. To perform the analysis, each row was assigned an integer value between 1 and 5 (increasing from A to E) and each column an integer value between 1 and 6 (increasing from caudal to rostral). Greek columns were associated with an adjacent row (α: A, β: B, γ: C, δ: D) and assigned a column integer value of 0. For parameters that had only the row or the column as a significant predictor (p<0. 001), a least squares regression analysis was used to test four different underlying relationships between the parameter and the position variable of interest. The four types of models tested were: polynomial, rational, power, and exponential functions (see Supplementary Information Text S1 and Table S2 for a comprehensive list of equations). For parameters that had both the row and the column as significant factors (ANOVA, p<0. 001), only multivariate regression was performed to generate the underlying relationship. These parameters were tested with polynomial relationships. To avoid overfitting, a higher-order or non-polynomial fit was chosen as the appropriate fit only if it was significantly better than the linear fit (F-test of the correlation coefficients, p<0. 001). The final equation types used in our analysis included linear, exponential and 2D linear (equations shown in Table 4). The average mystical pad ellipsoid was calculated from the right and left ellipsoids of three rats. Parameters from left-side ellipsoids were mirror-imaged across the yz-plane so that they could be averaged with parameters from right-side ellipsoids. After averaging all seven parameters, the resulting average ellipsoid was mirrored back across the yz-plane to generate symmetric, average ellipsoids on both right and left sides. Finally, these averaged ellipsoids were rotated about the global x-axis to ensure that the mean of the final base-point row-planes was parallel to the xy-plane (by an angle of 8. 72°), therefore following the convention set in Results for the head orientation. The final values of ellipsoid parameters after this rotation are shown in Table 2.

Animals move in order to sense the world. Sensing is thus an active process, constrained by muscle biomechanics and by the material, shape, and structure of the sensing organs. The rat vibrissal system provides an ideal model to examine how the physical embodiment of a sensory array shapes the sensing process. Rats have approximately thirty macrovibrissae (whiskers) arranged in rows and columns on each side of their face. They brush their whiskers against objects to tactually extract object features. To date, however, the three-dimensional shape of the whisker array has not been characterized. We scanned six rats to develop equations for the complete structure of the whisker array. Given only a whisker' s row and column identity, the equations establish the whisker' s two-dimensional shape and three-dimensional position and orientation. We used this equation-based model to simulate the whisker-object contact patterns that would be generated as a rat uses its whiskers to tactually explore objects with varying curvatures. Altering the shape of the array dramatically altered the relationship between the simulated sensory input and object curvature. The structure of the whisker array thus directly constrains spatiotemporal input patterns and thereby, the nature of the neural processing associated with extraction of particular object features.

Abstract Introduction Results Discussion Materials and Methods

neuroscience/behavioral neuroscience neuroscience/sensory systems computational biology/computational neuroscience evolutionary biology/animal behavior computational biology/systems biology

2011

The Morphology of the Rat Vibrissal Array: A Model for Quantifying Spatiotemporal Patterns of Whisker-Object Contact

Genome-scale metabolic models provide a valuable context for analyzing data from diverse high-throughput experimental techniques. Models can quantify the activities of diverse pathways and cellular functions. Since some metabolic reactions are only catalyzed in specific environments, several algorithms exist that build context-specific models. However, these methods make differing assumptions that influence the content and associated predictive capacity of resulting models, such that model content varies more due to methods used than cell types. Here we overcome this problem with a novel framework for inferring the metabolic functions of a cell before model construction. For this, we curated a list of metabolic tasks and developed a framework to infer the activity of these functionalities from transcriptomic data. We protected the data-inferred tasks during the implementation of diverse context-specific model extraction algorithms for 44 cancer cell lines. We show that the protection of data-inferred metabolic tasks decreases the variability of models across extraction methods. Furthermore, resulting models better capture the actual biological variability across cell lines. This study highlights the potential of using biological knowledge, inferred from omics data, to obtain a better consensus between existing extraction algorithms. It further provides guidelines for the development of the next-generation of data contextualization methods. Genome-scale metabolic models (GeMs) have been widely used for model-guided analysis of large omics datasets, since they provide cellular context to these data by establishing a mechanistic link from genotype to phenotype. GeMs include all reactions in an organism. Since not all enzymes are active in each cell type or culture condition, algorithms have been developed to build context-specific models using omics data to recapitulate the metabolism of specific cell types under specific conditions [1,2]. These algorithms have provided useful insights in the metabolism of specific cell and tissue types [1,3–10]. However, since each method uses different assumptions to guide reaction inclusion and removal, they result in considerable differences in size, functionality, accuracy, and ultimate biological interpretation, even when using the same data set [1,2, 11]. The poor consensus in generated models requires increased caution in the interpretation of model-derived hypotheses of how metabolism is used under specific environments. Indeed, most generated models, upon construction, will be missing known metabolic functions and this varies considerably for models built using different approaches [1]. To gain confidence in model predictions and reconcile the differences across approaches, users can enforce the inclusion of known metabolic capabilities in the model. In this regard, the tINIT extraction algorithm introduced the possibility to enforce the capacity of context-specific models to represent some cellular functionalities by using a list of metabolic tasks known to occur in all cell types [12]. However, this protectionist approach requires one to know and predefine the functionalities of a specific cell line, tissue, or context. To overcome this, we propose an approach to infer the functionalities of a cell or tissue from omics data, and then protect these functions to guide the construction of a context-specific model. To this end, we curated and standardized published lists of metabolic tasks [13,14], resulting in a collection of 210 tasks covering 7 major metabolic activities of a cell (energy generation, nucleotide, carbohydrates, amino acid, lipid, vitamin & cofactor and glycan metabolism). We also developed a framework to directly predict the activity of these functionalities from transcriptomic data and subsequently use these for a protectionist approach to several existing extraction algorithms. Models resulting from this approach should more comprehensively capture the unique metabolic functions of a given cell type. We evaluated the validity and variation across models built with this approach, coupled to existing context-specific extraction methods. Specifically, we constructed hundreds of models for 44 cancer cell lines in which we built the models using standard approaches or protected a list of metabolic functions that have been inferred from the original transcriptomic data of each cell line. We also varied the reference human reconstruction and algorithms employed for the generation of cell line specific models, using two different reference models (iHsa [13] and Recon 2. 2 [15]) and 6 different algorithms (mCADRE [16], fastCORE [5], GIMME [6], INIT [7], iMAT [4], and MBA [8]). We compared the sets of extracted models at the level of reaction content, metabolic functions, and capacity to predict essential genes identified in CRISPR-Cas9 loss-of-function screens. Through this study, we highlight the value of using experimental data to help infer the set of metabolic tasks that should be included in a model, in an effort to obtain greater consensus across existing extraction algorithms. We built models from Recon 2. 2 [17] and iHsa [13] using six model extraction methods (MEMs: mCADRE, fastCORE, GIMME, INIT, iMAT, MBA) for 44 different cell lines from the NCI-60 panel (S1 Table; 15 cell lines were not used due to the absence RNA-Seq data in [18] for these cell lines). Uptake and secretion rates of the input GeMs were quantitatively constrained using a list of experimentally measured metabolites (S2 Table) [19,20]. Furthermore, a biomass function, consisting of 56 metabolites required for growth, was added and constrained to the experimentally measured growth rate of the cell lines (S3 Table). The biomass function and constraints from exometabolomic data introduced in the GeMs were implemented as described in [1]. The extraction process of cell line specific models was done based on RNA-Seq data [18] to specify active genes in each cell line. Details on the implementation of MEMs tested and the preprocessing of gene expression data for the definition of gene activity are provided in the Methods section. To assess the relative impact of algorithm and data source on model content, we conducted a principal component analysis (PCA) of the reactions in all models for each reference GeM. As observed previously [1,11], the decisions regarding algorithm choice significantly impact the content of our cell line-specific models. The first principal (PC1) component explains 38% of the overall variance in model reaction content, with >60% of the variation in PC1 explained by the choice of model extraction method (Fig 1A and 1B). Indeed, the different algorithms yielded cell line-specific models that varied considerably in size, with few reactions common to all models extracted from either Recon 2. 2 or iHsa (Fig 1C, Fig A in S1 Text). Even among models extracted using the same algorithm, there is non-negligible variability in model reaction content (Fig 1D). This leads to the generation of models that are substantially different with respect to the cell line considered, while the transcriptomic data used to tailor the GeMs shows high consistency across most cell lines (Fig 1E). Model reaction content is often evaluated to compare context-specific algorithms. Recently, approaches to benchmark models with their functionalities have been proposed [1,12,21]. Current approaches use repositories of known cellular tasks to assess the capacity of models to achieve specific modeling goals or to enable the representation of specific metabolic functions. This idea of assessing the quality of a metabolic network reconstruction using biological knowledge was introduced in Recon 1 through the characterization of the “human metabolic knowledge landscape” [22]. However, the concept of “metabolic tasks” (Fig 2A) was clearly defined in 2013 by Thiele and coworkers [14] to benchmark the improvements of Recon 2 compared to Recon 1, wherein they stated that “a metabolic task is defined as a nonzero flux through a reaction or through a pathway leading to the production of a metabolite B from a metabolite A”. Since then, additional lists of tasks have been published. To standardize these and develop a framework for their easy use with GeMs, we curated the existing lists of metabolic tasks (13,14) and obtained a collection of 210 tasks covering 7 major metabolic activities of a cell (energy generation, nucleotide, carbohydrates, amino acid, lipid, vitamin & cofactor and glycan metabolism) (Fig 2B and 2C, S4 Table). We evaluated the task collection using genome-scale metabolic models for human [13,14,17,22,23], CHO cells [9], rat [13] and mouse [24] (Fig 2D, S5 Table). Details on our proposed formalism of the metabolic tasks and the associated computational framework for their use are presented in the Methods. Metabolic tasks can be used to compare the performance of models extracted from different reference GeMs. As observed at the level of the reaction content, the extraction method strongly influences the model functions (explaining >50% of the overall variance in the first PC; Fig 3A). However, the reference model is the most prominent factor in the second PC underlying a non-negligible influence of this variable in the extraction process. This is mainly due to differences in gene, protein, reaction association (GPR) annotations and reaction content between Recon 2. 2 and iHsa. Interestingly, Recon 2. 2 captures more metabolic functions with fewer reactions (Fig 3B). However, the number of successful tasks increases proportionally with the number of reactions in a model. Furthermore, as the extraction method used influences the number of reactions removed, distinct patterns are seen from the ratio of the number of metabolic tasks to the number of reactions introduced by the different algorithms (Fig 3C). As for the reaction content, the number of tasks retained in each model varies substantially, depending on the cell lines considered. Surprisingly, only 8% of the tasks are present in all models (Fig 3D, Fig B in S1 Text), thus highlighting the large variation in metabolic functions a model will have, depending on algorithm choice. We inferred active metabolic tasks directly from transcriptomic data using the whole genome-scale model. To this end, we computed the list of reactions associated with each task and used the GPR rules to determine the gene expression levels associated with each of these reactions. A metabolic score is attributed to each task by using the mean activity level of each reaction (Fig 4A; See Methods). We found that more than the half of the tasks should be conserved across all cell lines (Fig 4B), which is far more than those active using the algorithms in their standard format (i. e. , without protecting tasks). Therefore, we generated a new set of models, wherein we also enforced the inclusion of reactions associated with tasks inferred for each of the 44 different cell lines (S7 Table). We focused on MBA-like algorithms (i. e. , MBA, fastCORE and mCADRE), since they are directly amenable to use the protectionist approach with minor modifications to the algorithms (Fig 4C). Indeed, other algorithms do not ensure the inclusion of a reaction even if it is enforced. For example, iMAT relies on the definition of a core set of high-confidence reactions, but core reactions can be removed if it depends on many non-expressed non-core reactions (Fig 4C). See Methods for a detailed description of the implementation of the protectionist approach for each algorithm. For equivalent extraction setups (i. e. , same reference model, extraction method, and cell line), the number of reactions included in the extracted model was not considerably influenced by the protection of the metabolic task, while the number of active tasks clearly increases (Fig C in S1 Text). We performed PCA of the reaction content and the metabolic functions of the models with protected tasks. We observed that the protection of metabolic tasks inferred from data significantly decreased the influence of the extraction method on the final model content. However, the use of this approach remains sensitive to the choice of the reference model (Fig 5A and 5B). The reduction of the variability of model with respect to the extracted methods used can be explained by the increased number of shared tasks across cell lines, all supported by the transcriptomic data. This is seen in particular for tasks involved in amino acid and lipid metabolism (Fig D in S1 Text). Furthermore, we observed that the variation in model content was better explained by the cell lines (Fig E in S1 Text). Actually, the task protection increases the similarities between context-specific models with respect to the cell line (Fig F in S1 Text) but also with respect to the transcriptomic data (Fig G in S1 Text). Finally, all the models now share more than 64% of the metabolic tasks (Fig 5C). Beyond model content, we evaluated how the protectionist approach influenced model predictions. Thus, we analyzed the influence of protecting inferred tasks on gene-essentiality predictions (i. e. prediction of the genes whose knockdown leads to a growth impairment). We systematically deleted each gene in all generated models, and then used flux balance analysis to test models for normal or impaired growth. Gene deletions associated with impaired growth are considered as essential. We observe that task protection reduces the number of genes predicted to be essential for all thresholds considered (i. e. , percentage of the maximum wild type growth rate) for the various extraction methods used on both reference models (Fig 5D; Fig H in S1 Text). We further evaluated the accuracy of essentiality predictions by comparing these to CRISPR-Cas9 loss-of-function screens for 20 cell lines [25–27]. In these screens, essential genes are identified based on gene scores attributed using single guide RNA (sgRNA) abundance for each knockout before and after growth selection. Gene scores that are more negative have a higher probability of being essential. Therefore, the agreement between model predictions and the CRISPR screen data can be quantified as the percentage of predicted essential genes that have a negative gene score [28]. Furthermore, the significance of the improvement gained from protecting data-inferred metabolic tasks can be computed using a 1-tailed Wilcoxon test. Consistent with previous reports [1,29], we found that the models, without protecting metabolic tasks, correctly predicted many essential genes. However, overall, the protectionist approach provided a small but significant improvement to gene-essentiality predictions (up to 5% improvement; Fig 5E; Fig I in S1 Text). However, the task protection reduced the number of predicted essential genes, which increased the proportion of true positives and reduced the number of false positives. To further assess the identity of these true positives provided by task protection, we compared our model simulations to a collection of known anti-cancer drug targets (S11 Table) [30]. In this analysis, we found the protectionist approach better captured the gene essentiality related to known drug targets (Fig J in S1 Text). We further tested if the models contain cancer hallmark genes [31] and that models built using the protectionist approach increased the proportion of genes in the models that are associated with cancer hallmarks (Fig K in S1 Text; S12 Table). Here we generated hundreds of models for 44 cell lines from the NCI-60 panel using multiple MEMs and two reference GeMs (Recon 2. 2 and iHsa) using standard approaches or by protecting metabolic tasks that have been directly inferred from transcriptomic data. We presented a comparative analysis of these two sets of models. As previously observed, the analysis of the first set of extracted models (i. e. , models generated without protecting metabolic functions) indicated that the choice of model extraction algorithm significantly influenced the model content at the reaction level [1,2, 11]. This leads to considerable variability in context-specific model content, which dwarfed the biological variability across cell lines, otherwise seen in their transcriptomes. We provided here a curated list of 210 tasks that were used to compare the functionalities of the extracted models. The evaluation of metabolic tasks has emerged as a valuable practice in metabolic modeling studies [12–14,22,32–35]. Such an approach allows one to evaluate the capacity of models to achieve specific modeling goals by capturing known metabolic features. Here we also demonstrated that the approach allows one to objectively compare models that may not share the same structure, such as different reference network reconstructions or models that have been extracted using different methods or parameters. We demonstrated that the selection of a reference model can significantly impact the resulting metabolic functions captured by extracted models, thus possibly impacting the results and interpretations from modeling studies. Indeed, the comparison of the functions of models extracted from both Recon 2. 2 and iHsa demonstrated the non-negligible influence of these reference models. We found this is principally due to differences in the GPR annotations in both GeMs. However, these differences in GPR annotations do not considerably influence the inference of metabolic tasks from transcriptomic data. The functional similarity across cell lines captured using data-inferred metabolic tasks is highly consistent between both reference models (Fig L in S1 Text). While community initiatives to standardize the formal representation of GeMs will facilitate cross-comparison between diverse existing GeMs [36], these results highlight the potential of using the inference of functionalities directly from the transcriptome as a way to increase the consensus between extraction methods and reference models. One challenge in the evaluation of metabolic models is the difficulty of comprehensively defining metabolic functions from a manual search of the literature. Thus, another strength of our approach is that it decreases the need for a priori knowledge or assumptions of the metabolic functions that should be included when building a cell or tissue specific model. Therefore, this list of metabolic tasks provides a framework for modelers to develop more physiologically accurate models by inferring the activity of metabolic tasks directly from omics data. Thus, key reactions that need to be included in a model can be protected, without requiring one to know what the cell does. However, the resulting models should still be curated to evaluate expected functionalities, such as for example auxotrophies. Our protectionist approach can be implemented with diverse model extraction algorithms since it only requires the algorithms to prevent the removal of active metabolic tasks during the extraction process. However, some algorithms will require modifications to ensure the protection of all reactions related to a task. Current implementations of the GIMME-like and iMAT-like families do not favor this type of protection. By minimizing flux through reactions associated with low gene expression, GIMME-like extraction methods may remove low expression reactions one would want to retain for a validated metabolic task if there are high expression reactions that allow for growth. The iMAT-like methods are similar as they rely on finding an optimal trade-off between removing reactions associated with low gene expression, and keeping reactions whose genes/enzymes are highly expressed. Thus, modified implementations of these algorithms will be needed to allow the protection of reactions based on experimental observations. Finally, this approach can also be extended to any type of network complexity reduction that have been developed in the metabolic modeling field, such as the MILP-based approaches developed to tailor models based on exometabolomic data [37,38]. In our work, we also demonstrated that the models built with the protectionist approach are able to better capture cell-type specific metabolism and accurately predict many essential metabolic genes. Thus, these models may be invaluable for drug development strategies. The emergence of experimental techniques to assess the genetic vulnerabilities of a cell (e. g. , CRISPR-Cas9, RNAi) allows researchers to identify sets of genes that should be essential for growth maintenance. These essential genes can further be used to evaluate the capacity of models to represent the interdependence between down-regulation of a gene and the concomitant impairment of growth. Thus, models can be used for interpreting the mechanisms underlying metabolic vulnerabilities that may be invaluable for new drug discoveries. Furthermore, many of the metabolic changes occurring in certain diseases, such as cancer, can be captured by the current list of tasks. Since many of the metabolic tasks are shorter paths, cases where the metabolic flux is redirected due to disease-related metabolic perturbation might be captured by the specific collection of tasks computed using our method. Finally, for the rare cases where a mutation to a specific enzyme leads to a change in the metabolic reaction catalysed by the enzyme change, such as mutations in isocitrate dehydrogenase leading to the production of oncometabolites [39–41]. With such knowledge, researchers are able to define such changes as new metabolic tasks associated with the mutations and incorporate them into their models. Finally, the list of tasks presented in this study was constructed based on existing repositories. However, a community effort could be undertaken to extend the scope and the definition of these metabolic functions, including the development of tasks seen in plants and microbes and tasks associated with secondary metabolism and microbial gene clusters [42,43]. Furthermore enzymatic mutations leading to new metabolic functions [39] can be systematically defined and added, as currently efforts in the constraint-based modeling community do so on a model by model basis [41]. As these tasks are connected to their associated gene products, this repository of curated tasks would facilitate the description of genome-scale metabolic reconstructions as more than a network of reactions but rather as interconnected maps of cellular functions for diverse organisms. This would be invaluable for the development of algorithms using more relevant biological information and facilitate more comprehensive and accurate descriptions of metabolic adaptations that occur in cells facing a change of context. Conclusively, context-specific extraction methods are powerful approaches that provide insights in the metabolic state of a cell in specific environments. However, the underlying assumptions used to tailor the GeM based on omics data vary across algorithms, with the consequence that drastically different models can be obtained based on the same data. The poor consensus in generated models may limit the use of context-specific methods for data-driven hypotheses. The definition of metabolic tasks can help with these concerns. Our curated list of tasks and computational framework will allow users to infer metabolic functions directly from transcriptomic data using the whole genome-scale model, and drive the development of improved context specific models. Such models will pave the way toward a better consensus between existing context-specific extraction algorithms, and facilitate the application of models for novel biomedical and engineering applications. RNA-Seq data for the 44 cell lines from the NCI-60 panel were downloaded from [18]. We processed the gene expression data to attribute a gene activity score for each gene and define which genes are active in each cell line. A gene is defined as active in a sample if its expression value is above a threshold defined for this gene within the dataset considered. The threshold of a gene is defined by the mean value of its expression over all the samples coming from the same dataset with exceptions that the threshold needs to be higher or equal the 25th percentile of the overall gene expression value distribution and lower or equal to the 75th percentile. The gene score is computed as follows: GeneScore=5∙log (1+ExpressionlevelThreshold) These gene scores are mapped to the models by parsing the GPR rules associated with each reaction. The gene score for each reaction is selected by taking the minimum expression value amongst all the genes associated to an enzyme complex (AND rule) and the maximum expression value amongst all the genes associated to an isozyme (OR rule) [44]. Note that we have recently benchmarked the influence of preprocessing methods on the definition of the set of active genes and observed that this parameter combination presented the best performance [45]. Model extraction methods (MEMs) employ diverse algorithms to extract cell line- or tissue-specific models from a GeM. The MEMs we have considered can be categorized into three families: “GIMME-like” (i. e. , GIMME), “iMAT-like” (i. e. , iMAT and INIT) and “MBA-like” (i. e. , MBA, FASTCORE, and mCADRE), as proposed previously [2]. The GIMME-like family minimizes flux through reactions associated with low gene expression. The iMAT-like family finds an optimal trade-off between removing reactions associated with low gene expression, and keeping reactions whose genes/enzymes are highly expressed. In the MBA-like family, the algorithms use sets of core reactions that should be retained and active, while removing other reactions if possible. All the algorithms used in this study have been implemented using the function createTissueSpecificModel available in the COBRA Toolbox 3. 0 [46]. We describe below the list of required parameters needed to run the different methods, all optional parameters have been kept to their default setting. FASTCORE [5]—The core reactions set (options. core) is determined by all the reactions associated to a gene score superior to 5log (2). Note that the biomass reaction was added to the core reactions sets. GIMME [6]—The implementation of GIMME requires two parameters: the gene scores (options. expressionRxns) and a threshold value, the reactions associated with a gene score value below this threshold will be minimized (options. threshold = 5log (2) ). Note that we manually attributed a gene score of 10log (2) to the biomass reaction to ensure its inclusion. iMAT [4,47]—Three parameters need to be provided to run iMAT: the gene scores (options. expressionRxns), a lower threshold value (reactions with gene score below this value are considered as “non-expressed”) and a upper threshold value (reactions with gene score above this value are considered as “expressed”). To simplify the comparison across algorithms, we set both thresholds to the same value: options. threshold_lb = options. threshold_ub = 5log (2), as done in a previous benchmarking study (1). Note that we manually attributed a gene score of 10log (2) to the biomass reaction to ensure its inclusion. INIT [7]—The implementation of INIT requires attributing positive weights (options. weights) to each reaction with high expression and negative weights for the ones with low expression. All the reactions associated with a gene score below 5log (2) have been assigned a weight of -8 while the weights of remaining reactions were defined as the ratio between the gene score for each reaction and 5log (2). The weight associated with the biomass reaction was put to the maximum of obtained reaction weights. MBA [8]—The implementation of MBA requires the definition of two set of reactions: high confidence (options. high_set) due to their expression and others with medium confidence (options. medium_set). The set of reactions with high confidence is defined as reactions with a gene score above the 75th percentile of the distribution of all gene scores and the medium confidence set by all the reactions presenting score above 5log (2) and below the 75th percentile of the distribution of all gene scores. Note that the biomass reaction has been manually added to the high confidence set of reactions. mCADRE [16]—The implementation of mCADRE requires a score quantifying how often a gene is expressed across samples (options. ubiquityScore) and a literature-based evidence score (options. confidenceScores). Since the confidence score identification used in the original paper is difficult to transpose in this study, we did not define the confidence score as preformed in the tutorial presenting the implementation of mCADRE in COBRA Toolbox 3. 0 (46). Furthermore, as the gene scores are computed based on the knowledge of the gene expression of a gene across all samples, we used the gene scores as ubiquity scores. The curation has been done by first taking the union of previously published lists of metabolic tasks [13,14]. We removed duplicated tasks and lumped tasks that rely on the description of similar metabolic functions. Each remaining task without strong biological evidence was removed. We also created 9 new tasks that were essential for the acquisition of already described metabolic functions (i. e. , intermediate biosynthetic steps for the acquisition of other tasks). Doing so, we obtained a collection of 210 tasks associated with 7 systems (energy, nucleotide, carbohydrates, amino acid, lipid, vitamin & cofactor and glycan metabolism). For each task, we provided its original source (Recon and/or iHsa) and comments on the biological evidence of this metabolic function (S4 Table). In its original version, Thiele and coworkers (2013) [14] define a “metabolic task as a nonzero flux through a reaction or through a pathway leading to the production of a metabolite B from a metabolite A. The metabolic capacity of the network was demonstrated by testing nonzero flux values for these metabolic tasks. For each of the simulations, a steady-state flux distribution was calculated. Each metabolic task was optimized individually by choosing the corresponding reaction in the model, if present, as objective function and maximized the flux through the reaction”. In parallel, Agren and coworkers presented an alternative framework to compute the metabolic tasks present in a model within their RAVEN toolbox [48]. They defined a metabolic task through a list of inputs and outputs for which the pseudo-stationary assumption will be relaxed following a magnitude imposed by the user and assumed that a task successfully passes if the variation imposed to the inputs leads to the imposed variation of the outputs. We also propose to define a metabolic task as the capacity of producing a defined list of output products when only a defined list of input substrates is available. However, we modified the way to implement it from the RAVEN toolbox. Instead of relying on the relaxation of the steady-state assumption, we take an approach more similar to that proposed by [14] by imposing constraints only at the flux level. Therefore, a model successfully passes a task if the associated LP problem is still solvable when the sole exchange reactions allowed carrying flux in the model are temporary sink reactions associated with each of the inputs and outputs listed in the task. This framework allows the use of known stoichiometry to fix the ratio between the fluxes of the sink reactions associated with each input and output of the task. We implemented the code to compute the tasks in Matlab, and the code, checkMetabolicTasks, has been contributed to the COBRA Toolbox3. 0 [46]. We tested the list of tasks using published genome-scale models of human [13,14,17,22,23], Chinese hamster [9], rat [13] and mouse [24] cells (Fig 2D, S5 Table). All models successfully pass more than 90% of the tasks. For each failed task, we provided a reason of the failure (i. e. definition of the missing reaction to successfully pass the task) (S5 Table). As the definition of the metabolic tasks depends on the provision of the exact name of the metabolites in each model, we also provide a table of nomenclature compatibility between the different genome-scale models tested (S6 Table). We developed a computational framework for attributing a score to each metabolic task in order to extend the application of the concept beyond the model benchmarking scope. If a task successfully passes in a model, one can compute the list of reactions associated with this task and, in doing so, access the list of genes that may contribute to the acquisition of this metabolic function based on the GPR rules. To this end, we used the parsimonious Flux Balance Analysis (pFBA) algorithm to define the set of reactions and associated genes required to pass a task within a specified model [49]. Thanks to the availability of this information, metabolic functions can now be directly assessed from transcriptomic data. The proposed computation of a metabolic score relies first on the preprocessing of the available transcriptomic data and the attribution of a gene activity score for each gene (see associated Methods section). We further used the GPR rules associated with each reaction required for a task to decide which gene will be the main determinant of the enzyme abundance associated with this reaction and attribute the corresponding gene activity level (i. e. , selection of the minimum expression value among all the genes associated to an enzyme complex (AND rule) and the maximum expression value among all genes associated with an isoenzyme (OR rule) ). Therefore, each reaction involved in a task is associated with a reaction activity level (RAL) that corresponds to the preprocessed gene expression value of the gene selected as the main determinant for this reaction. Finally, the metabolic score can be computed as the mean of the activity level of each reaction: MTscore=sum (RAL) /numberofreactionsinvolvedinthetask Doing so, a metabolic task will be considered as active if its MT score has a value greater than 5log (2). The list of active metabolic tasks for each of the 44 cell lines from the NCI-60 panel is available in S7 Table. We used the list of active metabolic tasks (S7 Table) to determine the set of reactions that should be protected during the extraction process for each of the 44 cell lines. The protectionist approach has been implemented for each extraction method by using the same set of parameters as previously described with the following modification: FASTCORE—The set of reactions associated with the metabolic tasks defined as active based on the transcriptomic data has been manually added to the core reactions set (options. core). GIMME & iMAT—A gene score of 10log (2) (options. expressionRxns) has been attributed to all the reactions associated to the metabolic tasks defined as active based on the transcriptomic data. INIT—The weights (options. weights) for all reactions associated with the metabolic tasks defined as active based on the transcriptomic data were put to the maximum of obtained reaction weights. MBA—The reactions associated with the metabolic tasks defined as active based on the transcriptomic data have been manually added to the high confidence set of reactions. mCADRE—A ubiquity score (options. ubiquityScore) of 1 has been attributed to all the reactions associated to the metabolic tasks defined as active based on the transcriptomic data. For the reaction PCAs, a binary matrix is constructed in which each row represents an extracted model and each column represents a reaction, with each element representing the presence (1) or absence (0) of a reaction in a model. Reactions in all or no models were removed from the matrix. Similarly for the metabolic function PCA, the matrix had each row as an extracted model and each column as a metabolic task, with each element in the matrix representing if the task is present (1) or absent (0) in a model. For the PCAs, the matrix was centered to have zero mean within each row. PCA was done on this matrix. The variance explained by the different factors (MEM, cancer type and cell line) within each of the principal components is calculated as follows. Within one factor, the maximum Pearson correlation coefficient (R) of the component scores and categories is calculated across all possible orderings of the categories. Reported is the R2 scaled to percentages. The same procedure was used to perform the PCA on the model functionalities except that the binary matrix of reactions was replaced by the binary matrix representing the list of metabolic tasks that are successfully passed in each extracted model. The attributes of all extracted models (number of reactions and metabolites, number of successfully passed tasks and predicted growth rate) are available in S8 Table and the results of the extracted model benchmarking using the list of metabolic tasks is available in S9 Table. To predict gene-essentiality, FBA was used to optimize biomass production following the removal of each reaction in the cell line-specific models that would be affected by gene removal based on the GPRs. The function used to perform this deletion analysis is available in COBRA Toolbox 3. 0, singleGeneDeletion. m [46]. To test these essentiality predictions of the models against experimental data, we downloaded CRISPR-Cas9 loss-of-function screens data for 20 NCI-60 cell lines from depmap. org [25–27]. In these screens, essential genes are identified based on genes scores attributed using single guide RNA (sgRNA) abundance for each knockout before and after growth selection. A more negative gene score suggests a higher probability that the gene is essential. Therefore, the agreement between prediction and data can be analyzed by using the percentage of predicted essential genes that have a negative gene score [28]. A 1-tailed Wilcoxon rank sum test was used to test whether the percentage of predicted essential genes of the model extracted using the protectionist approach were significantly higher than the ones without protection. The results of the gene deletion study and prediction against CRISPR-Cas9 loss-of-function screens are available in S10 Table.

Genome-scale models of human metabolism have facilitated numerous exciting discoveries regarding human physiology and therapeutics. The accuracy of results from such studies requires that models capture the tissue or cell-type specific metabolism. In hopes to obtain accurate models, several algorithms have been developed to extract cell- or tissue-specific metabolic models. Each algorithm has provided useful insights into the metabolism of specific cell and tissue types. However, since each of these methods use different assumptions to guide reaction inclusion and removal, they result in considerable differences in size, functionality, accuracy, and ultimate biological interpretation, even when using the same data set. To overcome this, the enclosed research proposes an approach to infer the functionalities of a cell or tissue from omics data, and then protect these functions to guide the construction of a context-specific model. Through this study, we highlight the value of using experimental data to help infer the set of metabolic functions that should be included in a model, in an effort to obtain greater consensus across existing extraction algorithms. This study further provides guidelines for the development of the next-generation of data contextualization methods.

Abstract Introduction Results Discussion Methods

cell physiology carbohydrate metabolism medicine and health sciences applied mathematics enzymology cell metabolism simulation and modeling algorithms mathematics genome analysis enzyme metabolism pharmacology drug metabolism enzyme chemistry research and analysis methods gene expression pharmacokinetics biochemistry cell biology transcriptome analysis genetics biology and life sciences physical sciences genomics gene prediction metabolism computational biology

2019

Increasing consensus of context-specific metabolic models by integrating data-inferred cell functions

Tight control over gene expression is essential for precision in embryonic development and acquisition of the regulatory elements responsible is the predominant driver for evolution of new structures. Tbx5 and Tbx4, two genes expressed in forelimb and hindlimb-forming regions respectively, play crucial roles in the initiation of limb outgrowth. Evolution of regulatory elements that activate Tbx5 in rostral LPM was essential for the acquisition of forelimbs in vertebrates. We identified such a regulatory element for Tbx5 and demonstrated Hox genes are essential, direct regulators. While the importance of Hox genes in regulating embryonic development is clear, Hox targets and the ways in which each protein executes its specific function are not known. We reveal how nested Hox expression along the rostro-caudal axis restricts Tbx5 expression to forelimb. We demonstrate that Hoxc9, which is expressed in caudal LPM where Tbx5 is not expressed, can form a repressive complex on the Tbx5 forelimb regulatory element. This repressive capacity is limited to Hox proteins expressed in caudal LPM and carried out by two separate protein domains in Hoxc9. Forelimb-restricted expression of Tbx5 and ultimately forelimb formation is therefore achieved through co-option of two characteristics of Hox genes; their colinear expression along the body axis and the functional specificity of different paralogs. Active complexes can be formed by Hox PG proteins present throughout the rostral-caudal LPM while restriction of Tbx5 expression is achieved by superimposing a dominant repressive (Hoxc9) complex that determines the caudal boundary of Tbx5 expression. Our results reveal the regulatory mechanism that ensures emergence of the forelimbs at the correct position along the body. Acquisition of this regulatory element would have been critical for the evolution of limbs in vertebrates and modulation of the factors we have identified can be molecular drivers of the diversity in limb morphology. Forelimbs and hindlimbs are derivatives of the lateral plate mesoderm (LPM) that arise at fixed positions along the vertebrate body axis. Limb formation is initiated by limb induction signals from axial tissues [1]. The presumptive limb-forming regions initially express two T-box genes prior to overt limb bud formation, Tbx5 in nascent forelimbs and Tbx4 in hindlimbs [2]–[5]. Genetic studies in the mouse have shown that both genes are crucial for normal limb outgrowth by activating Fgf10 in the limb mesenchyme [6]–[8]. Fgf10 subsequently induces Fgf8 expression in the apical ectodermal ridge (AER) and Fgf8 produced from the AER, in turn, maintains Fgf10 expression in mesenchyme to establish a positive feedback loop of Fgf signalling that maintains limb growth. Mutations in human TBX5 cause Holt-Oram Syndrome (HOS OMIM142900), a disorder characterised by upper limb and heart abnormalities [9], [10] and mutations in TBX4 cause Small Patella Syndrome (SPS OMIM 147891), a disorder characterised by knee, pelvis and toe defects [11]. Tbx5 is the earliest marker of presumptive forelimb mesenchyme and because activation of this factor within a defined region of the LPM ultimately dictates the position at which the forelimbs will arise, identifying the factors that control activation of this Tbx5 expression domain will reveal the mechanisms employed that allowed the acquisition of limbs in vertebrates and that dictate forelimb position in the embryo. Tbx5 is initially expressed in the forelimb-forming region of LPM prior to the emergence of a bud and it is subsequently restricted to the forelimb region as development proceeds. Tbx5 is essential for forelimb formation and this exclusive requirement is limited to a short time window when limb bud initiation occurs [12]. Tbx4, the paralog of Tbx5, is able to rescue forelimb formation following conditional deletion of Tbx5 [13]. Furthermore, the ancestral Tbx4/5 gene represented by AmphiTbx4/5 of the limbless cephalochordate, amphioxus, can fully compensate for the loss of Tbx5 in the mouse [14]. This indicates the ancestral protein from a limbless organism has limb-inducing potential and supports a model in which evolution of a regulatory element sufficient to activate Tbx5 expression in the LPM was a critical step in the acquisition of limbs during vertebrate evolution. Hox genes are conserved homeodomain-containing transcription factors that are arranged in clusters in the genome. The chromosomal organization of the genes in the complex reflects their expression pattern along the rostro-caudal body axis to determine positional identity [15], [16]. As relative positions of limbs, axial vertebrae and Hox expression domains are conserved among vertebrates in spite of the variable numbers of each type of vertebrae (e. g. cervical, thoracic, lumbar and sacral vertebrae), Hox genes have been good candidates as determinants of limb position [17], [18]. Despite the unquestionable importance of Hox genes in patterning the developing embryo, very little is known about their direct targets and mechanisms of action. We have previously identified a Tbx5 regulatory element sufficient for early forelimb expression [19]. This element contains Hox binding sites that are required for the enhancer activity, thus implicating Hox genes in direct, positive regulation of Tbx5. However, since the ability to activate Tbx5 is not strictly restricted to Hox genes expressed only at forelimb level, the mechanism by which a rostro-caudal Hox code establishes forelimb-restriction of Tbx5 remained unknown. Here, we demonstrate how Hox paralogous group members act cooperatively to restrict expression of Tbx5 in the LPM, which ultimately determines the positions the forelimbs will emerge from the flank of the embryo. We show that mutations of a single Hox binding site in the Tbx5 forelimb regulatory element cause expanded reporter gene expression in caudal LPM. Rostral restriction in Tbx5 expression through repression in the caudal LPM is mediated by Hoxc8/9/10 genes and this repressive function is limited to Hox genes that are expressed in Tbx5-negative caudal LPM. We further map the Hoxc9 protein domains required to confer transcriptional repression that distinguishes these paralogs from other Hox proteins expressed throughout the flank of the embryo. Our results demonstrate how a nested, combinatorial code of Hox protein transcriptional activation and repression along the rostro-caudal embryo axis restricts Tbx5 expression to the forelimb and ultimately determines forelimb position. Previously, we identified a short regulatory element within intron 2 of the mouse Tbx5 gene that recapitulates the dramatic forelimb-restricted expression of this gene [19]. This 361 base pair (bp) sequence contains six Hox binding sites (Hbs) (Fig. 1A). To analyze which sites within this minimal element are required for Tbx5 expression, we generated a series of constructs in which each individual Hbs site1-6 is mutated and tested their ability to activate a LacZ reporter gene in transgenic mice. While the Tbx5 int2 (361) reporter construct drove forelimb-restricted expression of LacZ (Fig. 1B, H and [19]), mutation of either individual Hbs1, or Hbs3, or Hbs4, or Hbs5, or Hbs6 resulted in reduced reporter gene expression (Fig. 1C–G). Interestingly, in most cases residual expression was consistently detected in the anterior forelimb bud, however, mutation of Hbs5 produced mosaic expression throughout the limb. The six bp sequence (TGAGAG, bottom strand) situated 3′ of Hbs2 (6bp3′) is similar but not identical to both Pbx (TGAT) and Meis (TGACAG) canonical binding sequences [20]. Pbx and Meis are Hox cofactors that can bind DNA as heterodimers. Mutation of Hbs2 and 6bp3′ in the 361 bp core fragment produced a strikingly different result. Reporter expression was detected in the forelimb but was now also expanded throughout the interlimb and hindlimb-forming region (Fig. 1I). Since the number of transgenic embryos that showed expression with this construct was low, to study the effect of mutating the individual sites further, we used a 565 bp fragment (Tbx5 int2 (565) ) (Fig. 1J) that contains an additional 204 bp sequence 5′ to the 361 bp core element. The extra 204 bp sequence contains three putative Hox binding sites. However, these sites are not required to control the spatial restriction of expression since the 361 bp fragment produces forelimb-restricted expression equivalent to that observed with the 565 bp fragment (Fig. 1K and [19]). We have also previously shown that the fragment containing this 204 bp sequence and Hbs1 and Hbs2 cannot activate reporter gene expression indicating these sites are not sufficient for the enhancer activity [19]. As observed with the smaller fragment, mutations of Hbs2 and 6bp3′ caused caudally expanded LacZ reporter activity to include the LPM of interlimb and hindlimb-forming regions, which never normally express Tbx5 (Fig. 1L). These results suggest that these sites are required to restrict Tbx5 expression to the forelimb-forming region. The activity of the Hbs2+6bp3′-mutated construct is dependent on the presence of the other Hox sites since mutation of Hbs2 and 6bp3′ together with Hbs3-6 did not drive reporter expression at all (n = 0/6, data not shown). To distinguish the requirement for Hbs2 and 6bp3′, we next mutated either of these sites (Fig. 1M–N). Mutation of 6bp3′ did not affect the expression domain of the LacZ reporter (Fig. 1M), while mutation of Hbs2 caused caudal expansion (Fig. 1N) equivalent to that seen after mutating both Hbs2 and 6bp3′ (Fig. 1L). These results suggest that Hbs2 plays the predominant role restricting Tbx5 expression to the forelimb-forming region. These results demonstrate that the binding sites within this regulatory element can be divided into 2 distinct functional groups. Hbs1 and 3-6 act as ‘on’ switches important for the amplitude of activation, whereas Hbs2 determines spatial resolution by hosting repressive complexes that restrict the domain of activation. This element can therefore have a binary function, serving as a site for the formation of transcriptional activation or repression complexes. The presence of Hox binding sites in this element prompted us to search for candidate Hox genes that could be acting on this element as either positive or negative regulators of transcription. Previously, we have shown that PG 4 and 5 Hox genes can activate this regulatory element [19]. We now focused on Hox factors that could be mediating spatial resolution of this regulatory element by forming repressive complexes. We analysed the expression of Hox genes in chick and mouse embryos at stages when Tbx5 is first expressed in the forelimb-forming region. Tbx5 is first expressed at the level of somites 13–20 in chick and somites 4–11 in mouse embryos. As previously reported [19] the expression domains of Hox4 and Hox5 paralogs overlap with that of Tbx5 in both mouse and chick embryos (Fig. 2A–C, H–J). Since the expression patterns of HoxA, HoxB, HoxC and HoxD cluster genes are broadly similar, we show here the results of the HoxC cluster genes, as a representative example for simplicity. Hoxc6 is expressed within the caudal-most domain of Tbx5 as well as in more caudal LPM (Fig. 2D and K). Conversely, Hoxc8, Hoxc9 and Hoxc10 are exclusively expressed in caudal domains of the LPM that do not express Tbx5 (Fig. 2E–G and L–N) and are therefore candidates to repress Tbx5 expression. To determine whether caudally-expressed Hox genes can repress the Tbx5 forelimb-regulatory element, we compared the activities of Hoxc9 and Hoxc5 expression constructs when co-electroporated with the wild type Tbx5 int2 (361) (Fig. 3A) LacZ reporter into the forelimb-forming region of HH stage 14–15 chick embryos. As expected, following electroporation of the Tbx5 int2 (361) construct (with a dsRed reporter to assess electroporation efficiency (Fig. 3B) ), β-gal activity is detected in successfully targeted forelimb LPM (Fig. 3B′) indicating that this mouse Tbx5 regulatory element can also function in chick. Following co-electroporation of a Hoxc9 expression construct with the Tbx5 int2 (361) reporter, LacZ expression is repressed in the forelimb region (Fig. 3C′ white arrow). In contrast, performing the equivalent experiment with Hoxc5, which is expressed in the rostral, Tbx5-expressing LPM, does not negatively effect LacZ expression from the reporter (Fig. 3D′) demonstrating that the repressive activity is restricted to caudally-restricted Hox genes, such as Hoxc9. To determine whether Hoxc9 functions via Hbs2 to repress Tbx5 expression, we co-electroporated Hoxc9 with a Tbx5 int2 (361) reporter in which Hbs2 is mutated (Fig. 3F). In transgenic mice, this mutation caused LacZ expression throughout the forelimb, interlimb and hindlimb regions (Fig. 1N) and electroporation of this reporter alone in the forelimb-forming region produced LacZ expression (Fig. 3G′) where cells have been successfully targeted as shown by the dsRed reporter (Fig. 3G). Co-electroporation of Hoxc9 with the Hbs2 mutated reporter did not repress LacZ expression (Fig. 3H′ black arrow). As expected, no effect was observed following co-electroporation with the Hoxc5 construct (Fig. 3I′). Since the expression of Hoxc8 and Hoxc10 are also restricted in caudal LPM, we tested if they can also repress the Tbx5 reporter activity similar to Hoxc9. Ectopic expression of either Hoxc8 (Fig. S1B′) or Hoxc10 (Fig. S1C′) reduced LacZ expression. Together, these results demonstrate that Hoxc8/9/10, which are normally expressed in the caudal LPM, have the ability to repress the Tbx5 regulatory element and that this repression is mediated via the Hbs2. In contrast, Hoxc5 does not exhibit equivalent repressive activity. We next tested whether ectopic expression of Hoxc9 could repress endogenous Tbx5 expression in the forelimb-forming region. Electroporation of the right forelimb-forming region (Fig. 4A) with a Hoxc9 expression construct can repress the endogenous domain of Tbx5 (Fig. 4A′–A″). The electroporation protocol targets the proximal LPM most successfully and this is where the most profound repression of Tbx5 is observed consistent with Hoxc9 acting cell-autonomously. Although Tbx5 does not determine forelimb morphologies [13], its forelimb-restricted expression serves as a marker of forelimb identity. Since Hoxc9 is expressed in caudal LPM including the hindlimb region, we examined if, following ectopic activation of Hoxc9, hindlimb markers were activated in the forelimb region concomitant with down-regulation of Tbx5. Pitx1 is expressed in hindlimb, but not in forelimb, and determines some hindlimb morphologies [21]–[24]. Indeed, ectopic Pitx1 transcripts are detected in the forelimb (Fig. 4B′–B″) following electroporation of a Hoxc9 expression vector (Fig. 4B). The domain of ectopic Pitx1 is apparent in the proximal forelimb LPM consistent with the proximal bias in cells successfully targeted by electroporation and again consistent with a cell-autonomous mechanism of action. To understand the molecular mechanisms of caudal Hox-specific repressive activity on Tbx5 expression, we compared the DNA binding abilities of Hoxc5 and Hoxc9 since paralogous-specific functions of Hox can be explained by different DNA binding specificities [25]. We performed electrophoretic mobility shift assays (EMSA) with an oligonucleotide probe that contains Hbs2 (Fig. 5E). in vitro translated Hoxc5 can bind to the probe (Fig. 5A lane 2). Addition of non-labelled oligo as a competitor abolished the DNA-protein complexes showing their specificity (Fig. 5A lane 3–4). Non-labelled oligo in which Hbs2 is mutated (mut Hbs2) did not affect the complexes, confirming that the protein occupies Hbs2 (Fig. 5A lane 5–6). Similar to Hoxc5, Hoxc9 makes a complex with this probe (Fig. 5B lane 2) and the specificity was confirmed by a competition assay (Fig. 5B lane 3–6). We then performed a super-shift assay using an antibody against a flag epitope present in the C- terminal of our recombinant Hox proteins (Fig. 5C). Addition of this antibody resulted in super-shifts of DNA-protein complexes (Fig. 5C lane 3 and 5), indicating these complexes contain Hoxc5 or Hoxc9 proteins. These results suggest that both Hoxc5 and Hoxc9 can bind Hbs2 in vitro. To examine whether the occupancy of Hbs2 in forelimb forming, Tbx5-positive LPM and Tbx5-negative caudal LPM is different, we carried out EMSA analysis using nuclear extracts from rostral or caudal LPM. We observed two bands of the same size using both rostral and caudal extracts (Fig. 5D lane 2 and 7 arrows). We confirmed the specificity of Hox binding by competition assay. While the no mutation oligo disrupts both of the two bands (Fig. 5D lane 3–4 and lane 8–9) the mut Hbs2 oligo can only very weakly compete the complexes (Fig. 5D lane 5–6 and lane 10–11), suggesting that Hbs2 is required for these DNA-protein complexes. These results suggest that in vitro translated Hoxc9 and Hoxc5 can bind equivalently to Hbs2 and that the protein-DNA complexes from both rostral and caudal nuclear extract occupy Hbs2 specifically. Since the electroporation experiments demonstrate that the repression of the Tbx5 enhancer by Hoxc9 requires Hbs2 (Fig. 3), one of the Hox proteins forming a complex on Hbs2 using caudal nuclear extract as input is likely to be Hoxc9. In rostral LPM, since the repressive Hox genes, such as Hoxc8/9/10, are not expressed, the Hox proteins on Hbs2 using rostral nuclear extract as input are either activating Hox proteins, such as Hox PG4 and PG5 or Hox proteins with neutral function on Tbx5 expression. Thus, we propose a model in which HoxPG4 and PG5 protein complexes occupy Hbs2 in rostral forelimb forming LPM, while in Tbx5-negative caudal LPM the same site is occupied by Hoxc9 and/or Hoxc8/Hoxc10 containing-complexes that repress Tbx5 expression. Therefore, we conclude that a combination of restricted expression of Hox genes and the distinct activities of Hox proteins of different paralogous groups, which we demonstrate here, are harnessed to enable restricted expression of Tbx5 via the Hbs2. To further analyse the functional differences between Hoxc5 and Hoxc9, we generated chimeric forms of Hoxc5 and Hoxc9 proteins (Fig. 6A) and assayed their ability to repress the Tbx5 intron2 reporter construct (Fig. 6B–I). In both Hoxc5 and Hoxc9 the homeodomain is located in the C-terminus of the proteins. Paralog-specific DNA-binding properties have been reported to be determined by a specificity module spanning a Pbx-binding hexapeptide motif (W) present N-terminal to the homeodomain and the N-terminal arm of the homeodomain (NHD) [26] Fig. 6A). As would be predicted, a construct containing only the C-terminal half of Hoxc5 (Hoxc5C) cannot repress reporter gene expression (Fig. 6B–B′). Strikingly, addition of the N-terminal domain of Hoxc9 (N1N2) to the C-terminal half of Hoxc5 converts Hoxc5C into a chimeric protein (Hoxc9N5C) with Hoxc9-like repressor activity (Fig. 6C–C′). This supports our model that the opposing transcriptional activities of Hoxc5 and Hoxc9 do not lie in their distinct ability to bind Hox binding sites. To attempt to further refine the domain (s) responsible for transcriptional repression of Tbx5, we divided the N-terminus of Hoxc9 into two smaller domains, Hoxc9N1 and Hoxc9N2, and tested their function. Neither chimeric protein (Hoxc9N15C or Hoxc9N25C) showed clear repression of the reporter demonstrating that within the limits of this assay the entire N-terminus or the domain overlapping the junction between N1 and N2 is required for repressive activity (Fig. 6D–E). Although Hoxc9N5C reduced a reporter gene expression, this repression was weaker than that seen with full length Hoxc9. We, therefore, examined if there are other domains in the C-terminal half of Hoxc9 that can contribute to transcriptional repression. A chimeric protein that contains the N-terminal half of Hoxc5 and C-terminal half of Hoxc9 (Hoxc5N9C) can reduce LacZ expression (Fig. 6F–F′), suggesting that there is an additional repression domain (s) in the C-terminal region of Hoxc9. Replacement of a short C-terminal tail (Hoxc5 (9WNH) ) with equivalent regions of Hoxc5 did not affect its ability to repress the reporter (Fig. 6G–G′). Strikingly insertion of 18 amino acids spanning the hexapeptides and the homeodomain N-terminal arm from Hoxc9 (Hoxc5 (9WN) ) is sufficient to convert Hoxc5 to a transcriptional repressor (Fig. 6H–H′). To further test the requirement of these domains, we generated another chimeric protein in which all of the regions upstream from homeodomain N-terminal arm were replaced (Hoxc5 (9HC) ). This protein did not suppress LacZ expression (Fig. 6I–I′). To confirm that the loss of repressive activity is not caused by the disruption of the 3D-structure of the chimeric protein, we performed EMSA to demonstrate that this protein (Hoxc5 (9HC) and the chimeric proteins Hoxc5N9C and Hoxc5 (9WNH) can all bind a DNA probe containing Hbs2 (data not shown). These results suggest that repression of Tbx5 by Hoxc9 is mediated by two domains: one N-terminal and the other in the specificity module that contains Pbx-binding hexapeptides and the N-terminal arm of the homeodomain. Our results demonstrate that five of the six Hox binding sites (namely Hbs1 and Hbs3-6) within the Tbx5 forelimb regulatory element are required for the positive regulation of Tbx5 expression while a single site (Hbs2) is required for its repression in caudal LPM (Fig. 1). One possible mechanism to explain how these opposing transcriptional effects are mediated is that different Hox proteins have distinct binding preferences for these sites. For example Hox proteins that act as activators, such as Hox PG4 and PG5, have greater affinity for Hbs1 and 3-6 while repressive Hox proteins, such as Hoxc8/9/10, preferentially bind Hbs2. Our results do not support such a model. In this study, we show that both Hoxc5 and Hoxc9 proteins can bind repressive Hbs2 sites (Fig. 5), suggesting the repressive activity of Hbs2 is not mediated by preferential binding of repressive Hox proteins. An alternative model is that the transcriptional activity of the Hox complex bound at Hbs2 is determined by a co-factor (s). The sequence of Hbs2 is identical to the sequences of Hbs1 and Hbs3, therefore we compared the sequences surrounding these Hox binding sites. One distinguishing feature of Hbs2 identified using Mat Inspector (http: //www. genomatix. de) is the presence of a 6-bp sequence named Pbx1-Meis1 complexes site located 3′ of Hbs2 (6bp3′). Pbx is a Hox co-factor that can attenuate Hox-mediated gene transcription by recruiting histone deacetylases (HDACs) [27]. Therefore, a possible mechanism of the transcriptional repression through Hbs2 is the recruitment of HDACs to Hbs2/6bp3′ by Pbx. To examine this model, we mutated this 6 bp sequence while leaving Hbs2 intact. This mutation did not cause expansion of the reporter gene unlike mutation of Hbs2 or mutations of both Hbs2 and 6bp3′ (Fig. 1), suggesting that the repression is independent of 6bp3′. Thus, our results do not support a role for Pbx determining the transcriptional activities of Hox proteins bound to the Tbx5 forelimb regulatory element. The specificities of Hox proteins from different paralogous groups must be tightly regulated. One mechanism by which this is achieved is through distinct DNA binding specificity, for example homeodomains of Hoxc5 and Hoxc9 have different sequence preference in protein binding microarrays [25]. We found, however, that both Hoxc5 and Hoxc9 can bind Hbs2 (Fig. 5), suggesting specificity is not determined by distinct DNA binding abilities of Hox proteins. In addition, we also demonstrate that Hoxc9N5C chimeric protein, which contains the N-terminal repression domain of Hoxc9 fused to the homeodomain –containing C-terminus of Hoxc5, can repress Tbx5. Thus, the transcriptional repression specific to Hoxc9 is not mediated by DNA-binding specificity but rather achieved by transcriptional repression activities restricted to Hoxc9, which are mediated by two domains; the specificity module including the Pbx-binding hexapeptide and homeodomain N-terminal arm and a region N-terminal to the specificity module (Fig. 7). The mechanism by which these domains confer repressive activity remains to be elucidated. One possible model is by interacting with other transcriptional regulatory domain (s) in the protein. The hexapeptide of AbdA represses dpp expression by inhibiting the function of a glutamine (Q) -rich C-terminal activation domain [28]. Mutations in the hexapeptide converts AbdA from a repressor to an activator without affecting DNA-binding site selection. Although Hoxc9 lacks this Q-rich domain, the hexapeptide of Hoxc9 may block the activity of an unidentified activation domain. Another possibility is that the length of the linker region between the hexapeptide and homeodomain determines transcriptional activity. Several Antp isoforms are produced that have different linker sizes. Synthetic Antp protein with a long linker behaves as an activator, while the short-linker construct acts as a repressor, suggesting the importance of linker size [29]. As Hoxc9 has a shorter linker than Hoxc5, this may favour its function as a repressor. As it is unlikely that Hox protein itself directly represses Tbx5 transcription, we suggest the model that Hoxc9 supresses Tbx5 expression by interaction with co-repressor (s) (Fig. 7). One candidate is histone deacetylase (HDAC), which can bind Hox proteins directly [30], however, in EMSA we were unable to detect a HDAC/Hoxc9 complex on Hbs2, with in vitro translated proteins or nuclear extract from LPM (data not shown). Other potential collaborators are Smad proteins. In the Drosophila haltere, a Mad/Med/Shn complex works in combination with Ubx to repress Sal expression [31]. There is a potential Smad binding site proximal to Hbs2, however, we mutated this site and did not observe expansion in expression, rather it caused reduced expression in the distal limb bud, suggesting this Smad binding site may have a positive input on Tbx5 expression (). Other candidate repressors are engrailed (En) and sloppy paired (Slp) since, in Drosophila, they form a complex with Hox, Exd and Hth to repress transcription [32]–[34]. Neither of the two mouse En genes, Engrailed1 and Engrailed2, are expressed in LPM at pre-limb bud stages [35], [36]. The mammalian homolog of Slp, fork head box G1 (FoxG1) /brain factor 1 (BF-1) is also not expressed in LPM [37]. Therefore, the putative co-repressors enabling unique Hoxc9 repressive activity remain to be determined. We have shown that Hoxc8 and Hoxc10 have transcriptional repression ability similar to Hoxc9 (Fig. S1). To gain an insight of the mechanisms of their function, we compared the amino acid sequences of Hoxc8, Hoxc9 and Hoxc10 (data not shown). We could not, however, find any obvious conserved domains outside of homeodomains. It is possible that they use different mechanisms to repress Tbx5 expression or that they share similar 3D structure domains in spite of their distinct amino acid sequences. Our analysis of the Tbx5 forelimb regulatory element reveals a direct link between patterning of the rostro-caudal axis of the embryo by Hox genes and the programme that controls positioning of the forelimb forming territory. A clear correlation between Hox expression and establishment of the forelimb territory of the LPM has previously been suggested [17], [18], [38]. Application of Fgf to the interlimb flank adjacent to the normal wing induces a wing-like extra limb that expresses Tbx5 [5], [39]. Prior to the emergence of the ectopic wing the endogenous expression of Hoxc9 is reduced [38] consistent with downregulation of Hoxc9 as a repressor of Tbx5 (and the subsequent forelimb programme) being essential for emergence of an ectopic wing bud from this region. In the limbless python, Hoxc8 expression is rostrally expanded to the anterior limit of the trunk [18]. Hoxc8 is expressed exclusively in Tbx5-negative caudal LPM at pre-limb bud stages in chick and mouse (Fig. 2) and it can, like Hoxc9, repress Tbx5 (Fig. S1). Our results therefore, explain the mechanisms that lead to loss of forelimbs in snake through the repression of Tbx5 following expansion of Hoxc8 expression throughout the trunk. A previous study has demonstrated the presence of and a function for Hox9 genes in anterior-posterior patterning of the forelimb [40]. The complete loss of Hox9 paralogous group leads to the loss of Hand2 expression in posterior forelimb and a consequent reduction in Shh expression, while no effect on Tbx5 expression was reported. Failure to observe any caudal expansion of Tbx5 in this mutant can be simply explained by the redundant function of Hoxc8 and Hoxc10. The same study reported that Hoxc9 is expressed in the forelimb bud at E9. 5, but it is undetectable by E10. 5. Tbx5 expression is first initiated in the forelimb-forming region at E8. 5. We therefore examined the expression of Hoxc9 at stages E8. 5–E9. 5 (data not shown), however, we did not detect expression of Hoxc9 in the forelimb-forming region, in contrast to the strong staining in caudal tissues. We therefore conclude that Hoxc9 is not present in the forelimb-forming region at stages when Tbx5 expression is first initiated. Later expression of Hoxc9 is not sufficient to cause detectable repression of the domain of Tbx5 already activated by Hox4/5 paralogous genes. While we have shown that Hoxc8, Hoxc9 and Hoxc10 can repress Tbx5 expression, our study does not exclude the possibility that other caudally-expressed Hox genes have a similar repressive ability. We favour a model in which other caudally-expressed Hox paralogs have redundant functions in repression of Tbx5. Hoxc cluster null mice have no defects in the limb skeleton [41], however, the expression of Tbx5 in these mutants have not been reported and we predict that the ectopic expansion of Tbx5 in caudal LPM would not cause any skeletal defects. Further analysis will be required to uncover the requirement of caudally–restricted Hox paralogs, such as Hox8, Hox9 and Hox10 for Tbx5 repression in caudal LPM. In addition, while our results clearly demonstrate the importance of specific Hox inputs to generate the restricted expression of Tbx5 in the LPM, a similar Hox protein code is present in axial tissues (neural tube and somites) that do not express Tbx5. The activity of the forelimb regulatory element of Tbx5 is restricted to LPM and this LPM restriction is maintained following mutation of Hbs2 that leads to caudal expansion in expression. One possible explanation for LPM restriction is the presence of unknown repressors in axial tissues or alternatively additional factors, which are active exclusively in LPM, are required for Tbx5 expression. Odd-skipped related (Osr) genes are candidates as they are expressed in LPM, but excluded from axial tissues such as neural tube and somites [42]. We mutated a putative Osr binding site within the Tbx5 forelimb regulatory element to test if reporter activity was lost. The activity of the element was unaffected, however, suggesting Osr genes are not required for Tbx5 LPM expression (Fig. S3). Our analysis of the Tbx5 forelimb regulatory element has revealed a mechanism by which Hox genes regulate embryonic patterning and how recruitment of regulatory elements allow for the acquisition of novel structures and independent modulation of their morphology. Mechanisms that control PG-specific Hox functions have been described in Drosophila [26], [43]–[48]. Vertebrates, however, have a minimum of 2–4 Hox genes from the same PG and functional redundancy between Hox proteins from the same PG makes it difficult to examine their specific functions experimentally. Here we used a direct target of Hox activity, a regulatory element of Tbx5, to analyse the mechanism of Hox functional specificity and distinguished DNA binding specificity and transcriptional activity. Interestingly, the Tbx5 forelimb regulatory element contains both activating sites and a repressive site in a relatively short fragment of 361 bp. Active complexes are not spatially restricted and can be formed by a range of Hox PG proteins present throughout the rostral-caudal LPM. Instead, restriction of Tbx5 expression is achieved by superimposing a dominant repressive (Hoxc8, c9 and c10) complex that ultimately determines the caudal boundary of Tbx5 expression. Thus, the regulation of Tbx5 expression in the LPM represents an excellent system to understand the interactions between neighbouring Hox binding sites and how the consequent output is integrated. For reporter analysis in chick and mouse, we used the BGZA reporter vector [49]. Putative DNA binding sites were searched by MatInspector (http: //www. genomatix. de). Transgenic embryos were generated by the Procedural Service section, NIMR by standard pronuclear microinjection techniques. Mouse embryos were staged according to [50]. Noon on the day a vaginal plug was observed was taken to be E0. 5 days of development. Mice carrying the LacZ transgene were identified by PCR using specific primers (LacZfwd, 5′GGTCGGCTTACGGCGGTGATTT3′; LacZrev, 5′AGCGGCGTCAGCAGTTGTTTTT3′). Sequences surrounding putative Hox binding sites and the mutations induced are as followings, binding sites are shown in bold; Hbs1, ACATTATTGGA; mut Hbs1, ACATGCTTGGA; Hbs2, GACTCTCAATTATC; mut Hbs2, GACTCTCAACGATC; mut 6bp3′, GACTGCAAATTATC; mut Hbs2+6bp3′, GACGCTTAACGATC; Hbs3, AGATAATTC; mut Hbs3, AGATCGTTC; Hbs4, CCTTATTAAGG; mut Hbs4, CCTTGGCAAGG; Hbs5, CCATTTATCTTG; mut Hbs5, CCATTCGTCTTG; Hbs6, TGTTATTT; mut Hbs6, TGTCGTTT. Whole mount in situ hybridizations were carried out essentially as previously described [51]. Probe templates for chick Hox genes, Pitx1, Tbx5 and mouse Hox genes have been described previously [4], [17], [19], [52], [53] Embryos were sectioned by the Histology service, NIMR. Fertilized chick embryos (Henry Stewart Ltd, Winter Egg Farm) were incubated at 38°C and staged according to Hamburger Hamilton (HH) [54]. Reporter constructs and/or Hox expression constructs were mixed with fast green dye tracer and injected into the coelom located between the somatic and splanchnic LPM. Electric pulses (three pulses 30 v, 50 ms, with 200 ms intervals for tungsten electrodes or three pulses 20 v, 50 ms, with 200 ms intervals for platinum electrodes) were then immediately applied. Only those embryos showing robust expression of dsRed reporter (pCAβ-dsRed-Express) were processed for further analysis. In vitro translated proteins were produced using a TnT Coupled Reticulocyte Lysate System (Promega). Proteins were labelled with 35S-Methionine (PerkinElmer) to verify and quantify translation. LPM strips adjacent to somites 5–10 (rostral LPM nuclear extract) and lateral to somite 14 to its caudal extreme (caudal LPM nuclear extract) were dissected from E9 mouse embryos. Nuclear extracts were prepared using the NE-PER Nuclear and Cytoplasmic Extraction Kit (Pierce) following manufacturers instructions. Double-strand oligonucleotides were labelled with 32P by incubating with T4 polynucleotide kinase (NEB) for 30 minutes. 2 µl of in vitro translated protein or nuclear extract were blocked with 200 ng poly-dIdC, 2 µg of poly-dGdC or 2 µg of poly-dAdT in binding buffer (6. 7 mM Tris-HCl pH 7. 5,50 mM NaCl, 0. 67 mM EDTA, 0. 67 mM DTT, 2 µg BSA, 4% glycerol) in a total volume of 22 µl for 15 minutes on ice. For super-shift, 2 µl of the antibody recognising flag epitope (Sigma, F3165) was added to the binding reaction and incubated for a further 15 minutes. Then, 1 µl of 32P -labelled double-stranded oligonucleotides were mixed and incubated for 30 minutes. The protein∶DNA hybrids were resolved on 6% PAGE in 0. 5xTBE.

The acquisition of limbs during vertebrate evolution was a very successful innovation that enabled this group of species to diversify and colonise land. It has become clear recently that the primary driver behind the evolution of new structures, such as limbs, is the acquisition of novel regulatory elements that control when and where genes are activated rather than the proteins encoded by the genes themselves acquiring novel functions. We have identified the regulatory element from a gene, Tbx5. Activation of Tbx5 in the forelimb-forming region of the developing embryos is essential for forelimbs to form and disruption of human TBX5 causes limb abnormalities. We show that activation of Tbx5 in a restricted territory is achieved through a combination of activation inputs that are present broadly throughout the embryo flank and dominant, repressive inputs present only in more caudal regions of the flank. The sum of these inputs yields restricted activation in the rostral, forelimb-forming flank. Our results explain how the regulatory switches that were harnessed for the acquisition of limbs during evolution operate and how they can be turned off during the evolution of limblessness in species such as the snake.

Abstract Introduction Results Discussion Materials and Methods

developmental biology embryology organism development gene expression genetics organogenesis biology anatomy and physiology evolutionary biology pattern formation evolutionary developmental biology

2014

A Combination of Activation and Repression by a Colinear Hox Code Controls Forelimb-Restricted Expression of Tbx5 and Reveals Hox Protein Specificity

The evolution of drug resistance in fungal pathogens compromises the efficacy of the limited number of antifungal drugs. Drug combinations have emerged as a powerful strategy to enhance antifungal efficacy and abrogate drug resistance, but the impact on the evolution of drug resistance remains largely unexplored. Targeting the molecular chaperone Hsp90 or its downstream effector, the protein phosphatase calcineurin, abrogates resistance to the most widely deployed antifungals, the azoles, which inhibit ergosterol biosynthesis. Here, we evolved experimental populations of the model yeast Saccharomyces cerevisiae and the leading human fungal pathogen Candida albicans with azole and an inhibitor of Hsp90, geldanamycin, or calcineurin, FK506. To recapitulate a clinical context where Hsp90 or calcineurin inhibitors could be utilized in combination with azoles to render resistant pathogens responsive to treatment, the evolution experiment was initiated with strains that are resistant to azoles in a manner that depends on Hsp90 and calcineurin. Of the 290 lineages initiated, most went extinct, yet 14 evolved resistance to the drug combination. Drug target mutations that conferred resistance to geldanamycin or FK506 were identified and validated in five evolved lineages. Whole-genome sequencing identified mutations in a gene encoding a transcriptional activator of drug efflux pumps, PDR1, and a gene encoding a transcriptional repressor of ergosterol biosynthesis genes, MOT3, that transformed azole resistance of two lineages from dependent on calcineurin to independent of this regulator. Resistance also arose by mutation that truncated the catalytic subunit of calcineurin, and by mutation in LCB1, encoding a sphingolipid biosynthetic enzyme. Genome analysis revealed extensive aneuploidy in four of the C. albicans lineages. Thus, we identify molecular determinants of the transition of azole resistance from calcineurin dependence to independence and establish multiple mechanisms by which resistance to drug combinations evolves, providing a foundation for predicting and preventing the evolution of drug resistance. The evolution of drug resistance is a ubiquitous phenomenon that has a profound impact on human health. With the widespread deployment of antimicrobial agents in both clinical and environmental settings, the rate at which resistance evolves in pathogen populations far outpaces the rate at which new drugs are developed [1], [2]. Drug resistance threatens the utility of the limited arsenal of antimicrobial agents. The economic costs are staggering and exceed $33 billion in the United States alone to cover treatment of drug-resistant infections in patients, eradication of resistant pathogens in agriculture, and crop losses to resistant pests [3]. The evolution of resistance to antifungal drugs is of particular concern given the increasing incidence of life-threatening invasive fungal infections, and the limited number of antifungal drugs with distinct targets [4]. Unlike for antibacterials, fungal-specific drug targets are limited, in part due to the close evolutionary relationships of these eukaryotic pathogens with their human hosts, rendering most treatments toxic to the host or ineffective in combating infections [5]. Even with current treatment options, mortality rates due to invasive fungal infections often exceed 50%, and fungal pathogens kill as many people as tuberculosis or malaria [6], [7]. Thus, there is a pressing need to develop new strategies to enhance the efficacy of antifungal drugs and to minimize the emergence of drug resistance. A powerful strategy to extend the life of current antimicrobial agents is drug combination therapy [8]. Combination therapy has the potential to minimize the evolution of drug resistance by more effectively eradicating pathogen populations and by requiring multiple mutations to confer drug resistance [9]. Great success has been achieved with combination therapy in the treatment of HIV [10]–[12], and it is currently the recommended strategy for treatment of tuberculosis and malaria [13], [14]. Combination therapies have been less well explored in the clinic for fungal pathogens. However, targeting cellular regulators of fungal stress responses has emerged as a promising strategy to enhance the efficacy of antifungal drugs and to abrogate drug resistance [5], [15]. Two key cellular regulators that are critical for orchestrating cellular responses to drug-induced stress are Hsp90 and calcineurin. The molecular chaperone Hsp90 regulates the stability and function of diverse client proteins [16], [17], and controls stress responses required for drug resistance by stabilizing the protein phosphatase calcineurin [16], [18]–[21]. Compromise of Hsp90 or calcineurin function transforms antifungals from fungistatic to fungicidal and enhances the efficacy of antifungals in mammalian models of systemic and biofilm fungal infections [15], [22]–[24], suggesting that combination therapy with azoles and inhibitors of Hsp90 or calcineurin may provide a powerful strategy to treat life-threatening fungal infections. Targeting fungal stress response regulators holds particular therapeutic promise for enhancing the efficacy of the azoles, which are the class of antifungal drug that has been used most widely in the clinic for decades. Azoles block the production of ergosterol, the major sterol of fungal cell membranes, by inhibition of lanosterol demethylase, Erg11, resulting in a depletion of ergosterol and the accumulation of the toxic sterol intermediate, 14-α-methyl-3,6-diol, produced by Erg3 [25]. The azoles are generally fungistatic, causing inhibition of growth rather than cell death, and thus impose strong selection for resistance on the surviving fungal population [26]; as a consequence, resistance is frequently encountered in the clinic [27]. Azole resistance mechanisms fall into two broad classes: those that block the effect of the drug on the fungal cell and those that allow the cell to tolerate the drug by minimizing its toxicity [5]. The former class of resistance mechanisms includes upregulation of drug efflux pumps [28], or mutation of the azole target that prevents azole binding [29]. The latter class includes loss-of-function mutations in ERG3, which encodes a Δ-5,6-desaturase in the ergosterol biosynthesis pathway; Erg3 loss-of-function blocks the accumulation of a toxic sterol intermediate, conferring azole resistance that is contingent on cellular stress responses [16], [30]. Azole resistance acquired by loss of function of Erg3 or by many other mutations is exquisitely dependent on Hsp90 and calcineurin [16]; inhibition of these stress response regulators enhances azole sensitivity of diverse clinical isolates, and compromises azole resistance of isolates that evolved resistance in a human host [16], [18], [23], [31]. Inhibition of Hsp90 or calcineurin with molecules that are well tolerated in humans can impair the evolution of azole resistance [16], [20], though the potential for evolution of resistance to the drug combinations remains unknown. Azole resistance mechanisms have been studied most extensively in the opportunistic fungal pathogen Candida albicans and the model yeast Saccharomyces cerevisiae. C. albicans is the leading cause of death due to fungal infection [32], and the fourth leading cause of hospital-acquired infectious disease [7], [32]. It is a natural member of the mucosal microbiota of healthy humans, but can cause life-threatening illness in immunocompromised individuals, such as transplant recipients and those infected with HIV [7], [33], [34]. Drug resistance can readily evolve in C. albicans in the laboratory and the clinic, and molecular studies have revealed a diversity of resistance mechanisms [35]. Molecular studies with C. albicans are hindered by its obligate diploid state, lack of meiotic cycle, unusual codon usage, and inability to maintain plasmids [36], thus complementary experiments are often performed with its genetically tractable relative, S. cerevisiae, with which it often shares drug resistance phenotypes and underlying molecular mechanisms [37]. For both species, inhibition of Hsp90 or calcineurin reduces azole resistance acquired by diverse mutations [16], [18], [22], [38]. With short generation times and relatively small genomes, these organisms provide tractable and complementary systems to explore the dynamics and mechanisms underpinning the evolution of resistance to drug combinations. Here, we provide the first analysis of the genetic and genomic architecture of the evolution of resistance to drug combinations in fungi. To recapitulate a clinical context where Hsp90 or calcineurin inhibitors could be used in combination with azoles to render azole-resistant fungal pathogens responsive to treatment, we initiated an evolution experiment with strains that are resistant to azoles in a manner that depends on Hsp90 and calcineurin. We evolved populations of S. cerevisiae and C. albicans that were resistant to azoles due to loss of function of Erg3 with a combination of an azole and an inhibitor of Hsp90, geldanamycin, or calcineurin, FK506, to identify the mechanisms by which resistance evolves to the drug combinations. Of 290 lineages initiated, most went extinct, yet 14 evolved resistance. We identified mechanisms of resistance in the evolved lineages using a hypothesis-driven approach based on cross-resistance profiling and a complementary unbiased approach using whole genome sequencing. Resistance mutations in the drug target of FK506 or geldanamycin were identified and validated in five lineages. Non-synonymous substitutions conferring resistance were identified in a transcriptional activator of drug efflux pumps, Pdr1, and in a regulator of sphingolipid biosynthesis, Lcb1. Resistance also arose by premature stop codons in the catalytic subunit of calcineurin and in a repressor of ergosterol biosynthesis genes, Mot3. Several of the mutations conferred resistance to geldanamycin or FK506, while other mutations transformed azole resistance from dependent on calcineurin to independent of this stress response regulator. Genome analysis also identified extensive aneuploidy in four of the C. albicans lineages. Thus, we illuminate the molecular basis for the transition of azole resistance from calcineurin dependence to independence, and establish numerous mechanisms by which resistance to drug combinations can evolve, providing a foundation for predicting and preventing the evolution of drug resistance. Inhibition of Hsp90 or calcineurin has emerged as promising strategy to enhance the efficacy of azoles against resistant fungal pathogens, motivating our study to monitor the evolution of resistance to the drug combinations in azole-resistant populations. To do so, we used an experimental evolution approach starting with C. albicans and S. cerevisiae strains that harbour erg3 loss-of-function mutations or deletions, rendering them resistant to azoles in a manner that depends on the stress response regulators Hsp90 and calcineurin [5]. Propagation of these strains in the presence of azole and the Hsp90 inhibitor geldanamycin or azole and the calcineurin inhibitor FK506 at concentrations that exert selection pressure for resistance to the drug combination could lead to the evolution of resistance to geldanamycin or FK506, or the evolution of an azole resistance mechanism that is independent of Hsp90 or calcineurin among extant lineages (Figure 1A). Lineages were propagated by serial transfer for between 33 and 100 generations until robust growth in the presence of the drug combination was observed in extant lineages (Figure 1B). The effective population size per lineage was ∼4. 6×106, given that cultures reached saturation (∼107 cells/ml) between transfers. Of the 290 lineages initiated, the majority went extinct. Fourteen lineages evolved resistance to the combination of azole and inhibitor of Hsp90 or calcineurin (Figure 1C); seven of these lineages are C. albicans and seven are S. cerevisiae (Table 1). Six C. albicans lineages evolved resistance to azole and FK506 (Ca-F lineages), and only one evolved resistance to azole and geldanamycin (Ca-G lineage). Four S. cerevisiae lineages evolved resistance to azole and geldanamycin (Sc-G lineages) and three evolved resistance to azole and FK506 (Sc-F lineages). Resistance levels to the drug combinations of all fourteen evolved lineages were evaluated by performing minimum inhibitory concentration (MIC) assays in the presence of the inhibitors with which they were evolved, azole and FK506 (Figure 2A and 2B) or azole and geldanamycin (Figure 2C–2E). Because the azole resistance phenotypes of the starting strains were abrogated by geldanamycin or FK506, resistance of the evolved lineages was monitored with a fixed concentration of azole and a gradient of concentrations of geldanamycin or FK506. Resistance was monitored for a population of cells from each archived lineage, and for four clones isolated from the evolved population. In all cases, the clones reflected the resistant phenotype of the population (data not shown), suggestive of strong selective sweeps as mutations were rapidly fixed in the population. For each population, a clone was archived and further analyses were performed on that strain. The lineages evolved distinct levels of resistance to the drug combinations (Figure 2), indicating that they acquired different mutations conferring resistance. To gain insight into mechanisms of resistance to the drug combinations, we assessed cross-resistance profiles. Cross-resistance assays were performed in the presence of a fixed concentration of an azole and a gradient of concentrations of the structurally dissimilar counterpart to the Hsp90 or calcineurin inhibitor with which the population was evolved (native inhibitor), as well as with an azole and an inhibitor of the other stress response regulator not targeted in the evolution experiment (naïve inhibitor; i. e. Hsp90 inhibitor if the population was evolved with a calcineurin inhibitor). Cross-resistance profiles can be used to predict candidate resistance mechanisms based on an understanding of how these inhibitors bind to and inhibit their targets (Figure 3). Lineages evolved with azole and FK506 were assayed for resistance to azole and geldanamycin (a naïve inhibitor) as well as to azole and cyclosporin A (a structurally dissimilar calcineurin inhibitor) [39], [40] (Figure 3A). FK506 inhibits calcineurin by forming a complex with the immunophilin Fpr1, and it is the drug-immunophilin complex that binds to and inhibits calcineurin [41]. The structurally unrelated calcineurin inhibitor cyclosporin A binds to a distinct immunophilin, Cpr1, to form a complex that binds to calcineurin and inhibits its function [40]. Geldanamycin inhibits Hsp90 by binding directly to the unconventional Bergerat nucleotide-binding pocket of Hsp90 [42], [43]. The level of resistance to these specific drug combinations suggests several candidate mechanisms of resistance (Figure 3B). For example, resistance to the combination of azole and FK506 but not to azole and other inhibitors tested suggests an FK506-specific mechanism of resistance such as mutation of FPR1. If resistance was also observed to the combination of azole and cyclosporin A, this would suggest that calcineurin has been altered in a way that prevents the binding of both immunophilin-drug complexes, or that a calcineurin-independent mechanism of azole resistance has evolved. If resistance was also observed to the combination of an azole and the naïve inhibitor geldanamycin, this would suggest that resistance emerged by a mechanism that is independent of the stress response regulators Hsp90 and calcineurin; candidate mechanisms include those that block the effect of the azoles on their target, such as up-regulation of the drug efflux pump Pdr5 in S. cerevisiae [28], or alteration of the azole target Erg11 that prevents azole binding [29]. Lineages evolved with azole and geldanamycin were assayed for resistance to azole and radicicol, a structurally unrelated Hsp90 inhibitor. Like geldanamycin, radicicol binds to the unusual nucleotide-binding pocket of Hsp90, inhibiting its chaperone function [42] (Figure 3C). These lineages were also assayed for cross-resistance to azole and FK506, a naïve inhibitor to these strains. Resistance to azole and geldanamycin alone suggests that a mutation in HSP90 occurred that prevents the binding of geldanamycin (Figure 3D). This cross-resistance profile is also consistent with a specific increase in geldanamycin metabolism or efflux. Cross-resistance to azole and FK506 suggests that an azole resistance mechanism evolved that is independent of the stress response regulators Hsp90 and calcineurin. Variation in the patterns of cross-resistance to the distinct drug combinations was observed among the evolved strains (Figure 4, Figure 5, Figure 6, Figure 7), implicating a multitude of distinct resistance mechanisms. Even within a cross-resistance category variation was observed in the level of resistance to the drug combinations between strains, indicating that different mutations were responsible for resistance. This is consistent with the variation in levels of resistance with the native drug combination with which the population was originally evolved (Figure 2). Two S. cerevisiae lineages evolved with azole and geldanamycin, Sc-G-12 and Sc-G-14, displayed different levels of resistance to azole and geldanamycin relative to the ancestral strain (Figure 2C and 2D). Both Sc-G-12 and Sc-G-14 showed increased cross-resistance to azole and radicicol, although Sc-G-14 was able to grow with higher concentrations of azole and geldanamycin as well as azole and radicicol than Sc-G-12 (Figure 4A). Neither lineage showed any cross-resistance to azole and FK506. This suggests distinct mutations in HSP90 might confer resistance to the drug combinations in these lineages. In S. cerevisiae, Hsp90 is encoded by two genes, HSC82, which is expressed at constitutively high levels, and HSP82, which is induced by high temperature [44]. Sequencing of HSC82 and HSP82 in Sc-G-14 identified a non-synonymous point mutation that maps to the N-terminal domain of HSC82, T1350A. This leads to the amino acid substitution I117N, a residue located in the groove lining the nucleotide-binding pocket of Hsp90 to which geldanamycin and radicicol bind. This residue is highly conserved and thought to be 90–100% buried [43]. The impact of HSC82I117N on resistance to azole and geldanamycin was confirmed by performing an allele swap, where HSC82 was deleted from the ancestral strain and the evolved allele was introduced on a plasmid, and reciprocally, HSC82 was deleted from the evolved strain and the ancestral allele was introduced. Expression of the HSC82I117N allele in the ancestral strain conferred a level of resistance to the combination of azole and geldanamycin equivalent to the evolved Sc-G-14 lineage (Figure 4B). Reciprocally, expression of only the ancestral HSC82 allele in the evolved strain abrogated resistance to the drug combination (Figure 4B). This confirms that HSC82I117N confers resistance to the combination of azole and geldanamycin in the Sc-G-14 lineage, perhaps by blocking geldanamycin-mediated inhibition of Hsp90 function. Sequencing of HSC82 and HSP82 in Sc-G-12 identified a 4 bp insertion in HSC82 that results in a frameshift mutation and a premature stop codon in the middle of the coding sequence (HSC82K385*). This mutation is expected to render HSC82K385* non-functional [45]. Surprisingly, deletion of HSC82 in the parental strain confers a slight increase in resistance to azole and geldanamycin that phenocopies the resistance of Sc-G-12 (Figure 4C), suggesting that HSC82K385* is indeed non-functional and confers resistance to the combination of azole and geldanamycin in Sc-G-12. The C. albicans lineage Ca-G-10 exhibited increased resistance to azole and geldanamycin with no cross-resistance to azole and FK506 or azole and radicicol (Figure 4D). It was cross-resistant to azole and 17-AAG (Figure S1), a derivative of geldanamycin, suggesting a mode of resistance specific to ansamycin benzoquinone Hsp90 inhibitors. Sequencing identified a heterozygous, non-synonymous mutation in HSP90, G271T. This mutation causes a D91Y amino acid substitution at a residue in the Hsp90 nucleotide-binding pocket that is thought to be 60–90% buried. This residue is conserved in human Hsp90 although not in S. cerevisiae, where the native amino acid is glutamic acid [43]. To assess the impact of HSP90D91Y on resistance to the combination of azole and geldanamycin we performed an allele swap, replacing one allele of HSP90 in the ancestral strain with the HSP90D91Y allele from the evolved strain, and replacing the HSP90D91Y allele in the evolved strain with the ancestral HSP90 allele. Replacing HSP90D91Y in Ca-G-10 with the ancestral HSP90 allele abrogated resistance in two independent transformants (Figure 4E). Reciprocally, replacing a native allele of HSP90 in the ancestral strain with the HSP90D91Y allele conferred resistance that phenocopied that of Ca-G-10. This indicates that HSP90D91Y confers resistance to azole and geldanamycin and is responsible for resistance of Ca-G-10. Thus, distinct mutations in Hsp90 can block the impact of geldanamycin on azole resistance in both C. albicans and S. cerevisiae, providing a mechanism for resistance to this drug combination. Sc-F-2 and Sc-F-3 were evolved with azole and FK506, and demonstrate no cross-resistance to azole and cyclosporin A or azole and geldanamycin (Figure 5A and 5B), suggesting a mutation in FPR1 may confer resistance to azole and FK506 in these lineages. Sequencing identified a non-synonymous mutation in Sc-F-3 FPR1, G322T. This mutation leads to a V108F amino acid substitution that was responsible for the azole and FK506 resistance, as determined by an allele swap where the FPR1V108F allele was expressed from a plasmid in the ancestral strain in which the native FPR1 allele had been deleted, and reciprocally, the ancestral FPR1 allele was expressed in Sc-F-3 in which the FPR1V108F allele had been deleted. Expression of the FPR1V108F allele in the ancestral strain conferred resistance to azole and FK506, while replacing FPR1V108F with the ancestral allele in Sc-F-3 abrogated resistance (Figure 5C). This mutation likely reduces but does not completely block binding of FK506 to Fpr1 given that complete deletion of FPR1 confers an even greater level of resistance to azole and FK506. Consistent with this mutation conferring resistance to FK506 rather than altering the dependence of the azole resistance phenotype on calcineurin, deletion of the regulatory subunit of calcineurin required for its activation, CNB1, abrogated resistance of Sc-F-3 (Figure 5E). Sequencing FPR1 in Sc-F-2 revealed a tandem duplication of nine amino acids that maps to the middle of the coding sequence, FPRdupG53-D61. Expressing FPRdupG53-D61 in the background of the ancestral strain conferred increased resistance to azole and FK506 (Figure 5D). That resistance was not as strong as in Sc-F-2 is likely due to the difference in expression levels of the native gene and the plasmid borne allele, which is driven by the GPD1 promoter. It is unlikely that there are other mutations affecting resistance in Sc-F-2 given that the resistance phenotypes of the ancestral strain and Sc-F-2 with the plasmid borne FPRdupG53-D61 allele as the sole source of Fpr1 were identical. Further confirming the importance of FPRdupG53-D61 for resistance to azole and FK506, replacing FPRdupG53-D61 of Sc-F-2 with the ancestral FPR1 abrogated resistance (Figure 5D). As with the FPR1 mutation identified in Sc-F-3, the FPRdupG53-D61 mutation in Sc-F-2 likely reduces but does not block binding of FK506 to Fpr1 as deletion of FPR1 confers an even greater level of resistance to azole and FK506 (Figure 5D). As with Sc-F-3, deletion of the regulatory subunit of calcineurin required for its activation, CNB1, abrogated resistance of Sc-F-2, consistent with this duplication in FPR1 conferring resistance to FK506 rather than altering the dependence of the azole resistance phenotype on calcineurin (Figure 5F). Hypothesis driven approaches did not uncover any candidate resistance mutations for several evolved lineages. We therefore turned to whole genome sequencing to provide an unbiased approach to identify mutations that accompany the evolution of resistance to the drug combinations on a genomic scale. For example, S. cerevisiae Sc-F-1 was evolved with azole and FK506 and demonstrated robust resistance to the combination of azole and FK506 as well as azole and cyclosporin A (Figure 6A). This resistance profile suggested a possible mechanism of resistance involving alteration of calcineurin that prevents the binding of both protein-drug immunophilin complexes, or the emergence of a calcineurin-independent azole resistance mechanism. Calcineurin is encoded by the redundant catalytic subunits CNA1 and CNA2 and the regulatory subunit CNB1 in S. cerevisiae [39], [46]. Sequencing of CNA1, CNA2 and CNB1 did not reveal any mutations. Intriguingly, abrogating calcineurin function by deletion of CNB1 did not reduce resistance to azole and FK506 in Sc-F-1, indicating a calcineurin-independent mechanism of resistance had evolved (Figure 6B). Whole genome sequencing at high coverage (Table S1) identified two non-synonymous mutations (Table 2), as well as 58 mutations that were synonymous or in non-coding regions (Table S2 and Table S3); the best candidate for a mutation for affecting resistance was a mutation in MOT3, a transcriptional repressor of ergosterol biosynthesis genes [47]. The non-synonymous substitution in MOT3 resulted in a premature stop codon near the middle of the coding sequence, MOT3G265*, suggesting that this might be a loss-of-function allele. Deletion of MOT3 in the background of the ancestral strain or in Sc-F-1 phenocopied the level of resistance of Sc-F-1, which is consistent with MOT3G265* being a loss-of-function allele that confers resistance in Sc-F-1 (Figure 6C). S. cerevisiae lineage Sc-G-13 was evolved with azole and geldanamycin and demonstrates only a small increase in resistance to this combination, with no cross-resistance to either azole and FK506 or azole and radicicol (Figure 6D). This resistance profile is consistent with a mutation in HSC82 or HSP82 that partially reduces binding of geldanamycin, however, no mutations were identified upon sequencing HSC82 and HSP82. Genome sequencing of Sc-G-13 identified five non-synonymous mutations, as well as 130 that were synonymous or in non-coding regions (Table 2 and Table S3); the best candidate for a mutation affecting resistance was a C2593G mutation in PDR1, which encodes a transcription factor that regulates the expression of numerous multidrug transporters such as PDR5. Gain-of-function mutations in PDR1 are a well-established mechanism of azole resistance that is independent of Hsp90 and calcineurin [16], [30], [48]. The mild resistance phenotype of Sc-G-13 suggested that the PDR1P865R allele in Sc-G-13 confers only a slight increase in drug efflux pump expression. Cross-resistance to azole and FK506 was not observed, likely because FK506 inhibits Pdr5-mediated efflux [49]. To evaluate the importance of the PDR1P865R allele in resistance to azole and geldanamycin we deleted PDR1 from the ancestral strain and the evolved Sc-G-13 lineage and introduced the ancestral PDR1 allele on a plasmid driven by the GPD1 promoter. Replacing the PDR1P865R allele of Sc-G-13 with the ancestral PDR1 allele reduced resistance of Sc-G-13 (Figure 6E). Resistance remained slightly increased relative to the ancestral strain, likely due to higher expression of PDR1 from the GPD1 promoter relative to the native promoter; consistent with this possibility, simply replacing the ancestral PDR1 allele in the ancestor with the same allele on the plasmid conferred a small increase in resistance (Figure 6E). Since there was no difference in resistance phenotype between the ancestral and evolved strains when the plasmid provided the only allele of PDR1, there are likely no other mutations conferring resistance in Sc-G-13. For the six C. albicans lineages evolved with fluconazole and FK506 (Ca-F-4, Ca-F-5, Ca-F-6, Ca-F-7, Ca-F-8, and Ca-F-9), candidate resistance mutations were not identified by hypotheses-based cross-resistance profiles. These lineages shared the same cross-resistance profile of resistance to high concentrations of FK506 and increased resistance to cyclosporin A in the presence of azole (Figure 7A). This profile suggested that either a mutation in calcineurin preventing binding of both drug-immunophilin complexes occurred or a calcineurin-independent mechanism of resistance to azoles evolved. We sequenced the genome of all six lineages of this resistance class. Genome analysis revealed aneuploidies in four of these evolved lineages. For Ca-F-4, we identified extensive aneuploidies in the absence of any non-synonymous mutations (Figure 8). This lineage exhibited increased copy number of chromosomes 4,6 and 7 as well as an increase in copy number of the right arm of chromosome 5. Since approximately half the genome of Ca-F-4 had elevated copy number, resistance might be conferred by a combination of mechanisms including overexpression of the many relevant genes that were amplified including the gene encoding the drug transporter Mdr1, genes encoding ergosterol biosynthetic enzymes, the gene encoding the calcineurin regulatory subunit CNB1, or those encoding regulators of many other cellular pathways. We also identified increased copy number of chromosome 4 in three of the lineages, Ca-F-5, Ca-F-6 and Ca-F-7, as observed in Ca-F-4 (Figure 8). Ca-F-5 also had an increased copy number of chromosome 7. The remaining two lineages, Ca-F-8 and Ca-F-9, had no copy number variation other than variation in chromosome R, which was observed in all of the C. albicans lineages sequenced. Chromosome R contains the genes coding for rDNA, and extensive variation in size of the rDNA array has been observed in experimental populations of C. albicans [50], likely as a consequence of the highly repetitive nature of the genomic context. Two non-synonymous mutations were identified in C. albicans lineage Ca-F-9 (Table 3), and 7 mutations that were synonymous or in non-coding regions (Table S4). The best candidate for a resistance mutation is the C1201A mutation in CNA1, the gene encoding the catalytic subunit of calcineurin; this mutation leads to a premature stop codon, S401*. Truncation of C. albicans Cna1 at position 499 removes the autoinhibitory domain, resulting in a constitutively activated form of calcineurin [51]. Consistent with this mutation conferring resistance to the combination of azole and FK506 or azole and cyclosporin A, deletion of the evolved CNA1 allele in Ca-F-9 abrogates resistance to the combination of azole and calcineurin inhibitor (Figure 7B). Deletion of one allele of CNA1 in the ancestral strain has no effect on sensitivity to the drug combination. Thus, hyperactivation of calcineurin provides a mechanism by which resistance to azoles and calcineurin inhibitors can evolve. Five non-synonymous mutations were identified in the C. albicans lineage Ca-F-8 (Table 3), and 16 mutations that were synonymous or in non-coding regions (Table S4). The best candidate for a resistance mutation is the A1169T mutation identified in orf19. 6438 resulting in non-synonymous substitution, L390F. orf19. 6438 remains uncharacterized in C. albicans but is an ortholog of S. cerevisiae LCB1, which encodes a component of serine palmitoyltransferase that is responsible for the first committed step in sphingolipid biosynthesis, along with Lcb2 [52]. Sphingolipids are a necessary component of the fungal cell membrane and have known interactions with ergosterol [53], while inhibitors of sphingolipid biosynthesis can enhance the efficacy of azoles [54]. To test the model that LCB1L390F confers resistance to the combination of azole and calcineurin inhibitor we used the serine palmitoyltransferase inhibitor myriocin, which inhibits Lcb1 and Lcb2 [55]. Inhibition of Lcb1 with myriocin abrogated resistance to azole and FK506 of the evolved lineage Ca-F-8 but did not affect resistance of Ca-F-9 (Figure 7C), suggesting that LCB1L390F confers resistance to the drug combination. Notably, myriocin caused an increase in resistance of the ancestral strain to azole and FK506 suggesting that resistance phenotypes are exquisitely sensitive to the balance of sphingolipid biosynthesis. Our study provides the first experimental analysis of the evolution of resistance to drug combinations in fungi, illuminating the molecular basis of a transition of drug resistance from dependence on a key stress response regulator to independence, and a diversity of resistance mechanisms that can evolve in response to selection. This work addresses some of the most fundamental questions about the nature of adaptation. One key question is how many mutations underlie adaptive evolution. For all of the lineages for which we functionally tested the importance of mutations identified, we found that a single mutation was responsible for adaptation, in contrast to other experimental evolution studies with S. cerevisiae where multiple adaptive mutations were implicated [56], [57]. The small number of adaptive mutations identified in our study may reflect the short duration of the evolution experiment and the strength of the selection. Despite the limited number of adaptive mutations, we identified a larger number of total mutations in many lineages than reported in other studies [57]. The elevated number of mutations may be specific to the intense drug selection pressure, as bacterial mutation rates can increase in the presence of antibiotic selection [58], and antifungals have been associated with the rapid appearance of aneuploidies and genomic instability [59]. Another central question is how many genetic routes are there to adaptation. Among only 14 evolved lineages, we identified a diversity of adaptive mechanisms including target-based resistance to Hsp90 or calcineurin inhibitors and distinct mutations that render azole resistance independent of cellular stress response regulators, suggesting a complex adaptive landscape with multiple genotypes leading to high fitness adaptive peaks. Exploring the impact of the adaptive mutations on fitness in different environments, including in the absence of drug, will be key to understanding fitness costs of drug resistance, evolutionary trade-offs, and the limits of adaptation. By starting the evolution experiment with strains that are resistant to azoles in a manner that depends on Hsp90 and calcineurin, we provide relevance for a clinical context where Hsp90 and calcineurin inhibitors could be deployed in combination with azoles to render azole-resistant isolates responsive to treatment. There is some precedent for the evolution of resistance to these drug combinations, as clinical isolates recovered from an HIV-infected patient over the course of two years evolved increased resistance to the combination of azole and inhibitors of Hsp90 or calcineurin [16]. While this patient was not treated with Hsp90 or calcineurin inhibitors, fever may have provided the selection for Hsp90 independence given that febrile temperatures cause global problems in protein folding that can overwhelm Hsp90 function and reduce azole resistance in a manner that phenocopies Hsp90 inhibition [16]. In our experimental evolution study, most of the 290 lineages initiated went extinct, while the 14 lineages that evolved resistance to the combination of azole and inhibitor of Hsp90 or calcineurin acquired a diversity of resistance mechanisms. These resistance mechanisms included mutations that rendered erg3-mediated azole resistance independent of the stress response regulator calcineurin, mutations that blocked the effects of the Hsp90 or calcineurin inhibitor, and large-scale aneuploidies. This experimental evolution approach provides a powerful system to predict the mechanisms by which resistance to drug combinations may evolve in the clinic. Consistent with the relevance of our findings, the increased resistance to azole and inhibitor of Hsp90 or calcineurin in isolates that evolved in an HIV-infected patient was accompanied by mutations causing overexpression of multidrug transporters [16], [60], as expected for the PDR1 mutation identified in one of our lineages. One of the most prevalent mechanisms of resistance identified in our evolved populations was mutation in the target of the drug used in combination with azole during the evolution experiment. For Hsp90 inhibitors, it has been predicted that the probability of target-based resistance would be relatively low given that the amino acid residues in the nucleotide-binding site of Hsp90 family members are highly conserved from bacteria to mammals [61], suggesting that mutations that confer resistance would likely inactivate this essential molecular chaperone. This has helped fuel research on Hsp90 as a target for development of anti-cancer drugs, where inhibiting Hsp90 can impair the function of a multitude of oncoproteins [62]–[64]. Despite the constraints, there is precedent for point mutations in Hsp90 conferring resistance to Hsp90 inhibitors. One study engineered S. cerevisiae strains to be hypersensitive to drugs and expressed yeast or human Hsp90 as the sole source of the chaperone; introduction of a single point mutation (A107N for yeast, A121N for human Hsp90α, and A116N for human Hsp90β) conferred resistance to Hsp90 inhibitors [65]. Further, the fungus that produces radicicol, Humicola fuscoatra, harbours an Hsp90 with reduced binding affinity to radicicol but not geldanamycin [66]. Three of our evolved lineages acquired substitutions in Hsp90 that rendered erg3-mediated azole resistance more recalcitrant to the effects of Hsp90 inhibitors (Figure 4). For one S. cerevisiae lineage (Sc-G-14) and one C. albicans lineage (Ca-G-10), the mutations were in the nucleotide-binding domain, consistent with impairing drug binding. For S. cerevisiae lineage Sc-G-12, the mutation led to a premature stop codon (K385*); consistent with this HSC82 mutation causing a loss of function, deletion of HSC82 in the parental strain phenocopied resistance of Sc-G-12. Reducing dosage of a drug target often confers hypersensitivity to the drug rather than resistance [67]; this may suggest compensatory upregulation of the other S. cerevisiae gene encoding Hsp90, HSP82, which could confer elevated resistance. Target-based resistance to Hsp90 inhibitors has yet to emerge in Hsp90 inhibitor clinical trials, suggesting that these mutations may be associated with a fitness cost. Mutations in the drug target also emerged as a mechanism that renders erg3-mediated azole resistance recalcitrant to the effects of calcineurin inhibitors in our evolved lineages. Two S. cerevisiae lineages acquired mutations in FPR1, which encodes the immunophilin that FK506 must bind to in order to form the protein-drug complex that inhibits calcineurin function [41]. A V108F substitution was identified in Sc-F-3 and a nine amino acid duplication near the protein midpoint was identified in Sc-F-2 (dupG53-D61). These alterations likely reduce but do not block FK506 binding, given that deletion of FPR1 conferred a higher level of FK506 resistance (Figure 5). There is precedent for overexpression or disruption of FPR1 conferring resistance to FK506 in S. cerevisiae [68], as well as for a W430C amino acid substitution in one of the two redundant calcineurin catalytic subunits Cna2 [69]. One C. albicans lineage, Ca-F-9, acquired a mutation in the catalytic subunit of calcineurin, CNA1C1201A, which results in a S401* premature stop codon that confers resistance to azole and both FK506 and cyclosporin A (Figure 7), likely due to hyperactivation of calcineurin [51]. Despite the emergence of target-based resistance to calcineurin inhibitors in vitro, there may be significant constraints that minimize the emergence of resistance in the human host. FK506 (tacrolimus) and cyclosporin A are front line immunosuppressants broadly used in the clinic to inhibit calcineurin function, thereby blocking T-cell activation in response to antigen presentation and suppressing immune responses that contribute to transplant rejection [39], [70]. Invasive fungal infections occur in ∼40% of transplant recipients including those that receive a calcineurin inhibitor as an immunosuppressant [71], however, this immunosuppressive therapy does not select for resistance to calcineurin inhibitors in C. albicans or Cryptococcus neoformans recovered from these patients [72], [73]. That resistance has not been observed in the host suggests that the resistant mutants may have reduced fitness relative to their sensitive counterparts or that other selective constraints alter the evolutionary dynamics. Several of our evolved lineages took a distinct evolutionary trajectory, and evolved azole resistance mechanisms that are independent of the cellular stress response regulators. S. cerevisiae lineage Sc-F-1 evolved cross-resistance to azole and FK506 as well as azole and cyclosporin A (Figure 6). The azole resistance phenotype was independent of calcineurin but dependent on Hsp90 (Figure 6), suggesting a resistance mechanism that is contingent upon distinct Hsp90 downstream effectors, such as Mkc1 [74]. We identified an adaptive mutation in MOT3 (Table 2), a transcriptional repressor of ergosterol biosynthesis genes [47], which resulted in a premature stop codon, G265* and likely a loss-of-function allele (Figure 6C). Loss of function of Mot3 would lead to overexpression of ergosterol biosynthesis genes, which could minimize the impact of azoles on their target or could lead to a change in sterol balance that reduces the dependence of azole resistance on calcineurin. Changes in membrane composition may also explain the resistance of C. albicans lineage Ca-F-8 to azoles and calcineurin inhibitors, which was attributed to a mutation in the ortholog of S. cerevisiae LCB1 (Figure 7), encoding a regulator of sphingolipid biosynthesis. Notably, Mot3 is also a prion protein, which can convert between structurally and functionally distinct states, at least one of which is transmissible [75]; changes on Mot3 conformation and activity can modulate phenotypic variation in S. cerevisiae, and thus may influence the evolution of drug resistance phenotypes. S. cerevisiae lineage Sc-G-13 evolved a small increase in resistance to azole and geldanamycin associated with a mutation in PDR1, which encodes a transcription factor that regulates the expression of drug transporters such as PDR5 (Figure 6). Gain-of-function mutations in PDR1 are known to confer azole resistance that is independent of Hsp90 and calcineurin [16], [30], [48]. Cross-resistance to azole and FK506 may not have been observed because FK506 inhibits Pdr5-mediated efflux [49]. The weak resistance phenotype could reflect a small increase in transporter expression, or a fitness cost of the PDR1 mutation in an erg3 mutant background [76]. Several of the C. albicans lineages that evolved resistance to azole and calcineurin inhibitors demonstrated a complex genomic landscape of aneuploidies. The emergence of azole resistance in C. albicans has been associated with general aneuploidies as well as the formation of a specific isochromosome composed of two left arms of chromosome 5 (i5L) [77]. The isochromosome confers azole resistance due to increased dosage of two genes located on the left arm of chromosome 5: ERG11, which encodes the target of the azoles; and TAC1, which encodes a transcriptional regulator of multidrug efflux pumps [78]. Our lineages were resistant to azoles at the outset of the experiment, suggesting that the aneuploidies emerged in response to stress or were selected as a mechanism of resistance to the drug combination. Ca-F-4, Ca-F-5, Ca-F-6, and Ca-F-7 all had numerous aneuploidies relative to the parental strain (Figure 8). One aneuploidy that was common to all four lineages was increased copy number of chromosome 4, suggesting that an important resistance determinant might reside on this chromosome. While one might predict that such aneuploidies would be associated with a fitness cost, it is notable that a previous analysis of isolates carrying the i5L isochromosome demonstrated improved fitness in the presence and absence of azoles, relative to their drug-sensitive counterpart [59]. In contrast, many azole resistance mutations are associated with a fitness cost [79], though this cost can be mitigated with further evolution [80]. The prevalence of aneuploidies in the C. albicans lineages underscores the remarkable genomic plasticity of this pathogen [81], and the diversity of genomic alterations that can accompany adaptation. The landscape of genetic and genomic changes observed in our evolved lineages illuminate possible mechanisms by which resistance to drug combinations might evolve in the human host and suggest candidate targets to minimize the emergence of resistance. Despite optimizing our selection conditions to favour the evolution of resistance to the drug combination, the majority of lineages went extinct (Figure 1). Consistent with constraints that minimize the evolution of resistance to these drug combinations, treatment of organ transplant patients with calcineurin inhibitors has not yielded resistance to these drugs in fungal pathogens recovered from these patients despite the extensive use of these drugs in patient populations [72], [73]. While Hsp90 inhibitors remain at the clinical trial stage for cancer and other diseases [62], [63], [82], [83], resistance has yet to emerge in these patient populations. Although there are a multitude of mechanisms that can confer resistance to the drug combinations, they may not be favoured due to fitness costs in the complex host environments. The mechanisms by which resistance to the drug combinations evolved in our lineages suggest novel targets that could be exploited to block the evolution of drug resistance. Drug interactions have tremendous potential to influence the evolution of drug resistance [84]. Elegant studies with antibacterials emphasize that the impact of these interactions are often more complex than anticipated [8], [85]–[87]. While synergistic interactions that yield inhibitory effects larger than expected from individual drugs can maximize the rate at which infection is cleared, antagonistic interactions that yield inhibitory effects smaller than expected can suppress the evolution of multi-drug resistance. Ultimately, a systems biology approach incorporating experimental evolution, genetics and genomics, and clinical samples will be crucial for the development of effective strategies to enhance the efficacy of antimicrobial agents and minimize the evolution of drug resistance. All Saccharomyces cerevisiae and Candida albicans strains were archived in 25% glycerol and maintained at −80°C. Strains were typically grown and maintained in rich medium (YPD: 1% yeast extract, 2% bactopeptone, 2% glucose, with 2% agar for solid medium only), or in synthetic defined medium (SD, 0. 67% yeast nitrogen base, 2% glucose, with 2% agar for solid medium only), supplemented with amino acids, as indicated. Strains were transformed using standard protocols. Strains used in this study are listed in Table S5. Strains were constructed as described in Text S1. Plasmids were constructed using standard recombinant DNA techniques. Plasmids used in this study are listed in Table S6 and oligonucleotides used in this study are listed in Table S7. Plasmids were constructed as described in Text S1. All plasmids were sequenced to confirm the absence of spurious non-synonymous mutations. Evolution experiments were initiated with three ancestral strains of erg3-mediated azole resistant strains: two haploid S. cerevisiae strains (erg3Δ and erg3W148*) and one C. albicans strain (erg3Δ/erg3Δ; see Table S5). A founder colony was established for each ancestral strain and grown overnight in liquid, rich medium (YPD) without drug. From here, culture was transferred to a plate containing YPD with combinatorial drug concentrations of azole (fluconazole or miconazole) and geldanamycin, or azole (fluconazole or miconazole) and FK506 (i. e. treatments; see Table 1). Geldanamycin and FK506 were selected based on their specificity of target inhibition and their capacity to abrogate erg3-mediated azole resistance [16]; fluconazole and miconazole were selected as clinically relevant azoles of the triazole and imidazole class, respectively [4], [37]. Treatments were selected for the evolution experiment based on growth phenotype in the dose response matrices (Figure S2), such that strong directional selection for resistance would be applied. Concentrations were also varied to favour the emergence of distinct mechanisms of resistance. Lineages were then propagated in replicate in either 96-well plates (Sarstedt; 48 lineages initiated in this format) or 24-well plates (Becton Dickinson Labware; 242 lineages initiated in this format). The plates were formatted as described in Figure 1B. For propagation in 96-well plates, 1 µl of culture was transferred from the overnight culture to a final volume of 100 µl. Lineages were grown in a Tecan GENios plate reader and incubator at 30°C with constant agitation for three days. Subsequently, 1 µl of culture was transferred to a new plate containing YPD and treatment. Transfers occurred every 3 days to allow slow growing lineages to reach carrying capacity. This process was repeated until robust growth was present in some treatment wells. The experimental design for lineages propagated in 24-well plates was the same with the following adjustments: different drug combinations were selected for treatments; 10 µl of culture was transferred to 990 µl of YPD with treatment; plates were maintained at 30°C with constant agitation in a shaking incubator and transfers occurred every two days. With this dilution factor of 1/100, ∼6. 6 generations occurs between transfers. The effective population size per lineage of ∼4. 6×106 was estimated as described [88], given that cultures reached saturation of ∼107 cells/ml between transfers. Lineages that demonstrated reproducible resistance to the drug combination in which they were propagated were archived. Lineages unable to grow in the presence of the drug combination, either from when the cultures were initiated or over the course of the evolution experiment, were considered extinct. A summary of treatment concentrations, number of transfers and type of plate evolved in can been found in Table 1. Resistance to drug combinations was assayed in 96-well microtiter plates, as previously described [16], [21]. Minimum inhibitory concentration (MIC) assays were set up to a final volume of 0. 2 ml/well. MICs were performed in the absence of fluconazole (Sequoia Research Products) or with a constant concentration of fluconazole or miconazole (Sigma–Aldrich Co.), as indicated in the figures. All gradients were two-fold dilutions per step, with the final well containing no drug. The starting concentration of geldanamycin (Invivogen) gradients was 50 µM for S. cerevisiae strains and 5 µM for C. albicans strains. The starting concentration of FK506 (A. G. Scientific) gradients was 6 µM for S. cerevisiae strains and 100 µM for C. albicans strains. The starting concentration of radicicol (A. G. Scientific) gradients was 25 µM for both S. cerevisiae and C. albicans strains. The starting concentration of cyclosporin A (Calbiochem) gradients was 50 µM for both S. cerevisiae and C. albicans strains. The cell densities of overnight cultures were determined and diluted to an inoculation concentration of ∼103 cells/well. Plates were incubated at 30°C in the dark for the period of time specified in the figure legend. Cultures were resuspended and absorbance at 600 nm was determined using a spectrophotometer (Molecular Devices) and corrected for background of the corresponding medium. OD measurements were standardized to either drug-free or azole-only control wells, as indicated. Data was plotted quantitatively with colour using Java Treeview 1. 1. 3 (http: //jtreeview. sourceforge. net/). Resistance phenotypes were assessed on multiple occasions and in duplicate on each occasion with concordant results, validating that the phenotypes are reproducible and stable. Dose response matrices, or checkerboard assays, were performed to a final volume 0. 2 ml/well in 96-well microtiter plates, as previously described [74]. Two-fold dilutions of fluconazole were titrated along the X-axis from a starting concentration of 256 µg/ml, with the final row containing no fluconazole. Along the Y-axis, either geldanamycin or FK506 was titrated in two-fold dilutions with the final column containing no geldanamycin or FK506. The starting concentration of geldanamycin was 5 µM for checkerboards with either S. cerevisiae or C. albicans strains. The starting concentration of FK506 was 4 µM for checkerboards with S. cerevisiae and 40 µM for checkerboards with C. albicans strains. Concentrations were selected to cover a range that spanned from no effect on growth to near complete inhibition of growth. Plates were inoculated and growth assessed as was performed for MIC assays. Fluconazole was dissolved in sterile ddH2O. The Hsp90 inhibitors geldanamycin and radicicol and the calcineurin inhibitors FK506 and cyclosporin A were dissolved in DMSO. Myriocin (Sigma) was dissolved in methanol. C. albicans cell pellets were digested with R-Zymolase for 1 hour (Zymo Research, D2002), prior to genomic DNA extraction with phenol-chloroform (EMD Millipore, EMD6810), and sodium acetate precipitation. Whole genome libraries were prepared using Nextera XT kits (Illumina, FC-131-1096) according to manufacturer' s protocol. Libraries were sequenced on the Illumina HiSeq2000 platform using paired reads (101 bp) and version 3 reagents and chemistry. The yeast genomes were sequenced in a multiplexed format, where an oligonucleotide index barcode was embedded within adapter sequences that were ligated to genomic DNA fragments [89]. Only one mismatch per barcode was permitted to prevent contamination across samples. Next, the sequence reads were filtered for low quality base calls trimming all bases from 5′ and 3′ read ends with Phred scores < Q30. Trimming sequence reads for low quality base calls drastically lowered false positive SNV calls. De-multiplexed and trimmed reads from the S. cerevisiae strains were aligned to the S288c 2010 genome, a high fidelity sequence from an individual yeast colony (from F. Dietrich' s lab at Duke University; it is the SGD reference genome as of February 2011) [90]. Reads from the C. albicans strains were aligned to the SC5314 genome from CGD [91]. While C. albicans is an obligate diploid, the current build of the genome, assembly 21, is a haploid genome, and is more accurate than the original diploid genome, assembly 19 [92], [93]. The diploid assembly was not used because it features 412 supercontigs with non-obvious heterozygosity, whereas the haploid assembly has been curated and organized into 8 chromosomes [93]. Sequence reads were aligned with Bowtie2, which was chosen over other commonly used short-read aligners such as Illumina' s Eland, Maq, SOAP and BWA because it has been reported to be one of the fastest accurate aligners [94]–[98]. Additionally, it was chosen because it is updated frequently, supports variable read lengths within a single input file, is multi-threaded with a minimal memory and temporary file footprint and supports the standard Sequence Alignment/Map (SAM) file format [94], [98]–[100]. Alignments and all subsequent sequence data were visualized using the Savant Genome Browser [101]. The average coverage of each genome was calculated and was sufficient for confident variant detection (Table S1). Aligned sequence reads for S. cerevisiae were subsequently processed using the UnifiedGenotyper package of the Genome Analysis Toolkit (GATK), which features a comprehensive framework for discovering SNVs and calculating coverage across genomic data [102], [103]. Variants detected in the S. cerevisiae parental strains were subtracted from complete variant lists, yielding a set of novel variants that emerged during strain growth in the presence of drug. Since C. albicans is diploid, we processed the reads with a more accurate approach using the probabilistic framework JointSNVMix, which uses paired parental and evolved strain sequence data to determine significant novel variants [104]. After identifying candidate SNVs, the threshold for homozygous SNV calls for both haploid (S. cerevisiae) and diploid (C. albicans) systems was set to 85% alternate (non-reference) basecalls at a specific position. In a diploid system, 35% was the threshold set to identify heterozygous SNVs. All variant positions required a minimum coverage of 15× to be considered as a candidate SNV. The total number of high-confidence novel mutations agrees with mutation rates observed previously for S. cerevisiae (Table S3) [105]. To further verify that the sequence data are of high quality, we compared two distinct sequence runs from two different sequence library preparations of the same parent C. albicans strain CaLC660. The total number of diploid single nucleotide variants that exist between the parent strain and the reference genome (SC5314) is 3748, therefore there is 99. 99% concordance between both sequence replicates (Table S8). The software package CNV-seq was used to identify chromosomal regions that varied in copy number between parental strains and evolved lineages [106]. This analysis found no significant CNVs in the S. cerevisiae strains, but numerous large variants were observed in C. albicans. Sequence data is publicly available on the NCBI Short Read Archive with accession SRA065341.

Fungal infections are a leading cause of mortality worldwide and are difficult to treat due to the limited number of antifungal drugs, whose effectiveness is compromised by the emergence of drug resistance. A powerful strategy to combat drug resistance is combination therapy. Inhibiting the molecular chaperone Hsp90 or its downstream effector calcineurin cripples fungal stress responses and abrogates drug resistance. Here we provide the first analysis of the genetic and genomic changes that underpin the evolution of resistance to antifungal drug combinations in the leading human fungal pathogen, Candida albicans, and model yeast, Saccharomyces cerevisiae. We evolved experimental populations with combinations of inhibitors of Hsp90 or calcineurin and the most widely used antifungal in the clinic, the azoles, which inhibit ergosterol biosynthesis. We harnessed whole-genome sequencing to identify diverse resistance mutations among the 14 lineages that evolved resistance to the drug combination. These included mutations in genes encoding the drug targets, a transcriptional regulator of multidrug transporters, a transcriptional repressor of ergosterol biosynthesis enzymes, and a regulator of sphingolipid biosynthesis. We also identified extensive aneuploidies in several C. albicans lineages. Our study reveals multiple mechanisms by which resistance to drug combination can evolve, suggesting new strategies to combat drug resistance.

Abstract Introduction Results Discussion Materials and Methods

mycology model organisms fungal evolution microbial evolution microbial pathogens yeast and fungal models biology microbiology yeast

2013

Genetic and Genomic Architecture of the Evolution of Resistance to Antifungal Drug Combinations

During meiotic prophase, chromosomes display rapid movement, and their telomeres attach to the nuclear envelope and cluster to form a “chromosomal bouquet. ” Little is known about the roles of the chromosome movement and telomere clustering in this phase. In budding yeast, telomere clustering is promoted by a meiosis-specific, telomere-binding protein, Ndj1. Here, we show that a meiosis-specific protein, Csm4, which forms a complex with Ndj1, facilitates bouquet formation. In the absence of Csm4, Ndj1-bound telomeres tether to nuclear envelopes but do not cluster, suggesting that telomere clustering in the meiotic prophase consists of at least two distinct steps: Ndj1-dependent tethering to the nuclear envelope and Csm4-dependent clustering/movement. Similar to Ndj1, Csm4 is required for several distinct steps during meiotic recombination. Our results suggest that Csm4 promotes efficient second-end capture of a double-strand break following a homology search, as well as resolution of the double-Holliday junction during crossover formation. We propose that chromosome movement and associated telomere dynamics at the nuclear envelope promotes the completion of key biochemical steps during meiotic recombination. Meiotic recombination promotes the faithful segregation of homologous chromosomes at meiosis I (MI) by creating physical linkages between the homologs [1], [2]. Recombination produces two types of products: crossovers (COs) and non-crossovers (NCOs). Only COs mature into exchanges between chromosome axes called chiasmata, which together with arm cohesion ensure homolog separation. Recombination during meiosis is initiated by the formation of double-strand breaks (DSBs) at recombination hotspots [3]. A protein complex containing the Spo11 core catalytic subunit is involved in DSB formation. Resection of DSB ends results in the formation of single-stranded DNA (ssDNA), which is then used in the search for homologous DNA sequences. The homology search is catalyzed by two RecA homologs, Rad51 and Dmc1 with their accessory factors [4]–[7]. This homology search results in the invasion of ssDNA into duplex DNA, and the formation of a single-end invasion intermediates [SEIs; 8]. SEIs undergo second-end capture of the DSB to form a second prominent joint molecule, called the double-Holliday junction (dHJ), which is primarily resolved to form COs [9]. The intermediate required to form NCOs has yet to be identified. Importantly, the homology search resulting in SEI formation appears to be biochemically and temporally distinct from the second-end capture steps [8], [10]. CO formation is regulated by the action of a group of proteins called ZMM or SIC (synaptic initiation complex; hereafter called ZMM for simplicity). Members of the ZMM group include Zip1, Zip2, Zip3, Msh4, Msh5, Mer3, Spo16, and Spo22/Zip4 [11]–[16]. Mer3 and Msh4–Msh5 possess helicase and structure-specific DNA-binding activities, respectively [17], [18]. Zip3, together with the Zip2–Spo16–Spo22 adaptor complex, is thought to catalyze the post-translational modification of target protein (s), e. g. , sumoylation or ubiquitylation [15], [19]. Zip1 is a component of the synaptonemal complex [20]. The ZMM proteins ensure the formation of wild-type CO levels [12], [16]. In addition to the ZMM-dependent CO pathway, budding yeast has two additional pathways for recombination: a minor CO pathway and a NCO pathway, both dependent on the junction resolvase Mus81–Mms4 [21], [22]. One of the most notable features in meiosis is chromosome dynamics and morphogenesis. In most organisms, synapsis of homologous chromosomes is facilitated by the recombination. Synapsis culminates in the formation of SC, a tripartite structure seen in pachytene [23], [24]. In leptotene when DSBs are formed, sister chromatids form chromatin loops along a shared axis (the axial element). Leptotene is followed by zygotene, in which short patches of SC form between homologous axial elements. Elongation of SC occurs along entire chromosomes, resulting in the formation of full-length SC in pachytene. SCs are then disassembled in the diplotene. Importantly, SC formation is tightly coupled with CO formation. Formation of SEIs and dHJs occurs at the leptotene-zygotene and zygotene-pachytene transitions, respectively [8], [12]. Resolution of dHJs occurs during late pachytene. In the vegetative growth phase of S. cerevisiae, centromeres are present near the Spindle Pole Body (SPB), a fungal equivalent of the centrosome in other eukaryotes. In S. cerevisiae, the SPB is embedded in the nuclear envelope (NE), and telomeres are clustered and often associated with the NE in a dispersed distribution (Klein et al. 1992). This configuration of chromosomes in vegetative cells is referred to as the “Rabl” orientation. In meiotic prophase, cells undergo a drastic change in their chromosome configuration. Centromeres detach from the SPB, while telomeres cluster in one area of the nuclear membrane near the SPB. This chromosomal bouquet configuration is prominently seen only during zygotene. The bouquet is a conserved feature in the meiotic prophase of most eukaryotes, but its function remains unknown [23]. In S. cerevisae, a meiosis-specific telomere-binding protein, Ndj1 [25], [26], is involved in tethering telomeres to the nuclear membrane and promoting bouquet formation [27]. ndj1 mutation reduces spore viability and confers some defects in recombination [25], [26], [28]. In S. pombe, the Bqt1–Bqt2 complex promotes bouquet formation through interactions with a telomere-binding protein, Taz1 [29]. The bouquet is thought to facilitate pairing of homologous chromosomes by restricting the homology search to a smaller area. In this study, we found that a meiosis-specific protein, Csm4 [30], promotes efficient transition from SEIs to dHJs as well as resolution of dHJs in the CO-specific recombination pathway. These results suggest that Csm4 regulates various steps during meiotic recombination. Recombination-related phenotypes in csm4 mutants are very similar to those seen in ndj1 mutants [28]. Csm4 forms a complex with Ndj1 in vivo. We also found that similar to ndj1 mutants, csm4 mutants are deficient in bouquet formation, but unlike ndj1 mutants, they are proficient in tethering telomeres to the NE. These results suggest that chromosome architecture and/or dynamics, which are mediated by the tethering telomeres to the NE, control various biochemical steps during meiotic recombination. The accompanying paper by Wanat et al. (2008) shows similar and complementary results [31]. Previous analysis showed that csm4 mutants are defective in the segregation of chromosomes during meiosis [30]. However, little is known about the functions of Csm4 in meiosis. We re-analyzed the meiotic phenotypes of csm4 mutants in an SK1 background. Consistent with a previous study [30], the csm4 mutation reduces spore viability to 66%, as compared to 96% in the wild type. Interestingly, 4-, 2-, and 0-viable spore tetrads exceed 3- and 1-viable spore tetrads, suggesting non-disjunction of homologs at MI (Figure 1A) Similar results have been described by Wanat et al. in the accompanying paper [31]. Furthermore, csm4 mutation delays its entry into MI by 5 h (Figure 1B). This delay is suppressed by introducing a mutant allele of SPO11, spo11-Y135F, which abolishes the catalytic function [3], [32]. Similar results have been described by Wanat et al. [31], suggesting that the delay seen in the csm4 mutant is due to a defect in meiotic recombination. The delay is also suppressed by the introduction of a mutation of the RED1 gene (Figure 1B), which encodes a component of the axial element of the SC [33], is necessary for DSB formation [34], and acts as a barrier to inter-sister recombination [35], [36]. We then analyzed the turnover of meiotic DSBs at the HIS4-LEU2 recombination hotspot [37] in the csm4 mutant (Figure 1C). In the wild type, DSBs appear at 3 h after incubation in the sporulation medium (SPM) and then disappear at around 6 h (Figure 1D). The csm4 mutant accumulates DSBs up to a slightly higher level compared to the wild type. Formation of DSBs in the mutant is slightly delayed, and disappearance of the DSBs is delayed by 4 h. At 8 h, DSBs are still detected in the mutant. These results indicate that CSM4 is required for the efficient conversion of DSBs into later-stage recombination intermediates. Next, we examined the formation of crossovers (COs) in csm4. Consistent with delayed DSB repair, CO formation in csm4 is delayed by approximately 4 h compared to the wild type (Figure 1E). Similar results have been described by Wanat et al. in the accompanying paper [31]. However, the final level of COs at the HIS4-LEU2 locus is similar to the wild type (92% of the wild-type level). In addition to COs, meiotic recombination produces non-crossovers (NCOs). CO and NCO recombinants can be distinguished using restriction site polymorphisms around DSB site I in the HIS4-LEU2 locus [38]. As seen for COs, NCOs in the csm4 mutant are formed 5 h later than in the wild type (Figure 1F). In this assay, the final level of COs in the mutant is slightly higher (1. 2-fold) than the wild type. Wanat et al. show a slight reduction of NCOs using the same assay [31]. The level of NCOs in csm4 is reduced to 75% of the wild type. This suggests that Csm4 is required for timely and efficient formation of both types of recombinants. This was confirmed using a heteroduplex assay that detects CO and NCO at the same locus (Figure 1G). The final level of NCOs containing heteroduplex DNA at the MluI/BamHI site in the mutant is reduced to 50% that of the wild-type level, while the level of COs containing heteroduplex DNA is unaffected by csm4 mutation. Interestingly, the csm4 mutant increases ectopic recombination between HIS4-LEU2 and leu2: : hisG on chromosome III (Figure 1G; [39]). Meiotic recombination has been grouped into two CO pathways and a single NCO pathway [21]. One major pathway for COs depends on ZMM proteins [12] and the other depends on the junction-specific resolvase, Mus81–Mms4 [22]. To examine a possible role for Csm4 in these pathways, we constructed a csm4 mutant with a mutation in MSH4, which encodes a meiosis-specific MutS homolog that acts in the ZMM pathway [40]. csm4 and msh4 single mutants display reduced spore viability (66 and 29%, respectively; Figure 1A). The csm4 msh4 double mutant shows more severe defects in spore viability (18%) than either single mutant. In the CO/NCO assay, msh4 affects formation of both COs and NCOs (Figure 2A). As reported previously [12], at 30°C, msh4 mutation decreases the final amount of COs to 50% that of the wild type, but increases the level of NCOs to 1. 7-fold of the wild type. The csm4 msh4 double mutant shows more severe defects in CO formation; the final level of COs in the double mutant is significantly reduced compared to csm4 and msh4 single mutants. However, the amount of NCOs in the csm4 msh4 double mutant is only slightly reduced compared to the csm4 single mutant. This suggests that Csm4 functions in meiotic recombination independently of Msh4 and that Csm4 promotes CO formation in the absence of Msh4. Furthermore, Msh4 is not necessary for residual NCO formation in the absence of Csm4. Next, we constructed the csm4 mms4 double mutant. Unlike either single mutant, the csm4 mms4 double mutant cannot form spores. In the CO/NCO assay (Figure 2B), the mms4 single mutant exhibits a delay in formation of COs and reduces CO levels to 73% of the wild type [41]. Interestingly, NCOs in the mms4 mutant appear at the same time as in the wild type and at levels that are 1. 5-fold higher than the wild type. The csm4 mms4 double mutant shows an effect on NCO formation similar to the csm4 single mutant. Similar to csm4, CO formation in the double mutant is delayed, but reaches an almost wild-type level. These observations suggest that Csm4 works upstream of Mms4 in meiotic CO and NCO recombination pathways. We also examined the amount of DSBs formed in the csm4 mutant in the rad50S background, which blocks processing of DSB ends [42]. The csm4 rad50S double mutant accumulates DSBs like the rad50S mutant (Figure 2E). Similar results have been described by Wanat et al. in the accompanying paper [31]. DSB levels in the double mutant were slightly higher than those seen in rad50S. As shown above, Csm4 is necessary for timely CO formation recombination pathway, which mainly depends on ZMM proteins such as Msh4. In the ZMM-dependent CO pathway, single-end invasions (SEIs) and double-Holliday junctions (dHJs) have been identified as major recombination intermediates [8], [9]. We analyzed the effect of csm4 mutation on the formation of these intermediates, which can be detected at HIS4-LEU2 (Figure 3A) in 2D gel electrophoresis after cross-linking DNA samples with psoralen [8], [9]. In the wild type, SEIs begin to appear at 3 h, peak at 4. 5 h, and disappear at around 6 h (Figure 3B and 3D). In contrast, the csm4 mutant shows a slight delay in the onset of SEI formation, and SEIs persist at later times during meiosis (Figure 3C and 3D). At 8 h, a significant level of SEIs could be detected in the csm4 strains. Although delayed, SEIs are turned over in the mutant at around 12 h. dHJs in the wild-type cells start to appear at 4. 5 h, peak at 5 h, and then disappear (Figure 3B and 3E). In csm4, formation of dHJs is delayed by 3. 5 h compared to the wild type (Figure 3C and 3E). The maximum level of dHJs in the mutant at 8 h is slightly higher than in the wild type. Furthermore, the resolution of dHJs is clearly delayed in the mutant. These data suggest that csm4 mutation affects various steps of CO formation, likely during the SEI–dHJ transition and dHJ resolution. Similar results but with more quantitative analysis of recombination intermediates have been described in the accompanying paper by Wanat et al. [31]. We analyzed the localization of RecA homologs on meiotic chromosome spreads by immunostaining. Eukaryotic RecA homologs Rad51 and meiosis-specific Dmc1 both act in the homology search/strand exchange process that results in SEI and dHJ formation [6], [43], [44]. In the wild type, Rad51 as well as Dmc1 shows punctate staining, or foci [44], [45]. Rad51 foci begin to appear at 3 h, peak at 4 h, and then disappear at later times (Figure 4A). The kinetics of Rad51 focus formation is very similar to that of DSBs. In the csm4 mutant, the formation of Rad51 foci is slightly delayed compared to the wild type (Figure 4B and 4C), consistent with a delay in DSB formation in the mutant. Disassembly of Rad51 foci is clearly delayed in the csm4 mutant, indicating inefficient repair of DSBs. The average number of Rad51/Dmc1 foci in the csm4 mutant at 4 h is 42. 8 for Rad51 and 40. 5 for Dmc1 (per total nucleus), which is higher than that seen in the wild type (22. 6 and 24. 8 for Rad51 and Dmc1, respectively). At later time points, much brighter and larger Rad51 foci, possibly representing aggregates, are observed in the mutant (Figure 4B). These aggregates appear to be specific to csm4, since other mutants, which also accumulate Rad51/Dmc1 foci at later times (e. g. , tid1, mnd1, and hop2), do not accumulate these structures [45]–[47]. Dmc1 in csm4 shows a staining pattern similar to that seen for Rad51 (Figure 4B and 4C). These data suggest that CSM4 is necessary for a step after loading of Rad51 and Dmc1, e. g. , during the homology search. To examine the effect of csm4 on chromosomal synapsis, i. e. , formation of the synaptonemal complex (SC) during meiotic prophase, we stained chromosome spreads with an antibody against the Zip1 protein, which is a component of the central element of the SC [20]. In leptotene, Zip1 shows dotty-staining in the wild type (2–3 h; class I; Figure 4Di). In zygotene (3–5 h), short lines of Zip1 (class II; Figure 4Dii) are observed in addition to the Zip1 foci. At pachytene (5–7 h), Zip1 elongates along entire chromosomes (class III; Figure 4Diii), indicating full chromosome synapsis. The csm4 mutant shows a deficiency in SC formation. Similar to the wild type, Zip1 foci form in the mutant (Figure 4Ei). Zip1 starts to elongate, but full chromosome synapsis is rarely seen in the mutant (class II' ; Figure 4Eii). As a result, the csm4 mutant accumulates zygotene-like nuclei (Figure 4F). Consistent with a synapsis defect, most zygotene-like csm4 nuclei contain an aggregate of Zip1 called polycomplex. Although pachytene-like nuclei are rare in the mutant, Zip1 dismantles when further incubated with SPM (Figure 4F). These results indicate that CSM4 is required for efficient SC formation, particularly SC elongation. Similar results have been described by Wanat et al. using Zip1–Green fluorescent protein (GFP) fusion protein [31]. Expression of CSM4 mRNA is specific to meiosis [30]. Western blotting analysis using an antibody against Csm4 reveals that this protein is present in lysates from meiotic cells, but not from mitotic cells (Figure 5A). Our initial immunostaining analysis of both whole cells and chromosome spreads failed to localize the protein either in nuclei or on chromosomes (HK, unpublished results). However, when expressed in vegetative cells as a GFP fusion protein, Csm4 localizes to nuclear membranes and the endoplasmic reticulum [48]. We noticed that the csm4 and ndj1 mutants share similar recombination defects [28]. In particular, similar to csm4, the ndj1 mutant specifically decreases NCO formation in physical assays. When a csm4 ndj1 double mutant was constructed and analyzed for CO/NCO formation, the double mutant exhibited a phenotype similar to csm4 and ndj1 single mutants (Figure 2C). Although CSM4 and NDJ1 appear to function in the same recombination pathway, there are several phenotypic differences between two single mutants. In general, csm4 shows more severe defects than ndj1 and csm4 ndj1 double mutants show defects that are more similar to csm4. The spore viability of csm4 is lower than that of ndj1 (Figure 1A; 66% versus 77%), and the csm4 mutant enters into MI 2 h later than the ndj1 mutant (Figure 2C). Ndj1 is a meiosis-specific protein that binds to telomeres [25], [26] and is required to form the bouquet, where telomeres cluster near the SPB [27]. The similarity between csm4 and ndj1 phenotypes prompted us to examine the interaction of Csm4 with Ndj1. We used a strain in which Ndj1 protein is tagged with the HA epitope at its C-terminus. This strain exhibits wild-type spore viability. Immunoprecipitation (IP) using anti-HA antibody reveals the presence of Csm4 in precipitates of meiotic cell lysates from NDJ1-HA diploid, but not in those from the untagged strain (Figure 5B). Reciprocal IP using anti-Csm4 also detects Ndj1-HA in these precipitates (Figure 5B). These results demonstrate a physical association interaction of Csm4 with Ndj1 in meiotic cells. Since the csm4 mutant expresses Ndj1 (Figure 5A), the defect conferred by csm4 is not due to the inability of csm4 cells to express Ndj1. Next, we studied the localization of Ndj1-HA protein to the NE in csm4 mutants. Whole cells were fixed with formaldehyde and then stained with anti-HA antibody followed by fluorescent-conjugated antibody. The cells were then observed under an epifluorescence microscope. We also analyzed the localization of Dmc1 in intact cells as a marker for meiotic cells. As reported previously [27], in wild-type cells, Ndj1 shows several foci or patches near the nuclear periphery in the meiotic prophase (Figure 5C). The kinetics of accumulation and disappearance of Ndj1- and Dmc1-positive cells were very similar (Figure 5E). We sorted the staining patterns into three classes: rim, loose bouquet, and tight bouquet (Figure 5F). Loose and tight bouquets are only seen in the meiotic prophase (Figure 5F). Furthermore, a significant fraction of wild-type cells at 4 h shows clustering of Ndj1 foci (loose and tight bouquets) in one area of the NEs (Figure 5F). On the other hand, csm4 cells do not show clustering of Ndj1, but rather exhibit dispersed staining of Ndj1 patches at the nuclear periphery (Figure 5D and 5F). In csm4, Ndj1 patches persist in the periphery longer than in the wild type (Figure 5D and 5E). Importantly, Ndj1 in the csm4 mutant is still associated with the NE. These data indicate that Csm4 is required for efficient clustering of Ndj1 on the NE, but not for tethering, suggesting a role of Csm4 in Ndj1-mediated telomere clustering. Ndj1 promotes telomere clustering during meiotic prophase [27]. We examined bouquet formation by analyzing Rap1–GFP localization [49], [50]. Rap1 is concentrated at telomeres and is used as a marker for telomere localization [51]. As shown previously, Rap1–GFP is localized at the nuclear periphery as several foci in mitosis [49], [52]. Nuclei were visualized by deconvoluting Z-series images; one focal plane is shown in Figure 6. When diploid cells enter the meiotic prophase, after 3–5 h incubation with SPM, a small fraction of diploid cells in meiotic prophase show a polarized distribution of Rap1–GFP at the cell periphery. In S. cerevisiae, the bouquet appears unstable and is possibly dynamic during the meiotic prophase [50], [52]. In the wild type, clustering of Rap1 foci is prominently seen at 4 h (Figure 6A); however, due to the very transient nature of the clustering, only 15–25% cells show the clustering. On the other hand, the csm4 mutant shows a disperse distribution of Rap1 on NEs after 4 and 5 h incubation with SPM (Figure 6B). Very few cells show the clustering of Rap1–GFP in the mutant between 4 and 6 h (see Figure 5F). This indicates that Csm4 is necessary for clustering of telomeres but not for tethering telomeres to the NE. Furthermore, some Rap1–GFP foci in the ndj1 mutant are not localized at the nuclear periphery but are seen within the nucleus [Figure 6C; 27]. Similar results have been described by Wanat et al. [31]. We also noticed that most csm4 nuclei were round, while the wild type as well as the ndj1 mutant nuclei were irregularly shaped, suggesting a defect in nuclear deformation in the csm4 mutant. It was recently reported that a component of the SPB, Mps3, is necessary for telomere clustering during meiosis [53] and anchoring telomeres [54]. Mps3, which contains Sad1-Unc-84 (SUN) and trans-membrane domains, changes its localization from the SPB to the NE during meiosis [53], [54]. We examined the effect of csm4 mutation on Mps3 relocalization. Whole cells containing MPS3 tagged with HA were fixed with formaldehyde and stained with antibodies against the HA tag and Dmc1 protein. As reported previously [54], [55], at 0 h, Mps3 is seen as a single spot at the nuclear periphery (Figure 7A), which is consistent with its localization near the SPB. During the meiotic prophase (at 4 h in SPM), in Dmc1-positive nuclei, Mps3 relocalizes throughout the NE and occasionally exhibits patchy staining (Figure 7A). This NE localization of Mps3 is still observed after the MI division. In the csm4 mutant, Mps3 shows a distribution in the NE similar to the wild type, but remains longer than in the wild type (Figure 7A and 7B). This is consistent with a prolonged meiotic prophase in csm4. Therefore, the effect of the csm4 mutation on telomere clustering appears to be independent of Mps3 relocalization. In addition, csm4 does not affect Mps3 protein levels (Figure 7C). Previously, the csm4 mutant was isolated on the basis of its defect in chromosome segregation during meiosis [30]. In this paper, we show that Csm4 functions with the meiosis-specific telomere-binding protein Ndj1. Recombination defects conferred by a csm4 mutation are very similar to those caused by a mutation in NDJ1 [28]. Indeed, the csm4 ndj1 double mutant phenotype is similar to that seen for the single mutants. In addition, co-IP shows that Csm4 is physically associated with Ndj1 in vivo. Furthermore, Csm4 is required for efficient clustering of Ndj1 at the nuclear periphery. These results indicate that Csm4 and Ndj1 function in the same structural pathway. The csm4 mutant, however, shows more severe defects in meiosis than the ndj1 mutant. Spore viability is lower in csm4 compared to ndj1. The csm4 mutation delays its entry into MI to a greater extent than ndj1. These observations suggest that Csm4 has additional functions in meiosis or that ndj1 is not null for related functions. Csm4 is necessary for normal functioning of all three recombination pathways of meiosis: ZMM-dependent and -independent (MMS4-dependent) CO and NCO formation. Although the final level of COs in the csm4 single mutant is similar to that in the wild type, csm4 reduces the level of NCOs compared to the wild type, indicating the involvement of Csm4 in NCO formation during meiosis. Csm4 is a meiosis-specific protein; this suggests that NCO formation is under the control of a meiotic program and thus is likely to be mechanistically distinct from NCO formation during mitosis. When csm4 mutation is combined with a mutation in MSH4, the double mutant is almost completely deficient in CO formation. Therefore, Csm4 functions in CO formation independently of Msh4. CO formation in meiosis, about half of which depends on ZMM genes, is severely delayed in csm4, indicating that Csm4 is also required for efficient formation of COs in the major ZMM-dependent meiotic recombination pathway. Two intermediates, SEIs and dHJs, have been identified in the ZMM-dependent CO pathway [8], 9. The most severe effect of csm4 mutation is seen in SEI–dHJ transition and dHJ resolution, two distinct biochemical steps in the ZMM pathway. It is likely that SEI–dHJ transition is accompanied by the capture of SEI by the second end of the DSB [8], [56]. Therefore, Csm4 seems to promote the second-end capture during strand exchange. Generally, this capture is considered a simple annealing reaction between ssDNA of the second end and a displaced ssDNA in SEI [57]. However, our results strongly suggest that the second-end capture is not a simple biochemical reaction as believed previously; rather, it is a critical regulatory step in the CO pathway. Although the exact molecular nature of SEIs in the csm4 mutant is not known, they are likely to contain D-loop structures that can be converted into COs or NCOs (Hunter et al. 2002). Thus, the transition of SEIs to dHJs could be regarded as an irreversible commitment step towards CO formation. The transition can be independently governed by both ZMM- and Csm4-dependent functions. Furthermore, disassembly of the two RecA homologs Rad51 and Dmc1 is delayed in the csm4 mutant. This strongly suggests that disassembly of the RecA homologs occurs during the SEI–dHJ transition, and thus is somehow coupled with the second-end capture. How does Csm4 control the various steps during meiotic recombination? One notion is that Csm4 functions as an enzyme directly involved in recombination. However, it is very difficult to assign a biochemical activity to Csm4 (∼23 kD) with no apparent structural domains, since it is likely to be involved in various steps in the aforementioned three recombination pathways. One possibility is that Csm4 acts as a component of the meiotic chromosomes. Red1, a chromosome axis protein, is involved in various recombination steps [43], [58]. However, our initial attempt to localize the protein on DNA by chromatin IP failed to detect the binding of Csm4 to a recombination hotspot (unpublished results). Our initial attempt to localize Csm4 was also unsuccessful because both N- and C-terminal tagged genes are non-functional and our anti-Csm4 did not work for immunostaining (HK, unpublished results). However, Csm4 is predominantly enriched in the NE when overexpressed as a GFP fusion protein in vegetative cells [48], consistent with the fact that Csm4 contains a putative transmembrane domain. Furthermore, a Csm4 partner, Ndj1, is enriched at telomeres, that are tethered to the NE [25], [26]. These observations strongly suggest that Csm4 is localized in the telomeres. Indeed, similar to Ndj1, Csm4 binds to telomeres on nuclear spreads [59]. Thus, the Csm4–Ndj1 complex is likely to affect recombination indirectly through its function at telomeres and/or the NEs. In addition to Csm4–Ndj1, the Mps3 protein containing Sad1-UNC84 domain is also involved in the process [53]. During vegetative growth, Mps3 is localized to the SPB and then relocated to the NEs in the meiotic prophase [53]. Mps3 forms a complex with Ndj1 and Csm4 [59]. An allele of mps3 shows pairing defects in meiosis similar to those seen in ndj1 and csm4 mutants [59]. Given that Mps3 is an inner nuclear membrane protein, it is likely to tether Ndj1-bound telomeres to the NE. How do telomeres control recombination on the interstitial sites of chromosomes? The fact that ndj1 and csm4 mutants are defective in chromosomal bouquet formation [this study; 59], [60] suggests that a polarized configuration of chromosomes in zygotene might play a positive role in meiotic recombination. As proposed previously [23], [61], telomere clustering may restrict the arrangement of chromosomes in the nucleus, and in turn increases the probability that two allelic loci undergo colocalization. Although this could explain defects specific to zygotene, such as first end capture or SEI formation, those in second-end capture and dHJ resolution, occurring in the end of zygotene and pachytene, respectively, cannot be simply explained by telomere clustering during zygotene. Rather, we propose that chromosome dynamics accompanied by telomere movement facilitates meiotic recombination. Tethering telomeres to nuclear membranes followed by movement along the envelope might change the chromatin structure, which might indirectly promote various biochemical steps during recombination. Dynamic movement of chromosomes in the meiotic prophase has been recently described; it depends on actin polymerization [59], [60], [62], [63]. Furthermore, the dynamic nature of telomeres on the NE is somehow dependent on Ndj1 and Csm4 [59], [60]. It is likely that the global changes in the chromosome structure and/or movement of chromosomes, promoted by the anchoring of telomeres to the NE, control the biochemistry of recombination of meiotic chromosomes. Our analysis of csm4 provides new insights into the mechanism of telomere clustering in budding yeast. Both csm4 and ndj1 mutants are deficient in telomere clustering, but the nature of deficiency in these mutants is qualitatively different. While NDJ1 promotes tethering of telomeres to the NE, CSM4 facilitates clustering of Ndj1-bound telomeres in one area of the envelope. Csm4 may promote bouquet formation by directly clustering the telomeres and/or by stabilizing them. Given that telomere movement on the envelope is a dynamic process [52], [62], Csm4 might be involved in the movement of telomeres on the NE. However, the csm4 mutant exhibits some local movement of telomeres on the membrane, which is clearly different from the movement in the presence of an actin-inhibitor [60; HK and AS, unpublished results]. Thus, the movements of telomeres are either Csm4-dependent or Csm4-independent. Our results suggest that meiotic telomere clustering consists of different steps including telomere tethering, movement, and clustering. Consistent with this, nuclei in csm4 mutants are relatively round compared to the irregular shapes of meiotic nuclei seen in the wild type (Figure 6). Nuclear deformation may be induced by external physical forces on the nuclei. Therefore, Csm4 might be involved in the transduction of forces on the NEs. All strains described here are derivatives of SK1 diploids, NKY1551 (MATα/MATa, lys2/lys2, ura3/ura3, leu2: : hisG/leu2: : hisG, his4X-LEU2-URA3/his4B-LEU2, arg4-nsp/arg4-bgl) and NKY3230 (MATα/MATa, lys2/lys2, ura3/ura3, leu2: : hisG/leu2: : hisG, his4X-LEU2- (N/Bam) -URA3/HIS4-LEU2- (N/Bam) and its derivatives with csm4: : KamMX6 were used for the 2D analysis. Rap1-GFP was a kind gift from Dr. Y. Hiraoka. The genotypes of each strain used in this study are described in Table S1. csm4, ndj1, mms4, and msh4 null alleles were constructed by PCR-mediated gene disruption using either the URA3 gene or the KanMX6 [64]. NDJ1-3HA and MPS3-3HA were constructed by a PCR-based tagging methodology [65]. Primer details used for PCR amplification are available upon request. Anti-Csm4 antibody was raised against recombinant protein purified from E. coli. The open reading frame of Csm4 was PCR-amplified and inserted into pET15b plasmid (Novagen) in which the N-terminus of CSM4 gene was tagged with hexahistidine. Csm4 protein with the histidine tag was affinity-purified in accordance with the manufacturer' s protocol and used for immunization (MBL Co. Ltd). Primer details for PCR amplification are available upon request. Anti-HA antibody (16B12; Babco), anti-tubulin, guinea pig anti-Rad51 [45], and rabbit anti-Dmc1 [5] were used for staining. Antiserum against Zip1 was raised using a recombinant GST-fusion protein purified from E. coli [16]. Immunostaining of chromosome spreads was performed as described previously [45], [66]. Whole cell immuno-staining was preformed as described previously [27] with a slight modification. Cells were fixed with formaldehyde. Stained samples were observed using an epi-fluorescent microscope (BX51; Olympus) with a 100x objective (NA 1. 3). Images were captured by a CCD camera (Cool Snap; Roper), and processed using IP lab (Sillicon) and Photoshop (Adobe) software. For focus counting, more than 100 nuclei were counted at each time point. Rap1-GFP was observed as described previously [52]. Images were captured by a computer-assisted fluorescence microscope system (Delta Vision; Applied Precision) with an oil-immersion objective lens (100x, NA 1. 35). Image deconvolution was performed using an image workstation (SoftWorks; Applied Precision). Time-course analyses of DNA events in meiosis and cell cycle progression were performed as described previously [8], [12], [58]. IP assay was performed as described previously [5]. Each result presented in the figures is representative of several experiments. The number of experiments performed is shown in Table S2.

Meiosis is a specialized cell division that produces haploid gametes. Homologous recombination plays a pivotal role in the segregation of homologous chromosomes during meiosis I by creating physical linkages between the chromosomes. Drastic reorganization of chromosomes in the nucleus is a prominent feature of meiotic prophase I, during which telomeres get associated with the nuclear envelope and move within the envelope, culminating in the formation of telomere clusters, often referred to as “chromosome bouquets. ” The roles that telomere movement and clustering play in meiotic prophase I are largely unknown. In the budding yeast Saccharomyces cerevisiae, tethering of telomeres to the nuclear envelope is mediated by a meiosis-specific telomere-binding protein, Ndj1. We studied the functions of a meiosis-specific gene, CSM4, in telomere clustering and during meiotic recombination. CSM4 is necessary for the clustering of Ndj1-associated telomeres. Interestingly, csm4 mutants show pleiotropic defects during meiotic recombination. It is likely that the chromosome movement promotes various biochemical reactions during meiotic recombination.

Abstract Introduction Results Discussion Methods

genetics and genomics/chromosome biology

2008

Csm4-Dependent Telomere Movement on Nuclear Envelope Promotes Meiotic Recombination

Human mobility, presence and passive transportation of Aedes aegypti mosquito, and environmental characteristics are a group of factors which contribute to the success of dengue spread and establishment. To understand this process, we assess data from dengue national and municipal basins regarding population and demographics, transportation network, human mobility, and Ae. aegypti monitoring for the Brazilian state of Acre since the first recorded dengue case in the year 2000 to the year 2015. During this period, several changes in Acre’s transport infrastructure and urbanization have been started. To reconstruct the process of dengue introduction in Acre, we propose an analytic framework based on concepts used in malaria literature, namely vulnerability and receptivity, to inform risk assessments in dengue-free regions as well as network theory concepts for disease invasion and propagation. We calculate the probability of dengue importation to Acre from other Brazilian states, the evolution of dengue spread between Acrean municipalities and dengue establishment in the state. Our findings suggest that the landscape changes associated with human mobility have created favorable conditions for the establishment of dengue virus transmission in Acre. The revitalization of its major roads, as well as the increased accessibility by air to and within the state, have increased dengue vulnerability. Unplanned urbanization and population growth, as observed in Acre during the period of study, contribute to ideal conditions for Ae. aegypti mosquito establishment, increase the difficulty in mosquito control and consequently its local receptivity. The successful invasion of a parasite into a new place is the result of the interaction between two players: the invading species and the invaded community. In the context of dengue, the invaded community is composed of two main species: human hosts and mosquitoes of the species Ae. aegypti whose interaction is further affected by environmental factors (climate, urbanization, human behavior). The invading species under consideration, dengue viruses, cannot establish itself in new areas unless the receiving community presents certain conditions to sustain autochthonous virus transmission. More formally, the emergence of dengue fever in a new region depends on the occurrence of two major events: (i) the arrival of at least one infected individual in the city and (ii) the subsequent positive growth of autochthonous dengue cases (i. e. , disease establishment) [26]. The sustained growth of autochthonous cases depends on “the abundant presence of vectors and the existence of other ecological and climatic factors favoring transmission”, which is the concept of receptivity found in the malaria literature [27]. If an area is receptive, then the arrival of the virus by imported cases or vectors has a probability of triggering sustained local transmission. Transmission is then measured by the effective reproductive number, Rt, defined as the expected number of secondary human cases that would result from the arrival of a single infected individual in a susceptible population. A reproductive number greater than one suggests epidemic growth. The larger the influx of imported cases the higher the probability of establishment given the necessary environmental conditions. In the malaria literature, this exposure to imported cases is referred to as “Vulnerability” defined as “either proximity to endemic areas or…the frequent influx of infected individuals or groups and/or infective anopheles” [28]. The same concept can be applied to dengue epidemiology. Dengue influx might be measured directly from the number of cases that can be traced to outside areas (imported cases). In situations where the origin of the cases is difficult to ascertain, one can estimate the number of imported cases from statistics of transportation and prevalence in passengers’ origins [26]. The state of Acre (AC) is located in the north region of Brazil, bordering the countries of Peru and Bolivia, and the Amazonas and Rondônia states in Brazil (see Fig 1 for geographical reference). The total area is 164,122 km2 with a population of 733,559 inhabitants and a population density of 4. 47 hab/km2 [25]. The state has 22 municipalities, the most populous being the capital Rio Branco and Cruzeiro do Sul, with 336,038 and 78,507 inhabitants according to 2010 Census [25], respectively. The Acrean climate is warm and humid equatorial, which is characterized by high temperatures with small annual fluctuations, high pluviometric precipitation indices and high relative humidity. The temperature is approximately uniform throughout the state, with an annual average around 24,5°C, and maximum around 32°C [24]. Within our conceptual framework, vulnerability was defined in terms of the exposure of the focal population to imported dengue cases. Here, we measured vulnerability as the expected probability of dengue importation from other Brazilian states into Acre, by air transportation. The usage of national airline passengers database is of great interest for the study of human interaction dynamics within a country, particularly for the study of potential disease transmission routes. The flow of individuals between different regions allows agents that became infected in one area to carry the pathogen to another area within his/her path [8,11,17,18,26,31]. This mechanism serves as a driver for the (re-) introduction of pathogens throughout the national territory. In this context, we analyzed Brazilian airline grid to estimate passenger flow between Brazilian states and how this potential risk for the spreading of dengue to Acre varied since the year 2000. From the Brazilian airline database collected from 2001 to 2012, we estimated the daily probability of an individual from state i traveling to state j for each month m, πij, m (see S1 Appendix). Coupling this probability with the reported number of dengue cases in each state in the corresponding time window, we estimated the risk of each one state being the source of an imported dengue case to any other in the Brazilian territory. This framework allows us to estimate the most probable routes of case importation in each year to Acre. Say we have ki, m dengue cases reported in state i in month m. Each k of the ki, m infected individuals stays τk days as infectious, with an average of 4. 5 [1. 9–7. 9] days [32]. Given the daily probability of travel πij, m, τk is also the number of monthly trials for the probability of an infected individual boarding a flight during his infectious period. Combining this information, the monthly case importation probability pij, m can be estimated as p i j, m = 1 - (1 - π i j, m) ∑ k = 1 k i, m τ k. (1) Following the same rationale, the yearly probability of dengue importation, Pij, s, at dengue epidemiological year s, can be estimated by aggregating over all months of the corresponding epidemiological year, that is P i j, s = 1 - ∏ m (1 - p i j, m). (2) This construction allows us to provide a ranking of states by their probability of exporting a dengue-infected individual to Acre, that is, the most probable sources of case importation at any given epidemiological year. It also allows us to estimate the source-independent yearly dengue importation probability, that is, the probability of at least one importation to Acre in a given year regardless of its state of origin, which can be written as: P j, s = 1 - ∏ i (1 - P i j, s) (3) This quantity describes the temporal dynamics of the Acre’s overall exposure to dengue-infected individuals. To take into account the uncertainty regarding the length of the infectious period, the results shown correspond to the average over 1000 simulations of this process using a gamma distribution with mean of 4. 5 and 95% confidence interval of [1. 9–7. 9] for the infectious period of each reported case. The earliest evidence of the onset of disease establishment is the occurrence of autochthonous cases, which are ascertained based on investigations of the travel history of the patients. Secondly, the presence of clusters or transmission chains is indicative of further transmission. Thirdly, if the transmission is sustained, the incidence curve will increase at an exponential (or subexponential) rate, characterizing an epidemic [33]. The available dengue notification data do not have information on travel history or clustering of the reported cases. Here, evidence of transmission is indirectly obtained from the notification data via the calculation of the effective reproductive number, Rt. This index is interpreted as the average number of secondary cases generated by a primary case at time t [34], calculated as the ratio of secondary to primary cases. A value of Rt > 1 indicates a sustained growth of incidence. Since no information is available on who infected whom, the primary cases are inferred from the generation interval of dengue, that is, primary cases are those that could be the source of transmission for the current cases given the time between the onset of their disease symptoms. This approach formally described by Wallinga & Lipsitch [34] results in the following expression for Rt: R t = b (t) ∑ a = 0 ∞ b (t - a) g (a), (4) where b (t) is the number of new cases reported at week t and g (a) is the probability distribution of the dengue generation interval (time between onset of symptoms in a primary case and onset of symptoms in a secondary case) taken as a Normal distribution N (μ = 3weeks, σ = 1week). Confidence intervals for the ratio of two Poisson counts were calculated using the method described in [35] and Rt > 1 was ascertained if p (Rt > 1) > 0. 95. The effective reproductive number Rt was calculated weekly for each city, from 2000 to 2015. For each city, an epidemiological year was classified as epidemic if Rt > 1 for at least three consecutive weeks. A period of 3 weeks was chosen because it represents one generation of dengue transmission. We further defined T3 as the time in weeks from July 2000 to the first observation of three consecutive weeks with Rt > 1 in that city. In 2001, Acre witnessed its first dengue epidemic, in Rio Branco. Since then, dengue cases have been recorded almost continuously and with great intensity in this municipality. Here we investigate if the spread of dengue from Rio Branco to other municipalities in Acre was associated with the transportation network. First, we investigate if more central cities were invaded earlier than peripheral ones, where centrality is defined as a property of the transportation network (defined below). Secondly, we assessed whether cities closer to Rio Branco were invaded before distant ones. A structural transportation network was constructed as a graph with 22 nodes representing the municipalities and 29 edges representing the connections between them by (paved, unpaved or under maintenance) roads, waterways or airways, as described in the data subsection. In this network, the edges were considered as unweighted for most of the centrality metrics defined below. The only exception made was for the distance in kilometers from Rio Branco, in which case each edge received a weight equal to the corresponding distance between the connected nodes. We also constructed a mobility network with the same 22 nodes but with a total of 87 edges. As described in the mobility data subsection, in this network the edges are given weights equal to the sum of residents moving from one municipality to another for work and/or study in both directions. There are several measures of node centrality, the literature being rich with proposals for both weighted and unweighted networks [36]. Each of those proposals focuses on a different aspect of information flow and node properties, therefore producing potentially different node ranks for each measure [37,38]. Here, we considered the following centrality measures that provide different interpretations in relation to dengue exposure: To access how vulnerability to dengue importation changed in the study time at the Acrean municipalities in response to changes in the transportation network, we measured, using Spearman correlations, the intensity of the relationship between T3 to the network descriptors and compare to the structural state of the transportation network during the period of study. While the descriptors for the mobility network are static since we only have the flow of individuals for the year of 2010, the analysis of the transportation network status is dynamic since there were considerable interventions such as construction and paving of roadways during the period of study. We hypothesize that T3 and the descriptors should be correlated since the road changes in the Acrean network is very evident, and the invasion of dengue in the state occurred first in the nearest municipalities in terms of distance and with better road structure. The analysis were performed in software R 3. 3. 2 (R Core Team, 2016, cran. r-project. org/) and in Python 3. 5 (Python Software Foundation, www. python. org) using the libraries Networkx [45], Pandas [46], Numpy [47], and igraph [48]. Receptivity depends on the presence, abundance and vectorial capacity of the local population of Ae. aegypti. Moreover, dengue receptivity has been linked to unplanned urbanization, fast population growth, poor infrastructure as lack of urban services and effective mosquito control, and globalization [49]. Here, we review the available information on these topics in order to provide the best description possible for the evolution of dengue receptivity in the study period. First, we investigated the available records of Ae. aegypti presence in the region using LI/LIRAa. From this data, we computed the year when Ae. aegypti was first recorded in each municipality. Secondly, we computed how much each municipality changed from 2000 to 2010, in terms of population growth, urbanization, garbage collection, water supply and sanitary sewage. We also reviewed the scientific literature searching any biological information regarding the vector populations found in Acre. The literature is scarce, and a single study was found on the competence of Acrean Ae. aegypti to DENV-2 [50]. From 2001 to 2012, the overall flow of passengers to Acre increased from about 50,000 to more than 150,000 passengers per year. During this period, the main states of origin for travelers to Acre were Distrito Federal (DF), Rondônia (RO), Amazonas (AM) and São Paulo (SP) (Fig 1). Meanwhile, the states of Rio de Janeiro (RJ), São Paulo, Minas Gerais (MG) and Bahia (BA) had the highest dengue cases recorded. As discussed in the methods section, the probability of case importation to Acre is a combination of both the number of travelers and dengue activity. Fig 2 shows the probability of dengue case importation into Acre from 2001/2002 to 2011/2012. It reached ca. 70% in 2001/2002, reducing to its lowest level (38%) in 2003/2004, steadily increasing to ca. 100% from 2004/2005 to 2007/2008, remaining so until the end of the study period. The north region was found as the main source of dengue cases to Acre, in particular, the neighboring state of Rondônia (RO), followed by the center-west (Distrito Federal-DF and Mato Grosso-MT), southeast (Rio de Janeiro-RJ and São Paulo-SP), and northeast (Ceará-CE). The contribution of Rondônia, Distrito Federal and Mato Grosso to the exportation of dengue to Acre was mainly associated with the flow intensity between these states while that of Rio de Janeiro and São Paulo was a result of the combination of moderate flow intensity and high rates of dengue activity. Others states such as Tocantins (TO), Roraima (RR) and Amapá (AP) in the north, Maranhão (MA), Piauí (PI), Paraíba (PB) and Alagoas (AL) in the east and all states in the south region were unlikely exporters of dengue cases to Acre due to the combination of low flow and low dengue activity. The first autochthonous dengue case was reported in Acre in 2000. From this year until 2008, the annual incidence did not exceed 900 cases per 100,000 inhabitants, a level of activity similar to those found in the other northern states of Brazil, except for Roraima. However, in 2009 there was a significant increase in dengue incidence in Acre when it tripled to an alarming level of 2,800 cases per 100,000. In 2010, dengue incidence was even higher: 4,793. 3 cases per 100,000 [51]. According to the Ministry of Health, in 2011, Acre was among the states classified as at moderate risk, and within the state, the capital Rio Branco was classified as at the highest risk [52]. On the other hand, the Alto Juruá region, in the northwest of Acre, which until 2014 had not yet registered autochthonous cases of dengue, witnessed its first dengue epidemic in Cruzeiro do Sul. In 2014, Cruzeiro do Sul was among the municipalities with the highest number of dengue cases between the epidemiological weeks 01 and 47 of 2014. Compared to 2013, the number of cases jumped from 30 cases to 23,130 [53,54]. In Fig 3 we show the weeks with Rt > 1, from 2000 to 2015 at each Acrean municipality. A sequence of weeks with Rt > 1 is an indication of epidemic growth. All municipalities presented this event at least once during the study period. The municipalities that presented the lowest percentages of weeks with Rt > 1 were Porto Walter (PW) and Jordão (Jrd) (1. 15% for both) throughout the study period. These are the two most peripheral in the state, and whose access includes a stretch traveled by river. The ones with the highest percentages of “epidemic weeks” were Rio Branco (RB, 9. 32%), Senador Guiomard (SG, 6. 39%), Epitaciolândia (Epcd), and Brasiléia (Brl) (6. 00% for both). In Fig 3, colors indicate the main type of access to each city. There are two main types of access, those cities in which access is by road (black and orange) and those in which the final access is by waterway (blue). Before 2008/2009, only cities in which access is by road registered Rt > 1. However, these records were mostly intermittent, as can be seen for example in Bujari (Bjr), Capixaba (Cpb), Assis Brasil (AB), Porto Acre (PA), Acrelândia (Acld), Plácido de Castro (PC) and Xapuri (Xpr). Cruzeiro do Sul (CZS) in gray, which is the only city served by an airline to Rio Branco, also registered few weeks with Rt > 1 before 2008. After this year, when several paving works and road maintenance were concluded, the frequency increased eventually becoming more intense culminating with CZS recording the first sustained epidemic in 2014. The time frame of infrastructural changes and its potential impact on dengue cases notification is discussed in more detail in the section Dengue spread in Acre. The municipalities in orange, Mâncio Lima (ML), Rodrigues Alves (RA), Tarauacá (Trc) and Manoel Urbano (MU), also have access by road, but part of the final access was under construction or maintenance during all or almost all the period of study. In these municipalities, as well as in those with final access by waterway, the first records of Rt > 1 occurred in the beginning of 2008, in a slow way, while in those cities with access by waterway, the records were even rarer than those accessible by roads. Public investments since 2000 in the development of the Acrean transportation infrastructure, including the construction of the Pacific Highway and the pavement of intermunicipal roads, have brought more connectivity between the municipalities, in particular within southeast Acre, composed by Rio Branco, Senador Guiomard, Capixaba, Xapuri, Epitaciolândia, Plácido de Castro, Brasiléia, and Acrelândia. In contrast, transportation and mobility between the southeast and other regions of Acre (Alto Juruá) is still very difficult due to road conditions, although improvements have occurred during this period. The accessibility to the northwest area of Acre began to improve after the year 2008, when paving of the highway BR-364 in the region of Manuel Urbano (MU) and Feijó (Fj), in the central area of Acre. Highway BR-364 is the only one that crosses the state from Acrelândia (Acl, in the southeast) to Mâncio Lima (ML, in the northwest) and connects Acre to the rest of Brazil via Rondônia state. Highway BR-364 has been under construction since the beginning of the last decade. In southeast Acre the paving and restoration of this highway were completed quickly, and only the stretch from Acrelândia to Rio Branco was still under construction until 2007/2008. The stretch from the capital to Cruzeiro do Sul, at the northwest part of the state, passes through very unstable soil, making it difficult to maintain during the monsoon season. Only in 2012/2013, this stretch was completely paved, but in the following years repairment works were needed. Besides this highway, an extremely important link between the southeast and northwest Acre is the daily flight between Rio Branco and Cruzeiro do Sul, which increased by 18% from 2001 to 2012. The most remote municipalities are Santa Rosa do Purus, Jordão, Marechal Thaumaturgo and Porto Walter with very limited access and only by river, most of them only by small and medium-sized vessels depending on the time of year. On the other hand, air travel alone was responsible for an average of 134 passengers to Acre per day during the epidemiological year 2001/2002, steadily increasing to almost 500 per day in 2013/2014, most of them to Rio Branco, responsible for approximately 90% of all air travel to Acre from other Brazilian states. To illustrate the impact of interstate airflow in this population, the 2010 Census registered 3. 6 thousand commuters to Rio Branco, so that the average daily out-of-state air passengers corresponded to almost 4% of daily commuters in 2001/2002, up to almost 14% in that of 2013/2014. This data contrasts with the common misconception regarding air travel to Acre and its impact to information inflow, which is believed to be negligible due to its relatively low volume with respect to the rest of the Country. The connectivity between municipalities is shown in Fig 4. We used Cytoscape software [55] to build a representation of the Acrean municipalities’ mobility network, with edge width and color proportional to the natural logarithm of the daily number of travelers and node size proportional to the natural logarithm of each municipality’s population (node color indicate access mode, as in Fig 3). Nodes position are defined based on the Fruchterman–Reingold algorithm [56], which is a force-directed layout widely used in network visualization. In this layout, all nodes have a repulsion force between them, while edges weight are used as an attractive force between connected nodes. Nodes that share stronger connections, i. e. , with more travelers, are placed closer to each other than those that share weaker or no connection. Therefore, groups of nodes sharing stronger bonds among them form regional clusters in this abstract space of connectivity. This particular visualization facilitates the detection of municipalities that share relatively strong bonds even when geographically far apart. On the other hand, nodes that are placed on the periphery with this algorithm are those that have relatively few and/or weak connections. We can see that the two most populous municipalities, the capital Rio Branco (RB) and Cruzeiro do Sul (CZS), are important hubs in this network. The access mode (waterway, paved and unpaved roads, airway) confers a natural clustering on the mobility network, suggesting an intuitive relationship between ease of access (transportation network) and the number of travelers (mobility network). Municipalities accessed by paved roads are more densely connected with each other, while the strong connection between Rio Branco and Cruzeiro do Sul, which are connected by airports, considerably shortens the effective distance between east and west side of the state. Another example of this relationship is the fact that municipalities which are mainly accessed via waterways are peripheral in the mobility network. In fact, Marechal Thaumaturgo (MT), Jordão (Jrd) and Santa Rosa dos Purus (SRP) have the lowest Ew, Cw and Sw centralities, with Porto Walter (PW) also among the five municipalities with lowest values for those centralities. The capital Rio Branco is the most central node in the Acrean network with respect to all descriptors used (see in S2 Appendix) being the most present node in all shortest paths (as measured by B and Bw); the municipality with the smallest average shortest path to all other nodes (C and Cw); the strongest attractor (Ew) in the network; and the node with most travelers (Sw). Therefore, it is both the most vulnerable to exposure from a pathogen present in the state at a random municipality, as well as the node which poses the highest impact in other municipalities’ vulnerability once infected. Porto Acre, Bujari and Senador Guiomard, all with geographical borders with Rio Branco, also present high centrality measures, therefore sharing similar relevance to disease transmission. All of those cities are located on the east side of the state, and all of them can be accessed via paved federal highway. Another municipality with high centrality for most indicators is Cruzeiro do Sul (CZS), on the west side of Acre. Although geographically distant from Rio Branco, this municipality shares an airport connection with the capital, drawing a significant amount of travelers and therefore shortening its network distance with respect to individual flow. Being the second most populous municipality, it also shares strong connectivity with its geographical neighbors. In fact, it has the third highest strength centrality, that is, the third municipality with highest number of travelers. Due to its connectivity to Rio Branco and its role as western hub, Cruzeiro do Sul acts as an important bridge between the capital and the western region. In Fig 5 we can observe a sequence of panels showing how dengue invaded and established itself in Acre, following the transportation network. The nodes represent the cities, colored according to the number of consecutive weeks with Rt > 1 per year. In this study, we consider “at least three consecutive weeks with Rt > 1” as a marker of (at least temporary) disease establishment. Of the 22 Acre counties, nine were positive for this marker at least once during the study period despite the fact that all of them registered at least one case of dengue during this time. Until 2008/2009, only Rio Branco e Sena Madureira had evidence of disease establishment according to this criterion. The two cities are well connected in terms of human mobility, close regarding geographical distance, and have paved roads connecting them. At the end of the study period, other seven cities presented at least three consecutive weeks with Rt > 1. Of those, six are in southeast Acre, in the most connected and best-accessed region. The other city is Cruzeiro do Sul, one of the most distant cities in kilometers from Rio Branco, but populous and with an air connection to the capital. We further investigate if there is evidence of correlation between the centrality of a municipality and the time it took for dengue to establish itself, measured by T3, as defined previously. A summary of the main results is shown in Table 1. This is an exploratory analysis only, since the sample size is small for modeling. Strength centrality was the network descriptor most negatively correlated with T3. This result suggests that municipalities in which the sharing of individuals with other municipalities was larger had a greater exposure to dengue and, consequently, smaller T3 than those with lower integration. The Eigenvector centrality of the municipalities, which is related to how strong is the connectivity of a node’s neighborhood, therefore its ability to concentrate population flow in the network, also presents negative correlation with T3. This is an intuitive result since this property enhances the probability of a node being invaded once a node in its vicinity presents an outbreak, even when not in direct contact. Finally, the distance from Rio Branco, measured by both the effective distance and in kilometers, also showed high correlation with T3. It is interesting to note that, geographical distance aside, topological properties of the unweighted structural network are not significantly correlated with T3. It is important to highlight that about half of the Acrean municipalities did not meet the criteria of dengue establishment, which has a significant impact on the correlation obtained. With that in mind, we also investigated if those network properties were correlated with the event of dengue establishment itself. In Fig 6, we show boxplots comparing the centrality of municipalities that witnessed dengue establishment during the study period and those which did not. This comparison allows for the evaluation of whether municipalities grouped by occurrence (YES/NO) are also characterized by significantly distinct network properties. It is possible to identify that the descriptors most correlated to T3 were also those that differed the most between municipalities with and without dengue establishment, which are the geographical and effective distance to Rio Branco, Eigenvector and Strength centralities. The first record of Ae. aegypti in Acre occurred in 1995 in Rio Branco [57] (Table 2). Gradually the nearest municipalities in the southeast of the state also started recording it, and in 2001, eight municipalities had confirmed the presence of Ae. aegypti. These municipalities are the same municipalities where Rt > 1 was detected earlier in time, being municipalities along BR-317 (SG, Cpb, Xpr, Epcd and Brl) or with a direct connection to RB (PA) and two with good access to BR-364 since the beginning of the study (SM and Acld). From 2002 to 2006, another three municipalities reported the presence of Ae. aegypti: Plácido de Castro, Bujari and Assis Brasil. Bujari is the nearest neighbor of Rio Branco, but Ae. aegypti presence was not confirmed until 2006. It is difficult to ascertain the specific causes for this. On the one hand, this was a predominantly rural municipality in the early 2000’s. On the other hand, there was not a well-structured vector surveillance program implemented in small towns such as Bujari until later. Plácido de Castro and Assis Brasil are frontier municipalities, the former at the border with Rondônia state, receiving a considerable traffic of vehicles from the highly dengue endemic neighbor state. It is reasonable to infer that Ae. aegypti could be present there previously but unnoticed. However, dengue only established itself in Plácido de Castro in 2012/2013, despite the consistent detection of cases since 2001. Assis Brasil is a small town at the triple frontier Brasil-Peru-Bolivia. In 2000, it had a considerably small population, which may have contributed to the delayed invasion of Ae. aegypti. Between 2007 and 2013 no other city reported the presence of Ae. aegypti. The only exception being Cruzeiro do Sul, in 2008. This invasion was interrupted, however, and the mosquito was detected again only in 2013 [58]. From the municipalities to the north of Sena Madureira along the BR-364 highway, four confirmed the presence of Ae. aegypti in 2015: Feijó, Tarauacá, Mâncio Lima and Porto Walter. The first three, together with Cruzeiro do Sul, are located along the BR-364 and are important stopping points for those traveling. For the same reason, it is not possible to affirm that the mosquito was actually absent before that, since monitoring was not implemented. Five municipalities remained Ae. aegypti free as of 2015, according to the LI/LIRAa dataset (Table 2). These are the most peripheral municipalities of Acre, and also the most rural. Structural characteristics of Acrean municipalities can be part of the explanation for the slow invasion and establishment of Ae. aegypti in the state. According to IBGE’s urban/rural classification, Acre is mostly a rural state, but the population of all municipalities and the percentage of urbanization increased from 2000 to 2010. The exceptions are two, Rodrigues Alves and Assis Brasil, where total population increased as well but the urban population decreased [25]. Both municipalities presented an increase in the number of rural settlements [59]. Population growth, coupled with the increase in urbanization without improving the coverage of general services—such as garbage collection, water supply and sanitary sewage (Table 2) –, contribute to the fact that, even in a more rural state, the ideal conditions for establishing the mosquito are guaranteed. Those characteristics increase the difficulty in controlling Ae. aegypti [50] and consequently the local receptivity. In the process of describing the introduction of dengue in Acre, we assessed the contribution of network metrics and reproductive numbers as indicators of vulnerability, disease establishment and receptivity in the study of disease emergence. In the landscape epidemiology of infectious diseases, the description of the temporal dynamics of the populations of hosts, vectors, and pathogens, all spatially interacting in a favorable environment, contributes to the understanding of what characteristics and factors favor disease transmission an establishment [60]. Even small differences in landscape composition, which are often unapparent, can alter the microhabitats of the vector and thus the conditions that allow the amplification of a pathogen [60]. In addition, the dispersion of pathogens and vectors is directly linked to the development of transportation networks and increased globalization [49]. In this context, human population expansion has significantly affected the epidemiology of vector-borne diseases, creating a large urban continuum, altering the landscape structure and providing rapid mechanisms for hosts and pathogens to disperse [60]. The results of this study suggest that the landscape changes that occurred in the last decade have created favorable conditions for the establishment of dengue virus transmission, bringing together all the fundamental factors for its occurrence: the man, the virus, the vector, and especially the environmental, political, economic, social and cultural conditions favorable for the establishment of the transmission chain [61]. In Acre, the revitalization of its major roads, as well as the increased accessibility by air both to and within the state, have increased dengue vulnerability. Notice that the increase in the flow of people and importation of dengue cases to Acre coincides with when disease dispersion becomes more pronounced within the state, suggesting a synergy between increasing vulnerability of the state at a global scale and increasing local vulnerability among municipalities, fueling viral spread in the region. Human mobility is responsible for viral spreading, as both asymptomatic humans or those with very mild symptoms continue to carry out their tasks and take the virus to other regions. Humans are also responsible for possibly transporting the mosquito itself (which can be infected or not) by air, road, and waterway [2,7, 62]. Some network descriptors used in this study were useful for characterizing the level of centrality/periphery of the Acrean municipalities and its relationship with dengue importation. The mobility network that connects the Acrean counties is small and dense—22 nodes and density 0. 38 –, so centrality measures such as betweenness (B and Bw) and closeness (C and Cw) were not as relevant as they would be in larger and sparser networks. However, the strength (S) was the most relevant of the descriptors, because, in a small network, the relationship between local and global characteristics are stronger. Rio Branco is the main attractor of the state, with highest and strongest connectivity within Acre, since most of the routes connecting municipalities in the northwest to the southeast region pass by the capital. Also, it is a reference center for the entire state. A node with high connectivity, that is, a node to which many others connect to and where there are many outputs, is an essential node for the propagation of infectious diseases [63]. Therefore, investing in dengue control and prevention in a systematic manner in Rio Branco has potentially a strong impact throughout the state, not only because of its connectivity within Acre but also because this city is the main port-of-entry from other Brazilian states. The structure of the network, the distribution of degrees between nodes and the main routes are known to impact the spread of diseases [64]. In fact, we can observe that dengue first spread in the vicinity of Rio Branco and then, after an extended period of time, approximately 13 years, the first event of sustained transmission (here defined as at least three consecutive weeks with Rt > 1) occurred in Cruzeiro do Sul. The flow of people by air between Rio Branco and Cruzeiro do Sul, and the improvement of the BR-364 were very significant predictors of this first epidemic in the northwest region of Acre. Although Cruzeiro do Sul confirmed the presence of the Ae. aegypti only in 2008 [58], ever since the beginning of the study there were records of cases in the municipality, even when there were no records in intermediate connections by road between both municipalities. This last result highlights the importance of the airline connection between them for dengue case importation. The mobility data used in the present work is from the 2010 Census [30] when the structural network around Cruzeiro do Sul and Rio Branco was already stabilized regarding paving and maintenance, as shown in Fig 5. Since dengue epidemiological year of 2009/2010, all roadways directly connected to those two hubs had their maintenance concluded. Also, the unpaved connections to and from Manoel Urbano, which is structurally in between the two Acrean hubs, had also improved by that time. Most of its paving work was concluded by the end of 2010, although maintenance works continued for the rest of the study period. Despite the unavailability of detailed movement for performing a rigorous analysis, it is to be expected that during the period with unpaved and with major maintenance works in the roadways the flow of individuals was lower than when those structural bottlenecks were resolved. Therefore, we expect that the 2010 mobility flux data is more representative of the work/study-related movement from 2009 onwards. Before that, we conjecture that this particular type of human mobility was more concentrated over those connections with paved roads and without major maintenance work. The structural network status evolution depicted in Fig 5 allows us to identify a few key factors regarding the temporal evolution of ease of access and dengue case notification. Up until epidemiological year 2009/2010, when paved roads (black edges in that network) were mainly present in the vicinity of Rio Branco, the occurrence of Rt > 1 was also mostly present in that area—which can be identified by node’s color in Fig 5 as well as by the markings in Fig 3. In the area surrounding Cruzeiro do Sul, which was mainly connected by waterways (dashed blue edges) and roads under maintenance (orange edges), before 2007/2008 the occurrence of Rt > 1 was registered only once at Mâncio Lima (ML) and a few times in Cruzeiro do Sul. These results reinforce the hypothesis that those cases notified in Cruzeiro do Sul were most likely imported cases from Rio Branco due to the airline connection. It is interesting to note that in terms of roadways, the path from Rio Branco to Cruzeiro do Sul had a structural bottleneck at Manoel Urbano (MU) since it was only accessed by means of unpaved road (red edges) up to 2007/2008. Improvements in the structural network started in 2008/2009, which included paving and maintenance work of the roadways around Manoel Urbano, and were consolidated in 2009/2010 with the conclusion of major maintenance work in the roadways directly connected to both Rio Branco and Cruzeiro do Sul. We can see that is was precisely after those improvements that dengue notifications started to occur throughout the state of Acre, as well as the number of consecutive weeks with Rt > 1 in each municipality. The combination of low occurrence of Rt > 1 during the years with unpaved and under maintenance roadways, and the increased observation of Rt > 1 after improvements were implemented reinforces the hypothesis that the human mobility network and the structural changes were relevant to the observed geographical spread of the virus in the state of Acre. As for demographic characteristics, Acrean municipalities are marked by having small populations (Table 2). The capital Rio Branco, the one with the largest population, has 336,038 inhabitants [25]. The second largest municipality in population size is Cruzeiro do Sul, with 78,507. The remaining ones do not exceed 40,000 inhabitants, the majority of them having less than 20,000 inhabitants. The peripheral municipalities according to the network descriptors are also those with the smallest populations and are, mostly, more rural than urban. These municipalities are less vulnerable and less receptive to dengue, which is seen by the low occurrence of Rt > 1 and no event of sustained transmission as defined in this study. Nonetheless, in the world, there has been increased reporting of Ae. aegypti invasion and dengue introduction in less urbanized areas. In Colombia, dengue-infected Ae. aegypti was found in rural regions [65] south of Bogotá. In Nicaragua, a study showed the unsuspected presence of dengue cases in those areas as well [66]. These results from the literature and the results shown in the present manuscript suggest that introduction and establishment of dengue in those peripheral areas of Acre could be just a matter of time, especially since the virus has already been able to establish itself in the state. Overall, with this study, we understand that the determination of vulnerable and receptive localities for dengue in the state of Acre is essential for an effective action of entomological and epidemiological surveillance. It is fundamental that all municipalities, especially those with a roadway link, systematically monitor mosquito infestation, which is not currently the case. In addition, favorable conditions for mosquito development are present in Acrean municipalities, with low coverage of services, climate and increased urbanization [25], which probably impacts on the availability of breeding sites for Ae. aegypti and its establishment. Acre has a humid equatorial climate, which may contribute to the transmission of the dengue virus occurring throughout the year. According to [50], the mosquitoes of the region have the vectorial competence to transmit the DENV-2. Unfortunately, we did not find more biological information on the mosquito populations of Acre. We can conclude that Acre is a state at risk of major dengue epidemics, with entire populations susceptible to all serotypes of the disease. Due to the limited data and number of municipalities, our analysis is more descriptive/qualitative than quantitative. Nonetheless, we believe that the descriptors discussed here as well as the methodological approach has proved a valuable way to assess invasion risk which can be easily expanded to larger areas—conditioned on data availability –, allowing for a more quantitative analysis. .

The present ecological study is aimed at characterizing introduction and spread of dengue in the state of Acre, located in Brazil’s north region. We use sociodemographic and entomological data from the year 2000, when the first case was registered in Acre, up to 2015, when it was already endemic in most of the state. We use a methodological framework based on well-established concepts in malaria literature, namely vulnerability and receptivity, to assess dengue introduction and propagation in the study area. We make use of effective reproduction number as an empirical indicator of dengue outbreak spatiotemporal patterns, as well as complex network theory concepts applied to human mobility and transport infrastructure changes to address dengue importation probability from other Brazilian states and its circulation within Acrean municipalities. On the one hand, we show that the increase of airline passengers, coupled with observed dengue incidence in Brazil, has steadily increased Acre’s vulnerability to dengue introduction, while transport infrastructure development has facilitated its spread within the state. On the other hand, socio-demographic changes aimed at urbanization and economic development have increased receptivity to its main vector, the Aedes aegypti mosquito. The combination of those two factors ultimately leads to the establishment of the virus in Acre.

Abstract Introduction Materials and methods Results Discussion

invertebrates dengue virus medicine and health sciences pathology and laboratory medicine engineering and technology transportation pathogens geographical locations microbiology social sciences human mobility animals transportation infrastructure roads viruses rna viruses network analysis insect vectors civil engineering human geography infectious diseases computer and information sciences geography south america aedes aegypti medical microbiology epidemiology microbial pathogens brazil disease vectors centrality insects arthropoda people and places mosquitoes eukaryota flaviviruses earth sciences viral pathogens biology and life sciences species interactions organisms

2017

The introduction of dengue follows transportation infrastructure changes in the state of Acre, Brazil: A network-based analysis

The amount of genetic variance underlying a phenotypic trait and the strength of selection acting on that trait are two key parameters that determine any evolutionary response to selection. Despite substantial evidence that, in natural populations, both parameters may vary across environmental conditions, very little is known about the extent to which they may covary in response to environmental heterogeneity. Here we show that, in a wild population of great tits (Parus major), the strength of the directional selection gradients on timing of breeding increased with increasing spring temperatures, and that genotype-by-environment interactions also predicted an increase in additive genetic variance, and heritability, of timing of breeding with increasing spring temperature. Consequently, we therefore tested for an association between the annual selection gradients and levels of additive genetic variance expressed each year; this association was positive, but non-significant. However, there was a significant positive association between the annual selection differentials and the corresponding heritability. Such associations could potentially speed up the rate of micro-evolution and offer a largely ignored mechanism by which natural populations may adapt to environmental changes. Predicting an evolutionary response to selection in a phenotypic trait requires knowledge of the strength of selection acting on the trait and its genetic basis. Although it has long been recognized that the strength, and direction, of selection may vary with environmental conditions (e. g. , [1]), widespread recognition of the fact that additive genetic variance (and thus heritability) may also change with environmental conditions has been more recent [2], [3]. Taken together, these observations generate an expectation of an environmentally driven association between the two parameters that, in theory, has the potential to either enhance (positive association) or constrain (negative association) any response to selection. Surprisingly, however, to our knowledge only one study to date has quantified the association between annual estimates of selection and expression of genetic variance (measured as heritability) in a heterogeneous environment [4]. In this article, we present data from a long-term study of a great tit (Parus major) population known to be experiencing substantial shifts in climatic conditions, and test for the effects of the novel environmental conditions on the expression of additive genetic variance, and the selection on, a key life history trait, breeding time. Many studies have found that selection is often strongest when environmental conditions are adverse (e. g. , [4]–[8]), and there is a clear indication that “perturbed or stressed” populations have larger standardized selection differentials than “undisturbed” populations ([9] p. 208). For example, Garant and co-workers [5] examined selection on fledgling body mass in a population of great tits and found that selection differentials were greater in years when average body mass was low and when the proportion of individuals surviving to recruitment was low, both indicative of poor/adverse environmental conditions. In general, therefore, selection is often stronger when environmental conditions are adverse. Unlike the general tendency for selection to be stronger in adverse environments, conclusions regarding the effects of good versus adverse environments on the expression of additive genetic variance are more mixed. Laboratory studies investigating the effect of environmental conditions have generally found a weak tendency for heritability to increase in stressful environments with this being caused by changes in both the expression of genetic variance as well as the environmental variance (reviewed in [10]). This pattern, however, is in contrast to most studies from natural populations that find, at least for morphological traits, that additive genetic variance and heritability is often relatively lower in unfavorable conditions [3], [10], [11]. It is important to realize that heritability (h2) may change under different environmental conditions either because of changes in additive genetic variance (VA) or other variance components (e. g. , permanent environmental variance (VPE) or residual variance (VR) ). However, changes in VA are of particular interest because they indicate a change in the “evolvability” [12], or the potential to respond to selection, of a trait. Furthermore, changes in VA can only be due to a change in the genetic architecture of a trait through mechanisms such as genotype-environment interactions, changes in mutation and recombination rates, and removal of alleles with low fitness by selection (reviewed in [10]). Depending on the direction and scale of these changes, both additive genetic variance and heritability may increase or decrease depending on the relative impact of each of the above factors [10]. The possibility that both the expression of additive genetic variance of a trait and the strength of selection acting on it may vary with environmental conditions is significant, as such environmentally induced variation may be important in determining the evolutionary dynamics of natural populations. In particular, the observation of a general increase in genetic variance of morphological traits [3], [10], [11] and a reduction in selection [4], [6] during favorable conditions in natural populations leads to the expectation of a negative relationship between genetic variance and the strength of selection, such that selection should be strongest in years in which the expression of additive genetic variance is least. This association could severely constrain a response to selection and provide one explanation for the frequently observed scenario of apparent stasis in natural populations [4]. However, in contrast to morphometric traits, life history traits do not appear to show a clear indication of increased heritability in stressful environments [3]. This makes it more difficult to predict how, or if, additive genetic variance and selection on life history traits may covary in a heterogeneous environment. Surprisingly, despite the potential importance of environmentally induced associations between the strength of selection and expression of genetic variance, we are aware of only one previous study that has tested for such an association. Wilson et al. [4] found that the strength of selection on body weight in a free-living population of Soay sheep (Ovis aries) in a given year was negatively correlated with the expression of total genetic variance (assessed via the heritability) of body weight, suggesting a possible constraint on the potential for evolution of body weight in this species. However, so far no study has, to our knowledge, examined the association between strength of selection and VA (or h2) in a life history trait. Hence, we do not know if such relationships are common in nature, and whether they are generally negative, which may constrain an evolutionary response, or whether there are examples of positive associations between strength of selection and VA (or h2), which would speed up an evolutionary response. Here we use data from an exceptionally long-term study population of great tits (Parus major) in the Hoge Veluwe, the Netherlands, to investigate how selection and expression of additive genetic variance of a key life-history trait (timing of breeding, or “laying date”) vary in relation to rapid changes in environmental conditions (spring temperature). The evolutionary response in a trait between generations can be predicted as R = VA * β [13], [14], where β is the selection gradient, defined as the covariance between relative fitness and trait value divided by the phenotypic variance in the trait (i. e. , β = cov (ω, trait) /VPtrait) [15]; we therefore test the association between VA and the selection gradients β under different environmental conditions. We also consider the alternative format for predicted response, R = h2 * S [16], where S is the selection differential, defined as the covariance between relative fitness and trait value (i. e. , S = cov (ω, trait) ) and test for an association between heritability and selection differentials. This system is particularly well suited to an exploration of the association between selection and VA in a variable environment because phenotypic data, pedigree data, and a thorough understanding of how environmental conditions influence laying date are available [17], . Previous studies in this population have reported a significant increase in spring temperature over the past four decades [18] and have also shown that warm spring temperatures lead to earlier laying dates [17]. Furthermore, warmer temperatures lead to reproduction being mistimed relative to the food peak [17], resulting in a decrease in both the number and size of fledglings [19], and in the proportion of females producing a second clutch [20]. Spring temperatures are thus not only directly related to observed variation in laying dates but can also be used as a measure of environmental quality in the population. In addition, spring temperatures are now significantly above those which the population has previously experienced [18], providing an ideal opportunity to study how novel environmental conditions may influence evolutionary dynamics. We therefore tested the temperature dependence of the selection gradients and differentials, how expression of additive genetic variance and heritability changed with temperature, and finally, how the measures of selection were associated with the amount of genetic variance present in the population. We found, firstly, strong selection on laying date, with early breeding birds having higher fitness than late breeding individuals (Table 1). Indeed, 29 out of the 35 estimates of annual selection gradients and differentials were negative (Figure 1, Table S2), reflecting general selection for earlier breeding, as has previously been shown in this population [17], [21]. Secondly, the interaction between laying date and standardized spring temperature was significantly negative (Table 1), indicating that with increasing spring temperatures the relationship (slope) between fitness and laying date became more negative (i. e. , slope steeper in warmer years). Consequently, selection for early breeding was significantly stronger (indicated by more negative values of β) in warm years than in cold years; i. e. , the strength of selection on lay date varied with environmental conditions (Figure 1). This result was confirmed by regressing the annual selection gradients (β) against temperature: there was a significant increase in the (absolute) magnitude of the strength of selection with increasing temperatures (regression slope = −0. 044, se = 0. 019, t33 = −2. 203, p = 0. 035, Figure 1a). The results were the same for selection differentials (regression slope = −1. 589, se = 0. 450, t33 = −3. 529, p = 0. 001, Figure 1b). Comparing a model in which the additive genetic and permanent environment components of variance (VA and VPE) in a given year were constant across different spring temperatures to one in which VA and VPE could vary with the temperature gave strong support for environmental dependence of VA and VPE (χ24 = 74. 90, p<0. 001). Consequently, we used the predictions from the model in which the two variance components varied with spring temperature to generate estimates of annual VA and h2 and to explore how these annual estimates corresponded to the observed changes in the strength of selection on laying date. The estimated environment-specific G-matrix predicted a substantial increase in VA with increasing standardized spring temperatures (Figure 2a, each point represents an environment-specific VA estimate). Similarly, there was a corresponding increase in the year-specific heritability estimates with increasing temperature (Figure 2b, each point represents a environment-specific h2 estimate). We then tested whether the effects of increasing temperature on selection and genetic variance generated an association between them. The relationship between the selection gradients (β) and additive genetic variance (VA) for laying date was negative but non-significant (slope = −0. 006, se = 0. 005, t33 = 1. 18, p = 0. 25; Figure 3a, dotted line). However, as random regression models are known to give upwardly biased estimates at the endpoints of the polynomials [22], we also tested this relationship after removing the extreme VA outliers (VA >10, see Figure 3a). This resulted in a near-significant relationship between the two (slope = −0. 014, se = 0. 008, t31 = 1. 84, p = 0. 075; Figure 3a, solid line). Furthermore, there was a significant negative relationship between the selection differentials S and heritability (slope = −10. 96, se = 4. 43, t33 = 2. 48, p = 0. 019, Figure 3b), which was robust to excluding the two extreme heritability estimates (excluding h2 >0. 25: slope = −13. 82, se = 6. 2, t31 = 2. 23, p = 0. 03). Finally, using standardized measures of selection, there was a negative although non-significant significant relationship between selection and additive genetic variance and a significantly negative relationship between strength of selection and heritability (see Text S1). Note that because there is selection for early breeding, selection gradients and differentials are negative, but there is a positive association between the absolute strength of selection and levels of additive genetic variance (or heritability). As a result, in years in which selection on laying date was relatively strong, estimated VA (and h2) was higher than in years when selection was weak (Figure 3). This association resulted in a highly significant relationship between temperature and the magnitude of the predicted response to selection (Figure 4). Our analysis of long-term records on an important life history trait in a wild bird population found evidence that in years when spring temperatures were highest, selection was strongest, and the magnitude of estimates of additive genetic variance VA (and hence heritability) was also highest. As a result, there was evidence of a positive association between the strength of selection and the expression of additive genetic variance, and heritability. A positive association such as this between the strength of selection and expression of genetic variance and heritability could make the magnitude of the response strongly environmentally dependent; in this case, warming temperatures would considerably enhance any expected response to selection. As has generally been found in studies of selection on laying date in birds [17], [21], [23], [24], [25], selection gradients and differentials were generally negative, indicating that early-breeding individuals had higher fitness than late-breeding individuals. Furthermore, the strength of selection was strongest when temperatures were highest (Figure 1). It has previously been shown that reproductive success [26] has declined in this population over the study period, most likely because, with increasing spring temperatures, there is evidence of increased “mistiming” of reproduction relative to the peak in food abundance [17]. This decline in reproductive success suggests that high spring temperature is generally associated with adverse environmental conditions. Hence, our results confirm the expectation in natural populations of stronger selection in adverse environmental conditions [9]. It is important to point out, however, that high temperatures are not necessarily associated with adverse environmental conditions in other systems. For example, a population of great tits in the U. K. has also experienced increasing temperatures, but recruitment rates in this population have increased over time [27]. Previous studies on natural populations have found that heritability decreased when environmental conditions are stressful [3], [10], although we know less about how VA changes. Here, we found instead that both additive genetic variance and heritability of laying date increased rather than decreased (Figure 2). Although there was substantial evidence that VA and VPE changed with environmental conditions (see Results), the change in VA alone was not statistically significant [18], something that is reflected in the large standard errors in Figure 2a. However, the statistical power to detect significant changes in additive genetic variance in relation to varying environmental conditions using a random regression animal model approach may be limited [18], [28]. Most importantly, the increase in VA is very large and represents 81. 4% of the total change in VP (Figure 2a). This increase in VA is, for example, much larger than the increase in maternal genetic variance (VM) for birth weight in Soay sheep [4]. Note also that in the Soay sheep analysis, maternal environmental effects were not fitted with the same order polynomials as the maternal genetic effects, so that some of the increase in maternal genetic effects variance estimates could potentially be driven by environmental rather than genetic effects (in the same way as permanent environment variance will inflate additive genetic variance if not fitted explicitly, [29]). One possible explanation for why VA may increase with higher temperatures is that high temperatures constitute not only a stressful, but also a novel, environment. For example, 2005 and 2007 had the highest recorded spring temperatures since this population study began back in 1955. It has been suggested that VA could increase in novel environments because selection has not yet had the possibility to remove the most deleterious alleles, as it will have in the ancestral environment, thereby causing an increase in the standing genetic variation [30]; a suggestion that has been confirmed in some empirical studies [31], [32]. More generally, our finding adds support to the idea [3] that predicting the direction in which VA should change with environmental conditions is complicated when environmental changes also leads to novel conditions, as is often the case with human-induced changes [3]. The increase in VA, heritability and strength of selection with increasing spring temperature meant that there was a positive association between the strength of selection on laying date and the heritability as well as expression of additive genetic variance of laying date (Figure 3a and b, respectively). The relationship between selection and amount of genetic variance was in the same direction whether using β as the measure of selection and VA as the measure of the potential for the population to adapt, versus using S and h2, but it was stronger (and hence statistically significant) between S and h2 (Figure 3b) than between β and VA (Figure 3a). One possible explanation for this may be that in the S and h2 comparison, both parameters depend on VP whereas in the β and VA comparison only β depends on VP and thus a change in VP may more quickly lead to a disassociation between β and VA than between S and h2. Nevertheless, we believe the fact that the relationships between β and VA and between S and h2 are in the same direction (as well as that between standardized selection and VA/h2; see Text S1) offers strong support for an environmental coupling between these two parameters. This conclusion is supported by a highly significant temperature dependence of the predicted response to selection (see below, Figure 4). Following traditional methodology we predicted the expected response to selection (see Text S1) using the Lande equation: R = VA * β [13], [14] but correcting for overlapping generations and the sex-limited expression of laying date, with the year-specific VA and β estimates (see Table S2), which amounted to an advance of 1. 81 days in total over the study period. Furthermore, using the average of the annual VA and β values gave a predicted response of 1. 46 days advancement, which corresponds to only 81. 1% of the predicted response using year-specific values. Thus, not incorporating environmental dependence of the expression of genetic variance and strength of selection may underestimate the predicted response by up to 20%, at least in this specific case. Failing to incorporate an environmentally dependent association between the strength of selection and genetic variance may further obscure our understanding of microevolution as the predicted response will be dependent on the environmental variable in question. For example, in our study the predicted response is strongly correlated with spring temperature, with a much larger predicted response in warmer temperatures compared to cold (Figure 4). We caution, however, that the breeder' s equation (and equivalent Lande equation) has particularly poor success when applied to studies in natural populations [33], presumably because many of its underlying assumptions are not met (see Text S1 for further discussion on this topic). Very few studies have simultaneously examined how environmental factors influence genetic expression and selection and the association between them. Indeed, we are only aware of this being examined in a Soay sheep population [4], where there was a negative association between the strength of selection and heritability of body size. Another example where there may be a negative association between the strength of selection and heritability is for juvenile growth rates in North American red squirrels (Tamiasciurus hudsonicus) [34]. Although this study did not explicitly consider the association between selection and genetic variance, they found that VA and maternal genetic variance increased in years with low cone abundance (poor environment) whereas viability selection was stronger in years of high cone abundance (due to competition for territories [34]). This should generate a negative association between selection and total genetic variance that may hamper a response to selection. Our results thus demonstrate a relatively unexplored mechanism that could potentially increase the speed of adaptation to climate change in this population. As temperatures are expected to continue to increase [35], a positive association between strength of selection on laying date and its potential to evolve may prove an important factor allowing at least this specific population to adapt to the rapid environmental conditions experienced. As it is ultimately this rate of adaptation that is crucial if species are to cope with climate change [36], our findings suggest that models linking population viability to climate change should incorporate such dynamic processes. The data were collected in the Hoge Veluwe National Park, the Netherlands (52°05′ N, 05°50′ E), during the period 1973 to 2007. Nest boxes were visited at least once every week during the breeding season (April–June). The laying date of the first egg of a female' s clutch (laying date, LD) was calculated from the number of eggs found during the weekly checks, assuming that one egg was laid per day. Both parents were caught and individually marked on the nest using a spring trap when the young were 7–10 d old. Laying dates are presented as the number of days after March 31 (day 1 = April 1, day 31 = May 1). We only used information on the first clutch, defined as any clutch started within 30 d of the first laid egg in any given year. Replacement and second clutches (which currently compromise <5% of breeding attempts, 21]) were thus excluded from the analysis. In total, therefore, we had information about 3,852 breeding records from 2,394 females. More details about the study population can be found in van Balen [37]. Temperature data were obtained from the De Bilt weather station of the Royal Dutch Meteorological Institute (www. knmi. nl/klimatologie/daggegevens) and used to calculate the daily average temperature over the period March 13–April 20, which is the period that best predicts the onset of laying using a sliding window analysis (see [18] for more detail). To test for a relationship between spring temperature and the strength of selection on laying date, we took two approaches. First, we used a generalized linear mixed effects model (GLMM) with a Poisson error link fitted in ASREML-R [38] to model the relationship between number of recruits a female produced for the given year (as the measure of fitness) and her laying date that year, and to test its dependence on spring temperature (as measured by the interaction term between laying date and spring temperature). Individual identity and year were included as random effects to account for repeated measures on the same individuals and on years. Second, we estimated the annual strength of selection using the number of recruits produced per year divided by the mean number of recruits produced in the given year as a measure of relative fitness (ω) for each individual. Selection was then measured as the selection gradients (β) defined as the covariance between relative fitness and observed laying date divided by the variance in observed laying date, i. e. β = cov (ω, LD) /VPLD. Using this measure of selection allows us to predict the response to selection using the Lande equation: R = VA*β [13]. Predicting the response to selection can also be done using the more familiar Breeder' s equation, R = h2*S [16], which uses an alternative measure of selection, the selection differentials defined as the covariance between a female' s relative fitness (ω) and her observed laying date (LD), i. e. S = cov (ω, LD) [15]. Because a previous study examining the association between strength of selection and expression of genetic variance used S and h2 as parameters [4], we also present our results using these parameters for comparison. We note, however, that using selection gradients may represent a better measure of selection when the phenotypic variance in a trait changes [14], which it does here. We then regressed the annual selection gradients (and differentials) against the environmental values using a least-squares regression (with 1/se2 as weights when considering the selection gradients) in R 2. 8. 0 [39]. Finally, to allow comparison with other studies [40], we repeated all selection analyses using variance-standardized laying dates (i. e. standardizing laying date values to have zero mean and unit variance within each year). This did not change our conclusions and we report the results from these analyses in the Supporting Information section (Text S1, Table S1). Yearly spring temperature values, standardized spring temperature values, sample size, mean laying dates, selection gradients (β), selection differentials (S), standardized selection differentials, estimated additive genetic variance, and heritability estimates along with annual predicted responses to selection (VA*β) are all reported in Table S2. Quantitative genetic analyses require knowledge about the relationships among individuals within a population. Here, a pedigree was constructed where all ringed females known to have bred were assigned a mother and father as determined from observational data. In cases where brood manipulation experiments had been carried out and chicks had been moved between nests, we assigned the genetic parent rather than the social parent. If only one parent was known, we “dummy coded” the missing parent to preserve sibship information (note that we did not assign a phenotype to this parent). The extra-pair paternity (EPP) rate is unknown in this population, but is generally found to be low (3%–9%) in other populations of great tits [41], [42] and as extra pair paternity rates of less than 20% have been shown to have a negligible impact on heritability estimates [43] using a social pedigree is unlikely to be problematic. Phenotypic trait variances can be separated into genetic and environmental causes of variation using an “animal model” [44]–[46]. By maximizing the information available in an extensive multi-generational pedigree, the “animal model” minimizes upward inflation of estimates of additive genetic variance (VA) due to shared environmental effects between relatives; this approach has been shown in simulation studies to perform well in partitioning environmental and genetic components of variance [29]. There are several additional reasons to believe that the genetic and environmental components have been well separated here. First, a previous study found no indication that common environment effects in the form of maternal effects are important for laying date in this population (VM/VP = 0. 0023 [47]). Second, although common environmental effects frequently play a major role in inflating covariances between relatives in nestling traits [48], this is rarely the case for traits that are only expressed as adults, like laying date which we study here. Third, we explicitly take common environmental effects into account by fitting a permanent environmental effect [45]. In summary, therefore, we believe that our estimates of VA and h2 are accurate and unbiased by inflation of common environment effects. Rather than only estimating the amount of genetic and environmental variance in laying date, we are interested here in whether the variance components changed with environmental conditions, and we therefore used a “random regression animal model” [49]. Random regression models use covariance functions [50] to explicitly fit variance components as a function of the environment and hence allow a detailed examination of how environmental heterogeneity—in this case, spring temperature—influences genetic architecture. Thus our model was: (1) where LDi is the vector of the individual (i) laying dates and X, Z1, Z2, and Z3 are the design and incidence matrices relating to the fixed and random effects of the additive genetic (ai), permanent environment (pei), and year (yri) observations, respectively. T is the spring temperature each year standardized to a (−1,1) interval (Table S2). Fixed effects (bi vector) included age as a two-level factor (first year breeder or older), to correct for the fact that laying date is generally later in young birds compared to older birds in great tits [51], and spring temperature to account for the population-level response in mean trait value. Year (yr vector) was included as a random effect in order to model variation between years not explained by spring temperature and a permanent environment effect (pei vector) was fitted because of the repeated sampling of the same individuals; this also reduces inflation of estimates of the additive genetic variance due to environmental factors [29]. The error term (e vector) was partitioned into three decade–specific (1973–1984,1985–1996,1997–2007) groups, thus allowing residual errors to vary between decades. φ (ai, n1, T) is the random regression function of order n1 of the additive genetic effect of individual i, which varies as a function of the temperature T in a given year, and similarly, φ (pei, n2, T), is the random regression function of order n2 of the permanent environment effect varying as a function of T. Because we were only interested in whether the two variance components (and particularly VA) changed with the environment, we only fitted two models. The first model was a zero order function (n1 = n2 = 0) for both VA and VPE in which variance components are constant across the environment. In the second model, we fitted a first order polynomial (n1 = n2 = 1) for both VA and VPE, thus allowing both additive genetic effects and permanent environment effects, and hence their corresponding variance components, to vary across the environment T. These two models were then compared using a likelihood-ratio test by calculating twice the difference in log likelihood, which is chi-squared distributed with degrees of freedom equal to the difference in degrees of freedom between the two models [52], which is here equal to 4 (variance in slopes and covariance between elevation and slope for both VA and VPE). As the model where both variance components were allowed to vary was significantly better than a model in which they were assumed to be constant (see Results) we used the estimates from the first order polynomial model to generate predictions of annual values of VA (and VPE) across varying temperatures. The environment-specific additive genetic covariance matrix, G, was then obtained as G = zQzT, where z is the vector of orthogonal polynomials evaluated at standardized temperature values and Q is the additive genetic variance-covariance matrix of the random regression parameters. Approximate standard errors for the (co) variance components of G as a function of the temperature values were calculated according to Fischer et al. [22], with confidence intervals defined as twice the standard errors. Finally, environment-specific heritability estimates were calculated as the environment-specific VA estimate divided by the environment-specific VP estimate from the model in which both VA and VPE varied with the environment. Because it has been found that random regression models can be sensitive to “edge effects” [22], [53], we repeated our analyses where we look at the association between strength of selection and expression of genetic variance to be conservative. For more information about the use of random regression animal models in natural populations, see [18] and [54]. All animal models were fitted using REML methods implemented in ASReml v 2. 0 [38]. In order to test for an association between the strength of selection operating on laying date and the expression of additive genetic variance in laying date, we used environment-specific (and thus annual) VA and h2 estimates generated from the random regression animal model and regressed the annual selection gradients on our annual estimates of VA; we then repeated the regression for annual selection differentials against h2. Regressions using selection gradients were weighted by the inverse of the square of the standard error. Because some individuals bred in multiple environments (i. e. years), estimates of selection will not be entirely independent, potentially violating some of the assumptions of least squares regression analyses. Although this is an inherent problem to all longitudinal studies, we assessed the potential for it to bias our conclusions by repeating our selection analyses using only a single record per individual (its first breeding attempt). Because this did not change the direction or significance of our analyses (regression of β on VA using 1/se2 as weights: b = −0. 009, se = 0. 005, t33 = −1. 70, p = 0. 099; regression of S on h2: b = −14. 54, se = 5. 08, t33 = −2. 86, p = 0. 007), we conclude that the potential violation of the non-independence criteria caused by multiple breeding events from the same individuals is not a significant issue here. Although annual estimates of VA and h2 are derived from the random regression model, note that in testing for a relationship between them and selection, we use them only as predictor variables in a regression, for which there need not be an assumption of independence of data points.

The speed of evolutionary change in a phenotypic trait is determined by two key components: the amount of genetic variance underlying the trait and the strength of selection acting on it. Many studies have shown that both selection and expression of genetic variance may depend on the environmental conditions the population experiences. However, the possibility that the strength of selection and the expression of genetic variance become positively or negatively associated as a result of this environmental covariance, so as to speed up or hamper an evolutionary response, has been largely ignored. Here we show that, in a wild bird population, the annual strength of selection on and the expression of genetic variance in timing of breeding (a key life history trait) are positively associated due to changing environmental conditions (warmer temperatures). Such a positive association should potentially speed up any microevolutionary response to selection (such as that imposed by climate warming). Our results illustrate the existence of substantial temporal variation in response to environmental heterogeneity, and thus highlight a so far neglected mechanism that may be important in determining the evolutionary dynamics in natural populations.

Abstract Introduction Results Discussion Materials and Methods

evolutionary biology/evolutionary ecology evolutionary biology/evolutionary and comparative genetics genetics and genomics/complex traits

2011

Speeding Up Microevolution: The Effects of Increasing Temperature on Selection and Genetic Variance in a Wild Bird Population

Hepatitis C virus (HCV) induces interferon (IFN) stimulated genes in the liver despite of distinct innate immune evasion mechanisms, suggesting that beyond HCV infected cells other cell types contribute to innate immune activation. Upon coculture with HCV replicating cells, human CD141+ myeloid dendritic cells (DC) produce type III IFN, whereas plasmacytoid dendritic cells (pDC) mount type I IFN responses. Due to limitations in the genetic manipulation of primary human DCs, we explored HCV mediated stimulation of murine DC subsets. Coculture of HCV RNA transfected human or murine hepatoma cells with murine bone marrow-derived DC cultures revealed that only Flt3-L DC cultures, but not GM-CSF DC cultures responded with IFN production. Cells transfected with full length or subgenomic viral RNA stimulated IFN release indicating that infectious virus particle formation is not essential in this process. Use of differentiated DC from mice with genetic lesions in innate immune signalling showed that IFN secretion by HCV-stimulated murine DC was independent of MyD88 and CARDIF, but dependent on TRIF and IFNAR signalling. Separating Flt3-L DC cultures into pDC and conventional CD11b-like and CD8α-like DC revealed that the CD8α-like DC, homologous to the human CD141+ DC, release interferon upon stimulation by HCV replicating cells. In contrast, the other cell types and in particular the pDC did not. Injection of human HCV subgenomic replicon cells into IFN-β reporter mice confirmed the interferon induction upon HCV replication in vivo. These results indicate that HCV-replicating cells stimulate IFN secretion from murine CD8α-like DC independent of infectious virus production. Thus, this work defines basic principles of viral recognition by murine DC populations. Moreover, this model should be useful to explore the interaction between dendritic cells during HCV replication and to define how viral signatures are delivered to and recognized by immune cells to trigger IFN release. Hepatitis C virus (HCV) infection constitutes a major global health problem since more than 140 million people suffer from chronic sequelae of the infection [1]. Once infected, approximately 80% of the individuals are not able to clear the pathogen and develop a chronic infection that often is associated with liver diseases such as fibrosis, cirrhosis and hepatocellular carcinoma, thus resulting in the need of liver transplantation [2]. It is believed that chronic infections are a consequence of a multi-factorial immune failure, due to delayed and weak T-cell responses, as well as dysfunctional B-cell, natural killer (NK) -cell and dendritic cell (DC) responses [3–9]. In addition, it has been shown that a strong pre-stimulation of interferon (IFN) stimulated genes (ISGs) during chronic HCV infection constitutes a marker of decreased responsiveness to IFN-based therapies [10]. Bridging the innate with the adaptive immune responses, DC have an important role in the establishment of a protective immune response and they are crucial for the production of interferons and the activation of immune cells [11]. Based on their distinct phenotype and functional characteristics, human peripheral DC can be classified in 3 major subsets. These include conventional DC (cDC) which encompass the myeloid CD1c+/BDCA1+ DC (mDC1) that are the largest mDC population in the blood and that are known for their antigen presenting capacity and cytokine expression, and the myeloid CD141+/BDCA3+ DC (mDC2) which produce IL-12 and type III IFN and have the ability to cross-present antigens to CD8 T-cells. The third subset is represented by the plasmacytoid dendritic cells (pDC), also known as natural type I IFN-producing cells, which upon activation produce high levels of type I IFN [12–15]. All three subsets have been implicated to be involved in responses to HCV infection [16–20]. However, not much is known about the interplay between the DC subsets among each other and with other immune cells. This is at least in part due to technical as well as practical difficulties of studying human DC. Studies of murine DC principally offer an attractive alternative for dissecting the requirements of DC stimulation. However, due to the restricted species tropism of HCV to humans and the lack of suitable immune competent mouse models [21] not much research has been performed using murine dendritic cells. A clear advantage of the murine system is the availability of genetically modified animals that allow in depth mechanistic studies. Furthermore, homologs of human dendritic cell subsets can be easily generated upon addition of FMS-like tyrosine 3 ligand (Flt3-L) to murine bone marrow (BM) cells, thus inducing high numbers of BM-derived DC [22–25]. Likewise, murine DC can be generated from BM using granulocyte/macrophage colony-stimulating factor (GM-CSF), whereas these cell cultures lack the pDC counterparts and are believed to yield DC which primarily resemble monocyte-derived DC [24,26,27]. To some extent Flt3-L DC cultures reflect the physiological DC development that gives rise to three major DC subsets for which orthologues can similarly also be found in the human peripheral blood: pDC as well as the two subsets of conventional dendritic cells, CD8α-like DC which resemble the human CD141+ (mDC2) subset and CD11b-like DC which are homologous to the human CD1c+ DC (mDC1) [23,28]. Given the homology between murine and human DC subsets we used murine Flt3-L differentiated DC from wildtype as well as genetically modified mice to dissect the requirements for DC stimulation by HCV. Murine bone marrow (BM) cells were isolated and in vitro differentiated into dendritic cells using medium enriched either with the cytokines Flt3-L or GM-CSF. Subsequently, cells were cocultured with HCV subgenomic replicon (SGR) or HCV full-length (Jc1) transfected human Huh7. 5 cells for 18 hours (as further described in the materials and methods section) and analyzed by flow cytometry. In parallel, DC populations were stimulated with VSV-M2 at a multiplicity of infection (MOI) of 1. To assess the activation of the respective DC populations SiglecH+ CD11c+ Flt3-L DC and CD11c+ CD11b+ GM-CSF DC were analyzed for the expression of the activation marker CD69 using flow cytometry (Fig 1A and 1B). Cocultivation of HCV replicon or virus-transfected hepatoma cells with DC cultures led to an upregulation of CD69 expression on both cell types analyzed. This is evidenced by the representative FACS histograms depicted in Fig 1B and by the mean values of CD69 surface expression (MFI) across at least three independent experiments given in Fig 1C. Collectively, these results indicated that HCV replication was sensed by murine DC subsets and led to their activation. Since activation occurred also upon co-culture with replicon-transfected cells, which do not produce infectious viral progeny, DC activation seems to be independent of virus assembly and release. To analyze the cytokine production, supernatants of DC cocultures were harvested and analyzed for the production of type I as well as type III IFNs by using commercially available ELISAs. Only Flt3-L derived DC cultures produced significant amounts of IFN-α (Fig 1D), IFN-β (Fig 1E) and IFN-λ (Fig 1F) in response to HCV replication, whereas GM-CSF-derived DC did not (Fig 1G–1I). Stimulation with VSV-M2 as control led to a significant type I IFN production from both cell types (Fig 1D, 1E, 1G and 1H) indicating that only Flt3-L derived DC are activated by HCV replication to produce significant amounts of type I and type III IFN. To further characterize the requirements for IFN release by HCV-stimulated Flt3-L derived DC, cells were cocultured with HCV cells transfected with the HCV subgenomic replicon (SGR), the full length virus (Jc1), or a replication incompetent mutant full length virus which encodes an in-frame deletion of 10 amino acids spanning the GDD motif that abolishes the RNA polymerase activity (ΔGDD). Both replication of the HCV SGR and the full length virus stimulated IFN-α production by Flt3-L derived DC, whereas coculture with transfected cells harboring the replication incompetent HCV construct did not result in detectable IFN production (Fig 2A). These experiments indicated that virus replication but not infectious virus production is necessary for the murine Flt3-L derived DC to sense HCV. Treatment with RNAse and DNAse during coculture confirmed that the activation of the DCs was not due to residual exogenous nucleic acids in the supernatant of the cultures (Fig 2B). To further define the mode of stimulation, either cell-free HCV virus (Jc1) containing culture fluid or concentrated supernatant (SN) of SGR transfected cells was used to stimulate the DC. Cell-free HCV preparations (Jc1 MOI 10) were incubated with 2x105 DC for 18 h and secretion of IFN was quantified. As control, DCs were left untreated. As is evidenced in Fig 2C treatment with cell-free Jc1 slightly upregulated IFN production by dendritic cells compared to unstimulated cells. However, stimulation of Flt3-L derived DC with the culture fluid derived of SGR-transfected cells that was concentrated led to a statistically significant IFN production by dendritic cells (Fig 2D), indicating the cell-free stimulation of Flt3-L derived DC by HCV signatures released from replicon cells. Extracellular vesicles, like exosomes, have been implicated in the stimulation of human pDC [29]. To analyze whether extracellular vesicles could mediate stimulation of murine DC, we isolated vesicles from mock, replication incompetent HCV subgenomic replicon mutant (pUCΔGDD) ΔGDD or SGR transfected cells employing a commercial exosome purification kit and used these preparations to stimulate murine DC (Fig 2E). Western blot analyses confirmed enrichment of extracellular vesicles that contain typical markers of exosomes like CD81, CD63, Hsp70 and AnxII [30] (Fig 2F). Only vesicles derived from HCV replicating cells were able to stimulate the murine DC to produce type I IFN. These data indicate that cell-cell contact is not necessary for murine Flt3-L derived DC stimulation and that extracellular vesicles could contribute to DC mediated type I IFN production. As human DC have been described to recognize HCV replication in vitro in an endosome-dependent manner [17,31] the cocultures were treated with the endosomal acidification inhibitors bafilomycin A1 or concanamycin A at a dose of 25 nM or 5, respectively, throughout the duration of the coculture. These treatments with inhibitors of endosomal acidification resulted in an abolishment of the IFN-α production, indicating that endosomal acidification is required for the production of type I IFN by murine DC (S1 Fig). Taken together, type I IFN production by murine Flt3-L derived DC is dependent on RNA replication but independent of infectious virus production and of direct cell-to-cell contact. Human pDC have been described to sense HCV RNA via Toll-like receptor (TLR7) which signals via the adaptor molecule MyD88 [19,32] whereas human CD141+ DC (mDC2) mainly use TLR3 and TRIF signaling [17,18]. To define the immune sensing pathways of the murine DC counterparts, DCs from different knock out animals harboring genetic lesions in the type I IFN receptor (IFNAR), type III IFN receptor (IL28R) or the adaptor molecules CARDIF (also known as MAVS, IPS-1, VISA), MyD88 or TRIF were analyzed. As observed before, DC generated from wild type animals supported a significant amount of type I IFN production upon coculture with Jc1 transfected cells (Fig 3A). Genetic lesions in IL28R (Fig 3C), CARDIF (Fig 3D) and MyD88 (Fig 3E) had no influence on type I IFN production in these coculture assays. However, the lack of IFNAR (Fig 3B) or TRIF (Fig 3F) abrogated type I IFN production, indicating that the type I IFN feedback loop as well as TLR3-dependent signaling are important for the IFN production by Flt3-L derived DC after coculture with HCV transfected cells. To explore which cell subset within the Flt3-L DC culture is able to sense and respond to HCV replication additional surface marker were included for subsequent FACS analysis (Fig 4A–4C). Cells were stained with antibodies targeting SiglecH and CD11c to detect pDC, SIRPα and CD11b to detect CD11b-like DC or Clec9A and CD24 to analyze CD8α-like DC. Gating on the respective cell populations revealed that all three DC subsets, pDC, CD11b-like DC and CD8α-like DC, significantly upregulated the activation marker CD69 after coculture with HCV transfected cells (Fig 4D) indicating an involvement of all three DC subsets. To further define the role of each subset, cells were FACS sorted into pDC, CD11b-like DC, and CD8α-like DC and independently cocultured with HCV transfected cells. As observed before, cocultivation of the complete Flt3-L derived DC population with HCV transfected cells led to a significant type I IFN production (Fig 5A). Separation of the DC subsets revealed that pDC (Fig 5B) and CD11b-like DC (Fig 5C) were not able to respond to HCV replicating cells. However, CD8α-like DC cocultured with HCV transfected cells produced significant amounts of type I IFN (Fig 5D). In line with this finding, depletion of pDC (non pDC, Fig 5E) or CD11b-like DC (non CD11b-like DC, Fig 5F) had no impact on the IFN production after coculture, whereas depletion of CD8α-like DC abrogated type I IFN production (non CD8α-like DC, Fig 5G) indicating that the CD8α-like DC are the main IFN producing cells within the Flt3-L culture after co-culture with HCV transfected cells. Consistent with previous observations, stimulation with infectious VSV-M2 triggered mainly the pDC to produce large amounts of IFN-α (Fig 5B), confirming the role of pDC as the main IFN producing cells after viral infection. Since human pDC have been described to induce a robust type I IFN production after coculture with HCV transfected cells, we analyzed whether the xeno-situation, namely adding murine DC to human liver cells, might be responsible for this lack of response. To this end, murine liver cells harboring a genetic lesions in MAVS (also known as CARDIF, IPS-1, VISA) and containing all minimal factors required for HCV replication and infection (MLT-MAVS−/−miR-122/mmmmm [33]), were transfected with HCV SGR RNA and cocultured either with the mixed Flt3-L culture or purified murine Flt3-L derived DC subsets. To increase viral replication in these cells, the HCV subgenomic replicon SGR2 (pFK; i341NeoEI-NS3-NS5B/JFH1) displaying a more robust replication, was utilized in addition. As seen before, the complete Flt3-L DC culture produced type I IFN after co-culture with either human or murine HCV transfected cells (S2A Fig and S2E Fig, respectively). Among the purified subpopulations, only CD8α-like DC were able to respond to HCV replication in human or murine cells (S2D Fig and S2H Fig, respectively), whereas all other cell types did not secrete elevated levels of IFN upon coculture with these HCV replicating cells (S2B and S2C Fig and S2F and S2G Fig, respectively). These experiments indicated that murine Flt3-L derived pDC are not able to sense HCV replication. In summary, murine CD8α-like DC, but not murine pDC or murine CD11b-like DC, sense HCV transfected cells and produce type I IFN. Since Flt3-L derived DC only resemble physiological DC subsets but do display some differences to in vivo derived DC, we analyzed whether HCV replication can be sensed and IFN be induced also in vivo. To this end, IFN-β+/Δβluc reporter mice, encoding a luciferase gene under the control of the IFN-β promotor [34] were utilized. Mice were subcutaneously (s. c) injected with 5x106 SGR2 transfected Huh7. 5 cells (right flank) or 5x106 mock transfected Huh7. 5 cells (left flank) and the luciferase activity was measured at different time points post injection. The SGR2 transfected cells induced a significant luciferase signal and therefore IFN-β promotor activity 4 h post injection which decreased moderately at later time points (Fig 6A and 6B). Importantly, the IFN stimulation was dependent on active viral replication, as injection of cells harboring a replication incompetent HCV subgenomic replicon mutant (pUCΔGDD) or treatment with an HCV protease inhibitor abrogated IFN-β promotor activity (Fig 6C). Taken together, adoptive transfer of HCV transfected cells stimulates IFN-β production in replication-dependent manner in vivo. Given the subcutaneous route of injection it is possible that not DC but rather Langerhans cells or tissue-resident macrophages are responsible for the observed IFN induction in vivo. To address whether murine macrophages are able to sense HCV replication and produce type I IFN we isolated bone-marrow derived macrophages and cocultured them with HCV replicon or virus-transfected hepatoma cells. Coculture did not lead to an IFN-α or IFN-λ production by macrophages and only a moderate IFN-β production (S3 Fig). Most importantly we could show that this IFN-β production was also dependent on TRIF signaling as TRIF knockout rendered the cells unable to respond to HCV coculture (S3 Fig). This demonstrates that HCV replication can be sensed also by other cell types in vitro in a TRIF dependent manner, however that Flt3-L derived DC are the main producers of type I IFN. In this study, we describe basic principles of murine DC activation by HCV replicating cells. In the majority of cases, HCV exposure results in chronic infection with a high risk to develop severe liver diseases like cirrhosis, fibrosis and hepatocellular carcinoma [2]. Despite stimulation of early immune responses within the liver as well as high IFN-stimulated gene expression the virus is able to persist in its host [10,35]. Dendritic cells act as sentinels of the immune system and constitute a first line of defence against invading pathogens. Although human hepatocytes have been reported to produce type I and type III IFNs in response to HCV infection [36,37], the virus has evolved evasion mechanisms including the cleavage of the adaptor molecules MAVS and TRIF, thus preventing signal transduction and IFN induction in these cells [38–40]. Therefore, dendritic cells that are not infected by HCV constitute an important defence mechanism to establish an antiviral status. Intrahepatic studies of dendritic cell interactions are technically and practically challenging and the isolation of certain dendritic cell subsets from human peripheral blood mononuclear cells (PBMC) can be difficult. So far only few studies have been conducted using murine derived dendritic cells mainly due to the lack of a suitable immunocompetent murine animal model, which supports HCV replication in vivo [41,42]. Given the fact, that murine Flt3-L derived dendritic cells comprise subsets that are equivalent to physiological DC subsets found in humans, pDC, CD1c+ DC (with the murine counterparts CD11b-like DC) and CD141+ DC (with the murine equivalents of CD8α-like DC) [23,43], we used this system to study virus host interactions. Interestingly, only Flt3-L but not GM-CSF derived DC were able to produce type I and type III IFN after co-culture with HCV transfected cells. Similar to what has been described for human pDC and CD141+ DC [17,19,31], murine IFN-α production was dependent on active viral replication and endocytosis, as co-culture with a replication incompetent virus mutant (ΔGDD) or treatment with acidification inhibitors prevented type I IFN production of DC. Interestingly, cell-free virus as well as purified supernatant harvested from HCV SGR transfected cells was able to stimulate IFN production in vitro, indicating that viral signatures are sensed independent of direct cell-to-cell contact and independent of infectious virus production. A recent study suggested that exosomes are released from HCV infected cells that stimulate human pDC to produce IFN-α [44]. Other studies confirmed that SGR RNA can be transferred from one cell to another via exosomes resulting in a productive replication in susceptible cells [45–47]. Indeed, after stimulation of DC with preparations of extracellular vesicles we could observe type I IFN induction. Therefore, our data are consistent with these previous reports and indicate that exosomes and not the virus particle itself deliver signals to murine immune cells and thus trigger IFN production. However, further studies are needed to reveal the detailed mechanism and the features of these exosome-like vesicles. Sensing of HCV-transfected cells was dependent on TRIF and IFNAR signalling and sorting of individual DC subtypes showed that the CD8α-like DC, which are the murine counterpart of human CD141+ DC, are the main IFN producers in the murine culture after stimulation with HCV-transfected cells. Interestingly, there are striking parallels between the human and the murine system with the human CD141+ DC being able to sense HCV replication in a TRIF dependent manner, to produce IFN-λ and to amplify IFN-α responses [17,18]. Furthermore, ex vivo isolated murine CD8+ DC have been described to produce type III IFN in an IPS-1 dependent manner upon stimulation with HCV [42], further supporting our data. However, unlike human pDC [19,20], isolated murine pDC were not able to sense HCV replication. The reason for this remains elusive and needs to be further studied. One possible explanation is that pDC, in contrast to the other DC subtypes, do not express TLR3 [48]. However, similar to human pDC, recognition of HCV RNA could still be mediated via TLR7 in these cells [19]. Since co-culture with both human and murine HCV replicating liver cells did not stimulate murine pDC it is unlikely that species-specific recognition mechanisms are responsible for the inability of murine pDC to recognize HCV replicating cells. FACS analyses revealed that in the whole Flt3-L DC culture all cell types were activated after coculture with HCV transfected cells. This is supported by our finding that all cell types (including the pDC cell population), upregulated the activation marker CD69. However, if all cell types contributed to IFN production remains to be determined since the mode of activation may differ between cell subsets. Further experiments are needed to define the contribution of each cell type. Nevertheless, this interplay of cell subsets could resemble the physiological situation within the liver of the infected hosts, where several immune cells are likely to communicate to establish an antiviral state. Taken together, we have demonstrated the usage of murine BM-derived dendritic cells, which can be easily generated to study virus host interactions and to elucidate signalling pathways which are important for the establishment of an antiviral state during HCV infection. With the lack of an immune competent mouse model to study immune responses to HCV, alternative methods are needed to further elucidate virus-host interactions. We also show that there are conserved functions and similarities between the murine DC and their human counterparts, with respect to HCV sensing, but we could also show differences between both systems. Therefore this model should be useful to further define virus host interactions during HCV infection and to dissect basic principles of intercellular communication between infected cells and DC subpopulations. All mice were bred under specific-pathogen-free conditions at the mouse facility of the Helmholtz Centre for Infection Research, Braunschweig, Germany, or at the Twincore, Centre of Experimental and Clinical Infection Research, Hannover, Germany. All animal experiments were performed in compliance with the German animal protection law (TierSchG BGBI S. 1105; 25. 05. 1998) and were approved by the responsible state office (Lower Saxony State Office of Consumer Protection and Food Safety) under the permit number 12/1025. The mice were handled in accordance with good animal practice as defined by the Federation for Laboratory Animal Science Associations (FELASA). C57BL/6 mice were purchased from Harlan Winkelmann. IFNAR-/- mice [49], IL28R-/- mice [50], MyD88-/- mice [51], CARDIF-/- mice [52,53], TRIF-/- mice [54] and IFN-β+/Δβluc reporter mice, encoding a luciferase reporter gene under the control of the IFN-β promoter [34] have been described recently. MLT-MAVS−/−miR-122/mmmmm cells have been described recently [33]. Please note that MAVS and CARDIF are referring to the same gene. Huh-7. 5 cells were kindly provided by Charles Rice, Rockefeller University, New York, USA [55]. Huh-7. 5 cells and MLT-MAVS−/−miR-122/mmmmm cells were cultured in Dulbecco’s modified minimal essential medium (DMEM, Life Technologies) supplemented with 2 mM L-glutamine (Invitrogen), non-essential amino acids (Invitrogen), 100 μg/mL streptomycin (Invitrogen), 100 IU/mL penicillin (Invitrogen), and 10% fetal bovine serum (DMEM complete) at 37°C and 5% CO2. Femur and tibia of mice were flushed with DC medium (RPMI medium supplemented with 10% FCS, 10 mM HEPES (Gibco), 1 mM sodium pyruvate (Gibco), 2 mM Glutamax (Gibco), 100 U/mL penicillin (Gibco), 10 mg/mL streptomycin (Gibco), and 0. 1 mL β-mercaptoethanol (Sigma) ) to isolate bone marrow (BM) cells. Red blood cell lysis was performed using red blood cell lysing buffer (Sigma). Cells were washed, seeded at appropriate densities and incubated for 8 days in the case of dendritic cells and for six days in case of macrophages. For Flt3-L derived DC cultures, cells were seeded at a density of 2x106 cells/mL in DC medium supplemented with 100 ng/mL Flt3-L (R&DSystems) and incubated at 37°C and 5% CO2. A medium change was performed at day 4 by replacing two-thirds of the cell culture volume with fresh medium supplemented with Flt3-L. To obtain GM-CSF derived DC cultures, BM cells were seeded at a density of 1x106 cells/mL in DC medium supplemented with 100 ng/mL granulocyte-macrophage colony-stimulating factor (GM-CSF) (R&D Systems), and the medium was changed at days 4 and 6. For M-CSF derived macrophages cultures, cells were seeded at a density of 5x105 cells/mL in 100 μl/mL LCCM (L929 cell conditioned medium) supernatant. A medium change was performed at day 3 by replacing half of the cell culture volume with fresh medium supplemented with LCCM supernatant as described before [56]. The plasmids pFK-Jc1 (Jc1) and pFK; i389-LucNS3-3’ (SGR/SGR1) have been described recently [57,58]. The plasmid pFK-Jc1ΔGDD (ΔGDD), which has an in-frame deletion of 10 amino acids spanning the GDD motif in the NS5B polymerase catalytic domain, has been constructed by cloning the deletion into the pFK-Jc1 plasmid. The cloning strategy is available upon request. The plasmids pFK; i341NeoEI-NS3-NS5B/JFH1 (SGR2) and pUC-pSGR-JFH1/_GDD (pUCΔGDD) have been described elsewhere [59,60]. For preparation of infectious HCV or subgenomic replicon (SGR) supernatant stocks, Huh7. 5 cells were transfected with 10 μg in vitro transcribed viral RNA or no RNA (Mock) using electroporation [57]. After 48 h and 72 h incubation, cell-free supernatant (SN) was filtered through 0. 45 μm filters, ultrafiltrated through Amicon centrifugal filters (Millipore), and 100x concentrated. For the production of HCV subgenomic replicon (SGR) SN and Mock SN, medium was changed after 4 h to serum-free AEM medium and the supernatant harvested and concentrated as described above. SGR SN contained the following HCV genome equivalents: 1–9 x 1010 copies/mL as determined by quantitative real time PCR (qRT-PCR) as described before [61]. For extracellular vesicle isolation, Mock, SGR or pUCΔGDD supernatant was harvested as described above. 1 mL concentrated supernatant was incubated with 250μL ExoQuick-TC solution (SBI) overnight at 4°C and extracellular vesicles recovered according to the manufacturer’s instructions. Vesicle isolation was controlled via western blot according to standard procedures using the following antibodies: anti-Hsp70 (SBI), anti-Annexin II (clone 5, BD), anti-CD81 (JS-81, BD), anti-CD63 (MEM-259, Biolegend) and anti-actin (MAB 1501R, Chemicon) and Horseradish-peroxidase–conjugated secondary anti-rabbit or anti-mouse antibodies (Amersham; SBI). Vesicles contained the following HCV genome equivalents: ΔGDD [2–8 x108 copies/mL], SGR [4–10 x109 copies/mL] as determined by qRT-PCR as described before [61]. For RNAse and DNAse treatments of the coculture, transfected Huh7. 5 cells were washed and Flt3-L derived DC were added immediately before 0. 5 μg/mL RNAse (Roche) or 1 unit DNAse (Roche) were added and the cells were cocultured for 18 h. For bafilomycin A1 (Bafilo; Sigma) and concanamycin A (ConA; Sigma) treatments, transfected Huh7. 5 cells were washed, Flt3-L DC were added and the coculture was treated with 5 nM ConA or 25 nM Bafilo for 18 h. Huh7. 5 cells were transfected with HCV RNA constructs, seeded in 96 well plates at a density of 1x105 cells/200 μl, and incubated for 72 h. After 72 h, the cells were washed with 1xPBS before 2x105 cells/200μl DC were added and co-incubated with the cells for 18 h. For stimulation experiments 2x105 cells/200μl DC cultures were seeded in 96 well plates and stimulated with VSV-M2 at a MOI 1, HCV full length virus (Jc1) at a MOI 10 or with 5 μl of concentrated supernatant or isolated extracellular vesicles for 18 h. The cell-free supernatant was harvested and analysed by ELISA and the cells were stained for fluorescent activated cell sorting (FACS) analysis. For FACS analysis as well as FACS sorting of single-cell suspensions, cells were first incubated with CD16/CD32-specific antibody (2. 4G2; BD) for 10 min to block non-specific Fc-receptor interactions. A live death staining was performed using zombie aqua (eBioscience) according to the manufacturer’s instructions. To differentiate between DC subsets, cells were stained for 15 min at 4°C with combinations of antibodies specifically binding to CD11b (M1/70. 15; Invitrogen), CD11c (HL3; BD), Siglec-H (eBio440c; eBioscience), CD69 (H1. 2F3; BD), Clec9a (42D2; eBioscience), CD172 (SIRPα) (P84; Biolegend) and CD24 (M1/69; Biolegend). The cells were subsequently washed with 1 mL FACS buffer (PBS, 1% FCS) and then re-suspended in FACS buffer supplemented with 3% paraformaldehyde (PFA). Samples were measured using a FACS LSR II, and data were analyzed with FlowJo 7. 6. 5 software (TreeStar). Dead cells as well as cell duplets were excluded prior to subsequent analysis. DC subsets were sorted using the FACS Aria or FACS XDP and sorting efficiency ranged between 71. 9–86. 2% for the pDC, 81. 4–94. 4% for the CD11b-like DC and 44. 54–83. 2% for the CD8α-like DC. For the determination of IFN-α, IFN-β and IFN-λ levels in cell-free supernatants, enzyme-linked immunosorbent assay (ELISA) methods were applied (eBioscience, PBL Biomedical Laboratories) following the manufacturer’s instructions. Supernatants were harvested and stored at -20°C until they were tested for IFN. Background lines (dashed lines) depict the lowest standard detected of IFN-α (31. 25 pg/mL), IFN-β (15. 6 pg/mL) and IFN-λ (62. 5 pg/mL). Huh. 7. 5 cells were either mock transfected or transfected with the plasmids pFK; i341NeoEI-NS3-NS5B/JFH1 (SGR2) or pUC-pSGR-JFH1/_GDD (pUCΔGDD). For inhibitor treatments, 4 h after transfection a medium change was performed and 1 μM Telaprevir (Selleckchem) was added to the cells. After 72 h, the cells were harvested, washed once with PBS and resuspended as 5x107 cells/mL in PBS. IFN-β+/Δβluc reporter mice were anesthetized with isoflurane, and 5x106 cells were subcutaneous (s. c.) injected into the flanks of the mice. 100 μl luciferin (PerkinElmar) (30 mg/mL in PBS) /20 g mouse weight was intravenously (i. v.) injected at the indicated time points and the mice were immediately analyzed using the IVIS Spectrum CT (PerkinElmer). The acquired images were analyzed and quantified using the Living Image 4. 3. 1 software (PerkinElmar). Statistical analyses were done using GraphPad Prism 5. The one-tailed non parametric Mann-Whitney or 2-way ANOVA tests were used for the analysis of significant differences between groups with unmatched pair values. Error bars in graphs indicate standard deviation.

HCV is an RNA virus that, following exposure, in most cases establishes chronic infection. The virus has evolved numerous immune evasion strategies, including direct interference with interferon production. Nevertheless, HCV infection activates interferon-stimulated genes in the liver, implying that non-infected cells secrete IFN. Several DC subsets have been implicated in HCV sensing and production of IFN; however, the molecular mechanism resulting in HCV sensing is poorly understood. Using murine bone marrow derived DC, we dissected basic principles of HCV innate immune recognition and activation of dendritic cells. We show that HCV recognition by murine DCs depends on TRIF and IFN receptor signalling. This indicated the involvement of TLR3 and of the IFN receptor dependent amplification loop. Infectious virus production is dispensable since cells carrying subgenomic HCV replicons are also recognized. Moreover, specific DC subtypes, i. e. CD8α-like DC, are responsible for recognition of HCV. These findings highlight that specific murine DC subpopulations are uniquely capable of recognizing HCV replicating cells independent of infectious virus production. These observations open novel opportunities to explore the mechanisms of inter-cellular communication that mediate activation and IFN production of non-infected immune cells and to dissect the role of DC subsets in immune control.

Abstract Introduction Results Discussion Materials and Methods

medicine and health sciences immune cells vesicles pathology and laboratory medicine enzyme-linked immunoassays hepacivirus antigen-presenting cells pathogens immunology biological cultures microbiology hepatoma cells viruses animal models dendritic cells model organisms rna viruses cell cultures immunologic techniques cellular structures and organelles research and analysis methods animal cells proteins medical microbiology microbial pathogens immunoassays hepatitis c virus hepatitis viruses mouse models viral replication biochemistry cell biology flaviviruses virology viral pathogens interferons biology and life sciences cellular types cultured tumor cells organisms

2016

Hepatitis C Virus Stimulates Murine CD8α-Like Dendritic Cells to Produce Type I Interferon in a TRIF-Dependent Manner

The responses of neurons in sensory cortex depend on the summation of excitatory and inhibitory synaptic inputs. How the excitatory and inhibitory inputs scale with stimulus depends on the network architecture, which ranges from the lateral inhibitory configuration where excitatory inputs are more narrowly tuned than inhibitory inputs, to the co-tuned configuration where both are tuned equally. The underlying circuitry that gives rise to lateral inhibition and co-tuning is yet unclear. Using large-scale network simulations with experimentally determined connectivity patterns and simulations with rate models, we show that the spatial extent of the input determined the configuration: there was a smooth transition from lateral inhibition with narrow input to co-tuning with broad input. The transition from lateral inhibition to co-tuning was accompanied by shifts in overall gain (reduced), output firing pattern (from tonic to phasic) and rate-level functions (from non-monotonic to monotonically increasing). The results suggest that a single cortical network architecture could account for the extended range of experimentally observed response types between the extremes of lateral inhibitory versus co-tuned configurations. The firing properties and receptive fields of neurons in sensory cortex are heterogeneous and can vary both quantitatively and qualitatively with changing stimuli. The diverse responses are well exemplified in the primary auditory cortex (A1), where firing ranges from phasic (only at stimulus onset) to tonic (for the duration of a stimulus [1], [2], changes monotonically or non-monotonically with intensity [3], [4], and is often evoked selectively with complex stimuli [5]. Many receptive field properties are not simply inherited from presynaptic input from the thalamus but are shaped by interaction of local excitatory and inhibitory neurons in cortical circuits [6], [7], [8]. The processes governing these interactions are under some debate but are postulated to depend on the network architecture, which may range from the lateral inhibitory network configuration where excitatory inputs are more narrowly tuned than inhibitory inputs, to the co-tuned configuration where both are tuned equally. A1, because of its tonotopic organization, is an ideal system for examining how sensory stimuli are represented in the temporal and spatial interaction of principal cells and interneurons [9], [10]; c. f. [11]. Intracellular recordings in vivo have begun to explore mechanisms underlying the diversity of neuronal receptive field properties (for review, see [12]). Though some studies indicate that evoked excitatory and inhibitory conductances are co-tuned [13], [14], others using very similar conditions have found that co-tuning is only approximate or that there is significant lateral inhibition [15], [16], and that the balance may shift during postnatal development [16], [17]. Moreover, many of the response properties such as two-tone suppression and intensity tuning are more consistent with some degree of lateral inhibition [4], [5], [8], [12], [18], [19]. There is a similar debate in the visual system as to the extent to which lateral inhibition in cortex underlies extrareceptive field properties [20], [21]. The cortical circuitry and synaptic properties that underlie co-tuning and lateral inhibition are poorly understood. Lateral inhibition could result from greater sensitivity of inhibitory cells to input, greater convergence of presynaptic (e. g. , thalamic) input onto inhibitory versus excitatory cells, or broader spread of intracortical inhibition versus excitation. Canonical circuits typically consist of excitatory pyramidal (P) cells and fast-spiking (FS) interneurons. In layers 2/3 and 4, both P and FS cells receive direct thalamic inputs [22], [23], [24]. FS cells are distinguished from other interneurons by non-adapting high frequency firing [25], morphology [26], [27], [28], and synaptic dynamics [29], [30], suggesting a distinct functional role in sensory information processing. Here, we performed simulations in a large network model and with rate models to determine how the excitatory and inhibitory drive to P cells changes with stimulus. The patterns of connections and synaptic properties between excitatory P and FS inhibitory cells were based on experimental data. The simulations indicated that the same network generated both lateral inhibition and co-tuning: shifting between them is accomplished simply by changing the spatial distribution of the thalamic input. Therefore, a single, hardwired network potentially is consistent with many of the diverse response patterns previously reported in vivo. The broad goal was to build a detailed model of the pattern of stimulus-driven cortical activity. For simplicity, the parameters of the model were taken from published [31], [32], [33] and unpublished (Levy and Reyes, in preparation) experimental data. We stress, however, that the main result - that lateral inhibition and co-tuning coexist in the same network - is robust and does not depend on details of the network connectivity, provided that the inhibition is strong enough (as shown below). Neurons in sensory cortex are connected to their neighbors with a relatively low probability. The connection rates decrease with distance between cells, as the local axonal and dendritic arborizations are confined within several hundred µm of the soma. In cortical layers 2/3 and 4, the thalamorecipient layers of auditory cortex, our experimental data on connection probability versus distance between somata for each connection type (P→P, P→FS, FS→P) were fitted with a Gaussian function (Fig. 1A, right; Methods, equation 1), which was chosen for computational efficiency and to put our findings in the context of the theoretical literature (c. f. [34]). The radial spread of connectivity (σ), i. e. , the euclidean distance between cell bodies in the plane of the slice, was 145 µm (P→P), 92 µm (P→FS, FS→P) with peaks of 0. 10,0. 30,0. 39, respectively (Fig. 1A, right). Using these connection profiles, we constructed a network sheet of 160×160 P and 32×32 FS cells. The synaptic and intrinsic membrane properties were also based on experimental data [31], [32], [33], [35]. Measured values for synapse strength (peak amplitude of excitatory and inhibitory postsynaptic potentials (EPSPs/IPSPs) and short-term plasticity were used as parameter values for the phenomenological model of short-term synaptic plasticity in the simulations (Table S2; [36], [37]). The synaptic pathways examined here exhibit strong short-term depression (governed by U and τrec, Supporting Text S1 and Table S2). Unitary synapse strength, unlike connection probability, was not correlated with distance between cells (c. f. [32], [38]). Therefore the unitary synaptic amplitudes (A, Table S2) for each cell type applied at all distances. The intrinsic membrane properties of P and FS cells used in the simulations are listed in Table S1. We used the adaptive exponential integrate and fire model with variables adjusted appropriately to reproduce the P and FS cell firing behaviors (Fig. 1C; [39]; see Methods). The external input to the network was from the auditory thalamus, i. e. , the ventral division of the medial geniculate body (MGv). For simplicity, each input was modeled as a sequence of spikes arriving at a specified frequency (train duration = 100 ms; Fig. 1B). The firing of the thalamic neurons was phasic-tonic (Fig. 2B, bottom; see Methods) as was observed in intact animals [6] though other patterns produced qualitatively similar results (data not shown). The firing rate of thalamic neurons and the number of thalamic cells synapsing onto each cortical cell in the network were Gaussian distributed in space (parameterized by the maximum number of inputs Nmax each neuron can receive and by the standard deviation σ (in µm; see Methods). Cells in the center of the network received the most inputs (Nmax = 50–150); i. e. , the thalamic afferents were maximally active here because the stimulus was assumed to represent the preferred frequency for the center cell. The amplitudes, time course, and short-term dynamics of thalamic inputs were adjusted separately for the P and FS populations (Fig. 1C; Table S2), based on experimental data from auditory cortex [8], [40]; Schiff and Reyes, unpublished data). The thalamic inputs to both P and FS cells exhibited short-term depression (Table S2). Several salient features emerge from the simulations. In general, the P and FS cells tended to fire most at the onset of the stimulus (Fig. 2A, top, middle; rasters are compiled from 50 sweeps and arranged according to neurons' radial distance from the center of the 2D network). As observed in awake animals, the P cell firing was more transient than that of thalamic inputs (Fig. 2B; [6]). The transient nature of the P and FS cell firing was due to depressing thalamic inputs (Fig. 1C). The firing pattern of P cells was further shaped by the spatiotemporally complex combination of excitatory input from the thalamus, inhibitory input from FS cells, and excitatory input from neighboring P cells. The normalized contour plot of the synaptic conductances (Fig. 2C, top) shows that the spatial distribution of input to the P cells is variable: the thalamic input (gray) was restricted to the neurons near the center, while the inhibitory inputs from FS cells (red) and recurrent excitatory inputs (cyan) appeared later and were broader. The timing and relative magnitude of the synaptic components can be seen by focusing on the inputs to the P neuron at the center of the network (* in Fig. 2C, top & middle). A slice through the center of the contour plot shows that the (non-normalized) thalamic input arrives first, causing the initial firing in both the P (Fig. 2A, top) and FS (Fig. 2A, middle) cell populations. After a short delay, the inhibitory input (red, Fig. 2C, middle) from the FS cells appeared followed by the excitatory input from other P cells (cyan). Inhibition was transient due to a combination of transiently firing FS cells (Fig. 2A, bottom) and depressing IPSPs [33]. The recurrent excitatory input from neighboring P cells (cyan) was considerably weaker than the thalamic input (Fig. 2C, middle, cyan vs. gray) largely due to the small unitary EPSPs [31] and low probability of connection between P cells [31], [35]. Within approximately 50 ms after firing onset, only the thalamic inputs remained. Each synaptic component increased with the number, Nmax, of thalamic inputs (Fig. 2E). The excitatory input from the thalamus (gray) grew nearly linearly from the origin whereas both the inhibitory input from the FS cells and recurrent excitatory input from neighboring P cells appeared only when Nmax was sufficiently large to cause the neurons to fire (Nmax ∼20). The FS inhibitory component rose steeply owing to the FS cells' high frequency firing responses to input. The magnitude of the intracortical synaptic inputs to the P cells also depended on the spatial extent of the thalamic drive. With Nmax constant, widening the thalamic input from σ = 40 µm (Fig. 2C) to σ = 110 µm (D, middle) activated more FS cells with the result that the inhibition to P cells population increased (compare to Fig. 2C, middle). As σ increased, FS inhibition (Fig. 2F, red) was initially smaller than but then exceeded thalamic excitation (gray). The P cells still fired, albeit more transiently, because the inhibition was delayed with respect to excitation (data not shown). There was only a modest increase in the recurrent excitatory input (cyan). For the remainder, analyses will be restricted to the first 50 ms after the stimulus onset. The firing rates and the changes in the conductances were greatest in this interval. In addition, in vivo recordings from auditory cortex using brief tone pips show that neuronal firing is dominated by thalamic drive: contributions from reverberatory network activity and other cortical areas are substantially smaller [41]. Finally, inputs from low threshold spiking (LTS) interneurons, another major class of inhibitory cells, are unlikely to contribute significantly to firing. Experiments suggest and simulations confirm (not shown) that the weak, facilitating synaptic drive to LTS from the thalamus and local P neurons do not increase substantially during brief stimuli to affect overall network activity [29], [30], [42], [43], [44]). As mentioned above, the spatial profiles of inhibitory and excitatory inputs to P cells in the co-tuned network are comparable whereas in the case of lateral inhibition, the inhibitory input is broader. Lateral inhibition and co-tuning could represent separate circuits that differ in e. g. , the axonal spread of their associated inhibitory neurons and/or the degree of convergence of inputs from the thalamus. The following simulations suggest, however, that the same circuit can produce both configurations, depending only on the spatial width of the thalamic inputs. Lateral inhibition predominated when the spatial distribution of thalamic input was narrow. For σ = 40 µm (Fig. 3A, top), both the inhibitory (red) and recurrent excitatory (cyan) inputs were broader than the thalamic input. When combined and normalized (3A, bottom), the excitatory inputs (thalamic + recurrent, black) are substantially narrower than the inhibitory inputs (red), consistent with lateral inhibition. The recurrent P input (Fig. 3A, cyan) was small and contributed only to the foot of the total excitatory distribution. In the same network, co-tuning was generated when the thalamic input was broad. When σ was increased to 110 µm, the distribution of both inhibitory and recurrent excitatory inputs widened (Fig. 3B, top). However, the rate of widening was proportionately less than the change in thalamic spread (see below for mechanism). As a result, the width of inhibitory profile became nearly equal to that of the composite excitatory width (Fig. 3B, bottom), consistent with the co-tuned configuration. A plot of the ratio of the inhibitory to excitatory profile widths (Winh/Wexc, measured at half the peak; circles) shows that the network shifted progressively from the lateral inhibitory configuration (Winh/Wexc >1) to the co-tuned configuration (Winh/Wexc ∼1) (Fig. 3C). These configurations generally changed very little with Nmax (Fig. 3D), except for small values where inhibition was not fully activated. When the network was co-tuned, the balance between excitatory and inhibitory inputs was maintained for neurons at different distances from the center of the network where the thalamic input peaked. The proportions of synaptic conductances of cells in the periphery (Fig. 2D, bottom) were similar to those of cells in the center (middle). When lateral inhibition was predominant, the excitatory-inhibitory balance shifted so that inhibition dominated in cells at the periphery (Fig 2C, compare bottom to middle); for the most distant neurons, only inhibition was present. The input dependent transition between lateral inhibition and co-tuning was robust under in vivo -like conditions. To simulate background synaptic barrages, white noise current was injected into each P and FS neuron to produce membrane fluctuations with a standard deviation of up to +/− 8mV, similar to what has been observed in vivo [45], [46]. The transition between lateral inhibition and co-tuning still occurred, though the region of lateral inhibition was increased slightly (Fig. S1). Background firing would also cause tonic depression of the synaptic potentials, the degree of which differs between thalP, thalFS, PP, PFS, and FSP connections. However, performing the simulations with the calculated steady-state values for the depression at different background firing rates [36], [47] preserved the transition (Fig. S2). The transition also did not depend on the exact details of the connectivity scheme, because simulations using a rate-based model (see below) with a number of non-Gaussian connectivity profiles gave comparable results (Fig. S6). It should be noted that inhibition must be stronger than the recurrent excitation, as is the case for auditory cortex [33]. The recurrent excitatory input has a spatial profile that is similar in width to that of the inhibitory input (Fig. 3A). Hence, if the amplitudes are comparable, excitation and inhibition cancel in the periphery and lateral inhibition is not possible (data not shown). Missing from the simulations are the inhibitory connections and electrical coupling between FS cells, both of which have not yet been fully characterized. Mutual inhibition among FS cells would be expected to reduce their firing and hence decrease the net inhibitory input to P cells. To mimic these effects, simulations were performed where the threshold of the FS cells was set at −37 mV, which is 10 mV above the control (Fig. S3). Though firing was reduced by 50%, the shift between lateral inhibition and co-tuning still occurred. The electric coupling between FS cells is likely to have complex effects on the timing of action potentials [48], [49], [50] and cannot be readily predicted without more information about the patterns of connections and coupling strengths. Extensive analyses and simulations are needed to fully characterize the effects, which are beyond the scope of this paper. Firing, quantified as the average counts in a 50 ms interval, was greatest in the lateral inhibitory configuration (Fig. 4A, left). The spatially narrow thalamic input (σ = 40 µm) recruited few FS cells such that the net inhibition was small (Fig. 2C). As the input width increased to approach the co-tuned state (σσ = 110 µm), the firing decreased (Fig. 4A, left) due to increased inhibition (Fig. 2D, F). Peak firing, which corresponds to that of the center cell, decreased systematically with σσ (Fig. 4A, right). The firing sensitivity of neurons to input was correlated with the extent of lateral inhibition. Increasing the number, Nmax, of thalamic inputs from 60 to 100 evoked more firing, with a modest change in the overall width of the profiles (Fig. 4B, left). To a first approximation, firing increased nearly linearly with Nmax (Fig. 4B, right). The slope of the relation was steep when σσ was small (40 µm) but became shallower with increasing σσ (75,110 µm). Note that the slope change was not accompanied by horizontal shifts in the curves, consistent with a pure divisive gain change [51], [52]. This modulation of sensitivity by σσ is a novel form of gain control. To examine the interaction of multiple inputs, two Gaussian stimuli (S1 (x), S2 (x) ) separated by different distances (Δx, Fig. 4C left) were delivered to the network. This simulates e. g. , two-tone stimuli, which produces side-band suppression of firing rate [18], [19]. The positions of the Gaussians represent the tone frequencies along the tonotopic axis [9], [10], [53]. In the lateral inhibitory configuration (middle), the evoked firing was greatest at Δx = 0 and decreased with Δx to a level below that evoked with a single stimulus (dashed line). When the network was co-tuned, the firing decreased monotonically with Δx, eventually converging with single stimulus firing (right). The mechanism underlying the shift between co-tuning and lateral inhibition can be understood by using a reduced model to examine the shifts in excitatory and inhibitory balance. The PFS connections and the weak recurrent excitatory inputs from local P cells were omitted. Under these conditions, the network reduced to a feedforward network where the thalamic afferents synapsed onto both excitatory and inhibitory cells, with the inhibitory cells in turn synapsing onto the excitatory cells (Fig 5A). For the following, the presynaptic thalamic synaptic current Ithal was Gaussian distributed in space and was parameterized by the peak current, Imax, and spatial spread σ (Fig 5B) as in the above simulations. The spatial profile of the excitatory inputs (Iexc) to both excitatory and inhibitory cells was inherited directly from, and was therefore identical to, Ithal (Fig. 5B). On the other hand, generation of the inhibitory input profile (Iinh) to excitatory cells involved several steps since the inhibitory cells must be activated. First, Ithal was transformed with a threshold-linear function (Methods, equation 6; analogous to firing rate – current relations) to obtain the spatial profile of inhibitory cell firing rate, Finh, . The presence of the threshold precluded recruitment of weakly driven inhibitory cells far from the center so that Finh, . was narrower than Iexc (the so-called ‘iceberg’ effect, c. f. [13]). Second, to account for the axonal spread of inhibitory cells to neighboring excitatory cells [54], Finh was convolved with Pinh, the connection probability profile between inhibitory and excitatory cells (Fig. 1A, right). Finally, multiplying by a constant that has units of nA/Hz gave Iinh. The network shifted between co-tuning and lateral inhibition as the spatial profile (Nmax and σ) of Ithal changed. Plotting the ratio of the widths of Iexc and Iinh (winh/wexc) as a function of Nmax (normalized by rheobase current) and σ (normalized by the standard deviation, σ inh, of Pinh) revealed the regimes (Fig. 5C) observed in the full network (Fig. 2C, D). When σ was small, winh was broader than wexc, indicating the lateral inhibitory configuration. As σ increased, winh/wexc reached a regime where Iexc and Iinh were perfectly co-tuned (intersection of the surface with unitary plane) followed by a regime where Iexc was slightly wider than Iinh. With increasing Nmax (here, Nmax is the maximum synaptic current from the thalamus), the ratio was initially less than one but grew as the inhibitory cells became more active. Taking a slice through the winh/wexc surface (dotted line in Fig. 5C, top) reproduces qualitatively the plot in Fig 3C. The model also predicted the relative changes in the peak excitatory and inhibitory current magnitudes (Fig. 5D) observed in the full network simulations. A slice through the surface (dotted line) reproduces qualitatively the plot in Fig. 2E. The transition still occurred in absence of the I cell threshold (Fig. S4). However, the presence of the threshold minimized the input σ needed for the transition. Without the threshold, even cells far from the center were activated so that the width of Finh is equal to Ithal. After the convolution with the axonal spread, wI/wE asymptoted toward but was always >1 (Fig. S4). Larger values of σσ were needed to achieve the near-co-tuned regime. For simplicity, the analyses assumed that the thalamic afferents were distributed equally to excitatory and inhibitory cells. Whether or not the P and FS cells have the same tuning properties in intact animals is unclear. To determine the effects of unequal tuning, the relative widths of thalamic inputs to P and FS were varied (Fig. S5). The shifts still occurred though a broader (narrower) input to inhibitory cells shifted the location of exact co-tuning (winh/wexc = 1) toward larger (smaller) σ values. To determine the changes in firing rate associated with the changes in network configuration, we performed simulations using a rate-based model [55], [56]. Input was a constant current with a Gaussian spatial profile parameterized by Nmax (maximum thalamic current) and σ. In the LIN regime (σ = 40 µm), firing was confined to a narrow band (Fig. 6A, top) and was sustained (bottom; firing rate profile shown for neuron in center). With relatively broad input (σ = 120), the spatial profile of firing rapidly narrowed within 50 ms (Fig. 6B, top), reflecting the fact that firing was transient (bottom). As above (Fig 3C, D; Fig. 5C, D), increasing the input width (σ) shifted the network from lateral inhibition toward co-tuning (Fig. 6C, left) and produced an associated increase in the magnitude of inhibition (right). To document the changes in firing with stimuli, the average rate (calculated over the first 50 ms of the firing profiles of the center neurons, Figs. 6A, B, bottom) is plotted against Nmax and σ (Fig. 6D). As expected, firing was greatest when lateral inhibition predominated, and decreased for increasing σ and decreasing Nmax. This surface essentially describes the change in firing behavior as the input to the network changes. Physiologically, Nmax and σ of the Gaussian may represent the change in e. g. thalamic input to primary auditory cortex as the stimulus intensity increases (Fig. 7A, top). To illustrate, we use sigmoid functions to simulate the changes in Nmax (Fig. 7A, middle left) and σ (right) that may occur with graded increases in stimulus intensity [8]. The resultant Gaussians become taller and wider as intensity increases (bottom). The solid and dotted curves in the σ vs. intensity plot (Fig. 7A, middle, right) represent two scenarios where thalamocortical afferent spreads are relatively narrow (solid curve) and broad (dotted curve), respectively [57], [58]. The changes in σ and Nmax with intensity resulted in the trajectories shown by the curves traversing the average firing rate surface (Fig. 7B). When the input spread was narrow (solid black curve, Fig. 7B, top), the peak firing rate climbed the steep region of the response contour before rolling off toward the end, resulting in a non-monotonic stimulus-response curve (bottom, solid; c. f. [3], [4]). With broader input (dotted curve), the trajectory was shifted to the flatter part of the contour, yielding a monotonically increasing response (bottom, dotted). Thus, monotonic and non-monotonic rate intensity functions were obtained when co-tuning and lateral inhibition, respectively, were predominant, as predicted previously [8] and observed experimentally [59]. The results of the simulations are robust under a variety of conditions (see Supporting Text S1). The main observation is maintained with different models (Fig. 3 vs Fig. 5 vs. Fig. 7) and with different network and input parameters (Figs. S1-S3, S5). Nevertheless, several caveats must be considered. One is that the model applies only to cortical regions where graded changes in response properties reflect the orderly arrangement of thalamic afferents and FS cell arbors. These conditions appear to be met in rodent A1; frequency tuning of neurons in the middle layers of A1 has been found to vary systematically, with substantial changes in tuning over a few hundreds of µm on the rostrocaudal axis, and thalamic afferents showed a comparable degree of tonotopic organization in a recent study [60]; but see [11]. Likewise the model could potentially apply to phenomena such as the graded changes in orientation tuning within a single ocular dominance column of primary visual cortex (V1 [61]). Another proviso is that the FS-mediated inhibition must be stronger than the recurrent, intracortical excitation. This condition is met in auditory and somatosensory cortices where connection probabilities and the unitary synaptic potentials are much larger for FS to P than for P to P connections [33], [52], [62], [63]. Prominent inhibitory synaptic conductances are also evoked in vivo during auditory stimulation [13], [14], [64]. In primary visual cortex, the role of inhibition in shaping the tuning of cells remains controversial [20], [65]: strong transient inhibition has been evoked by electrical stimulation of the thalamus [66] though generally not with visual stimuli [67], [68], [69]; but see [70]. In addition, the model while incorporating many of the measured parameters is necessarily incomplete because several parameters - notably the spatiotemporal profile of thalamic inputs to P and FS, and sources of noise under in vivo conditions - are yet poorly characterized. Nevertheless, simulations where these parameters are varied yield qualitatively, if not quantitatively, similar results (Figs. S1–S5). The simulations also do not consider other types of interneurons and potentially important inputs from other cortical areas. Many of these issues can be circumvented if only the synaptic events and firing within the first 50 ms after the stimulus are considered. In the auditory system, as noted above, brief tone pips (25–50 ms) are often used to characterize the tuning and firing properties of neurons. As argued in the Results, the synaptic inputs to P cells would be dominated by excitatory inputs from the thalamus and inhibitory inputs from the FS cells. Finally, in vivo whole-cell recordings from the auditory system of mice and rats [13], [71] have suggested that excitation and inhibition are co-tuned, though there is some evidence that the co-tuning is only approximate [15], [16]. The apparent lack of lateral inhibition may mean that the spatial extents of the thalamic afferents from the medial geniculate are relatively broad so as to bias the network towards co-tuning. Alternatively, the experimental conditions may bias the network towards co-tuning. All of the in vivo experiments thus far have been with anesthetized animals. The evoked responses in awake animals are markedly different [5], [72], and are more consistent with the presence of lateral inhibition [8]. Conservatively, our findings are most comparable to the in vivo studies in anaesthetized animals using brief stimuli. However, the transition between lateral inhibition and co-tuning persists in the presence of background noise (Fig. S1) and synaptic adaptation (Fig. S2), more similar to conditions that obtain in the awake state or with prolonged and/or natural stimuli. With these caveats, the results of the simulations have several implications. First, because the recurrent excitatory connections are weak, the firing of P cells is determined primarily by a feedforward configuration (thalamic to P, FS; FS to P). There is some experimental support for this finding in auditory cortex in vivo, because blocking recurrent excitation was found not to grossly affect the tuning of neurons responding to brief stimuli [41]. During prolonged stimuli, there may be a greater contribution from recurrent excitation; non-FS cells may also play a larger role with prolonged stimuli, because physiological studies have shown that PSPs to and from some non-FS cells facilitate [29], [30], [42], in contrast to the depressing synapses between P and FS cells. A second implication of the model is that the magnitude of inhibition increases in parallel with the spread of thalamic input (Fig. 5D and 6C). The resultant decrease in excitatory cell firing (Fig. 4A) resembles that observed when auditory stimuli become more broadband [59] or when the size of visual stimuli expands beyond the classical receptive field [73], [74], [75], [76]. A third implication is that the thalamocortical terminal field widths will determine whether a cortical area is biased towards co-tuning or lateral inhibition. In rodent somatosensory and primate visual cortices, the spatial distribution of thalamic axons varies 6 fold in the thalamorecipient layers [57], [58]. In auditory cortex there are few reports on the widths of thalamocortical terminal fields, and none to our knowledge in rodents. However, frequency layer organization in rat [77] and gerbil [78] auditory thalamus shifts from narrow at the caudal end to broad at the rostral end; similar shifts in laminar organization are observed in the cat [79], [80]. These thalamic pattern differences have been postulated to underlie heterogeneous response properties of neurons within an isofrequency band [59]; c. f. [77], [81], [82], [83]. The model predicts that if the distinct sets of inputs differ in their axonal arborization widths, this variation alone is enough to support a wide range of receptive field properties seen experimentally in rodent A1. Fourth, whether the network is biased toward lateral inhibition or co-tuning will also depend on the dendritic and axonal arborizations of FS cells [62], [84], [85], [86], [87], which may vary along with other properties of the specific sub-circuits targeted by broad vs. narrowly distributed thalamic axons [88]. In somatosensory cortex, the spatial distribution of FS cell processes is well conserved, suggesting systematic shifts in network configuration that parallel the changes in the distribution of thalamic afferents. Finally, the fact that the relation between excitation and inhibition is malleable in a single network potentially provides a mechanism for modulating the response of the network to a variety of inputs and behavioral states. Recently, for example, the magnitudes of feedforward and lateral inputs in visual cortex were found to be modulated by stimulus contrast [89]. It would be of interest to determine experimentally if the relative degree of co-tuning vs. lateral inhibition can be triggered by changing the stimulus characteristics, experimental conditions, or state of the animal. It will also be interesting to see whether the reduced model presented here can account for response properties obtained with complex sound stimuli in awake animals [2], [5], [72], or whether additional elements such non-FS inhibitory neurons, input from other cortical areas, and state-dependent effects of neuromodulators are important. The network was a 2 dimensional sheet of 160×160 P and 32×32 of inhibitory FS cells (Fig. 1A). The density of neurons was 91125 neurons/mm3 and network was assumed to correspond to a volume (x by y by h) of 1185×1185×200 µm before compression to 2 dimensions (1185×1185 µm). In 2 dimensions, the probability that a reference neuron at x0, y0 is connected to its neighbors xi, yj is given by: (1) where d represents radial distance A represents peak probability, and σ represents spread of connectivity (see Results for values); fits to the experimental data (not shown) were done using nonlinear-least squares regression. Probability values predicted from the fitted curves did not differ significantly from the experimental values (p>0. 05, χ2 goodness-of-fit tests). The neurons were adaptive exponential integrate and fire units (aEIF; [39]; see Supporting Text S1). The aEIF accurately reproduces the firing patterns of cortical neurons with relatively little computational cost, thereby allowing modeling large networks efficiently. The parameters governing the firing behaviors (Table S1) were adjusted so as to produce firing patterns and firing rate-current (F/I curves) that resembled those of the experimentally recorded P and FS cells. The postsynaptic conductance, gsyn, was described with an alpha function: (2) The parameters k and α were adjusted so that the amplitudes and time courses of resultant EPSPs and IPSPs matched experimentally measured synaptic potentials evoked in the two cell types (not shown). Short-term depression and facilitation were implemented using a phenomenological model [36], [47]. For connections between cortical neurons, model parameters (Table S1) were adjusted to match the experimentally measured synaptic dynamics. The amplitude and short-term dynamics of thalamic EPSPs evoked in P and FS neurons were taken from data obtained in mouse thalamocortical slices [40] Schiff and Reyes, unpublished) and obtained from somatosensory thalamocortical slices [43], [44]. The thalamic inputs to the network were drawn randomly from a set of simulated spike trains (Fig. 1B, inset). For each train, the spikes occurred repetitively at a specified rate F; the latency of the first spike was Gaussian distributed (mean = 1/F; standard deviation = 0. 25/F) as was the latency of each successive spike. With this procedure, the population spikes tended to cluster at the onset of the stimulus but over time became more evenly distributed as the spikes became less synchronous, resulting in a histogram with a phasic-tonic firing profile (Fig. 2B, bottom). Both the number (Nin) of inputs and frequency Fin of each input to the network were Gaussian distributed in space (Fig. 1B) so that cells in the center received the maximum number of inputs (Nmax = 25–150) firing at the maximum specified rate (Fmax = 50 Hz): (3) (4) where xctr and yctr are the center of the network. Nin and Fin were adjusted so that the evoked firing rates of the P and FS cells were in the midrange of their respective F/I curves. The spike trains were used to calculate the composite synaptic currents generated in each P and FS neuron. Representative synaptic conductances and associated firing are shown in Figure 1C. The evoked firing was quantified by calculating the number of spikes that occur within a 50 ms time window after the onset of the stimulus. This corresponded approximately to the peak of the post-stimulus time histogram (PSTH) compiled from the spike trains (Fig. 1D). The simplified network shown in Figure 5 consisted of inhibitory cells that synapsed onto excitatory cells; both excitatory and inhibitory cells received spatially distributed inputs from the thalamus. The spatial connection profile of the inhibitory to excitatory cells (Pinh (x) ) and that of the thalamic synaptic inputs (Ithal) to the excitatory and inhibitory cells are each Gaussian: (5a) (5b) where σinh and σthal are the standard deviations of the inhibitory to excitatory connection profile and the thalamic input, respectively. The spatial profiles of the excitatory (Iexc) and inhibitory input (Iinh) were calculated as described in Results. The transformation of thalamic input to obtain inhibitory firing was given by: (6) where θ is the threshold current for firing and m is the slope of the firing-current relation. Additional simulation details, including parameter values, are in the Supporting Information. The simulations in Figures 6 and 7 were based on an established model [53], [54]. The model assumes that the cell population is large and firing is random, so that the calculation of individual spike trains and cell-by-cell responses can be replaced by a simplified expression for excitatory and inhibitory population firing rates (Fexc, Finh, in Hz) in terms of position (x) and time: (7a) (7b) The time constants τexc, τinh (Table S3) for the population firing rates reflected the relative membrane time constants measured for P and FS in vitro (c. f. [8]). The synaptic weight functions wee, wei, wie were each the product of three terms: (8a) (8b) (8c) where ρ is presynaptic cell density (Supporting Information, Table S3), A is average unitary synapse strength in pA/Hz (table S4), and P (x) is the Gaussian connectivity profile (eqn 1, for parameter values, see Results). Thalamic firing rate Fthal was fixed at 20 Hz, while the thalamic input weight functions we (thal), wi (thal) were the corresponding average unitary synaptic strengths (Table S4) multiplied by the Gaussian input profiles, parameterized by Nmax and σ as detailed for Figure 6. Nmax and σ for thalamic input did not differ between E and FS cells. Because only transient responses were examined, the model did not incorporate short-term synaptic plasticity, which influences network dynamics on longer time scales [36]. The relationship of firing rate to total input current for each cell type (Sexh, Sinh,), was modeled as a Naka-Rushton function: (9) M, θ, and n (Table S3) were fitted to average plots of non-adapted firing rate versus injected current obtained from cortical P and FS cells in vitro (not shown), using an iterative search procedure.

The cerebral cortex contains a network of electrically active cells (neurons) interconnected by synapses, which may be excitatory (tending to increase activity) or inhibitory. Network activity, i. e. , the ensemble of activity patterns of the individual cells, is driven by input from the sense organs, and creates an internal representation of features of the outside world. In auditory cortex, sound frequency (pitch) is encoded by the physical location of activity in the network. Thus, connections among cells at various distances may blur or sharpen the frequency representation. Recent work in living animals has yielded conflicting results: sharpening of responses via lateral inhibition in some cases, versus balanced excitation and inhibition (co-tuning) in others. It was previously unknown whether a single cortical network architecture could account for this spectrum of findings. Here, computer simulations based on experimental data reveal that this is indeed the case. Varying input to the network causes smooth transitions between lateral inhibition and co-tuning, accompanied by changes in the strength and timing of the responses. Diverse input-dependent response patterns in a single network may be a general mechanism enabling the brain to process a wide range of sensory information under various conditions.

Abstract Introduction Results Discussion Methods

circuit models auditory system central nervous system synapses computational neuroscience biology sensory systems neuroscience neurophysiology

2011

Coexistence of Lateral and Co-Tuned Inhibitory Configurations in Cortical Networks

Neurons communicate primarily with spikes, but most theories of neural computation are based on firing rates. Yet, many experimental observations suggest that the temporal coordination of spikes plays a role in sensory processing. Among potential spike-based codes, synchrony appears as a good candidate because neural firing and plasticity are sensitive to fine input correlations. However, it is unclear what role synchrony may play in neural computation, and what functional advantage it may provide. With a theoretical approach, I show that the computational interest of neural synchrony appears when neurons have heterogeneous properties. In this context, the relationship between stimuli and neural synchrony is captured by the concept of synchrony receptive field, the set of stimuli which induce synchronous responses in a group of neurons. In a heterogeneous neural population, it appears that synchrony patterns represent structure or sensory invariants in stimuli, which can then be detected by postsynaptic neurons. The required neural circuitry can spontaneously emerge with spike-timing-dependent plasticity. Using examples in different sensory modalities, I show that this allows simple neural circuits to extract relevant information from realistic sensory stimuli, for example to identify a fluctuating odor in the presence of distractors. This theory of synchrony-based computation shows that relative spike timing may indeed have computational relevance, and suggests new types of neural network models for sensory processing with appealing computational properties. Neuronal synchronization is ubiquitous in the nervous system [1], [2]. In the retina, neighboring cells are often synchronized at a fine timescale [3], [4], and relative spike timing carries information about visual stimuli [5]. Visual and somatosensory stimulation also elicits synchronized activity in the thalamus [6]–[8], which impacts target cortical neurons [9]–[12]. In olfaction, fine odor discrimination relies on transient synchronization between specific neurons [13]. In the auditory system, phase locking in brainstem neurons [14] produces fine stimulus-driven correlations in spike timing which are determinant for sound localization [15]. At cellular level, modeling and experimental studies show that correlated inputs are more likely to make neurons fire [16]–[19], and synaptic plasticity mechanisms favor correlated synaptic inputs [20], [21], so that developed neural circuits should be very sensitive to correlations. These findings suggest that neural synchronization is functionally important in early sensory pathways, but it is not clear what it implies in terms of computation. In many theoretical studies of spiking neural networks, spike timing and neural heterogeneity are treated as noise to be averaged out in the activity of “neural masses” [22]. One theory, reservoir computing, assigns a computational role to neural heterogeneity, that of representing sensory stimuli in a high-dimensional space where decoding is easier [23], but it does not assign a specific role to spike timing or synchrony. Thus, although many authors have advocated the idea that the brain may use precise spike timing to process sensory information [24], [25], there are few general theories of spike-based computation. One such theory postulates that the rank order of spikes carries information [26]. This is supported by experimental evidence in the retina [5], but physiologically decoding this information is not entirely straightforward, as it would involve rather specific circuits of inhibition and excitation. In addition, although it seems to be a metabolically efficient way of processing information, the advantages in terms of computational power are not obvious. On the other hand, synchrony can be easily decoded by neurons, by means of coincidence detection [27], and is compatible with Hebbian learning theories, in which correlated inputs tend to be strengthened [20]. In this article, I focus on synchrony induced by the stimulus (rather than by coupling between neurons [28]–[31]) and I address the two following questions: what does synchrony mean? how and what can neurons compute with synchrony? It appears that neural heterogeneity, which is considerable in the nervous system [32], is the key ingredient that makes synchrony computationally interesting, because synchrony then reveals sensory invariants, which play a central role in psychological theories of perception. For synchrony to be computationally useful, it must be stimulus-dependent. To illustrate this idea, let us consider neurons which spike after being hyperpolarized (“rebound spiking”), because of the presence of voltage-activated conductances (Fig. 1; simple neuron models are used in this and all other figures; see Methods for details). Neurons with rebound spiking have been found for example in the superior paraolivary nucleus of the auditory brainstem, a structure involved in encoding the temporal structure of sounds [33]; and in the pyloric network of lobsters, involved in the generation of rhythmic motor patterns [34]. Fig. 1 shows a minimal neuron model with this property (but it is only meant as an illustration). The model includes a slow outward current, modeling K+ channels, which activates at low voltages (half-activation voltage −70 mV). This current prevents the neuron from spontaneously spiking. When the neuron is inhibited for a few hundred ms (Fig. 1A, top), the K+ channels slowly close (the conductance decreases, Fig. 1A, bottom). When inhibition is released, the negative K+ current is smaller than at rest, which makes the neuron spike. The latency of the rebound spike depends on the value of the K+ conductance when inhibition is released, and therefore on the duration of inhibition: if the neuron is inhibited for a shorter duration, K+ channels are still partially open when inhibition is released and the neuron spikes later. If inhibition is very short, the neuron may not spike. Thus, the timing of the rebound spike is negatively correlated with the duration of inhibition. Fig. 1B shows this relationship for two different model neurons A and B, which have the same rebound spiking property but quantitatively different parameter values (spike threshold and time constant of K+ channels). The receptive field of a neuron can be defined as the set of stimuli which elicit a response in the neuron: in this example, stimuli are inhibitory pulses with duration varying between 0 and 1000 ms, and the receptive fields of neurons A and B are inhibitory pulses lasting more than 200 ms. Therefore, the individual receptive fields of the neurons convey little information about duration. I now define the synchrony receptive field (SRF) of two neurons as the set of stimuli which elicit synchronous firing in these two neurons. For neurons A and B in Fig. 1B, the SRF is found at the intersection of the duration-latency curves: the two neurons fire in synchrony when the stimulus lasts about 500 ms. At this point, I make three remarks. First, the SRF reveals information about the stimulus that may not be available from individual receptive fields (here, both neurons fire one spike to all stimuli lasting more than 200 ms). Second, this additional information can only be available when neurons have heterogeneous properties (otherwise, the SRF is the set of all stimuli). Third, the SRF is specific of a pair (possibly group) of neurons: the duration-latency curve of neuron A will generally intersect that of another neuron C at a different point, or may not intersect it at all (and the SRF is empty). Therefore, in a heterogeneous population of neurons, any given stimulus will trigger a specific synchrony pattern. How can this synchrony pattern be decoded? Consider a postsynaptic neuron receiving excitatory inputs from neurons A and B (Fig. 1C). The neuron also receives inputs from other sources, which are modeled as background noise. If this neuron is sensitive to coincidences, then it will fire more when the two inputs are synchronous, that is, when the stimulus is in the SRF of A and B. As a result, the firing rate of this neuron will be tuned to the duration of the stimulus, although its inputs are not (Fig. 1D). The model used in Fig. 1D is a simple integrate-and-fire neuron with background noise (time constant τ = 5 ms). As shown in [19] (elaborating on ideas proposed by Abeles [16]), the key ingredient for the neuron to be sensitive to coincidences is that the average background input is subthreshold. In this regime, the neuron is said to be “fluctuation-driven”: it fires to large fluctuations above the mean potential. This property can be understood in terms of signal detection theory [35]. In vivo intracellular recordings show that in many areas, the membrane potential distribution p (v) peaks well below threshold, indicating that neurons are indeed fluctuation-driven (e. g. in auditory cortex [36], visual cortex [37], barrel cortex [38], frontal cortex [39]). This distribution is represented in Fig. 1E (“background”), which we consider as noise with standard deviation σ. When coincident spikes depolarize the neuron by an amount Δv (= nw for n coincident postsynaptic potentials (PSPs) of size w), this probability distribution is shifted by Δv (Fig. 1E, “signal”). The neuron spikes when the membrane potential exceeds the spike threshold θ, which implements the decision threshold to detect these coincidences over the background. The neuron will respond to coincidences (hits) but also to background activity (false alarms), with some probability called the “hit rate” (HR) and “false alarm rate” (FR). Both rates decrease when the threshold increases. For a given value σ of the noise, HR and FR are linked by a curve named the receiver-operating characteristic (ROC), obtained by varying the threshold. ROC curves are shown in Fig. 1F for a noisy integrate-and-fire neuron with exponentially decaying PSPs, with three noise levels (black curves). The rates are calculated as the probability of firing within one integration time constant τ when the neuron receives a PSP of size Δv (HR) and when it does not (FR). That is, the FR is the product Fτ, where F is the spontaneous firing rate. Each ROC curve is calculated (with numerical simulations) by varying the spike threshold while keeping the same noise level. Higher thresholds correspond to lower rates. When the noise is very high, this ROC curve is a diagonal (dashed), meaning that coincidences cannot be distinguished from background. As the noise decreases, the ROC curve shifts toward the upper left corner, meaning that spikes indicate coincidences more reliably. In signal detection theory, this relationship between hit rate and false alarm rate is quantified by the sensitivity index d′, which, for normal distributions, is the spread between the distributions in units of the noise standard deviation: . Red curves in Fig. 1F show the theoretical ROC curves for the noise values used in the simulations. Thus, d′ quantifies the ability to detect coincidences while the value of the spike threshold corresponds to a particular trade-off between hit rate and false alarm rate. For example, in the case of two coincident spikes, one simple choice is θ = 2w (relative to the mean membrane potential), which ensures a HR of 50% when the two input spikes are synchronous, and a lower FR for background activity (Fig. 1F, horizontal dashed line). More generally, the ratio between HR and FR increases when the FR decreases: this implies that, to detect coincidences, the false alarm rate should be set to a low level. For an integrate-and-fire neuron with spontaneous firing rate F and integration time constant τ, we have defined the FR as F. τ. Some experimental evidence indicates that this quantity is indeed low in vivo: the membrane time constant is short in vivo (e. g. around 5 ms in the frontal cortex [39]), because of the large total synaptic conductance [40]; average firing rates are low, possibly smaller than 1 Hz [41]. Although the latter point is controversial, the product F. τ remains small even with larger estimates of F. In addition, we note that the temporal window of integration is in fact shorter than the membrane time constant, because of spike threshold adaptation [42], [43], and because of coordinated inhibition [44]. This ensures that the ratio HR/FR is high, even for small d′ (small depolarization). Thus, neurons can detect coincidences above background noise, but an important question is the temporal resolution of coincidence detection. We can use signal detection theory again to address this question. Consider two input spikes delayed by a time δ, each one producing an exponential PSP of size w and decay time τ (Fig. 1G). When the spikes are synchronous (Fig. 1G, left), the membrane potential at peak time is 2w, plus the background noise. When they are delayed (Fig. 1G, right), the peak membrane potential is, plus the noise. To detect between these two possibilities, we need to distinguish between two random variables with means differing by and standard deviation σ (Fig. 1H). This corresponds to a sensitivity indexand for short delays (): . This can be described as the product of the signal-to-noise ratio (w/σ) with the delay in units of the time constant. We can now define the temporal resolution of coincidence detection using the concept of “just noticeable difference” (JND), defined as the delay for which spikes can be correctly distinguished from synchronous spikes with 75% probability (assuming 50% correct answers for). This corresponds to a d′ of 1. 35 [35], which gives for short delays: Thus, the temporal resolution of coincidence detection is proportional to the integration time constant, and inversely proportional to the signal-to-noise ratio. Note that the approximation corresponds here to, i. e. , low noise. The precise expression using the original formula for d′ is: This expression is only defined with relatively low noise, when: this is because above this value, it is not possible to correctly distinguish between synchronous and asynchronous spikes () with 75% probability. Let us come back to the specific example of duration selectivity I have presented above. The postsynaptic neuron receives input spikes from neuron A and neuron B, at latencies LA (D) and LB (D), where D is the duration of the stimulus. The latency curves intersect at some duration D* (500 ms in Fig. 1B). The timing difference between the two spikes is. We approximate it near the intersection point as, and we obtain this approximate expression of the JND in duration: The term quantifies how different the latency curves are near the intersection point. This formula indicates that the detection of duration is more accurate when the properties of the presynaptic neurons are heterogeneous. We can now apply these principles to decode synchrony patterns at the population level. Consider a population of neurons with rebound spiking properties but heterogeneous parameters. For example, in Fig. 2A, the membrane time constant varies randomly across neurons between 10 and 50 ms, and the K+ channel time constant varies between 300 and 500 ms (see Text S1 for a justification of this choice of parameter values). For a given stimulus, for example an inhibitory pulse with duration 300 ms, we can look at the synchrony pattern in the neural population. In Fig. 1E (top left), neurons represented with the same color produce a rebound spike at the same time (with a 2 ms precision), that is, the stimulus is in the SRF of neurons with the same color. Thus, the neuron population can be divided in groups of synchronous neurons (possibly containing just one neuron). I call this partition of the neural population the synchrony partition (mathematically, it is the neural partition defined by synchrony, which is an equivalence relation). This definition mirrors the definition of the SRF: the SRF describes the set of stimuli for which a given group of neurons are synchronous, the synchrony partition describes the groups of neurons that are synchronous for a given stimulus. Fig. 2A shows the synchrony partition in a population of 25 heterogeneous neurons for three stimuli: inhibitory pulses of 300 ms, 400 ms and 500 ms. Each stimulus produces a different synchrony partition: for example, the three neurons colored in green for the 300 ms stimulus are not synchronous for the 400 ms stimulus. Decoding synchrony patterns is now straightforward (Fig. 2B). For each synchrony partition (each stimulus), we assign a population of postsynaptic neurons, one neuron for each group in the partition (colored neurons in Fig. 2B). Presynaptic neurons in the same group (same color) make excitatory synapses onto the same postsynaptic neuron. In this figure, the peak size of PSPs was set as the difference between threshold and mean potential divided by the number of neurons in the presynaptic group: this choice means that the hit rate should be 50% (only approximately, since input synchrony is not perfect). Therefore, the postsynaptic neural assembly maximally fires for a specific synchrony partition, that is, for a specific stimulus (Fig. 2C). In this way, synchrony partitions are mapped to patterns of postsynaptic activity, and SRFs are mapped to standard receptive fields. We note in Fig. 2C a few deviations from the ideal scenario described above. First, the maximum firing probability is generally lower than 0. 5. This is because a synchronous group was defined as a group of neurons that fire within 2 ms of each other, rather than at the exact same time. With more encoding neurons, groups could be defined with a better precision (i. e. , finer synchrony partitions). Second, the duration selectivity curves are non-symmetrical, with more spikes produced at longer durations. This is because there is more heterogeneity in spike latency at short durations (where latency curves diverge, see Fig. 1B) than at long durations (where latency curves are constant). Making the integration time constant of coincidence detectors shorter would reduce this phenomenon. As a consequence of these two facts, selectivity curves do not peak exactly at the expected duration. The ideal scenario corresponds to the limit case where stimuli are encoded by many neurons (allowing fine synchrony partitions) and synchrony patterns are decoded with a fine resolution (short time constant of coincidence detectors). Decoding synchrony patterns requires that neurons are sensitive to coincidences (in the sense that they fire more when their inputs are coincident), but it does not rely on specific neural properties, as is shown in Fig. 3. Varying the amount of internal noise quantitatively changes the neuron sensitivity to coincidences (the sensitivity index d′ in the signal detection theory perspective) but it does not change the qualitative properties (Fig. 3A). In Fig. 3B, inputs to the neurons were modeled as excitatory synaptic conductances (exponentially decaying with time constant τe = 2 ms). The main difference is that the size of PSPs now depends on the driving force (synaptic reversal potential minus membrane potential). However, as argued in [19], for an excitatory synapse, the driving force is restricted to a rather small range below spike threshold (50–80 mV), so that it has little impact on PSP size and on coincidence detection properties. In Fig. 3C, the coincidence detector neurons are modeled in the same way as the presynaptic neurons, with rebound spiking (with time constants τ = 10 ms and τKLT = 400 ms, see the Methods for details). That is, neurons of the same type encode the stimuli and decode the synchrony patterns. The results are qualitatively unchanged. I have shown an explicit construction of the decoding circuit, but how can this circuit spontaneously emerge? As explained above, a simple condition for a neuron to be sensitive to coincidences is to ensure that its firing rate is low. This can be implemented by a homeostatic principle. Two physiologically plausible mechanisms are intrinsic plasticity, where excitability (e. g. spike threshold or membrane resistance) changes with activity [45], and synaptic scaling, where synaptic weights change with pre- and/or post-synaptic activity [46]. In the context of signal detection theory (Fig. 1E–H), homeostasis can be seen as the process of setting the decision threshold so as to maintain a low false alarm rate. I consider a simple synaptic scaling mechanism in which synaptic weights continuously increase, independently of pre- and post-synaptic activity, and each postsynaptic spike reduces all synaptic weights: This multiplicative form corresponds to experimental observations [47] and it also has theoretical advantages: 1) it is equivalent to a change in spike threshold, 2) it leaves the relative strengths of the synapses unchanged and 3) it keeps the weights positive, without imposing a hard boundary. Weights are stable when (where F is the postsynaptic firing rate), that is, when. Thus this mechanism maintains a target firing rate F. Homeostasis acts on the decision threshold but is not synapse-specific (that is, it does not improve the sensitivity index d′). In the circuit shown in Fig. 2, the postsynaptic neuron fires when the presynaptic neurons belong to the same (stimulus-specific) synchronous group. To develop such circuits requires a synaptic plasticity mechanism that selectively strengthens synapses that are co-activated with the postsynaptic neuron, in a short temporal window corresponding to the precision of the synchrony partition. This is consistent with the properties of long-term potentiation in spike-timing-dependent plasticity (STDP) seen at excitatory synapses onto excitatory neurons [48], and theoretical studies have shown that STDP favors correlated inputs [20], [21], [49]. In addition to homeostasis, I consider an STDP rule in which the synaptic weight modification depends on the difference in timing tpost-tpre of a pre- and post-synaptic spike (Fig. 4A): The synaptic modifications induced by all pairs of pre and post spikes are added, but in this context where firing rates are low (around 1 Hz), the precise way in which pairs interact does not make a difference. The time constant is set equal to the membrane time constant τ. I also choose a small value for aLTP, so that the average firing rate is mainly determined by the homeostatic mechanism while the relative strengths of synapses are determined by the correlations between the synaptic inputs and the neuron output. It is not necessary to impose a boundary on the synaptic weights, because stability is ensured by the homeostatic mechanism. In the same way, long term depression (LTD) is unnecessary here, and it is ignored for simplicity. I consider a group of presynaptic neurons (100 were simulated) and postsynaptic neurons as in Fig. 2, connected by random synapses, with an average of 5 synapses per postsynaptic neuron (Fig. 4B). The synaptic weights are initially random between 0 and 1 (1 is the spike threshold), and they evolve through homeostasis and STDP while 5000 stimuli with random duration are sequentially presented. Fig. 4C shows the selectivity curves of 5 postsynaptic neurons, before (top) and after learning (bottom), as in Fig. 2C. Initially, neurons tend have high-pass properties, that is, they fire when the stimulus is longer than a given duration. This mirrors the properties of the inputs (Fig. 1B). In one case (green curve), the neuron almost never fired to any stimulus. After learning, most neurons have a peaked selectivity curve, with a preferred duration. Fig. 4D shows the evolution of synaptic weights during learning for the postsynaptic neuron corresponding to the blue curves in Fig. 4C. It appears that most synaptic weights decay, except two of them which stabilize at 0. 5 (half the distance to spike threshold) and one weaker synapse. The properties of these synapses are shown in Fig. 4E. Each curve represents the spike latency of the presynaptic neurons for the neuron considered in Fig. 4D (as in Fig. 1B), and are the two strongest synapses are displayed in red. It appears that these two curves intersect at a duration of about 430 ms, which is the best duration of the neuron shown in blue in Fig. 4C. This illustrates the idea that the postsynaptic neuron fires when the stimulus is in the synchrony receptive field of its presynaptic neurons. Fig. 4F shows that the learning mechanism selects synapses in the same way as I described in Fig. 2, that is, it selects synapses that are synchronously active for a specific stimulus duration. Each color corresponds to a postsynaptic neuron (same color code as in Fig. 4C) and each dot represents the weight of one synapse vs. the spike latency of the corresponding presynaptic neuron, at the best duration of the postsynaptic neuron. For example, for the green neuron, the two strongest synapses are synchronously active (same spike latency) at the best duration (about 420 ms, Fig. 4C), while the other synapses are activated at diverse latencies. Similar observations can be made for the two other neurons. An interesting point is that the blue and green neurons have the same best durations (about 420–430 ms, Fig. 4C) but respond at different latencies (about 25 ms and 55 ms; strongest synapses in Fig. 4F). This corresponds to two different groups of the synchrony partition in Fig. 2 (neurons shown with two different colors in the same column). Thus, the proposed decoding circuit (Fig. 2) can emerge in an unsupervised way, through a combination of homeostasis and STDP. I introduced the concepts of synchrony receptive fields and synchrony partition with an elementary example, duration selectivity, where stimuli are one-dimensional. Real world stimuli, on the other hand, vary along many dimensions, which makes computation much more difficult [50]. To understand synchrony patterns in this more general setting, I describe neuron responses in the following simplified way (Fig. 5A, top): a stimulus S is transformed through a linear or non-linear filter N, which represents the (standard) receptive field of the neuron, then the filtered stimulus N (S) is mapped to a spike train through a non-linear spiking transformation (for example, N (S) is the input to a spiking neuron model). Note that although this description appears to be feedforward, the computation of the filter N may or may not rely on a feedforward circuit. Assuming that two neurons A and B fire in synchrony when they receive the same dynamic input NA (S) and NB (S), the SRF of A and B is the set of stimuli S such that NA (S) = NB (S). In mathematical terms, this is a manifold of stimulus space; if the neural filters are linear, it is a linear subspace of stimuli. For example, in two dimensions, the SRF is a line (Fig. 5B, left). In contrast, a neuron fires when the filtered stimulus exceeds some threshold, N (S) >θ, that is, in two dimensions, when the stimulus is on one side of a line (Fig. 5B, right). In higher dimension, a neuron fires when the stimulus is on one side of a hyperplane, while two neurons fire in synchrony when the stimulus is close to a hyperplane (assuming linear filtering). This makes computation with synchrony qualitatively different from rate-based computation, with interesting computational properties, for example SRFs are unchanged by linear scaling of the stimulus (i. e. , intensity change). I will describe these qualitative differences in more details in the next section, but first I will comment on the hypothesis that two neurons fire in synchrony when they receive the same dynamical input. First, this should not be true if the neurons have different intrinsic properties (for example, spike threshold or resistance). Therefore I consider that the heterogeneity in intrinsic properties is implicitly included in the description of the receptive field (or filter) N. For example, the membrane resistance can be included as a gain applied to the filter N (N→R. N) rather than in the spiking transformation; the membrane time constant can be included as a low-pass filter. Thus the hypothesis really means that two identical neurons fire in synchrony in response to identical time-varying stimuli. In vitro experiments have demonstrated that a single cortical neuron responds identically (at a millisecond timescale) to repeated time-varying currents [51]. As for coincidence detection properties, the main condition is that the neuron is in a fluctuation-driven regime, with a subthreshold average input [52], [53]. This property is illustrated with neuron models in Fig. 5C–E, which shows the response of a spiking neuron model to a fluctuating input (Fig. 5C) over repeated trials, with a subthreshold mean. The same current is presented in all trials, with an additional independent noise (red). This noise represents both the intrinsic noise and the difference in inputs between trials. If the noise level is low enough, spike timing is reproducible at a fine timescale, as shown by the shuffled autocorrelogram (SAC, see [54]) (Fig. 5D, right). A very important property is that the precision of synchrony between trials, as estimated by the width of the SAC (Fig. 5E; see Methods), reflects the similarity of the input signals (measured by the signal to noise ratio), rather than the intrinsic timescale of the signal fluctuations (seen in the autocorrelation of the signal in Fig. 5C, right). In particular, when noise level goes to 0, precision converges to 0 ms rather than to the timescale of input fluctuations (Fig. 5E, left). Therefore, when two identical neurons receive inputs NA (S) and NB (S), their degree of synchrony reflects the degree of similarity between NA (S) and NB (S). This is related to the mechanism used by Brody and Hopfield [55], [56] in a previous model of odor recognition based on spike timing, where constant inputs are added to an external oscillation, but it is more general. That oscillation-based mechanism works only in a limited input range (see Fig. 1 in [55]) because it relies on 1∶1 phase-locking (one spike per period of the oscillation) in a mean-driven regime (average input above threshold), which is less robust than the mechanism shown here [53] (phase locking is also more robust in the fluctuation-driven regime [57]). This reproducibility of spike timing has been demonstrated in vitro [51] and in vivo in early sensory pathways such as the retina [5] and the auditory brainstem [58], but it could be argued that it is an unrealistic assumption in other neural structures. However, synchrony-based computation does not critically rely on reproducible spike timing but rather on reproducible synchrony. Specifically, network activity may introduce inter-trial variability that is shared by neurons, as seen in the auditory cortex [59], degrading the reproducibility of absolute spike timing but not of relative spike timing. This is shown in Fig. 6, where three model neurons receive a stimulus-driven input, identical in all trials, and a shared external input, variable between trials. In addition, each neuron has a private source of noise. Neurons A and B receive the same stimulus-driven input, meaning the stimulus is in the SRF of A and B, and neuron C receives a different input (Fig. 6A). It appears that spike-timing reproducibility is low for all neurons (Fig. 6B, C), but that A and B are reliably synchronized in all trials (Fig. 6D, cross-correlogram). The peak of the cross-correlogram depends on the signal-to-noise ratio, defined between the shared and private components of the noise (Fig. 6E, F). This dependence can be quantified in exactly the same way as in Fig. 5E, where the signal is the sum of the stimulus and of the shared noise, while the noise corresponds to the private noise. Therefore, the mechanism used here does not critically rely on reproducible spike timing, but rather on reproducible stimulus-dependent synchrony. In this framework, a random stimulus cannot produce tightly synchronous responses in neurons with different receptive fields. Therefore, synchrony must reflect some non-randomness or “structure” in the stimulus. Fig. 7 illustrates the relationship between synchrony and structure with a few sensory examples. A classical example is binaural hearing (Fig. 7A). Leaving sound diffraction aside for the moment (see last section of the Results), the sound S (t) produced by a source on the left of the animal will arrive first at the left ear, then at the right ear, with propagation delays dL and dR, respectively. Therefore the two monaural signals are SL (t) = S (t−dL) and SR (t) = S (t−dR), respectively. The interaural time difference ITD = dR−dL depends on the azimuth of the source. The binaural stimulus (SL, SR) has a structure, in that both SL and SR are transformations of the same signal. That structure is specific of a particular ITD. Consider two monaural neurons A and B on opposite sides, which project to a binaural neuron with axonal delays δL and δR. From the postsynaptic point of view, the SRF of A and B should include the axonal conduction delays. It is the set of stimuli (SL, SR) such that SL (t−δL) = SR (t−δR), that is: SL (t) = SR (t− (δR−δL) ). Therefore, the SRF of A and B is the set of all binaural signals produced by a single source with ITD δL−δR, and it is independent of the source signal. Thus, the SRF indicates the structure of the stimulus, an information that is not present in the individual responses of the monaural neurons. The binaural neuron depicted in Fig. 7A fires when the stimulus is the in SRF of A and B, that is, at a specific source location. This is in essence the Jeffress model of sound localization [60]. Similar concepts apply to pitch perception (Fig. 7B). Pitch is the perceptual correlate of the periodicity of sounds, such as vowels or musical notes (to a first approximation). A periodic sound S (t) can be described as the repetition of a signal defined on one cycle (red curve). The repetition rate f0 determines the pitch, while the original signal determines the timbre. As for the binaural example, this produces a specific structure in the signal S (t), and the structural information is associated with the pitch of the sound. Consider two neurons A and B with the same properties but different axonal delays δA and δB. The SRF of A and B (again from the postsynaptic side, including axonal delays) is the set of signals such that S (t−δA) = S (t−δB). These are all the periodic signals with repetition rate f0 = 1/|δB−δA|, or a multiple of it. Again synchrony reflects a structural property of the stimulus. In essence, this is Licklider' s model of pitch perception [61]. The third example is olfaction (Fig. 7C). There is considerable heterogeneity in the properties of olfactory sensory neurons: there are about 1000 receptor types in rats, and neurons which express the same olfactory receptor type respond to the same odorants but vary in global sensitivity, up to 100-fold [62]. Odor plumes are highly turbulent [63], so that their concentration c (t) in the olfactory epithelium varies very quickly (Fig. 7C, left). The receptor coverage is defined as the probability that the receptor is bound to the odorant molecules. It can be expressed as aO. c (t), where aO the affinity of the receptor type to the presented odor O. Thus the olfactory stimulus can be represented as 1000 time-varying signals (receptor coverage for all types), but these signals have a strong structure since they are all scaled versions of the same signal (odor concentration c (t) ). Olfactory neurons of the same type differ in their global sensitivity s, so the activation of an olfactory neuron is essentially determined by the product of concentration c (t), odor affinity aO (type-specific and odor-specific) and global sensitivity s (neuron-specific): c (t). aO. s (the transformation of this signal to spike trains is highly nonlinear). Fig. 7C (right) schematically represents the value aO. s as a function of odor identity for three neurons: neurons A and B respond to the same odors (same receptor type), but A has higher global sensitivity than B; neuron C responds to different odors (different type). Tuning is broad, so a given odor elicits responses in many different olfactory neurons. The SRF of A and C is the set of olfactory stimuli such that c (t). aAO. sA = c (t). aCO. sC: the product of odor affinity and sensitivity is the same for neurons A and C. Although odor concentration c (t) varies very quickly, the identity aAO. sA = aCO. sC, which defines the SRF of A and C, does not depend on it. In Fig. 7C, the SRF of A and C is the single odor at the intersection of the two tuning curves. Neurons B and C have a different SRF since their tuning curves intersect at a different odor. I will make this example more specific in the next sections. In all these examples, synchrony patterns reflect the structure of the stimulus. This idea can be formalized by describing a structured stimulus S as the image of a lower-dimensional object X through some transformation T: S = T (X) (Fig. 7D). In the binaural hearing example, X is the source signal, S = (SL, SR) is the binaural signal, and T is the acoustical transformation: T (X) = (X (t−dL), X (t−dR) ). In the olfactory example, X is the time-varying concentration c (t), S is the time-varying coverage of all receptors (a time-varying 1000-dimensional vector), and T is the transformation c (t) →aO. c (t), where aO is the vector of affinities of all receptor types to the presented odor. Table 1 describes other examples in this framework. This structure introduces synchrony in all neurons whose receptive fields match when combined with the transformation T: NB ∘ T = NA ∘ T (composition of mappings), where NA and NB are the receptive fields of the two neurons. In the olfactory example, this means that the product of odor affinity and sensitivity is the same for neurons A and B (aiO. si = ajO. sj); in the binaural hearing example, this means that the combination of acoustical and neural delays match on both sides. This identity defines a synchrony partition that reflects the structure of stimuli (induced by the transformation T), independently of the source X (e. g. the time-varying concentration). This is an appealing property from a computational point of view, because stimulus structure has natural invariances: for example, binaural structure depends on source location, but not on source signal; in olfaction, structure is independent of concentration. These invariances appear in the synchrony partitions, even though neurons have heterogeneous properties and their individual responses vary with many aspects of stimuli. We now look at the computational properties of these structural codes, taking the example of olfaction. Fig. 8 shows odor-specific synchrony in a simple olfactory model, corresponding to the situation shown in Fig. 7C, with an odor in the SRF of neurons B and C. Odor concentration c (t) varies randomly with turbulences, and receptor coverage depends on concentration and receptor type: receptor type 2 (neurons A and B) is more sensitive to the presented odor than receptor type 1 (neuron C). The odor is then transduced into a current, which produces spikes. The transduction current is modeled as a Hill function of receptor coverage: I = Imax*Hn (s. c) (Fig. 8, middle), where Imax is the maximum current, c is the odor coverage, s is the global sensitivity (inverse of the half-activation coverage) and n is the Hill coefficient, related to the slope of the curve [64]. The Hill coefficient is not very variable, but s can vary 100-fold among olfactory sensory neurons expressing the same olfactory receptor: here, neuron B has a higher sensitivity than neuron A. Thus, the transduction current is essentially determined by the quantity a. s, where a is the affinity of the receptor type to the presented odor. Here, neuron B has a higher affinity to the odor than neuron C, but its global sensitivity is lower, so that the transduction current is the same. As a result, the neurons fire in synchrony (black traces in Fig. 8, bottom; neurons were modeled as integrate-and-fire models). On the other hand, neuron A has the same global sensitivity as C but different affinity and thus does not fire in synchrony (red dashed trace). Synchrony is independent of odor concentration. Let us now consider a population of olfactory neurons (Fig. 9). Each odor is represented by a random vector of affinities and odor concentration is modeled as a half-wave rectified low-pass filtered noise. Receptors and postsynaptic neurons are noisy integrate-and-fire models with random global sensitivity (see Methods). Each odor induces a specific synchrony partition in receptors (Fig. 9A, top). The color represents the product of their odor affinity and global sensitivity, therefore as in Fig. 2, two receptor neurons with the same color fire in synchrony to the presented odor. These patterns can be decoded by postsynaptic neurons, which receive inputs from neurons in the same synchrony group (Fig. 9A, bottom). When odor A is presented (Fig. 9B, first column), postsynaptic neurons wired to the specific synchrony pattern of odor A fire. When odor B is presented, the corresponding postsynaptic neurons fire (Fig. 9B, second column), but neurons tuned to A do not fire, because they do not see synchronous inputs. On the other hand, most receptor neurons fire in both cases, because of their broad tuning. One interesting aspect of mammalian olfaction is that mammals can recognize odors at concentrations that they were not previously exposed to [65]. This invariance to odor intensity is also a natural property of synchrony-based computation, because synchrony receptive fields are invariant to intensity (Fig. 9B, third column), that is, the synchrony partitions (Fig. 9A, top) do not change when intensity varies, even though individual neural responses may change. This simply reflects the fact that the structure of the stimulus (constant ratios of time-varying coverage of different receptor types, as shown in Fig. 7F and 8), which is encoded by synchrony partitions, is intrinsically concentration-invariant. Another interesting computational property is noise tolerance. When a distracting odor is presented at the same intensity as the target odor, postsynaptic responses are reduced but still odor-specific (Fig. 9B, fourth column). The firing rate is reduced because noise reduces the probability of coincidences, but noise does not increase firing in other odor-specific assemblies, because these neurons receive incoherent inputs. Indeed, by construction, postsynaptic neurons fire when they see coincidences that are unlikely to be caused by chance. Therefore, false alarms (firing of neurons tuned to B) are rare, while neurons tuned to A fire when the signal-to-noise ratio is high enough. This is similar to a strategy described as “listening in the dips” in speech recognition in noise [66]. When two known odors are simultaneously presented, both can be recognized by this principle (Fig. 9B, last column). It should be stressed the reduction in firing rate of the neurons tuned to A when A+C or A+B is presented is not due to an inhibitory mechanism. There is no inhibition in this model. Instead, it is due to a desynchronization of the inputs caused by the distracting odor. A neuron tuned to an odor responds less when another odor is added because the operation that the neuron performs is detecting similarity between sensory signals rather than adding them. For example, suppose the stimulus produces two sensory signals x1 and x2, and the postsynaptic neuron fires when x1 = x2. If another stimulus (y1, y2) is added and the neuron is not tuned to it (y1≠y2), then x1+y1≠x2+y2 and the neuron does not respond. This reduction in firing rate occurs even though all presynaptic receptors fire more (i. e. , x1+y1>x1 and x2+y2>x2). Fig. 9C shows the distribution of firing rates and coefficients of variation in the receptors and postsynaptic neurons, when odor A is presented. The peak in the firing rate distribution indicates that many receptors saturate. As previously discussed (Fig. 1), the sensitivity of postsynaptic neurons to coincidences depends on the level of intrinsic noise. In Fig. 9B, the noise level was σ = 0. 15 (relative to the spike threshold). In Fig. 9D, the noise level was varied between 0 and 0. 5 and the model was presented with odor A (as in Fig. 9B, first column). The top graphs show the average firing rate of the postsynaptic neurons tuned to A (blue) and to B (red) as a function of noise level. It appears that only neurons tuned to A respond when the noise level is lower than about σ≈0. 3. The bottom panel shows the responses of both groups for the highest noise level (σ = 0. 5). We note that, although the firing rates of both groups are similar, the temporal structures are very different - that is, the responses of neurons tuned to A are more coherent. In Fig. 10A, I consider a mixture of two odors A and B and a postsynaptic assembly tuned to the equal mixture (50% A, 50% B). The average firing rate varies with the concentrations of both odors in the mixture and in contrast with Fig. 9B, the presented mixture is always highly correlated with the target mixture. Fig. 10A shows that the assembly responds best when there is an equal proportion of A and B in the mixture, at all concentrations (varying by a factor 100). Although selectivity is broader at the highest concentration, the assembly still responds more to the target mixture at the lowest concentration than to either odor A or B at the highest concentration (×100). Odors A and B are bound into a single mixture because their fluctuations are coherent. If the same odors are simultaneously presented but as a mixture of two independent plumes with their own fluctuations (two different turbulent flows representing two different odor sources), then the network does not bind them together and the assembly does not respond (Fig. 10B). Thus the model implements the idea of binding by synchrony, where precise spike timing acts as a “signature” of an object [24]. More precisely, since neural responses follow the temporal structure of the stimulus, precise coincidences can only detected between neurons that respond to the same stimulus. This is a weak version of binding by synchrony, in the sense that the temporal “signature” is intrinsic to the stimulus rather than created as a result of object formation. One interesting aspect of synchrony-based computation is that it assigns a functional role to the variability of individual neural properties – here, the variability in sensitivity of olfactory neurons. I have not considered variability in other neural parameters, which may be an issue. In Fig. 9, all receptors had the same membrane time constant (20 ms). If it is made heterogeneous (Fig. 10C, τ = 15–25 ms) and the synaptic projections are unchanged, then postsynaptic neurons see fewer coincidences and fire less (Fig. 10C, initial wiring). This affects the rate but not the specificity of the responses. We may redefine the synaptic projections to take this heterogeneity into account: for example, in Fig. 10C (new wiring), for each postsynaptic neuron, we only choose presynaptic neurons with membrane time constants that differ by less than 5 ms (as well as similar sensitivity to the target odor). As a result, the firing rate is increased and the specificity is unchanged. Finally, the specific wiring I have described can be learned by synaptic plasticity mechanisms, as explained for the duration model (Fig. 4). In Fig. 11, the two odors A and B were randomly presented to the olfactory model, with random synapses between receptors and postsynaptic neurons (50 synapses per postsynaptic neuron). The presented odor is updated every 200 ms, for a total duration of 40 s. The synaptic weights evolve according to the same homeostatic and synaptic plasticity mechanisms as for the duration model (Fig. 4). At the end of the stimulation, a tuning ratio is calculated for each neuron, as the proportion of spikes in response to odor A, over the second half of the stimulation. That is, a tuning ratio of 0 means that the neuron only responds to odor A, while a tuning ratio of 1 means that it responds only to odor B. Fig. 11A shows the distribution of tuning ratios of the postsynaptic neurons. All neurons but one have tuning ratios clustered near 0 or 1, that is, they are tuned to a single odor. The neurons are then ordered by tuning ratio, and they are presented with odor A with an increasing concentration, then with odor B (Fig. 11B). The concentration varies between 0. 1 and 10 (bottom), where 1 is the concentration during the learning phase. It appears that odor selectivity is preserved at all tested concentrations. Fig. 11C shows the voltage traces of a neuron tuned to odor B, when odor A (left) and B (right) are presented (spikes are added for readability). The membrane potential has standard deviation 0. 17 (odor A) and 0. 18 (odor B, calculated without the spikes), and mean 0. 08 (A) and 0. 07 (B). Thus, the membrane potential distributions are similar for the preferred and non-preferred odors: the increased firing is due to transient synchrony events rather than changes in input statistics. This olfactory model shares a few ideas with a spike-timing-based model previously proposed by Brody and Hopfield [55], in particular, odor-specific neurons detect an equality between different quantities by means of synchrony detection. There are a few differences: 1) in the Brody-Hopfield model, the input to encoders is a constant signal, 2) this constant is a logarithmic function of concentration, 3) it is translated into phase by an intrinsic oscillation. The model I have presented has conceptual similarities, but makes weaker hypotheses. First, the input is time-varying instead of constant. Second, the transformation between concentration and input current must be similar across receptors, but it can have an arbitrary form. Third, the transformation from signal to spike times does not rely on an intrinsic oscillation but on the input signal itself. It is less restrictive, because 1∶1 phase-locking occurs under specific conditions. However, adding an internal oscillation to the stimulus-locked signal would also work in the present model, if it is shared across encoding neurons (as shown in Fig. 6). Finally, I will show how the concepts I have exposed apply to a few auditory and visual examples (Fig. 12). In Fig. 7A, I illustrated the notion of structured stimulus in a simplified description of binaural hearing, where the sound arrives at the two ears with an interaural delay that depends on the source direction. In reality, bina ural cues are more complex because the sound is diffracted by the head and pinnae, and even the body (Fig. 12A). The correct physical description is that the two monaural signals are two linearly filtered versions of the original signal S: SL = FL*S, SR = FR*S (* is the convolution). These location-specific filters are called head-related impulse responses and are more complex than pure delays (in particular, ITD is frequency-dependent [67]). I consider two monaural neurons A and B on opposite sides with different receptive fields NA and NB. These neural filters represent basilar membrane filtering around a characteristic frequency (CF), and include an outgoing axonal delay. Thus, they may differ both in CF and in axonal delay. In the framework I have described, these two neurons have synchronous responses when NA*FL*S = NB*FR*S, that is, their SRF includes all acoustical filter pairs (FL, FR) such that NA*FL = NB*FR, meaning that the combination of neural and acoustical filtering match on both sides. Therefore this is a spatial field, and synchrony signals source location independently of source signal. A spiking neural model based on these properties can accurately estimate the location of previously unheard sounds in a realistic virtual acoustic environment [68]. This corresponds to the idea that the tuning properties of binaural neurons may come not only from mismatches in axonal delay but also in the preferred frequency of their monaural inputs [69]–[71]. A prediction from this theory is that the preferred ITD of a binaural neuron can depend on sound frequency, because ITDs depend on frequency when diffraction is taken into account [67]. This property has indeed been observed in binaural neurons of many species [72], [73]. More specifically, the theory predicts that the frequency-dependence of preferred ITDs should match the corresponding quantities in the acoustical filters, which can be measured. We may also look at the SRF of two monaural neurons on the same side (Fig. 12B). In Fig. 7B, I only considered neurons with identical auditory filters but different delays. If the two neurons have different filters NA and NB, then synchrony occurs when NA*S = NB*S (S is the sound). Looking at this identity in the frequency domain, this means that the two filters must agree at all frequencies where the sound S has power. The synchrony identity means that both the phase and the amplitude spectrum of the filters must agree at all frequency components of S. If the two neurons have different CFs, then this can only occur at the frequency f0 where the two amplitude spectra agree. In addition, the phases agree if the difference in delays exactly compensates the difference in phase delays of the filters. Therefore, the SRF is a pure tone with frequency f0 - or a resolved partial harmonic of a complex sound (that is, only one frequency component falls in the bandwidth of filters NA and NB). In summary, only one specific type of sound elicits patterns of synchrony in monaural neurons: periodic sounds, which are associated with pitch in humans. This produces a new theory of pitch perception, generalizing temporal models, according to which pitch is represented by the pattern of synchrony across frequency (CF) and time (axonal delay). It offers a solution to the two major problems of temporal models of pitch: 1) that they require large axonal delays (as large as the maximum period of a pitch-evoking sound, about 30 ms), 2) that they do not distinguish resolved and unresolved harmonics, while there is a perceptual difference between these two types of pitch-evoking sounds [74]. In the synchrony pattern hypothesis, large axonal delays are not necessary, because mismatches in CF can play this role, and resolved harmonics produce wider synchrony patterns, as they also include neurons with different CFs. The analog of binaural hearing in vision is binocular disparity (Fig. 12C). Consider two retinal ganglion cells in different eyes, which move with microsaccades (tremor). The two cells fire in synchrony when they see the same dynamic stimulus through their receptive field (retinal ganglion cells are known to fire with millisecond precision [5], [75], [76]). This occurs when there is an object at the convergence point of their fixation lines (connecting their retinal position to the pupil). Thus the synchrony receptive field is a three-dimensional spatial receptive field. Synchrony patterns across the two eyes reflect the structure of the binocular stimulus, which comes from the fact that a single object produces the two retinal images. The hypothesis that depth perception is mediated by the detection of synchrony between two retinal ganglion cells (presumably by a neuron in V1) predicts that decorrelating the images in the two eyes should disrupt depth perception. In a similar way, the SRF of two monocular visual neurons with circular receptive fields (e. g. neurons in the lateral geniculate nucleus of the thalamus) is the set of images that are unchanged by translations of the vector connecting the centers of the two receptive fields (Fig. 12D). These are edges with the same orientation and spatially periodic textures with the period given by that vector. To understand the functional role of synchrony, I introduced the concept of synchrony receptive field: the set of stimuli that produce synchronous responses in a given neuron pair or group. In a heterogeneous population of neurons, synchrony reflects the structure of stimuli, for example a constant activation ratio between two olfactory receptors responding to an odor. This structure can then be detected by postsynaptic neurons which are sensitive to synchrony. This framework applies to many perceptual tasks, such as recognizing an odor or locating a sound source. I will first comment these results from a computational perspective, and then discuss the biological plausibility of this proposition. Over the last century, the operating function of neurons has been mainly described in terms of firing rates, and this point of view has led to important developments in computing, from the perceptron [77] to modern artificial neural network theories for pattern recognition [78]. More recently, experimental evidence and theoretical studies, showing the importance of the temporal coordination of spikes [5], [13], have triggered considerable interest for spiking neuron models in computational neuroscience [79]. However, few theories of computation are specifically spike-based, as I have proposed here. Fig. 5B illustrates a fundamental difference between synchrony-based computation and traditional neural network theory: a formal neuron (e. g. perceptron) fires when the stimulus is on one side of a hyperplane, while two neurons fire in synchrony when the stimulus is close to a hyperplane. In this framework, synchrony in neurons with heterogeneous receptive fields reflects some structure in the stimulus (for example, the periodicity of a pitch-evoking sound). The computational interest stems from the fact that structure is invariant to many aspects of the stimulus: for example, receptor coverage ratios are invariant to odor concentration in a turbulent odor plume, and binaural cues in sound localization are invariant to the signal produced by the source. This computational principle applies to many perceptual tasks in all sensory modalities, and it may also apply to the exploration of sensorimotor contingencies [80]. Robustness to noise stems from the fact that incoherent signals result in an absence of response (no synchrony) rather than in a false response. This relates to the idea that meaningful structure in an image (e. g. edges) is what could not occur by chance in a random image, a principle called “Helmholtz principle” that was recently successfully applied in computer vision [81]. At behavioral level, invariance is a striking aspect of perception: translation invariance in vision [82], concentration invariance in olfaction [65], acoustic scale invariance in hearing [83]. On the other hand, neural responses often vary with many aspects of sensory stimuli. This theory agrees with these observations because the spatial structure of synchrony is invariant but individual neural responses are variable. It relies on two main assumptions: that neurons can synchronize to a similar signal, and that postsynaptic neurons can detect this synchrony. Both properties are seen when neurons are fluctuation-driven (rather than mean-driven), which is in agreement with the temporal irregularity of spike trains in vivo [84] and with direct intracellular measurements [40]. Spike timing reproducibility has been observed in vivo in early sensory areas [5], [58], but also more recently in the sensory cortices, although at longer timescales [85], [86]. The sensory examples I have chosen are all thought to be processed in subcortical areas, at least for the neurons for which SRFs are defined: odor recognition in the olfactory bulb, sound localization and pitch perception in the auditory brainstem, binocular disparity in the retina and thalamic relay cells (with coincidence detection in the primary visual cortex). There is stronger evidence for the reproducibility of spike timing in these subcortical areas. However, for this theory, stimulus-locked reproducibility is a sufficient but not necessary condition: as I previously remarked (Fig. 6), there may be stimulus-specific synchrony without trial-to-trial reproducibility, if there is a shared source of variability (e. g. activity of the local network, or feedback from other areas). Finally, I have shown that the neural circuits that detect structure-specific synchrony can spontaneously emerge under the effect of spike-timing-dependent plasticity. This was expected because modeling studies have shown that STDP tends to select correlated inputs [20], [21]. In many theories of spiking neural networks (with the notable exception of liquid state machines [23]), neural heterogeneity is seen as a source of noise to be averaged out: the unit of computation is a neural population or “neural mass” [22]. On the contrary, in this theory, it is specifically because of neural heterogeneity that synchrony carries meaningful information. This has some interest for neuroengineering. Indeed, a major problem in low-consumption neuromorphic circuits is that there is substantial variability in neuron properties, which makes it difficult to specify a precise neuron model [87]. In the framework I have described, this variability can be exploited. How can this theory be experimentally tested? I have mentioned a few predictions in the specific cases of ITD processing and binocular disparity. More generally, a straightforward approach is to measure synchrony receptive fields, by examining how the cross-correlogram of a given neuron pair varies with stimuli, and in particular with the structure of the stimulus, using multielectrode recordings. Previous studies in the olfactory and visual systems support the idea of stimulus-specific synchrony [88], [89], but new experiments should specifically test whether synchrony is related to the structure of the stimulus, for example whether it is robust to changes in intensity. Neurons with rebound spiking are modeled with the following membrane equation: where v is the membrane potential, τ is the membrane time constant (20 ms in Fig. 1, random between 10 and 50 ms in Fig. 2–4), gKLT is the low-threshold K+ conductance (in units of the leak conductance), gKD is the delayed-rectifier K+ conductance, gin (t) is the inhibitory synaptic conductance, El = −35 mV is the leak reversal potential and EK = −90 mV is the K+ reversal potential (note that the resting potential is smaller than El because of the low-threshold K+ conductance). The low-threshold K+ conductance depends on voltage through the following equation: where τKLT is the time constant (100 and 400 ms for neurons A and B in Fig. 1, random between 300 and 500 ms in Fig. 2), Va = −70 mV is the half-activation voltage, ka = 5 mV is the activation Boltzmann factor and gKLT* is the maximal conductance (1 in Fig. 1, random between 1 and 1. 4 in Fig. 2–4). Thus this hyperpolarizing conductance increases with voltage. A spike is produced when v reaches vt = −55 mV, then the membrane potential is reset to −70 mV and the delayed-rectifier K+ conductance is set to gKD = 2. This conductance then decays exponentially: where τKD = 300 ms. This prevents the neuron from producing bursts of spikes. Synaptic conductances are pulses of amplitude gin = 5 (in units of the leak conductance) and variable duration. This choice of parameter values is explained in Text S1. Coincidence detectors are modeled as noisy integrate-and-fire models: where τ = 5 ms is the membrane time constant, n (t) is a filtered noise with standard deviation σ = 0. 2 (ξ (t) is white noise). The resulting standard deviation of the membrane potential v is. In Fig. 2, each presynaptic spike increases v by an amount 1/N, where N is the number of presynaptic neurons, and a spike is produced when v = 1. The 1/N scaling factor ensures that the postsynaptic neuron fires with probability 1/2 when inputs are synchronous. After spiking, the membrane potential is reset to 0. In Fig. 3B, inputs are modeled as synaptic conductances rather than currents, i. e. , where τe = 2 ms is the excitatory time constant and Ee = 4. 7 is the excitatory reversal potential (relative to the threshold; this corresponds to Ee = 0 mV for a threshold vt = −55 mV and El = −70 mV). Each presynaptic spike increases ge by an amount α/N, where α is calculated so that the PSP produced by a conductance increase of size α reaches the spike threshold vt = 1, with the approximation that the synaptic driving force is Ee−1/2 (1/2 being the average of the resting potential 0 and the spike threshold 1). This gives the following formula: In Fig. 3C, the coincidence detectors are described by the same equations as the model with rebound spiking, with τKLT = 400 ms, gKLT* = 2. 1, τ = 10 ms, and inputs are also modeled as synaptic conductances, with Ee = 0 mV. The noise is scaled so as to represent the same proportion of the difference between resting potential and threshold (which gives 1 mV). Each presynaptic spike increases ge by an amount α/N, where α is calculated as above, but taking into account the total conductance of the cell at rest (leak plus K+) and the resting potential (empirically determined as mV). The resulting formula is: where is the K+ conductance at rest (calculated from the activation curve) and. Synaptic weights w evolve with homeostasis and spike-timing-dependent plasticity (STDP). Homeostasis is defined by: STDP is defined by a modification of synaptic weight that depends on the timing of pre- and postsynaptic spikes: The time constant is ms for the duration model and 3 ms for the olfaction model. In Fig. 4, (where I is the duration of a stimulus presentation), and. In Fig. 11, , and. The fluctuations of concentration in an odor plume are described by a half-wave rectified Ornstein-Uhlenbeck process: with time constant τ = 75 ms [63], where the odor concentration is proportional to [x]+ (= max (x, 0) ). Each of the N = 5000 olfactory receptor neurons has an odor-specific affinity which depends on its type, and a global sensitivity, which is neuron-specific. Thus, each odor can be represented as an N-dimensional vector of binding coefficients bi, combining affinity and sensitivity (bi = ai. si). To generate an odor, we draw random binding coefficients logarithmically distributed between 10−3 and 103. The transduction current of a receptor cell is a Hill function of odor concentration: where c is the (time-varying) concentration, n = 3 is the Hill coefficient, related to the slope of the curve [64], Imax is calculated to produce a maximum firing rate of 40 Hz, and K1/2 is the half-activation concentration, which is the inverse of the binding coefficient: K1/2 = 1/bi. The concentration varies in time as c (t) = c0. x (t) (x (t) are the random fluctuations defined above). Note that this latter parameter depends on both the neuron and the odor. Currents are transformed into spike trains through an integrate-and-fire model: where τ = 20 ms is the membrane time constant (except Fig. 7C, where it is uniform between 15 and 25 ms), and I (t) is the transduction current. A spike is produced when v = 1, then the membrane potential is reset to 0. Coincidence detectors are defined as for models of duration selectivity, with τ = 8 ms and σ = 0. 15. In Fig. 6,400 such postsynaptic neurons are split in two groups tuned to either odor A or B. Each postsynaptic neuron receives excitatory synapses from presynaptic neurons with similar binding coefficients for the target odor. Specifically, the range of binding coefficients is divided in 200 equal layers (in logarithmic scale), and each layer is associated with one postsynaptic neuron, which receives inputs from all receptors with binding coefficients in that layer. The synaptic weight is 1/n, where n is the number of presynaptic neurons (12. 4±3. 6). In Fig. 10C (new wiring), odd index neurons only receive inputs from receptors with τ<20 ms and even index neurons from those with τ>20 ms. We compensate by doubling the size of layers for binding coefficients, so that the average number of presynaptic neurons is unchanged. Precision and reliability measures (Fig. 6) are obtained from shuffled auto-correlograms (SAC) [14] - the average cross-correlogram between distinct trials. These are normalized by, where is the time bin and D is the duration of trials. After removing the baseline (equal to r2, where r is the firing rate), the precision is defined as the half-width of the SAC, and the reliability as the normalized integral of the peak: which gives a number between 0 and 1, where 0 is obtained for independent spike trains and 1 when comparing a spike train with a jittered copy (i. e. , perfect synchrony if the timescale is 0 ms). This corresponds to the total correlation coefficient in [91]. In Fig. 7, stimuli, shared input and private noise are generated as Ornstein-Uhlenbeck processes with time constant 10 ms. Stimuli and shared inputs have the same standard deviation and that of the private noise is set by the signal-to-noise ratio. Neurons are modeled as integrate-and-fire units: with τ = 10 ms and I (t) is the total input. The spike threshold is 1 and the reset is 0. Shuffled and cross-correlograms are calculated as in Fig. 6 (previous paragraph), averaged over many trials.

How does the brain compute? Traditional theories of neural computation describe the operating function of neurons in terms of average firing rates, with the timing of spikes bearing little information. However, numerous studies have shown that spike timing can convey information and that neurons are highly sensitive to synchrony in their inputs. Here I propose a simple spike-based computational framework, based on the idea that stimulus-induced synchrony can be used to extract sensory invariants (for example, the location of a sound source), which is a difficult task for classical neural networks. It relies on the simple remark that a series of repeated coincidences is in itself an invariant. Many aspects of perception rely on extracting invariant features, such as the spatial location of a time-varying sound, the identity of an odor with fluctuating intensity, the pitch of a musical note. I demonstrate that simple synchrony-based neuron models can extract these useful features, by using spiking models in several sensory modalities.

Abstract Introduction Results Discussion Methods

computational neuroscience biology sensory systems neuroscience

2012

Computing with Neural Synchrony

Entamoeba histolytica is responsible for causing amoebiasis. Polyamine biosynthesis pathway enzymes are potential drug targets in parasitic protozoan diseases. The first and rate-limiting step of this pathway is catalyzed by ornithine decarboxylase (ODC). ODC enzyme functions as an obligate dimer. However, partially purified ODC from E. histolytica (EhODC) is reported to exist in a pentameric state. In present study, the oligomeric state of EhODC was re-investigated. The enzyme was over-expressed in Escherichia coli and purified. Pure protein was used for determination of secondary structure content using circular dichroism spectroscopy. The percentages of α-helix, β-sheets and random coils in EhODC were estimated to be 39%, 25% and 36% respectively. Size-exclusion chromatography and mass spectrophotometry analysis revealed that EhODC enzyme exists in dimeric form. Further, computational model of EhODC dimer was generated. The homodimer contains two separate active sites at the dimer interface with Lys57 and Cys334 residues of opposite monomers contributing to each active site. Molecular dynamic simulations were performed and the dimeric structure was found to be very stable with RMSD value ∼0. 327 nm. To gain insight into the functional role, the interface residues critical for dimerization and active site formation were identified and mutated. Mutation of Lys57Ala or Cys334Ala completely abolished enzyme activity. Interestingly, partial restoration of the enzyme activity was observed when inactive Lys57Ala and Cys334Ala mutants were mixed confirming that the dimer is the active form. Furthermore, Gly361Tyr and Lys157Ala mutations at the dimer interface were found to abolish the enzyme activity and destabilize the dimer. To our knowledge, this is the first report which demonstrates that EhODC is functional in the dimeric form. These findings and availability of 3D structure model of EhODC dimer opens up possibilities for alternate enzyme inhibition strategies by targeting the dimer disruption. Amoebiasis is an infectious disease caused by single-celled parasitic protozoan Entamoeba histolytica. Parasitic amoeba infects liver and intestine, which may cause mild diarrhea and serious dysentery with bloody and mucoid stool. If untreated, the parasite can cause life-threatening hemorrhagic colitis and/or extraintestinal abscesses. E. histolytica is responsible for over 50 million infections in tropical and temperate regions, and nearly 100,000 deaths worldwide each year [1], [2]. The parasite mainly affects primates and humans, and is transmitted by ingestion of water and food contaminated with feces containing E. histolytica cysts. First-line amoebiasis treatment is anti-amoebic therapy that relies on a very small number of drugs such as metronidazole, emetine, tinidazole and chloroquine [3]–[5]. These drugs target different stages of the life cycle of E. histolytica. Frequent and widespread usages of these drugs have led to the increase in the minimum inhibitory concentration (MIC) values and also development of clinical drug resistance in pathogen. Some of these drugs have been reported to have significant side effects. For instance, metronidazole, an effective drug for amoebiasis, has been reported to be tumorigenic and mutagenic [6]–[8]. Nitrazoxanide, a broad spectrum anti-parasitic drug used for amoebiasis treatment, is found to be associated with many side effects [9], [10]. Consequently, development of alternate strategies and discovery of new anti-amoebic agents targeting polyamine synthesis is necessary to combat the disease. Ornithine decarboxylase (ODC), a Pyridoxal 5′-phosphate (PLP) dependent homodimeric enzyme catalyzes the first rate-limiting step of polyamines biosynthetic pathway by decarboxylation of L-ornithine to form putrescine (Figure 1). Polyamines have an eminent role in various cell growth and differentiation processes [11], [12]. Consequently, ODC being the key enzyme of the polyamine biosynthetic pathway is a promising therapeutic target for anti-protozoan therapy. The ODC enzyme has been reported to be present in various protozoa including Leishmania, Trypanosoma, Giardia, and Plasmodium and is a validated drug target in Trypanosoma brucei for treatment of African sleeping sickness [13]–[18]. ODC enzyme has a very short half-life due to its ubiquitin-independent 26S proteasome mediated degradation which is stimulated by the binding to antizyme [19]. Besides increase in ODC proteolysis, interaction of antizyme with ODC leads to catalytic inactivation of the enzyme by disrupting the enzymatically active ODC dimers [19], [20]. In addition, the antizyme binding loop which is accessible in ODC monomer is found to be buried in the dimers of ODC that ultimately prevents it from degradation. Thus, dimer formation is not only important for its catalytic function but also for its protection against antizyme-dependent endoproteolysis. Crystal structures of ODC enzyme from T. brucei (PDB ID: 1QU4), human (PDB ID: 2OO0), and mouse (PDB ID: 7ODC) have revealed that the monomeric subunits interact in head to tail manner and form two catalytic active sites at the dimer interface [21]–[23]. The structure of ODC in complex with substrate and product analogues including ornithine analog α-difluoromethylornithine (DFMO) have been investigated [21]. DFMO is a suicide inhibitor of ODC and has been reported to inhibit growth of various pathogenic protozoan parasites such as Giardia lamblia [14], Trichomonas vaginalis [24], Plasmodium falciparum, and various Trypanosoma species [13], [18]. In E. histolytica, the only enzyme of polyamine biosynthesis reported to exist is ODC. E. histolytica ODC (EhODC) has been reported to form homopentamers [25]. Interestingly, EhODC is insensitive to DFMO and DFMO has no inhibitory effect on the cell growth of the parasite [25]–[27]. Therefore, it is necessary to develop an alternate method for inhibition of EhODC enzyme for targeting the polyamine biosynthetic pathway to curb the disease. In the present work, we have re-investigated the oligomeric state of EhODC using biochemical, mutational and in silico methods. Previously, it has been reported that the EhODC enzyme exists only as a homopentamer [25]. However, our studies evidently demonstrate that EhODC is functionally active in the dimeric form. In the absence of crystal structure of EhODC, we have generated 3D model of EhODC homodimer to structurally characterize the dimer interface containing two active sites and have performed molecular dynamics simulations to verify the dimer stability. Our investigation yields that disruption of dimer disrupts the active site pocket and renders the enzyme inactive. 3D structure model of EhODC homodimer may be beneficial in designing structure based anti-amoebiasis peptides or agents that would disrupt enzyme dimerization. We propose that a compound having the capability to disrupt the dimer could be a good candidate for amoebiasis treatment. The E. coli expression vector pET 30a (Novagen) containing full-length gene of EhODC having N-terminal Histidine tag (6× His) followed by enterokinase cleavage site was used for over-expression of the enzyme [26]. Oligonucleotides for site directed mutagenesis were ordered from Imperial Life Sciences (India). Restriction endonuclease DpnI and Phusion DNA polymerase were acquired from New England BioLab Inc. For protein purification, 5 ml HisTrap HP and HiLoad 16/60 Superdex 200 gel filtration columns were obtained from GE Healthcare. Imidazole (low absorbance at 280) was obtained from Acros. ÄKTA Prime plus system from GE Healthcare was used for protein purification. Putrescine, 4-aminoantipyrine, diamine oxidase (DAO), horseradish peroxidase, L and D-ornithine were procured from Sigma Aldrich. Amicon ultra protein concentrators were purchased from Millipore. All other chemicals were of analytical grade and obtained from commercial sources. The expression and purification of EhODC enzyme was done by following the published procedure with minor modifications given below [26]. The plasmid pET30a having the full-length EhODC gene insert (pET30a-EhODC) was transformed into freshly prepared E. coli BL21 (DE3) competent cells and plated on Luria-Bertani (LB) agar plate containing kanamycin (50 µg/ml). Plates were incubated overnight at 37°C and colonies were obtained. Single colony was picked and cells were seeded in 5 ml LB broth containing 50 µg/ml of kanamycin and culture was grown overnight at 37°C with agitation. Overnight culture was used for inoculation of 1 L LB broth. Expression was induced with 1 mM isopropyl β-D-thiogalactoside (IPTG) when optical density (A600) reached 0. 6. After induction, culture was moved to 18°C and was grown for ∼14 h. Cells were harvested by centrifugation at 5,000 rpm at 4°C for 10 min and cell pellets were stored at −80°C until further processing. Expression and solubility of the protein was confirmed by analysis of lysed cell supernatant and pellet on 12% sodium dodecyl sulfate-polyacrylamide gel electrophoresis (SDS-PAGE). The histidine-tagged EhODC was purified using a two step procedure that employed metal ion affinity chromatography followed by gel filtration chromatography. All purification steps were performed at low temperature (4°C–6°C). Briefly, frozen cell pellets from a 1 L culture were thawed on ice and re-suspended in buffer A [50 mM Tris-HCl (pH 7. 5), 40 mM imidazole, 250 mM NaCl and 5% glycerol (v/v) ] containing lysozyme (0. 7 mg/ml) and 0. 2 mM phenylmethanesulfonyl fluoride (PMSF). Cells were disrupted by sonication on ice with a pulse of 20 s on and 1 min off for 10 times. The obtained cell lysate was clarified by centrifugation at 18,000 g for 45 min at 6°C and supernatant was applied on HisTrap HP column (5 ml) pre-equilibrated with buffer A. Unbound proteins were removed by washing the column with ∼40 ml of buffer A. Bound protein fractions were eluted using a linear gradient of 40 mM to 1 M imidazole of 60 ml at a flow rate of 1 ml/min. Eluted fractions were examined on 12% SDS-PAGE and fractions containing pure protein were pooled together. To remove the N-terminal His-tag, enterokinase was added to pure protein (∼0. 02 units/mg protein) and incubated for ∼12 h at 4°C and simultaneously dialyzed against buffer A without imidazole. To remove uncleaved tagged protein and the cleaved His tags, the sample was reloaded onto HisTrap HP column and the flow-through containing untagged EhODC was collected and concentrated using a 10 kDa cutoff Amicon Ultra-15 concentrator (Millipore, Bedford, Massachusetts, USA). For removal of enterokinase, the concentrated sample was loaded onto HiLoad 16/60 prep grade Superdex 200 size-exclusion chromatography column pre-equilibrated with buffer B [50 mM Tris-HCl (pH 7. 5), 250 mM NaCl, 0. 2 mM dithiothreitol (DTT) and 5% glycerol (v/v) ]. Fractions of the major peak containing pure protein were pooled and concentrated. Homogeneity of the concentrated enzyme preparation was analyzed by 12% SDS-PAGE. The yield and concentration of purified EhODC was measured using the Bio-Rad protein-assay kit with bovine serum albumin (BSA) as a standard. EhODC mutant proteins were expressed and purified using the same protocol. Ornithine decarboxylation activity of EhODC was spectrophotometrically determined by the method developed by Badolo et al. [28]. This method is based on the reaction between DAO and putrescine, the product of the ODC-catalyzed reaction. For EhODC enzyme assay, the purified protein was buffer exchanged with 20 mM sodium phosphate buffer (pH 7. 5) and concentrated to final concentration of 0. 3 mg/ml. The reaction mixture of 180 µl containing 20 mM sodium phosphate buffer (pH 7. 5), 0. 1 mM EDTA, 0. 1 mM PLP, 0. 2 mM DTT, and 1 mM of L-ornithine was prepared to which 20 µl of protein solution was added to make up the final volume of 200 µl. The reaction mixture was incubated at 37°C for 5 h. Further, 100 µl of the above EhODC reaction mixture was added to 900 µl of diamine oxidase (DAO) reaction mixture composed of 50 mM Tris-HCl (pH 9. 8) containing 100 µg/ml phenol, 100 µg/ml 4-aminoantipyrine (4-AAP), 0. 02 U of DAO, and 7 U of horseradish peroxidase (HRP). The reaction was incubated at 25°C for 60 min and then terminated by heating the solution at 90°C for 4 min. The concentration of putrescine formed by ornithine decarboxylation catalysis was determined by measuring the absorbance at 492 nm for the colored complex formed as a result of the reaction of H2O2 with 4-AAP and phenol catalyzed by HRP. For negative controls, purified protein or substrate L-ornithine were substituted with buffer in the ODC enzyme reaction mixtures. Effect of stereoisomer of substrate was observed by incubation of L and D-ornithine at 37°C. To obtain preliminary information on the oligomeric association of EhODC, glutaraldehyde crosslinking experiment was performed using the method described by Fadouloglou et al. [29]. Purified protein solution was exchanged with 20 mM sodium phosphate buffer (pH 7. 5) for cross-linking studies. Experiment was carried out using 24 well crystallization plate (Hampton research) and a siliconized coverslip in a manner similar to a hanging drop crystallization method. For cross-linking EhODC, 40 µl of 12. 5% glutaraldehyde solution (v/v) acidified with 1 µl 5 N HCl was added in the well of crystallization plate. Then, 15 µl of protein solution (1 mg/ml) was loaded onto the coverslip, which was inverted on the reservoir well and sealed with vacuum grease (Hampton Research). The entire setup was incubated at 37°C for 10 min and then the sample was mixed with an equal volume of 2X SDS-PAGE loading buffer and boiled for 4 min on a dry bath. Cross-linked oligomers were analyzed on 12% SDS-PAGE followed by Coomassie Blue R-250 staining. The molecular mass of recombinant EhODC was determined by running purified protein on 12% SDS-PAGE with standard molecular weight protein marker (Bio-Rad). To analyze the oligomerization state, 500 µl of purified and concentrated (∼10 mg/ml) protein was applied onto a HiLoad 16/60 Superdex 200 gel filtration column pre-equilibrated with buffer B using 500 µl sample loop at a flow rate of 0. 5 ml/min on ÄKTA purifier chromatographic system (GE Healthcare) and protein elution profile was monitored by measuring the absorbance at 280 nm. The size-exclusion column was calibrated with blue dextran (2000 kDa), and Gel Filtration HMW Calibration kit containing ferritin (440 kDa), aldolase (158 kDa), Conalbumin (75 kDa) and ovalbumin (43 kDa) (GE Healthcare) for determination of the void volume, construction of the standard curve and estimation of the molecular weight of purified protein. The oligomerization state of EhODC was also analyzed by matrix-assisted laser desorption/ionization time of flight mass spectrometry (MALDI/TOF MS). The purified protein sample was dialyzed against 50 mM Tris buffer (pH 7. 5) containing low concentration of NaCl (25 mM) and 0. 2 mM DTT to avoid any instrumental interference and was concentrated to ∼2 mg/ml using 10 kDa cutoff Amicon ultra 15 (Millipore). The MALDI/TOF MS analysis was carried out at Proteomics Facility, TCGA (New Delhi, India) using Ultraflex mass spectrometer (Bruker Daltonics, Germany). The protein ionization spectra were analyzed on FLEX-PC2 mass spectrometer and data was acquired across the range of about 0 to 250 amu. To study the effect of urea and NaCl on oligomeric state of protein, purified and concentrated EhODC was pre-incubated with variable concentration (2 M or 4 M) of above chemical agents separately at 4°C for 4 h. The protein was further loaded onto Hi-load 16/60 superdex 200 gel filtration column equilibrated with Buffer B containing the same concentration of urea or NaCl and elution profiles were analyzed. For estimation of secondary structure elements, purified EhODC was subjected to circular dichroism (CD) analysis using Chirascan Circular Dichroism Spectrometer (Applied Photophysics Ltd. , Surrey KT22 7PB, United Kingdom). CD spectra were collected using a 1 mm quartz cell under constant nitrogen purge between 190 to 260 nm in 0. 5 nm wavelength steps and an average time of 3. 0 s at 25°C. The protein solution was buffer exchanged with 20 mM potassium phosphate buffer (pH 7. 5) at 4°C. Protein samples at concentration 0. 35 mg/ml were analyzed and three scans were collected, averaged and the baseline corresponding to the above buffer was subtracted to obtain the final values. The obtained data were analyzed using the software K2d (http: //www. embl. de/~andrade/k2d. html) [30]. The pET30a-EhODC plasmid containing EhODC gene was mutated using the QuikChange XL mutagenesis kit by following the instructions of manufacturer (Stratagene, La Jolla, CA). Mutations were introduced into the synthetic mutagenic oligonuclotide primers and were used for construction of mutant plasmids. Mutations and respective mutagenic primers are listed in table (Table 1). The pET30a-EhODC plasmid was used as a template in the primer extension reaction for constructing the mutants. The reaction mixture used for PCR amplification contained 10 µl of 5X HF phusion buffer supplied with the enzyme, 300 µM of dNTP mix, 6. 25 pmol of each primer, 10 ng of template DNA, 2. 5 U of phusion polymerase, and water was added to make up the final volume of 50 µl. PCR reaction was performed by subjecting the samples to 20 cycles of 30 s denaturation at 95°C, 1 min at annealing temperature as given in Table 1, and 6 min 50 s elongation at 72°C, and finally reaction was completed by doing extension for 15 min at 72°C. PCR products were analyzed on 1% agarose gel electrophoresis. The parent methylated template plasmids were digested with DpnI restriction enzyme at 37°C for 1 h. Digested product was directly used to transform XL-1 Blue competent cells. Transformed cells were plated on LB agar plate containing 50 µg/ml of kanamycin and plates were incubated at 37°C for ∼16 h. The presence of the mutations in the constructed plasmids were confirmed by DNA sequencing using T7 promoter or terminator universal primers at genomic and proteomic facility of TCGA (New Delhi, India). The ODC sequence of E. histolytica was retrieved from NCBI database. Blast and PSI-blast search were performed using AAX35675. 1 as query against the non redundant protein sequence database to identify and analyze orthologous sequences. These homologous sequences were retrieved from the NCBI database and multiple sequence alignment was generated using ClustalW and compared for phylogenetic analysis [31]. Three-dimensional (3D) homology model of EhODC homodimer was generated by comparative modeling using MODELLER 9v8 [32]. To obtain an effectual model, five sequential steps were performed: template selection from Protein Data Bank (PDB), sequence-template alignment, model building, refinement of the obtained model and validation. Template search was done using NCBI BLAST search tool for PDB database [33]. BLASTP algorithm was run with BLOSUM62 as a scoring matrix. Crystal structure of human ODC (PDB ID: 2OO0) which has 34% sequence identity with EhODC was selected as a template for structure modeling [23]. The graphically enhanced alignment with secondary structures were obtained using ESPript 2. 2 server [34]. MULTALIN server was used to align the query sequence with the template sequence [35]. Some manual corrections were done in the alignment file for missing residues in the template sequence. The cofactor, PLP was incorporated into the modeled structure of EhODC from the template structure and five preliminary models were generated using MODELLER 9v8. All models were selected on the basis of lowest DOPE scores and assessed sterio-chemically by PROCHECK [36]. Energy minimizations of the best chosen models were performed using Swiss-PDB Viewer4. 01 (http: //www. expasy. org/spdbv/). Loop refinement module of the MODELLER was used for the refinement of the disorganized residues in loops and refinement process was considered for structure validation. Each refined model was verified using ERRAT plot which gives the measure of structural errors in each model at residue level in the protein (http: //nihserver. mbi. ucla. edu/SAVES/). The refined model was further validated by ProSA energy plot and VERIFY-3D of the SAVES server [36], [37]. All the graphical visualization and image production were performed using PyMOL [38]. Molecular dynamics (MD) simulation of dimeric model of EhODC was performed using GROMACS (v 4. 5. 4) package [39]. GROMOS96 43a1 force field and 47324 SPC water molecules for solvation of protein were used for simulation. The molecule was solvated in a cubic box at a distance of 1. 0 nm between the proteins and the box edge. Electrostatic interactions were calculated using the Particle-mesh Ewald method. Van der Waal and coulomb interactions were truncated at 1 nm. Molecule was neutralized by adding 24 Na+ counter ions to the surface and was allowed to undergo 1000 energy minimization steps. All bond lengths including hydrogen atoms were constrained by the LINCS algorithm. To maintain the system at isothermal and isobaric conditions of 300 K and 1 bar, a V- rescale and Parrinello-Rahman barostat coupling was applied for 100 ps. Following to the equilibration, MD simulation was initiated for 1 ns and then extended to 8 ns, with all trajectories sampled at every 1. 0 ps. The completion of genome sequence project of E. histolytica headed by the Institute of Genome Research (TIGR, Rockville, USA.) opened up the possibilities of new therapeutic targets as well as detailed mechanisms of various biosynthetic pathways [40]. The polyamine biosynthesis in E. histolytica is an essential pathway required for the existence of the pathogen [11], [12]. In present study, the sequence of EhODC, the first and rate-limiting enzyme of polyamine biosynthetic pathway, has been retrieved from NCBI database with accession number AAX35675. The protein consists of 413 amino acids with predicted molecular weight of 46. 43 kDa. In E. histolytica, the gene encoding ODC is of 1242 bp, thus it implies that there is no intron present in the gene. The enzyme has been previously characterized by Jhingran et al. [26]. The amino acid sequence alignment of EhODC with representative ODCs from different sources revealed that the active site residues along with dimer interface residues responsible for dimerization are highly conserved (Figure 2). EhODC showed overall 36 to 39% identity with plants, 15 to 25% with bacteria, 35 to 38% with fungi and 32 to 38% with animals. Interestingly, E. histolytica, being a protozoan was expected to show high sequence identity, but surprisingly it shows same range of identity with other protozoa including T. brucei, Dictyostelium dasciculates and P. falciparum, etc. i. e. 32 to 35%. From phylogenetic tree, the ODC from plants, fungi, and bacteria make different clusters on the basis of sequence homology where as the protozoan ODCs do not cluster together, instead are distributed throughout showing resemblance with bacteria, fungi and plants (Figure 3). However, EhODC shows maximum homology with plant ODCs and the evolutionary origin of EhODC or protozoan ODCs on the basis of phylogenetic analysis is not conclusive. Nevertheless, sequence analysis shows conservation of dimer interface residues which specify the possibility of EhODC enzyme dimerization similar to other ODCs. Further sequence analysis revealed that the substrate binding motif having a consensus sequence WGPTCDGL (I) D is highly conserved in human, mouse and T. brucei and Cys plays a critical role in catalysis. However, in EhODC, though Cys is conserved, but the sequence exists as 330 YGPSCNGSD 338 (Figure 2). The regulation of ODC activity is partially modulated by antizyme-induced, ubiquitin-independent degradation by the 26S proteasome, mainly found in mammals [20], [41]–[43]. Antizyme binds to the inactive ODC monomer forming a hetero-dimer complex which promotes proteolysis degradation [20], [44]. In human ODC, the antizyme binding locus consists of 30 residues at N-terminal ranging from 115Lys to 144Arg residues. The same locus is also highly conserved in mouse. However, this locus in EhODC which corresponds to 105Tyr to 132Lys having 23% identity is not conserved. In this locus, three residues 121Lys, 141Lys and 144Arg (in human ODC) are highly conserved and responsible for antizyme binding [22]. However, in EhODC, 121Lys and 144Arg are substituted by 109Ile and 132Lys respectively. Thus, it may be possible that these differences in sequence makes EhODC insensitive or poorly sensitive to antizyme binding as antizyme dependent ODC degradation has not been reported in E. histolytica till date. Addition to this, in mouse ODC two basal degradation elements (376 to 424 and 422 to 461) at C-terminal are reported which are rich in proline (P), glutamic acid (E), serine (S), and therionine (T) called PEST sequence [23]. In this region, C441 (in both mouse and human ODC) is identified to be a critical residue that promotes polyamine-dependent proteolysis [20], [45]. Similar pattern of sequence arrangement is also observed in EhODC where it ranges from 395 to 413, and conserved Cys400 corresponds to Cys441 in mouse ODC. The recombinant EhODC protein was purified to homogeneity using two step procedure consisting Ni2+ affinity chromatography and size exclusion chromatography. The crude containing over-expressed EhODC from E. coli having N-terminal His-tag was loaded onto HisTrap Ni2+ column and eluted using a linear gradient of imidazole. The N-terminal His-tag from eluted protein sample was removed using enterokinase and sample was re-loaded onto HisTrap Ni2+ column. Then, the flow-through containing EhODC without His-tag was collected, concentrated and loaded onto HiLoad 16/60 superdex 200 gel-filtration column for further purification. Homogeneity of pure protein sample was estimated on 12% SDS-PAGE, which exhibited a single band of ∼46 kDa corresponding to the molecular weight of EhODC protein (Figure 4). The yield of the purified protein was estimated to be ∼3. 0 mg/L of culture and protein was concentrated to ∼6 mg/ml. The enzymatic activity of purified protein was demonstrated using the simple and rapid colorimetric ODC activity assay [28]. The decarboxylation activity of purified enzyme was assayed in 200 µl reaction containing 20 mM sodium phosphate buffer (pH 7. 5), 0. 1 mM EDTA, 0. 1 mM PLP and 1 mM of L-ornithine. The reaction was assayed in terms of the formation of product, putrescine by its oxidation using DAO enzyme which releases H2O2 that forms a colored complex as described in materials and methods. His-tagged and untagged protein showed no difference in the enzymatic activity. Furthermore, the purified EhODC actively catalyzed the conversion of L-ornithine to putrescine, while it showed no activity when D-ornithine was used as a substrate in enzyme reaction. This reveals that EhODC enzyme is stereospecific in binding to L-ornithine substrate suggesting that substrate based stereospecific inhibitors may be designed for EhODC. An effort was made to elucidate the secondary structure of EhODC by using Far-UV circular dichroism (CD). CD spectrum analysis of EhODC exhibits two negative peaks at 211 and 219 nm and a positive peak in the range of 192-203 nm, as expected for a protein with α/β content, indicating that purified protein has a well defined structure (Figure 5). The deconvolution of CD data with K2d program indicates a secondary structural content of 39% α-helix, 25% β-sheet, and 36% random coil (http: //www. embl. de/~andrade/k2d. html) [30]. For comparative secondary structure analysis, the server SOPMA was used for the prediction of secondary structural elements in EhODC sequence [46]. K2d results were found to be in agreement with the result of SOPMA showing 33% α-helix and 25% β-sheet content (Figure 5). These estimations are in accordance with the available crystal structures of ODCs and also with the molecular model for EhODC, which was generated by homology modeling in the present study. These results reveal that EhODC contains an α/β tertiary structure and has the overall folding pattern similar to the other ODCs from mammals, plants and protozoa. ODC purified from E. histolytica has previously been reported to exist in a pentameric state [25]. Three dimensional crystal structure studies of ODCs from different sources have shown that the enzyme exists as a homodimer and association of monomeric subunits directs the formation of two equivalent catalytic pockets at the dimer interface. Structural analysis revealed that each active site at the dimer interface is assembled by amino acid residues contributed from each monomer subunit, which has also been confirmed by mutational studies [21]–[23]. Therefore, we were interested in characterizing the functional oligomeric form of EhODC. To accomplish this, we purified recombinant EhODC enzyme and first confirmed that the purified protein is enzymatically active. Cross-linking agent, glutaraldehyde is used for obtaining crude information about the quaternary structure of proteins [29]. Previously, the crosslinking experiment has been performed to reveal the dimeric form of mouse ODC [47], [48]. Therefore, EhODC was cross-linked using glutaraldehyde in a closed setup similar to protein hanging drop crystallization method. After incubation for 10 min, the protein sample was analyzed using SDS-PAGE. The cross-linked sample showed two bands of ∼90 kDa and ∼46 kDa corresponding to the molecular weight of EhODC dimer and monomer (Figure 6) indicating the possibility of EhODC dimerization. To further analyze EhODC oligomerization, the molecular weight of purified protein was estimated by applying the sample onto a HiLoad 16/60 prep grade Superdex 200 gel-filtration column using ÄKTA purifier. Purified protein showed a major peak with the elution volume 71. 3 ml (Figure 4). Using a standard curve based on molecular weight markers, the molecular weight of major elution peak was calculated and was estimated to be approximately ∼90 kDa, which corresponds to the molecular weight of EhODC dimer (Figure 4). This suggests that EhODC exists in the dimeric form. Furthermore, MALDI/TOF MS analysis of the purified protein was carried out to verify and confirm the dimerization of protein. MS data showed two narrow peaks having average intensity of 44558. 430 m/z and 90667. 295 m/z and these correspond to the monomeric and dimeric state of the protein respectively (Figure 6). Thus, it was established that EhODC enzyme exists in dimeric state. The study of effect of chaotropic agents on oligomeric state is critical to evaluate the stability of quaternary structure of proteins. The behaviour of ODC in presence of such agents differs from species to species and dissociation of oligomeric state is dependent on the concentration of chaotropic agents [49], [50]. In T. brucei, ODC dissociates into monomers in presence of high concentration of salt and urea [51]. This provoked us to examine the effect of different concentrations of NaCl and urea on oligomeric state of EhODC. Incubation of protein sample with 2 M and 4 M of NaCl resulted in partial dissociation of dimeric enzyme to monomeric state (Figure 7). Two peaks were observed in gel filtration chromatogram: one at 71 ml elution volume followed by a smaller peak at 81 ml elution volume which correspond to the molecular mass of the dimeric and monomeric forms of EhODC respectively (Figure 7). With increased concentration of NaCl from 2 M to 4 M, the small peak corresponding to monomer becomes more distinct demonstrating that higher concentration of NaCl partially disrupts the dimerization. This also suggests the role of inter-molecular salt-bridges and weak polar interactions in EhODC dimerization. Similar results were observed when the protein was treated with 2 M and 4 M urea (Figure 7). Destabilization of EhODC dimers in higher urea concentration points to the presence of inter-molecular hydrophobic interactions at the dimer interface. The molecular structure and subunit interactions in EhODC were investigated by constructing a dimeric model of the enzyme using homology modeling approach. The sequence homology search for EhODC gave the hits of 29 sequences against PDB database. The crystal structure of human ODC was the first hit with 34% sequence identity (PDB ID: 2OO0) followed by TbODC (33%, PDB ID 1QU4). For comparative homology modeling, it could be significant to select a template for ODC from protozoan source i. e. TbODC. However, too much variations in the sequences of ODC within protozoa (Figure 2) and higher sequence identity of EhODC with plant and mammalian ODC, give an indication of caution required in the interpretation of template selection. Here, we have selected human ODC as template for a reliable model generation considering two major facts: firstly, the N-terminal loop region consisting of approximately eight amino acids is missing in all crystal structures of ODC except human ODC. Secondly, multiple sequence alignment analysis showed a PEST like sequence in the C-terminal region of EhODC sequence that has maximum similarity with human ODC (Figure 2). The model for EhODC along with its cofactor PLP was generated from PDB 2OO0 as a template using Modeller 9v8 and model with lowest DOPE score was considered for further loop refinement using Modeller loop refinement tool. The model was subjected to energy minimization where PROCHECK, ERRAT plot and ProSA energy plot were used for validation and quality assessment of the model. The root-mean-square deviation (RMSD) of Cα atoms between the modeled EhODC dimeric structure and the template structure was 0. 744Å. Ramachandran plot of the model generated by PROCHECK showed 90. 3% residues in the core region, 7. 8% in allowed region, 0. 6% in generously allowed region and 0. 3% in disallowed region. The generated models have been submitted to Protein Model database (PMDB) with PMDB id: PM0077698 (monomer) and PM0077699 (dimer). The molecular model of EhODC dimer that was generated using the crystal structure of human ODC dimer as a template was MD simulated for 8 ns in equilibration with water molecules. Evaluation of the dimer stability was made by monitoring the root-mean-square deviations (RMSD) of the Cα of the dimer which was computed against the starting structure. Analysis of MD trajectory of EhODC homodimer revealed that RMSD value increases to 0. 327 nm in about 1. 2 ns and this plateau value is stable till the end of the simulation indicating a stable conformation of the dimer (data not shown). Structure of EhODC monomer subunit is comprised of two major domains i. e. β/α-barrel and β-sheet domain which is a distinct characteristic of ODC structure (Figure 8). In human ODC, N-terminal starts with a β-strand while in EhODC, it starts with α-helix. The N-terminal emerges from β-sheet domain and enters the barrel through a coil connecting both the domains. The barrel contains eight parallel strands each followed by a helix in the order α2β2, η1α3β3, α4η2β4, α5β5, α6β6 α7β7, α8β8 and α9η3β9. One important feature observed in EhODC is the presence of turns in a pattern at the N-terminal barrel secondary structures. Such pattern has been observed in ODC like antizyme inhibitor proteins that have structures similar to ODC, but do not possess decarboxylation activity [52]. The sheet domain is subdivided into two clusters of sheets S1 and S2 as observed in all ODC structures. These sheets S1 and S2 remain perpendicular to each other having four helices with one turn (α1, α10, α11, α12 and η4) around it. Sheet S1 includes three parallel β-strands (↓β11, ↑β12 and ↑β13) which extends into S2 containing four parallel β-strands (↓β10, ↑β14, ↑β15 and ↑β1) (Figure 8). In the dimeric structure of enzyme, two active site pockets rest at the dimer interface involving the interactions of residues from both the subunits. β/α-barrel domain is the main site for cofactor PLP binding where as residues from the sheet domain of other subunit interacts with the substrate L-ornithine to form the complete catalytic pocket for enzymatic activity. The subunits associate in a head to tail manner (Figure 9). The dimeric structure is stabilized by various polar interactions present between the two subunits at the dimer interface as shown in figure 9. However, four major salt bridges K157-D338′ and D122-R277′, D338-K157′ and R277-D122′ are observed and these have been reported to play a vital role in the dimer formation of human, mouse, and T. brucei ODCs [22]. These interface residues are partially hydrophilic and are highly conserved in human, mouse and EhODC. Furthermore, the most prominent feature observed near C-terminal domain is presence of a stack of aromatic rings i. e. F371′/H296′/F305 and F305′/H296/F371 which is anticipated to function as an amino acid zipper. Distal amino acid residues of the zipper participate in active site pocket formation. Further, the structural analysis revealed that the close packing of dimers shields the putative N-terminal antizyme binding loop (residues 105Tyr-132Lys) as well as the C-terminal PEST like sequence because these are concealed in between the two subunits of the dimer. Thus, it is expected that the dimerization of EhODC may be responsible for protecting EhODC enzyme from proteolytic degradation. Molecular model of the EhODC dimer evidently shows that the conserved catalytic residues from both monomeric subunits form two equivalent active sites at the dimer interface (Figure 2, Figure 9). Consequently, it can be hypothesized that the dimeric state of EhODC enzyme is the active form. Therefore, 3D structure based site-directed mutagenesis approach was used to examine the functional role of EhODC dimerization. Conserved residues of the catalytic pocket present at the dimer interface and also the conserved residues of the dimerization interface were mutated. The conserved catalytic residues Lys57 and Cys334 present in the active site were selected for mutational studies, because the structure model of EhODC as well as the sequence alignment of EhODC with human ODC revealed that Lys57 of one subunit (Lys69 in human) and Cys334′ of other subunit (Cys360 in human) jointly play critical role in catalysis and substrate specificity in a single active site pocket (Figure 2, Figure 9) [53]–[55]. The residue Lys57 plays crucial role in PLP binding by forming Schiff base to aldehyde group with its –NH2 group, thus serves as a proton donor during catalysis [56]. The interaction of Lys57 with PLP governs its position and correct orientation at active site. Gel filtration analysis indicates that K57A mutant exists in the dimeric form indicating that this mutation does not disrupt dimerization (data not shown). However, when enzyme activity was examined, K57A mutation was found to abolish enzyme activity with ∼2% activity as compared to the wild type (Figure 10). Moreover, Cys residue in the same active site from other subunit in the active site is involved in substrate binding and stabilizes the quinonoid intermediate by using its carbonyl group [54], [57]. This residue is crucial for decarboxylation of L-ornithine and release of decarboxylated product towards the interface to exit from active site. The C334A mutant was also found to be a dimer indicating that mutation does not affect dimerization (data not shown). However, C334A was also found to be inactive with ∼2% enzymatic activity as compared to wild type (Figure 10). Interestingly, when the two mutant proteins K57A and C334A were mixed in equal concentration, the enzyme activity was partially regained having 29% activity as compared to wild type (Figure 10). The recovery of enzyme activity on mixing these two mutants is only possible when the two mutants associate to form a heterodimer. The formation of heterodimer is anticipated to restore one of the two active sites at the dimer interface as depicted in figure 11. Three types of enzyme population are expected in mutant mixture i. e. homodimers of K57A, homodimers of C334A and heterodimers of K57A and C334A. Therefore, restoration of approximately one-third of the wild-type enzyme activity in the mixture of mutants is due to the dimerization of K57A and C334A which possesses a catalytically active site pocket at one end of the heterodimer. These mutagenesis results evidently demonstrate that dimeric state is the functional form of ODC enzyme in E. histolytica. In mouse, 19 conserved residues at the dimer interface were mutated to identify the key residues responsible for dimerization [48]. It was noted that substitution of conserved Gly387 to any amino acid except alanine abolished the enzymatic activity. The same result is also observed in case of Lactobacillus and hamster, where the corresponding glycine was mutated to any bulky amino acid resulted in inactivation of the enzyme [58], [59]. Crystal structure of mouse ODC revealed that this mutation could position β/α-barrel at a different angle to β-sheet so that in the mutant protein these domains have different orientations in the dimer compared to the wild type which makes the enzyme inactive [22]. In the present study, EhODC Gly361 (Gly387 in mouse) was mutated to bulky Tyr residue and its influence on dimerization was assessed by gel filtration analysis. The chromatogram showed partial destabilization of dimer with two distinct peaks corresponding to the molecular weight of monomer and dimer (Figure 12). The examination of enzyme activity showed that the Gly361Tyr mutant is functionally inactive (Figure 10). These results suggest that Gly361 in EhODC is not involved in direct interaction between the two subunits of dimer, however it plays an indirect role in the dimer stability through long range molecular interactions. Additionally in the structure model and sequence alignment analysis, Lys157 of EhODC is conserved and forms a salt bridge with Asp338′ connecting the two monomeric subunits. At the same position in the crystal structure of human ODC, Lys169 of one subunit is involved in the salt bridge formation with Asp364′ of other subunit near the active site [21], [22]. Thus, Lys157 of EhODC plays a critical role in spatial arrangement of active site residues from both the subunits in a proper orientation along with its role in dimer formation. Mutation of Lys157 to Ala (K157A) leads to inactivation of enzyme (Figure 10). Moreover, partial disruption of the dimer as compared to the wild type protein was observed for K157A mutant, because a peak corresponding to the monomeric state of EhODC along with the dimer peak was observed in the gel filtration chromatogram (Figure 12). These results suggest that Lys157 plays a direct role in dimerization that eventually leads to the active site formation. Furthermore, a double mutant of EhODC having two mutations i. e. G361Y and K157A was expressed in E. coli. The protein was over-expressed using high IPTG concentration of ∼2 M for induction. This double mutant was found to be unstable and susceptible to protease degradation during purification. Therefore, it could not be purified for further analysis. The instability of the double mutant G361Y and K157A could be due the dimer disruption making the protein insoluble as well as proteolytically unstable. Our current study, evidently demonstrates that EhODC enzyme exists in the dimeric form. The role of dimerization with respect to functionality was investigated by comparative structure modeling and mutational studies. Molecular structure reveals a sharp complementary arrangement of interface and active site residues to support the proper spatial arrangement. Thus, it contributes both the subunits in generation of two equivalent active sites. The partial recovery of the enzyme activity on mixing the two mutants, C334A and K57A which were individually inactive, shows that dimer is the active form of EhODC. Additionally, a single substitution at G361Y resulted in partial destabilization of the dimer and renders the enzyme inactive. Further, K157A mutation expected to disrupt a salt bridge K157-D338′ between two subunits didn' t completely disrupt the dimer but inactivates the enzyme. These results signify that various long and short range forces play a crucial role in the dimerization and the geometry of the dimer interface is ideal for enzyme activity. Based on these observations, it can be proposed that disruption of functional EhODC dimer could be a novel target for anti-amoebiasis drugs. Molecular 3D model of EhODC dimer may support and open possibilities to find new structure based inhibitor molecules for the enzyme.

E. histolytica genome sequence divulged the existence of ornithine decarboxylase enzyme that performs the first-rate limiting catalytic step of polyamine biosynthetic pathway. ODC enzyme is a potent therapeutic target in many eukaryotic disease causing pathogens. DFMO, a potent substrate analogue inhibitor, is widely used for the treatment of various diseases including Trypanosoma brucei infections. However, DFMO does not inhibit E. histolytica ODC. As ODC is a validated drug target for protozoan disease, an alternate strategy to inhibit the EhODC enzyme may be developed. In our study, we have evidently proved that the purified recombinant EhODC is functional as an active homodimer. Molecular modeling and simulation studies indicate that two independent active sites are present at the dimer interface. Our mutational studies indicate that the enzyme activity can be abolished by targeting the dimer interface and this in turn suggests the alternative inhibitory mechanism for the enzyme. Our investigation yields that disruption of dimer disrupts the active site pocket and renders the enzyme inactive. As EhODC crystal structure is unavailable, the 3D structure model of EhODC homodimer may assist in designing structure based anti-amoebiasis peptides or agents that disrupt the active site by destabilizing the dimer.

Abstract Introduction Materials and Methods Results and Discussion

biology computational biology microbiology biophysics macromolecular structure analysis

2012

Biochemical, Mutational and In Silico Structural Evidence for a Functional Dimeric Form of the Ornithine Decarboxylase from Entamoeba histolytica

Increasing volumes of data and computational capacity afford unprecedented opportunities to scale up infectious disease (ID) mapping for public health uses. Whilst a large number of IDs show global spatial variation, comprehensive knowledge of these geographic patterns is poor. Here we use an objective method to prioritise mapping efforts to begin to address the large deficit in global disease maps currently available. Automation of ID mapping requires bespoke methodological adjustments tailored to the epidemiological characteristics of different types of diseases. Diseases were therefore grouped into 33 clusters based upon taxonomic divisions and shared epidemiological characteristics. Disability-adjusted life years, derived from the Global Burden of Disease 2013 study, were used as a globally consistent metric of disease burden. A review of global health stakeholders, existing literature and national health priorities was undertaken to assess relative interest in the diseases. The clusters were ranked by combining both metrics, which identified 44 diseases of main concern within 15 principle clusters. Whilst malaria, HIV and tuberculosis were the highest priority due to their considerable burden, the high priority clusters were dominated by neglected tropical diseases and vector-borne parasites. A quantitative, easily-updated and flexible framework for prioritising diseases is presented here. The study identifies a possible future strategy for those diseases where significant knowledge gaps remain, as well as recognising those where global mapping programs have already made significant progress. For many conditions, potential shared epidemiological information has yet to be exploited. Maps provide an essential evidence-base to support progress towards global health commitments [1]. For example, they provide important baseline estimates of disease limits [2–7], transmission [8–10] and clinical burden [11–14]; underpin surveillance systems and outbreak tracking [15,16]; help target resource allocation from the macro- [17,18] through the meso- [19–22] to the micro-scale [23]; and inform international travel guidelines [24–26]. Significant developments in mapping techniques have occurred over the last few decades, particularly through the use of species distribution models and model-based geostatistics [1,27]. Similarly, disease data has become more widespread and easier to share [28]. Despite these advances however, a recent review of 355 clinically-significant infectious diseases (IDs) indicated that of the 174 IDs for which an opportunity for mapping was identified, only 4% had been comprehensively mapped [1]. For many of these conditions, there is a significant shortfall between existing maps and what can be achieved with contemporary methods and datasets. Traditional mapmaking has focussed on a vertical, single-species approach, requiring highly labour intensive, and therefore expensive, manual data identification and assembly [13,21,29–31]. The present era of open-access big data, high computational capacity, and rapid software development offers new opportunities for scaling-up the spatial mapping of IDs, primarily through the automation of data gathering and geopositioning but ultimately also to mapping. The Atlas of Baseline Risk Assessment for Infectious Disease (abbreviated ABRAID, as in, “to awake”) is a developing software platform designed to exploit this opportunity and has the ambition to produce continuously updated maps for 174 IDs globally [28]. Realising automation of data retrieval and positioning at this scale is a practically non-trivial but conceptually simple, logistic scaling exercise. In order to automate mapping for each ID so that it is continuously updated and improved as new information becomes available, the spatial inference methods used need to be tailored to each unique ID epidemiology [28]. In some cases this will require disease-specific methodological developments. This requires substantial investment, so an objective and systematic approach is required to determine the order in which IDs are to be mapped. The first stage in this process is to organise all IDs using a schema based upon shared biological and epidemiological traits; for example, “the mosquito-borne arboviruses”. Such groups will likely have similar mapping requirements, enabling synergies in data collation, covariate selection, increased efficiency (i. e. in software development), and more robust validation of outputs [32,33]. We refer to these disease groups as “mapping clusters” and they form the basic architecture of the prioritisation process. To rationally prioritise mapping of these conditions, the diseases within each mapping cluster were evaluated based upon their global burden (both morbidity and mortality), as well as the disease’s importance amongst public health stakeholders. Data inputs are quantitative in nature and reliant on either independently derived data or data sourced from entire communities rather than selected expert individuals. Therefore, this proposed framework is unaffected by much of the subjectivity associated with other prioritisation studies, and also provides a platform for rapidly incorporating changes to existing diseases, as well as emerging novel public health threats. The prioritisation exercise helps to guide the order in which diseases are mapped to best support public health priorities; we argue that all relevant diseases can and should eventually be mapped. A comprehensive atlas of IDs is of central importance in providing geographical context to the understanding of tropical disease and global health [34–36]. Moreover, as the atlas becomes more complete the overlay of maps will provide opportunities for investigating patterns of global disease diversity [37,38] and the process of disease emergence [39]. In order to generate disease prioritisation standards, diseases with shared taxonomy and transmission characteristics were grouped together to create clusters. Diseases within each cluster were evaluated based upon two factors reflecting their importance from a public health perspective: (a) the global burden of the disease and (b) the current public health focus on the condition. Both metrics were assessed simultaneously in order to rank the clusters, and specific diseases were then identified for prioritisation. This study aimed to be comprehensive in its scope of IDs. All diseases identified in a previous review as meriting mapping were included [1]. This earlier study categorised 355 diseases into five classes: Option 1, indicating that the disease was unsuitable for occurrence based mapping methods; Option 2, mapping the observed occurrence of the disease; Option 3, mapping the maximum potential range of the disease using knowledge of vector, intermediate host and reservoir species; Option 4, using niche mapping methods such as boosted regression trees; and Option 5, where sufficient data exist to allow for global maps of variation in prevalence of infection and/or disease. Option 1 diseases included those that showed no sustained spatial variation in occurrence (i. e. had a cosmopolitan distribution) and had insufficient evidence to allow for the global mapping of variation in prevalence using advanced statistical methods such as model-based geostatistics. In cases where such information does exist, these diseases were promoted to Option 5 status. Revisions to the Hay et al. (2013) paper have led to the inclusion of tuberculosis, ascariasis, trichuriasis and trachoma—all previously listed as Option 1—as Option 5 diseases. Further revisions included the exclusion of New and Old World Spotted Fever Rickettsiosis and New and Old World Phlebovirus because their constituent diseases were included. In addition, Plasmodium knowlesi was included due to the increasing appreciation of its significance to human health in Southeast Asia [40,41]. The new revised total of diseases that warrant mapping was therefore 176. Those diseases not considered for mapping due to Option 1 classification are outlined and justified in the Supporting Information. Diseases were grouped into clusters based on characteristics relevant to spatial epidemiology. Diseases were placed in the same cluster if they had the potential to mutually reinforce each other in terms of data assembly, mapping requirements and cross-validation of data by comparison of outputs. Clustering classifications were therefore based on the key factors influencing the approach taken for mapping. At the coarsest level, pathogens were grouped by agent type (virus/bacteria/fungus/other) and the larger agent groupings were split into specific phyla (e. g. Nematoda and Platyhelminths) [42]. These relatively coarse groups reflect fundamental differences in life histories and epidemiology as well as the most basic taxonomic divisions. Within these broad groupings, the mode of transmission was used to create the final disease clusters. This is an important factor when mapping IDs, as the mode of transmission has a large influence on which abiotic correlates are relevant to the mapping process. For instance, the transmission limits of vector-borne diseases are restricted in part by the environmental suitability for the vector species in question, thus diseases spread by similar vectors will share covariates [43]. Similarly, sexually-transmitted diseases are likely to share mapping methods linked to human distribution and behaviour, whilst pathogens spread by water contact would share common traits linked to the environment; these groupings can therefore be logically considered together within a mapping framework. The mode of transmission classifications are defined in the previous publication [1]. The burden of each disease was assessed using the disability-adjusted life year (DALY) estimates from the 2013 Global Burden of Disease Study (GBD 2013) [44,45]. DALYs quantify both morbidity and mortality attributed to each disease and therefore better capture the total impact of a disease than do clinical cases or mortality alone [44,45]. The GBD’s systematic approach across a wide spectrum of diseases provides an extremely valuable resource from which to compare the relative impact of diseases on human health. Wherever possible, direct links were made between the GBD estimates and diseases in the mapping list. The GBD disease categories, which are based upon the International Classification of Diseases and Related Health Problems (ICD-10) [46], do not always specify particular infectious agents, but rather focus on the clinical symptoms of infection, or non-specified disease groups. These aggregated DALY estimates had to be split across the relevant causative diseases in the mapping list, therefore the Hay et al. (2013) study was reconciled with the ICD-10 codes and then GBD categories in order to disaggregate the broader classifications such as “other diarrheal diseases” and “other neglected tropical diseases”. The full process is outlined in the associated Supporting Information, S1 Text. Overall, 11 of the 176 mapping diseases could not be reconciled to the GBD categories. Some were not considered due to having unknown pathogenic agents (e. g. tropical sprue) and others were very rare and fell into ICD-10 categories that were assigned over various groupings (e. g. pentastomiasis). These diseases were allocated a nominal DALY of 100; this value, while arbitrary, is low enough to avoid skewing the analysis. For each cluster, the total DALYs for all diseases was calculated and contributed to the final analysis. Of equal importance is the need to produce maps for those diseases where there is the greatest demand, whether from international organisations or from local public health authorities. Measuring this factor was achieved by surveying a representative subset of potential end-users, to identify which diseases have been prioritised by major public health stakeholders: state-funded public health agencies, private companies (e. g. vaccine developers), political bodies, non-governmental advocates and practitioners, as well as the scientific research community. For each disease, the final policy score was the sum of three component scores: public health, stakeholder interest, status as a notifiable disease, and h-index. Cases from the different categories of public health stakeholders were included to capture the spectrum of interest groups (see S1 Text for full listing). Each organisation’s mission statement and project pages were reviewed to identify the diseases contained in their public health portfolio. Depending on the type of stakeholder, this would indicate that the organisation would, for example, dedicate funding and effort towards control of that disease, advocate for the disease to governments or public health agencies, or dedicate research funding to the disease. Each disease was allocated one point per stakeholder reporting an interest in it. An inclusive approach was followed, whereby diseases were considered to be of interest to a stakeholder, irrespective of any hierarchy within the agency’s prioritisation system. Another point was allocated to diseases which were notifiable to national disease control agencies. In order to mitigate spatial bias in the notifiable disease listed by different agencies, a search for countries which had readily-accessible and clearly defined domestic policy relating to named pathogens was performed, and one country from each of the main GBD defined regions was selected: USA (High Income), Brazil (Latin America and the Caribbean), Zambia (sub-Saharan Africa), United Arab Emirates (North Africa and the Middle East), India (South Asia), Malaysia (South East Asia, East Asia and Oceania) and Croatia (Central and Eastern Europe and Central Asia). Interest in these diseases at a domestic level suggests that there will be interest in maps of these diseases, as demonstrated by the presence of subscription-only online databases of maps including GIDEON [47] and the rapid expansion of real-time maps to which physicians are encouraged to contribute [15,28]. Academic output, a proxy of funder agency awards, but also of high-quality data availability [1], was quantified based on the h-index of each disease [48], as reported by Scopus [49]. More commonly used to assess a scientist’s productivity and impact, the h-index is used here to quantify the level of active interest across the academic community in each disease [50]. The h-index is the number of published papers (referring to a particular disease) that have been cited by at least as many other papers. In other words, an h-index of 7 signifies that 7 published papers including that disease name have been cited at least 7 times. For each disease in turn, Scopus citation numbers were generated for all publications referring to the disease (document search for" Disease Name" in" Article Title, Abstract, Keywords" ). This Scopus search generates a Citation Tracker file showing the number of citations to each publication referring to the" Disease Name" . Diseases were then categorised according to their h-index. Those for which there was evidence of very high scientific output scored 2 (h-index >100), those with intermediate h-index (>50–100) scored 1. 5, while diseases with h-index of <50 scored 1. The diseases classified as Option 4 (use niche modelling methods) and Option 5 (model prevalence or incidence) have the most epidemiological data available and have the greatest potential to benefit from a dedicated mapping exercise, but also require the most resources. Option 2 and 3 diseases are data-poor and both require mapping of occurrence data only [1], and therefore are significantly less time-intensive to map, limited to more simplistic analyses, than those diseases categorised as Option 4 and 5. Option 3 disease mapping relates potential transmission limits to aspects of vector biology. In cases where Option 4 and 5 diseases also have the same vector, the Option 3 disease will be considered as part of mapping these complementary diseases; where this is not the case, a disease’s transmission limits can be assessed through a mixture of literature surveys and occurrence data overlap. Option 4 and 5 diseases within the disease clusters were therefore prioritised and for each cluster, the average policy score for the Option 4 and 5 diseases was calculated and contributed to the final analysis. These diseases should be the primary focus of future cartographic efforts as these require the most attention and bespoke inputs to be generated. The final step in the process was to combine these assessments to produce a ranking of disease clusters and therefore recommend diseases to prioritise for mapping. Each cluster was plotted on a graph based on its total DALYs and the average policy priority of its Option 4 and 5 diseases. Option 2 and 3 diseases were included in the cluster DALY scores in order to reflect the relative importance that each cluster represented in terms of burden of disease. One cluster may consist of a large number of minor diseases which, as a collective grouping, represent a significant problem—by retaining the DALY score, this burden is reflected, With the policy priority score however, the opposite is the case; inclusion of multiple low scoring diseases would down-weight the cluster as a whole. In scenarios where clusters consist of a diverse grouping of pathogens, averaging policy score across all conditions misrepresents those with a high policy priority and therefore masks these diseases in comparison to clusters that only consist of those diseases with high policy priority scores. Each cluster was then evaluated based upon its distance from a hypothetical cluster which had the highest DALYs (i. e. that of HIV) and the highest policy score (i. e. that of Malaria) relative to a line drawn from this cluster to the origin; those closer to this hypothetical cluster, along this axis, were prioritised higher. As a result, the relative influence of burden and policy priority could be considered both simultaneously and independently. Within each cluster, the diseases to be prioritised (i. e. Option 4 or 5) were then reported (Table 1). The code to replicate this methodology is freely available from: https: //github. com/SEEG-Oxford/prioritisation. The 176 diseases identified as having a rationale for mapping were organised into 33 clusters, based upon the biological and taxonomic classifications of the causative pathogen, modes of transmission and the mapping method recommended in a previous review [1] (Fig 1). Seven of these clusters included only a single disease due to their unique transmission within their broader taxonomic grouping (HIV, poliomyelitis, avian influenza, pythiosis, South American bartonellosis, tuberculosis and babesiosis). Conversely, the mosquito-borne arbovirus cluster was the largest cluster, consisting of 26 diseases, many of which have the potential to benefit from modelled maps. Fig 2 brings together the two indices selected to prioritise diseases for mapping—disability adjusted life-year (DALY) burden and relative stakeholder interest. These plots demonstrate that the HIV, malaria and tuberculosis clusters are exceptional in representing an overwhelming share of DALY burden [51] and being of highest priority to the global health community with their placement in the top right quadrant of the graph. These three clusters contain five individual diseases that are a mapping priority, malaria (Plasmodium falciparum, P. vivax, and P. knowlesi), HIV and tuberculosis. Table 1 shows the top 15 disease clusters (i. e. those in the top right of Fig 2), representing 44 individual diseases, with their associated scores. Fig 3 demonstrates that there exists a group of approximately 45 diseases that are the collective focus of public health agencies. The 44 diseases prioritised by this study include all those diseases that represent a significant cartographic challenge (i. e. those diseases requiring either species distribution modelling approaches to produce occurrence maps or model-based geostatistics to produce prevalence maps, n = 33) identified by these public health agencies, save rabies and avian influenza. The clusters are ranked in order, whilst the diseases within each cluster are alphabetical and should be considered equal on the basis of this prioritisation. The top ten priority clusters account for over 92% of all DALYs for those IDs which require mapping (i. e. the 176 IDs identified); if this is expanded to the top 15 clusters containing 44 diseases to map, this value increases to 95% (Fig 4). Within these 44 diseases, 19 of the 29 neglected tropical diseases (NTD) highlighted by the WHO are represented. Within the top ten prioritised clusters, 14 individual diseases relate to these same NTDs [52,53]. The top 15 prioritised clusters include some diseases, such as the picornaviridae (polio), that have a low DALY burden but a high public health ranking because they are high on the eradication agenda. It was possible to establish a direct correspondence with GBD estimates for 34 of the 176 diseases with a strong rationale for mapping as listed by Hay et al. (2013) [44,45]. DALY estimates were allocated to a further 132 diseases by linking diseases with ICD-10 codes [46] and their respective GBD category definitions. Whilst these burden values are not accurate absolute values, and should not be interpreted as such, this DALY allocation does allow relative burdens to be determined. The remaining 11 diseases were given the baseline DALY allocation of 100, a value not intended to represent an estimate of the “true” DALYs associated with these diseases, but rather to distinguish them from diseases which were considered to cause a major burden in the GBD analysis. It is safe to assume that if such diseases were not assigned a specific GBD classification, their global impact on mortality and morbidity is relatively small. In total, the 176 diseases with a strong rationale for mapping [1] represent over 230 million DALYs, approximately 10% of the global DALY burden and 47% of the global ID DALY burden. At the cluster level, HIV, malaria and tuberculosis represent 80% of the overall mapping-disease DALY burden (Fig 5A). Apart from these three conditions, the only other IDs in the top 50 highest DALYs globally are not currently recommended for mapping because they do not show spatial variation in their occurrence and have insufficient data to map variation in disease prevalence with model-based geostatistical analyses. The high-burden diseases not currently considered for mapping include respiratory diseases, meningitis, and many diarrhoeal infections. Alternative approaches to mapping broader symptom groupings (severe pneumonia, severe diarrhoea and severe febrile illnesses) and then differentiating constituent disease components, are being developed. Together, this would map 80% of all DALYs caused by communicable diseases. A higher resolution focus on the clusters excluding HIV, malaria and tuberculosis (Fig 5B) shows that over 60% of DALYs associated with the 176 IDs are accounted for by the other top ten prioritised clusters; approximately three quarters of the remaining DALYs are accounted for when the remaining prioritised clusters of diseases are included. The treemap in Fig 5C displays the repartition of interest from the global health community across the clusters. Interest was scored in terms of: 1) the stated priorities of a survey of assorted public health stakeholders who are expected to be end-users of the maps, 2) status as a notifiable disease, and 3) prominence in the academic literature. A total of 20 diverse stakeholders were surveyed. This was found to be a sufficiently large number to sample based on an analysis similar to a species accumulation curve that demonstrates the diminishing returns from increasing sampling effort [54]. The number of new diseases reported levelled off at around 15 organisations sampled (Fig 3) and so the 20 organisations used for this analysis was sufficient to capture the diseases of public health priority. Of the 176 diseases recommended for mapping [1], 24% were prioritised by at least one public health agency, and 55% were notifiable to at least one of the national disease control agencies. Of those diseases that represent the greatest cartographic challenge, all were prioritised by at least one public health agency and two thirds were notifiable diseases. Of the 176 diseases, thirty diseases (17%) had an h-index [48] above 100 (with HIV having the highest h-index of 461), while 64% of the diseases had an h-index of 50 or less. Of the occurrence mapping and prevalence mapping diseases, 30% had an h-index above 100 and only 37% had an h-index of 50 or less. Unlike the DALY burden, which was allocated at the disease level (S1 Text), the stated priority diseases were often grouped to the cluster level by the surveyed stakeholders. For instance, rather than specifying “Plasmodium vivax” or “visceral leishmaniasis” as a focus, “malaria” and “leishmaniasis” would be more commonly stated targets. Each component disease of these clusters would therefore be allocated a point, meaning that the number of component diseases in each cluster strongly inflated the overall interest score allocated at the cluster aggregate. Interest scores were calibrated in the final prioritisation assessment to the number of diseases classified as occurrence or prevalence mapping within each cluster (i. e. those requiring the more advanced geostatistical techniques, see Methods for more details), so as to avoid being unduly skewed by the size of the cluster. Overall, malaria, HIV and tuberculosis were the leading clusters of interest, with scores of 11. 8,11 and 11, respectively. A further seven clusters received repeated interest, including food-borne/water-borne bacteria (score = 8) and water-borne trematodes (7. 7), trypanosomiasis (7. 5), filariasis (6. 2), picornaviridae (6), avian contact viruses (6), soil-transmitted helminths (5. 3) and leishmaniasis (5. 2) all scoring highly, indicating their importance to the public health community. These scores are relative and intended to reveal general trends across the clusters rather than quantitatively reflect the weighting that any one institution places on a particular disease. A review of all clinically significant IDs identified 176 with a strong rationale for mapping, of which only 4% have been adequately mapped [1]. The current study was undertaken to define a ruleset for determining which diseases, from a cartographic and public health perspective, should be prioritised when sequentially addressing this shortfall. Diseases were clustered together based upon shared characteristics (such as basic taxonomic division and mode of transmission) in order to consider together those diseases that would synergise operationally in terms of data collection, covariate selection and methodology used. Given the large number of diseases identified, prioritisation is necessary; we addressed this by evaluating both within the context of disease burden as well as considering the diseases’ influence within public health organisations and the wider academic community. It is important to stress that the study was focussed on priorities for mapping, and was not a general prioritisation of IDs; this is particularly important to emphasise given that a number of high-burden diseases, including meningitis, pneumonia and some diarrhoeal diseases, were not included in the list of 176 diseases [1,44,45]. Malaria is the infectious disease for which the most detailed and robust global risk maps exist [13,29]. The work of the Malaria Atlas Project [33,55] along with a proliferation of national and local-scale studies [56] has established a mature and sophisticated methodological approach centred on the use of model-based geostatistics to generate continuous surfaces of risk. This has been possible, in part, due to the long history of population-based malaria infection prevalence surveys where researchers and control programmes have used microscopy or rapid diagnostic tests to establish the proportion of randomly sampled individuals testing positive for malaria parasitaemia [30,57]. Crucial for geospatial mapping, such data are increasingly georeferenced with a latitude and longitude for each observation established via gazetteer methods (recorded location names linked to digital atlases) or directly using Global Positioning System (GPS) technology at the time of survey [58,59]. The high prioritisation of HIV and tuberculosis shown in the current study brings into sharp focus the need for similar mapping activities to be established for HIV and tuberculosis. All three diseases have an established history of routine and survey-based data collection that, in comparison to many other diseases, is of relatively high quality and consistency, laying the foundation for similar statistical mapping approaches to those used for malaria to be applied. A cornerstone of HIV surveillance over the last several decades has been routine blood testing for HIV infection in mothers attending sentinel antenatal clinics. Such data provide rich longitudinal observations of prevalence in this demographic group and the potential exists to combine these with cross-sectional data from nationally representative household surveys [60] to generate optimal space-time models of the changing geographical pattern of infection across individual countries. Unlike HIV and malaria, population-based tuberculosis prevalence testing is not currently included as part of the major international survey programmes [58,61]. However, such surveys (reporting on the prevalence of bacteriologically-confirmed pulmonary tuberculosis) have been undertaken in a number of high-burden countries in recent years, with many more planned in the near future [62]. In a similar way to HIV, the prospect exists of a mapping methodology that could combine survey-based data with the rich health-system based data on new case notifications and other metrics, leveraging the respective strengths of community- and facility-based data. A longer-term goal must be the development of a data assimilation and modelling architecture for all three of these major global diseases to support robust and regularly updated global maps detailing their joint distribution and its evolution though time which can be used to assess the impact of control and international financing efforts [18]. The current analysis identifies a number of different NTDs as priority diseases for mapping, a finding which is consistent with the emphasis given to mapping by the global NTD community in order to geographically target NTDs interventions [63,64]. Specifically, for those NTDs where morbidity control is the goal, including soil-transmitted helminths (STH) and schistosomiasis, interventions are most cost-effective when they are targeted to areas of highest transmission [21]. For those NTDs which are identified for elimination, such as onchocerciasis and lymphatic filariasis, it is essential to know where transmission occurs and when it has been successfully halted following control measures. As a consequence of these operational requirements, large-scale mapping initiatives are underway for each of the main NTDs (Table 1). A challenge for mapping the NTDs, and indeed for mapping many IDs, is the need to continually update maps in order to help track the progress in control. As interventions reduce transmission levels and therefore distributions become more focalised, the need for mapping will only increase. Unsurprisingly, the top 44 diseases for prioritisation are dominated by those with the highest global burden. However, certain clusters stand out as having high public health attention without a high burden, particularly the picornaviridae cluster and its constituent disease, polio. Although cases are now restricted to a few hundred each year, polio has been identified as an eradication target and is a high priority for many public health stakeholders despite recent obstacles in the eradication schedule [65,66]. In these eradication and elimination scenarios, the role of mapping changes subtly to both identifying areas where cases continue to occur, and in highlighting potential future risks and improving surveillance [67]. Following a similar logic, diseases such as dracunculiasis, African trypanosomiasis and onchocerciasis, in spite of relatively low burdens, remain high policy priorities due to elimination efforts in various parts of the globe [68,69]. These examples demonstrate the utility of the approach used in this study of using assessments of the public health burden as well as metrics of public health attention. The disease prioritisation methodology used here differs from existing approaches, such as the “Delphi panel method”, in that it does not include a panel of experts scoring various criteria associated with the diseases being considered [70–74]. In contrast, this study uses a simplified methodology, placing importance in reproducibility and flexibility, using clearly defined rules to assess available evidence and remove potentially subjective expert-opinion. The methods employed are reliant on independent, third party information, and are assessed in a consistent manner, which can easily respond to changes either in burden or public health focus. The relative importance of these diseases will most likely change over time, so an approach that can easily accommodate this is preferable. Burden estimation using the GBD is crucial, since it is the leading globally consistent measure by which to compare these various diseases and the effects of their many different clinical manifestations. Any global assessment of 301 causes of mortality and morbidity, and associated sequelae, will be subject to the limitations of data availability and epidemiological understanding as well as model assumptions and implementation [53,75,76], and will require frequent updates in a rapidly changing world. The technique presented here has the advantage of being rapidly updateable, and we will reproduce these numbers with each new iteration of the GBD project. As a consequence, public health authorities can also easily create bespoke prioritisation lists based upon a selection of disease inclusion criteria (such as those endemic to their particular country or region). This can more easily be achieved with the availability of sub-national estimates of disease burden from the GBD study. Country specific estimates of the interest scores can also be generated with greater specificity, and can therefore avoid some of the potential biases resulting from the use of other countries as representatives of each GBD region used in this study. Additional factors that may influence the disease priority, such as potential economic impact [77–79], were not used in this analysis because insufficient information was available to include these metrics. The methodology outlined above benefits from two metrics that can be applied globally to quantify DALYs and public health priority. As and when measures of additional disease impacts become available, they can and should be incorporated into assessments such as this. The study also identifies some high DALY groupings that do not have high-level policy interest. Three groupings (Tick-borne (Bacterial), Tick-borne (Viral) and Mammal contact (Viral) ) have a cumulative high DALY burden, but relatively low policy rankings and therefore are just outside the top 15 cluster listing. This may reflect the large number of diverse pathogens that make up these groupings, many of which are relatively restricted in distribution and hence would not commonly be prioritised by globally focussed organisations. That said, the high DALY value indicates that these diseases are of international interest, particularly when secondary human-to-human transmission is a possibility such as with Lassa fever and Crimean Congo Haemorrhagic Fever [80]. These conditions further advocate the utility of regional and national level priority estimates. The exclusion of diseases not suited for occurrence based mapping, and therefore omitted from the prioritisation process (so called Option 1 diseases [1]), is entirely based on cartographic considerations. Some of these diseases are inherently linked to human-to-human interactions, others are endogenous in origin, with the pathogen essentially ubiquitous amongst humans and only occasionally causing opportunist infections in certain scenarios, whilst some have the potential to cause infection anywhere across the globe due to the cosmopolitan distribution of their sources of infection, whether they be environmental or human based. Many of these diseases can vary spatially, as evidenced by the African meningitis belt, although such variation, when considered relative to the rest of the world, is due to differences in prevalence or intensity, not presence or absence. Occurrence based mapping methods, such as boosted regression trees, rely on binary presence/absence data. For conditions such as the common cold, diphtheria or respiratory syncytial virus, which have the potential to occur across the globe, these mapping techniques are ineffective. It is only through using more advanced methods, such as model-based geostatistics, that maps analysing the variation in intensity of these diseases can be produced. The limitation of this methodology is the amount of prevalence survey data required, which for many diseases is not comprehensive or detailed enough to allow for global analyses. Basic human related covariates, such as population density, urban extent profiles and national vaccination statistics can be used to explain a degree of the global variation in these diseases, but fall short of the wealth of information that can be derived from comprehensive global prevalence datasets, such as those available for malaria. As we continue to explore additional data avenues, there will be an increasing number of diseases where such data become available. The disease prioritisation outlined in this study offers a logical framework for proceeding with disease mapping, which reinforces the necessity of existing programmes and identifies those diseases to focus on next (Table 1). Diseases which will form the initial focus of future study comprise both those with the highest-burden and those of greatest concern to the global health community. The initial top-priority diseases include a range of disease agents and transmission routes, and therefore present a variety of challenges for mapping. The prioritisation and clustering of these diseases presents a clear plan of action designed to maximise the effectiveness and value of future cartographic efforts.

Maps have long been used to not only visualise, but also to inform infectious disease control efforts, identify and predict areas of greatest risk of specific diseases, and better understand the epidemiology of disease over various spatial scales. In spite of the utilities of such outputs, globally comprehensive maps have been produced for only a handful of infectious diseases. Due to limited resources, it is necessary to define a framework to prioritise which diseases to consider mapping globally. This paper outlines a framework which compares each disease’s global burden with its associated interest from the policy community in a data-driven manner which can be used to determine the relative priority of each condition. Malaria, HIV and TB are, unsurprisingly, ranked highest due to their considerable health burden, while the other priority diseases are dominated by neglected tropical diseases and vector-borne diseases. For some conditions, global mapping efforts are already in place, however, for many neglected conditions there still remains a need for high resolution spatial surveys.

Abstract Introduction Methods Results Discussion

2015

Prioritising Infectious Disease Mapping

Chromosome replication in Escherichia coli is in part controlled by three non-coding genomic sequences, DARS1, DARS2, and datA that modulate the activity of the initiator protein DnaA. The relative distance from oriC to the non-coding regions are conserved among E. coli species, despite large variations in genome size. Here we use a combination of i) site directed translocation of each region to new positions on the bacterial chromosome and ii) random transposon mediated translocation followed by culture evolution, to show genetic evidence for the importance of position. Here we provide evidence that the genomic locations of these regulatory sequences are important for cell cycle control and bacterial fitness. In addition, our work shows that the functionally redundant DARS1 and DARS2 regions play different roles in replication control. DARS1 is mainly involved in maintaining the origin concentration, whether DARS2 is also involved in maintaining single cell synchrony. The circular chromosome of Escherichia coli is replicated bidirectionally from a single origin, oriC. The DnaA protein is responsible for replication initiation [1]. DnaA belongs to the AAA+ (ATPases Associated with diverse Activities) proteins and can bind both ATP and ADP with similar high affinities[1]. DnaA is active in replication when bound to ATP (DnaAATP) [2] and facilitates the unwinding of oriC [3–5]. Integration Host Factor (IHF) [6] and DiaA [7,8] stimulates initiation from oriC, while initiation is opposed by the binding of Fis to oriC [9–11]. After open complex formation, DnaA loads the DnaB helicase onto the single-stranded DNA, which promotes further duplex opening and assembly of the replisome [5]. After initiation DnaAATP is inactivated, i. e. converted to DnaAADP, by RIDA (Regulatory Inactivation of DnaA) [12–14] and DDAH (datA-dependent DnaAATP hydrolysis) [15] to prevent re-initiation. RIDA is more efficient in lowering the DnaAATP/DnaAADP ratio than DDAH [15]. At later cell cycle stages DnaAADP is reactivated at the two DnaA-Reactivating Sequences (DARS1 and DARS2) to allow for the next round of initiation [16,17]. DARS1 is not regulated by any known pathway [16], while DARS2 activity is modulated by both an IHF- and Fis-dependent pathway [17]. In E. coli, there is a selective pressure to maintain chromosome symmetry; i. e. two nearly equal length replication arms [18]. The datA, DARS1, and DARS2 regions have the same relative distance to oriC in all E. coli strains sequenced [19], but none of the loci alone or in combination are essential for cell viability. Loss of either region is however associated with a fitness cost [19]. The E. coli chromosome consist of four insulated macrodomains (MD) and two less constrained regions called nonstructured (NS) regions [20,21]. DNA recombination occurs preferentially within MDs, while DNA interactions between the different MDs are highly restricted. The two NS regions can however interact with both its flaking MDs [21]. The Ori MD is flanked by the two NS (NSRight and NSLeft) whereas the Ter MD is flanked by the Left and Right MDs [21]. Both oriC and datA are located within the Ori MD, DARS2 within the NSLeft, DARS1 within the Right MD, and terC in the Ter MD [19]. Chromosomal rearrangements resulting in a mixture of different macrodomains have deleterious effects of cell growth whereas rearrangements within them are better tolerated [22]. Chromosome organization is reported to affect gene expression, mainly at the transcriptional level. A recent study found that a reporter gene cassette, comprised of the lac promoter driving expression of gfp, varied ∼300-fold depending on its position on the chromosome, in a manner fairly unrelated to the replication-associated gene dosage [23]. Gene expression even varies between insertion sites within the same MD, and both MD- and NS-regions contain high and low activity regions [23], due to intrinsic properties of the region. However, for fast growing bacteria the replication-associated gene dosage determines the organization of the chromosome; i. e. genes involved in translation and transcription (but not other highly expressed genes) are located close to oriC [24]. The selective pressure that keeps these genes in the relative proximity to oriC could be due to the challenges that E. coli faces a very high growth rates, which are hardly observable at lower growth rates [24]. The replication-associated gene dosage could also be important for the activity of non-coding regions such as datA, DARS1, and DARS2. Here we show that the genomic location of datA, DARS1, and DARS2 are important for cell cycle control and for bacterial fitness. This provides a direct link between DNA replication control and genomic positioning. The conserved relative distances from oriC to the DARS1, DARS2 and datA regions in E. coli [25], suggest that their chromosomal positions are important for correctly controlled replication initiation. In order to construct strains carrying datA, DARS1, and DARS2 at different loci, eight Tn10 insertions from the Singer library were chosen [26,27] (Fig 1A). Strains had their respective chromosomal datA, DARS1, or DARS2 loci deleted, and the tetC gene of Tn10 replaced with the datA, DARS1, or DARS2 loci respectively, resulting in strains carrying a single copy of the respective region at a new chromosomal location. Mutant strains were evaluated by growth- and flow cytometry studies (Materials and Methods). Wild-type cells exhibited the expected synchronous initiation pattern with the majority of cells containing 2,4 or 8 replication origins (Fig 1B). Cells deficient in datA have an increase in the DnaAATP/DnaAADP ratio [15,28], resulting in an increased origins/mass, and a high degree of initiation asynchrony (Fig 1I) [25]. Rifampicin-resistant initiations, that could be suppressed by increasing the drug concentration, have previously been reported for cells deficient in datA [29]. However, a fourfold increased rifampicin concentration did not affect any of the parameters measured here. Relocating datA to any of the positions close to oriC (CAG18499 or CAG18496), datA (CAG18488), or DARS2 (CAG12173) resulted in cell cycle parameters similar to wild-type (Table 1). Relocation to the positions near DARS1 (CAG18493) or especially terC (CAG12151 and CAG18461) resulted in an increase in origin concentration but no asynchrony (Table 1; Fig 1K). This suggests that although datA is functional when located close to the termini the cells initiate replication at a decreased initiation mass relative to wild-type. This correlates with previous observations [28]. To test if the reduced datA function in the terminus region was related to gene dosage we integrated a 2nd datA region close to terC (ΔdatA CAG12151: : datA CAG18461: : datA; Table 1). A cell with two datA regions close to terC has an origin concentration slightly below wild-type, but initiates initiation in synchrony. The chromosomal wild-type position of datA has a gene dosage of 2, while the Tn10 positions CAG12151 and CAG18461 have gene dosages of 1. 1 and 1. 0, respectively (Materials and Methods). This suggests that the copy number is important for correct datA function; a single datA relocated to terC has a reduced function, which can be complemented by a second datA in terC that brings the overall datA copy number up to wild-type level. Cells deficient in DARS1 (Fig 1C) or DARS2 (Fig 1F) have a reduced DnaAATP/DnaAADP ratio compared to wild-type [16,17]; resulting in a reduced origin concentration compared to wild-type and our data are in agreement with this. Asynchrony of initiation was only observed in ΔDARS2, but not in ΔDARS1 [16]. The chromosomal position of DARS1 did not affect cellular doubling time or single cell initiation synchrony (Table 1). When DARS1 was relocated from its normal position to one of the two positions bordering oriC, the cells contained an elevated number of origins and an increased origin concentration (Fig 1D). We propose that DARS1 located close to oriC results in a phenotype similar to datA relocated to the termini (see above), where cells initiate replication at a decreased initiation mass relative to wild-type. Relocation of DARS2 influenced cell cycle parameters to a larger extent than DARS1. A wild-type phenotype was obtained for the five locations, CAG18499, CAG18496, CAG18488, CAG12135, and CAG18493 (Table 1). However when DARS2 were moved to positions close to terC, cells initiated asynchronously and had a decreased origin concentration relative to wild-type, although less than that of a DARS2 deletion (Fig 1H); indicating a still somewhat functional DARS2 in the terminus. In addition we tested the relocation of DARS2 to yafJ, a putative glutamine amidotransferase, located with the same relative distance to oriC as the wild-type DARS2 position, but on the other replication arm. Here DARS2 had a similar doubling time, origin concentration, and synchrony as the wildtype (Table 1). In order to assess whether the replication defect of cells carrying DARS2 near terC resulted solely from a gene dosage effect we cloned DARS2 into the F-based plasmid pALO277 [30]. The F plasmid has a copy number of 1–2 plasmids per genome equivalent [31], which is equal to or higher than the copy number of DARS2 at its chromosomal location (Materials and Methods). Replication of the F plasmid is limited by the availability of RepE [32], but DnaA is also required [33]. To ensure that the presence of DARS2 on pALO277 did not alter the copy number we determined the number of the F plasmid relative to the termini by qPCR analysis for the wild-type cell transformed with either pALO277 or pALO277: : DARS2 (S1 Fig). The data shows that the presence of DARS2 on the F plasmid does not alter the copy number. Wild-type cells with pALO277: : DARS2 showed an increased origin/mass, indicating a functional DARS2 locus on pALO277: : DARS2 (S1 Fig). ΔDARS2 cells were only partly complemented by pALO277: : DARS2 as their origin concentration remained below wild-type level and initiation synchrony was never restored. These observations suggest that the contribution of DARS2 to cell cycle control is not solely through copy number. As plasmid F replicates randomly in the cell cycle [34] the coordination of DARS2 replication relative to oriC and other cis-acting control regions is no longer present and this may explain why DARS2 deficiency cannot be complemented by a plasmid borne DARS2 copy. DARS1 and DARS2 both rejuvenate inactive DnaAADP to active DnaAATP [16,17]. We proceeded to investigate if loss of either locus could be complemented by an extra copy of the other. The effect of having two functional DARS2 loci were investigated in a ΔDARS1 cell, i. e. , retained the original wild-type DARS2 locus, with the addition of a DARS2 inserted into one of seven Tn10 positions. Loss of DARS1 did not result in initiation asynchrony, and an extra copy of DARS2 only changed this when inserted close to oriC in CAG18499 (although only marginally with and asynchrony index of 0. 17) but not in any of the remaining tested positions (Table 2). Why a 2nd copy of DARS2 in CAG18499 but not CAG18496 gives asynchrony is unknown. DARS2 reintroduced close to the termini, close to the wildtype DARS2 position and even close to datA, which is located near oriC, failed to fully complement loss of DARS1 with respect to origin concentration (Table 2; S2 Fig). An additional copy of DARS2 only fully complemented DARS1 deficiency, i. e. with respect to origin concentration and synchrony, when introduced at a position close to oriC (CAG18496) or precisely at the wild-type DARS1 position (replacing the chromosomal DARS1 copy with a 2nd chromosomal DARS2 copy). This suggests that a relatively high gene dosage of DARS2 is required to complement DARS1 deficiency and that the context of the DARS1 region might be favorable for rejuvenation of DnaAADP to DnaAATP, hence the 2nd DARS2 copy here can fully complement to wild-type even though the chromosomal position does not provide a high copy number. The ability of DARS1 to complement DARS2 deficiency was addressed by the same approach. The decrease in origin concentration observed for DARS2 deficient cells could be fully complemented by an additional DARS1 copy, irrespective of chromosomal position (Table 2; S2 Fig), although slightly elevated when DARS1 were reintroduced close to oriC. However, the additional DARS1 locus failed in all cases to complement the asynchrony phenotype of DARS2 deficient cells. These data suggests that DARS2 is a poor replacement for DARS1 and vice visa. They also indicate that DARS1 and DARS2 serves different functions that are both required for an efficient control of the cell cycle. The optimal chromosomal positions of DARS1 and DARS2 were determined by in-culture evolution using a novel transposon mediated approach. Here, the chromosomal DARS1 and DARS2 loci were cloned into the mini Tn10 based transposon NKBOR (on pNKBOR) [35], resulting in NKBOR: : DARS1 (pJFM3) and NKBOR: : DARS2 (pJFM1), respectively (Materials and Methods). pNKBOR is a R6K-based suicide vector [35]. Hence when pJFM1 or pJFM3 were transformed into the Pir deficient strain MG1655, random insertions of either NKBOR: : DARS1 or NKBOR: : DARS2 were obtained; these are for simplicity referred to as DARS1 or DARS2 insertions, respectively. Five different experiential set-ups were performed; NKBOR into wild-type, DARS1 into ΔDARS1 and ΔDARS1 ΔDARS2 cells, and DARS2 into ΔDARS2 and ΔDARS1 ΔDARS2 cells, where an estimated 70,000 different transposon insertions were obtained for each, corresponding to an insert pr. 65 base pairs. We feel this is more than adequate to evaluate the importance of the genomic position for the non-coding regions. The 70,000 different insertions were pooled for each set-up (t = 0) and continuously propagated in LB for a total of 700 generations (t = 700). The hypothesis is that the optimal position of DARS1/DARS2 would result in the fittest clone, which over time would out-compete the rest. A Southern Blot, probed for NKBOR, was performed to confirm that the output of the direct competition experiment would become more clonal over time (Fig 2). Representative insertion sites from the Input, selected time points and end (t = 700) transposon pools were mapped by full genome sequencing (Materials and Methods). In addition, single clones were isolated after 700 generations from each set-up, for further investigation, and their precise transposon locations were determined by easy gene walking ([36]; Materials and Methods). The mini-transposon NKBOR (Fig 2A), DARS1 (Fig 2B and 2C), and DARS2 (Fig 2B and 2C) were as expected inserted randomly throughout genome at t = 0. After 700 generations of growth, all NKBOR insertions into wild-type cells were mapped to three locations in the fimE gene (termed NKBOR Clone fimE #1, #2, and #3, Fig 3A, S1 Table). This indicates that under the used experimental settings the disruption of fimE resulted in a fitness advantage. Our Southern Blot showed two distinct bands for NKBOR insertion. However, we do not know the origin of the second band. The fimE gene encodes FimE that along with FimB are the two recombinases, which mediate inversion of the DNA element containing the promoter for fimA (Type I fimbriae) [37]. We investigated whether NKBOR insertion into the fimE gene altered the level of fimbriae transcription between wild-type and NKBOR fimE mutants using qPCR. The level of fimA transcript in NKBOR and DARS1 insertions into fimE was increased 7- to 31-fold relative to wild-type (S3A Fig), and remnants of pellicle formation were observed (S3B Fig). This indicates that the fitness advantage linked to loss of fimE was due to an increased expression of fimbriae although other yet to be discovered fimE-regulated pathways could be involved. Selection of DARS1 insertions from a random pool in DARS1 or DARS1 DARS2 deficient cells resulted in fewer bands over time on a Southern Blot; i. e. , six and five separate band at t = 700, respectively (Fig 2B and 2C). By genome sequencing 71. 4% of all DARS1 insertions in DARS1 deficient cells mapped to four different insertions in the fimE gene (termed DARS1 Clone fimE #1, #2, #3, and #4, Fig 3B, S2 Table). DARS1 was also inserted in the ydeS gene close to terC (DARS1 Clone ydeS) and in the intergenic region between yfbN and yfbO (DARS1 Clone yfb), that of interest has the same gene dosage as the wild-type DARS1 copy. The functions of putative proteins encoded by ydeS, yfbN and yfbO are not known. Also, DARS1 in DARS1 deficient cells was found twice in yghG (termed DARS1 Clone yghG #1 and #2) and in adeQ (termed DARS1 Clone adeQ). In enterotoxigenic E. coli YghG is an outer membrane lipoprotein that is required for the correct localization of the GspD secretin in the outer membrane [38], while AdeQ is a high-affinity adenine transporter in E. coli K-12 [39]. These results indicate that the gene dosage of DARS1 does not specify its location; i. e. , no unique location of DARS1 was selected for. Transposon insertions sites for DARS1 into DARS1 DARS2 deficient cells were resolved using easy gene walking only. We isolated 20 single clones and mapped their transposon insertion sites. Here, two different insertions were mapped to the fimE gene (termed DARS1 Clone fimE #5 and #6, Fig 3C, S3 Table). DARS1 was also inserted in the intergenic region between gspA and gspC (DARS1 Clone gsp) and in the intergenic region between tomB and acrB (DARS1 Clone tomB). tomB is in a toxin-antitoxin operon with hha, where expression of TomB diminish the toxicity of Hha expression [40]; while the gspCDEFGHIJKLMO (gspC-O) and gspAB operons encode homologs of type II secretion machinery involved in extrusion of folded proteins [41]. Selection of DARS2 insertions from a random pool in DARS2 deficient cells resulted in the selection of only one band at t = 700 (Fig 2B), indicating a single optimal chromosomal position for DARS2. This correlates well the data above; i. e. movement of DARS2 had a larger effect on the control of the cell cycle than movement of DARS1. Genome sequencing mapped 97. 8% of the DARS2 insertions to two locations approximately 650 bp from the wild-type DARS2 position, and separated by only 8 bp (S4 Table; Fig 3D). One insertion was located in the intergenic region between the ptsP and rppH gene (DARS2 Clone IR). This clone was already present at high frequency after 500 generations of growth (S4 Table). The second DARS2 proximal insertion was located inside the rppH gene (DARS2 Clone rppH), which encodes for RppH an RNA pyrophosphohydrolase that initiates mRNA decay [42]. The remaining 2. 2% of the mapped inserts were found in the intergenic region between nanS and nanM (termed DARS2 Clone nan). nanS and nanM are transcribed in an operon with nanC, which supports the efficient use of α-N-acetylneuraminate as the sole source of carbon [43]. The Southern Blot for DARS2 into ΔDARS2 showed an abrupt changed between 400 and 500 generations of direct completion (Fig 2B), therefore both time points were resolved by full genome sequencing (S4 Table). Several interesting DARS2 insertions were found in t = 400, which were not present when the experiment was terminated after 700 generations. One DARS2 insertion was found in chromosomal position 2. 316. 202 bp (S4 Table), which is close to the tested DARS2 mirror position on opposite replication arm (yafJ; Table 1). This suggest that these and other may have a fitness advantage over the majority of the initial 70. 000 insertions, but that they were overall less fit than the insertions immediately flanking DARS2 strongly suggesting that the wild-type position is optimal for DARS2 function. Two pairs of insertion sites for DARS2 into ΔDARS1 ΔDARS2 were resolved after 700 generations of direct competition. One pair included closely spaced insertions in the lgoR gene (DARS2 Clone lgoR) and in the intergenic region between lgoR and lgorT (DARS2 Clone IR lgor) and the other were two insertions in the gene ttdR (DARS2 Clone ttdR #1 and #2) (Fig 3E). lgoR is a predicted transcriptional regulator that is essential for growth on L-galactonate as the sole carbon source [44], while ttdR is an LysR-type transcriptional regulator for L-tartrate fermentation that is induced under anaerobic growth [45]. The two DARS2 insertions, lgoR and ttdR, have almost the same distance to oriC albeit located on each replication arm (S5 Table). Thus, indicating that the optimal position of DARS2 in ΔDARS1 ΔDARS2 cells could be linked to gene dosage. The abrupt change between t = 300 and t = 400 for DARS2 into ΔDARS1 ΔDARS2 (Fig 2C) was also investigated by full genome sequencing (S5 Table). Of interest we find none of the selected end-point at t = 300, but DARS2 Clone ttdR #1, DARS2 Clone lgo, and DARS2 Clone lgoR was found at t = 400. Representative transposon insertions identified by easy gene walking were moved into a fresh background by P1 transduction and analyzed by flow cytometry (S6 Table). None of the selected clones had a different doubling time compared to wild-type (+/- 2 minutes). All DARS1 insertions in ΔDARS1 cells (fimE #1, fimE #2, ydeS, and yfb) restored cell cycle parameters to those of wild-type cells, and all DARS1 insertions in ΔDARS1 ΔDARS2 cells (fimE #5, fimE #6, gsp, and tomB) restored the phenotype to that of DARS2 deficient cells. The DARS2 insertions next to the deleted DARS2 locus (IR and rppH) fully complemented loss of DARS2. Furthermore, the DARS2 insertions selected in ΔDARS1 ΔDARS2 cells (lgoR and tddR) restored the low origin/mass ratio to levels close to a cell deficient in only DARS1. We proceeded to investigate the fitness of representative selected clones. Here the DARS2 Clone IR, which was predominant after 700 generations culture evolution, was tested against wild-type and ΔDARS2 cells. The wild-type was found to be dominant to ΔDARS2 (1. 5 LOG differences after 80 generations of competition) (S4A Fig) as expected [19]. The wild-type was also slightly dominant over the DARS2 Clone IR (0. 7 LOG differences) (S4B Fig) whereas DARS2 Clone IR was slightly dominant to ΔDARS2 (0. 8 LOG differences) (S4C Fig). Thus, although DARS2 Clone IR did not have an identical fitness to the wild-type it was more fit than ΔDARS2 as expected. The pattern seen for DARS2 Clone IR was also observed for DARS1 Clone ydeS tested against the wild-type and ΔDARS1, while DARS1 Clone fimE #3 was dominant to both the wild-type and ΔDARS1; probably due to the fitness advantage of mutating fimE in the used competition experimental setup. Gene expression is known to be affected by local chromosomal context [23]. The effect of transcription on activity of cis-acting regions such as datA, DARS1 and DARS2 is not known. The three regions were cloned into the R1 based plasmid pNDM220 [46], downstream on the IPTG inducible pA1/O4/O3 promoter. The three resultant plasmids were transformed into wild-type cells and their effect on the initiation of replication, in the presence or absence of IPTG, assessed by flow cytometry (Fig 4). The vector plasmid pNDM220 did not alter synchrony or origin concentration irrespective of IPTG addition (compare Fig 4A and 4E). In the absence of IPTG, plasmid pNDM220: : datA reduced the cellular origin content and concentration without affecting synchrony (compare Fig 4A, 4B and 4I). Plasmid-carried DARS1 and DARS2 increased the cellular origin content and concentration (compare Fig 4A and 4C (for DARS1), Fig 4D (for DARS2), Fig 4I) as previously observed [16]. Only the presence of pNDM220: : DARS2 resulted in asynchronous initiations (Fig 4D). Strong transcription through either datA, DARS1 or DARS2, by addition of IPTG, restored both origin concentration and synchrony to wild-type level (Fig 4F, 4G and 4H). Therefore, transcription through datA, DARS1, and DARS2 is detrimental to their function and further enforce that the activity of either region could be affected by local chromosomal context. There are at least two explanations for the over-initiation resulting from datA relocation to terC; i. e. , gene dosage and local access to DnaAATP (compare Fig 5A to 5B). The chromosomal datA position is close to oriC and this was proposed to enhance interaction with DnaAATP released from oriC at initiation, resulting in a high DDAH activity [15]. Thus, when datA is relocated to terC, both gene dosage and interaction with DnaAATP is diminished resulting in an increased DnaAATP/DnaAADP ratio and increased origin concentration (Fig 5B). The Gram-positive bacteria Bacillus subtilis also contains DnaA box clusters (DBC) that analogue to datA can repress untimely initiation [47]. The DBC in B. subtilis is however, unlike datA, only shown to titrate DnaAATP [47]. Of interest, the genomic position of the DBC is important for regulation of initiation; i. e. relocation from close to oriC (wild-type position) to terC reduces its function [47], as shown here and previously [28] for datA. In addition, we find that the diminished datA function in terC can be complemented to wild-type by a second datA copy in the terminus, suggesting that gene dosage is an important parameter for an optimal datA function. Directed translocation of DARS1 indicated little preference for location, and only insertions very close to oriC resulted in an increased origin concentration, possibly due to a gene dosage effect or increased interaction with oriC associated DnaA (Fig 5C). In agreement with this, culture evolution revealed no optimal chromosomal position for the DARS1 region, although it´s presence is important [19]. The strong selection for loss of fimE in otherwise wild-type cells, suggest that the DARS1 insertions into fimE does not indicate the optimal position of DARS1 but results from a fitness advantage of disrupting fimE. The level of fimA transcription was increased in fimE mutants relative to the wild-type. As fimA encode the Type I major fimbrial subunit, fimE mutants are likely to carry an elevated number of fimbria. fimE mutant cells formed a pellicle at the top of the culture tubes suggesting, that increased fimbriation facilitates migration towards more aerobic conditions, which in turn could lead to a small growth advantage relative to wild-type cells. The selected DARS1 insertions into ydeS, yghG, adeQ or yfb in DARS1 deficient cells or in tomB or gsp in DARS1 DARS2 deficient cells does not immediately suggest why these locations should provide a fitness advantage, and it may well be that these would be outcompeted by fimE insertions had competition proceeded longer. As the majority of the insertions are distant from oriC, it may be a fitness disadvantage to have DARS1 next to the origin. Directed translocation of DARS2 indicated that many positions (CAG18499,18496,18488,12135,18493, and yafJ) restored origin concentration and synchrony to wild-type levels. Yet, only two DARS2 insertions were selected in ΔDARS2 cells after 700 generations of growth; 97. 8% were located next to the wild-type DARS2 position, while 2. 2% were located in the intergenic region between nanS and nanM. At t = 400, an insertion was found on the opposite replication arm with an almost identical distance to oriC as the wild-type DARS2 position, but this insertion was not recovered after 700 generations. Thus, gene dosage cannot be the single determinant for optimal position. This is corroborated by the inability of a low-copy DARS2 plasmid to complement DARS2 deficiency on the chromosome. The present data therefore implies that DARS2 and the immediate surroundings at its wild-type position specify the optimal genomic position. Even though we only investigate the effects of DARS2 translocation on the regulation of initiation of replication, we cannot exclude that the optimal DARS2 position, either directly or indirectly, contributes to fine tune other DnaA functions in the cell. A cell with two functional DARS2 regions only resembled wild-type cells when the additional DARS2 region was located close to oriC (high gene dosage) or directly exchanged the wild-type DARS1 region. This suggests that the local genomic context at DARS1 also provide an ideal setting also for DARS2 activity. Interestingly, RNA sequencing reveals very low transcriptional activity at the genomic DARS1 and DARS2 locations [48], and we observed that both DARS1 and DARS2 lost activity by transcription though the regions (Fig 4). Therefore both gene dosage and local transcriptional activity contribute to DARS activity. Recent reports also highlight the importance of a conserved genomic position for the function of key regulatory genes [49–51]; i. e. the local genomic environment is important for function. Thus, the importance of the local genomic environment may very well contribute to the selected positions of DARS1 and DARS2. The activity of DARS1 is neither cell cycle nor growth phase regulated; i. e. always active [16]. Thus, de novo synthesis of DnaA will along with the constitutive DnaAADP rejuvenation at DARS1 ensure a steady DnaAATP increase throughout the cell cycle of wild-type cells. Therefore increased DARS1 gene dosages lead to an increased origin concentration (Fig 5C). Rejuvenation at DARS2 is dependent on the binding of both Fis and IHF [17]. Therefore the activity at DARS2 is growth phase regulated by Fis, i. e. is only active during exponential growth [17], and cell cycle regulated by IHF [17]. Maximal IHF binding to activate DARS2, immediately precedes initiation [17]. Initiation of all cellular origins in synchrony was explained by a mechanism where DnaAATP released from the first origin initiated will trigger initiations on fully methylated not yet initiated origins, by a cascade-like mechanism [52]. However, new initiations will inevitably lead to more DNA loaded β-clamps that are instrumental in RIDA [12] which is therefore expected to accelerate. We propose that timely duplication and activation of DARS2 raises the DnaAATP/DnaAADP ratio to a sufficient high level at the onset of initiation, to ensure that all origins present in the cell are initiated in virtual synchrony even though RIDA is increased during the initiation period (for review see [53]). When DARS2 is located near terC the gene dosage may be too low to result in a sufficient pre-initiation burst in DnaAATP to ensure synchronous initiation at all cellular origins despite of its activity remaining cell cycle regulated (Fig 5D). This explains why a cell with two DARS2 loci will initiate in synchrony but fail to obtain a wild-type like origin/mass, while a cell with two DARS1 loci will be asynchronous (Table 2). Thus, DARS1 is primarily responsible for coupling replication initiation to cell mass increase, whereas DARS2 is primarily important for maintaining synchronous initiation of all origins contained within a single cell. DARS1 and DARS2-like sequences (including conserved IHF- and Fis binding sites) with genomic positions similar to E. coli have been identified in E. coli-like species [16,17]. This suggests not only a common mechanism to regulate initiation between species, but also that the genomic position of the regulatory regions are important for correct function in the related species. The variable size of the E. coli genome, between 4. 6 to 5. 7 mega base pairs (Mb), indicates that horizontal gene transfer and genome reductions frequently takes place [54]. Thus, its puzzling how new DNA is distributed along the genome to preserve the observed chromosomal symmetry in E. coli [19], which as shown here at least for DARS2, is important for correct fine tuning of initiation and fitness. Cells were grown in Luria–Bertani Broth (LB) medium or AB minimal medium supplemented with 0. 2% glucose, 10 μg/ml thiamine, and 0. 5% casamino acids. Cells were cultured at 37°C. When necessary, antibiotics were added to the following concentrations: kanamycin, 50 μg/ml; chloramphenicol, 20 μg/ml; ampicillin, 150 μg/ml; tetracycline, 10 μg/ml. All strains are found in Table 3. A set of strains containing the transposon Tn10 (encoding tetracycline resistance) at known positions on the chromosome were described previously [27] and generously provided by Dr. Martin G. Marinus (Fig 1A). None of the Tn10 insertions from the Singer library altered origin concentration (origins/mass), synchrony of initiation of replication, or doubling time of otherwise wild-type cells. The Tn10 insertions CAG18499, CAG18496, CAG18488, CAG12173, CAG1315, CAG18493, CAG12151, and CAG18461 were individually moved into BW25113 [55], MG1655 strR (ALO4292), MG1655 strR ΔDARS1 (ALO4314), MG1655 strR ΔDARS2 (ALO4312), and MG1655 strR ΔdatA: : kan (ALO4331) [19] by P1 transduction using established procedures [56] and by selection for tetracycline resistance. In the individual Tn10 constructed strains the tetC gene on the Tn10 was replaced by either the datA, DARS1, or DARS2 region linked to the cat gene by the lambda red procedure [55]. Briefly, datA was PCR amplified from pJFM9, DARS1 from pJFM5, and DARS2 from pJFM6 using primers tetC_FW and tetC_RV (S7 Table). tetC_FW and tetC_RV will PCR amplify both the cloned locus into pKD3 (see above) along with the cat cassette including the two FRT-sties. CAG18499: : datA: : cat, CAG18496: : datA: : cat, CAG18488: : datA: : cat, CAG12173: : datA: : cat, CAG18493: : datA: : cat, CAG18461: : datA: : cat, and CAG12151: : datA: : cat were individually moved from BW25113 into MG1655 strR ΔdatA: : kan by P1 transduction and by selection for cat and kan resistance. Hence, derivatives of the wild-type strain devoid of datA at the original locus and instead carrying Tn10: : datA at indicated chromosomal loci were created. In addition, a strain with two Tn10: : datA loci in the terminus region were created by removing the cat gene from CAG18461: : datA: : cat, by pCP20, according to a method described previously [57]. Thereafter the CAG12151: : datA: : cat were moved from BW25113 into MG1655 strR ΔdatA: : kan CAG18461: : datA by P1 transduction, selecting for cat resistance, resulting in a strain with datA inserted in both CAG18461 and CAG12151 (ALO5041). CAG18499: : DARS1: : cat, CAG18496: : DARS1: : cat, CAG18488: : DARS1: : cat, CAG12135: : DARS1: : cat, CAG12173: : DARS1: : cat, CAG18493: : DARS1: : cat, and CAG18461: : DARS1: : cat, CAG12151: : DARS1: : cat were individually moved from BW25113 into MG1655 strR ΔDARS1 and MG1655 strR ΔDARS2 by P1 transduction and by selection for cat resistance. Hence, derivatives of the wild-type strain devoid of DARS1 at the original locus and instead carrying Tn10: : DARS1 at indicated chromosomal loci were created (Table 2). Similarly, derivatives of the wild-type strain devoid of DARS2 at the original locus and instead carrying Tn10: : DARS1 at indicated chromosomal loci were created. CAG18499: : DARS2: : cat, CAG18496: : DARS2: : cat, CAG18488: : DARS2: : cat, CAG12135: : DARS2: : cat, CAG18493: : DARS2: : cat, CAG18461: : DARS2: : cat, and CAG12151: : DARS2: : cat were individually moved from BW25113 into MG1655 strR ΔDARS1 and MG1655 strR ΔDARS2 by P1 transduction and by selection for cat resistance. Hence, derivatives of the wild-type strain devoid of DARS2 at the original locus and instead carrying Tn10: : DARS2 at indicated chromosomal loci were created (Table 2). Also derivatives of the wild-type strain devoid of DARS1 at the original locus and instead carrying Tn10: : DARS2 at indicated chromosomal loci were created. Furthermore, DARS2 were relocated to yafJ. yafJ was replaced with DARS2 amplified from pJFM6 using primers yafJ_FW and yafJ_RV (S7 Table) in MG1655 StrR ΔDARS2 harboring pKD46, resulting in the MG1655 StrR ΔDARS2 yafJ: : DARS2: : cat mutant (ALO5042). DARS2 was also relocated to DARS1 in MG1655 StrR. DARS1 was replaced with DARS2 amplified from pJFM6 using primers D2toD1_FW and D2toD1_RV (S7 Table) in MG1655 StrR harboring pKD46, resulting in the MG1655 StrR ΔDARS1: : DARS2: : cat mutant (ALO5043). A spontaneous tonA mutant of E. coli MG1655 was isolated by Dr. Stanley Brown, resulting in E. coli MG1655 tonA (ALO4255). The DARS1 region was replaced with the cat gene in MG1655 harboring pKD46, resulting in the ΔDARS1: : cat mutant (ALO4075). Briefly, the cat gene was PCR amplified using primers DARS1_KO_FW and DARS1_KO_RV from pKD3. The resultant DNA fragments were introduced into ALO1825 bearing pKD46. Each deletion was verified by PCR. The DARS1 deletion was moved from ALO4075 to ALO4255 by P1 transduction, selecting for chloramphenicol resistance, resulting in MG1655 tonA ΔDARS1: : cat (ALO4256). The DARS2 deletion was moved from ALO4254 into ALO4255 by P1 transduction, selecting for chloramphenicol resistance, resulting in MG1655 tonA ΔDARS2: : cat (ALO4257). The cam cassette was removed from ALO4257 and the DARS1 deletion was moved from ALO4075 into the chloramphenicol sensitive MG1655 tonA ΔDARS2 by P1 transduction, selecting for chloramphenicol resistance, resulting in MG1655 tonA ΔDARS1: : cat ΔDARS2 (ALO4259). The replication-associated gene dosage was estimated for every position for growth in ABTG + CAA and LB. Eq (1) was used to calculate the replication-associated gene dosage Here the average no. of copies per chromosome of a gene with position x on the chromosome (x = 0 at the origin and x = 1 at the terminus), C is the replication period in minutes, D is the time following termination of replication until cell division, and τ is the doubling time [58]. At generation times below 60 min the C- and D period has been shown to be constant in E. coli K-12 strains; i. e. 42 minutes and 33 minutes, respectively [59]. Generation time for the wild-type grown in ABTG + CAA was shown to be 40 minutes. By Eq (1) the gene dosage of DARS2 is calculated to be 1. 66 copies/cell at the given growth rate. Hence, by Eq (2), the cellular DNA content (G) was found to be 1. 82 genome equivalents per cell. As the copy number of plasmid F is 1–2 per genome equivalent [31] it follows that the copy number per cell of the F plasmid pALO277 and derivatives is 1. 82–3. 64. pNKBOR is a R6K-based suicide vector that permits the random insertion of a mini-transposon (NKBOR) into a π protein deficient E. coli chromosome [35]. Here we transformed pJFM1 (NKBOR: : DARS2) into ΔDARS2 and ΔDARS1 ΔDARS2, pJFM3 (NKBOR: : DARS2) into ΔDARS1 and ΔDARS1 ΔDARS2, and pNKBOR (NKBOR) into wild-type selecting for kan resistance. Approximately 70,000 random insertions were obtained for each transformation. The 70,000 strains were pooled and inoculated in the same tube. They were grown in LB aerated by continuous shaking at 37°C. The populations were propagated by continuously transfers after estimated 10 generations. Samples for genomic DNA from each population were taken at 100-generation intervals, until direct competition for estimated 700-generations (used for southern blot; Fig 2). After 700-generations of direct competition 10 single clones was isolated from NKBOR: : DARS2 into ΔDARS2,20 single clones were isolated from NKBOR: : DARS1 into ΔDARS1, from NKBOR: : DARS1 into ΔDARS1 ΔDARS2, and from NKBOR: : DARS2 into ΔDARS1 ΔDARS2, and 15 single clones were isolated from NKBOR into wild-type. Total cellular DNA was prepared according to Løbner-Olesen and von Freiesleben [60]. DNA was digested with PvuI, and fragments were separated on a 0. 7% agarose gel, transferred by capillary transfer to a Hybond-N+ membrane (Amersham Pharmacia Biotech), and probed with an approx. 1 kb NKBOR fragment, which hybridize to NKBOR. The probe was prepared by PCR amplification using primers NKBOR_Probe_FW and NKBOR_Probe_RV (S7 Table) using pNKBOR as template and labeled with [α-32P]dATP (Amersham Pharmacia) using the Random Primer system (DECAprime II DNA Labeling Kit; Life Technologies). The chromosomal position of the NKBOR, NKBOR: : DARS1, and NKBOR: : DARS2 insertion was found as described by Harrison et al. [36], using nested primers specific to NKBOR (NKBOR_N3, NKBOR_N2, and NKBOR_1) and random primers Random_HindIII and Random_Sau3AI (S7 Table). Whole-genome sequencing was performed at the University of Copenhagen on an Illumina MiSeq benchtop sequencer. A total of 6 million paired-end reads were generated, with read length of 35 to 300 nucleotides. Reads were aligned to 150 N’s contiguous to NKBOR, AF310136. 1 1,904…2,204 using Bowtie2 [61] with interval between seed substrings = S, 1,1. 15 and maximum number of ambiguous characters = L, 0,0. 9. Aligned reads were then aligned to MG1655 ref|NC_000913. 3 and NKBOR gb|AF310136 using blastN [62] and sorted for contiguous alignment MG1655-NKBOR. Note that the coverage in the present deep sequencing was insufficient for a complete mapping of insertion sites. The data presented in S1–S5 Tables therefore contains a representative subset of the total number of insertions. Total RNA from bacterial samples was extracted using the GeneJET RNA Purification Kit (Thermo Fisher Scientific) according to the manual. Following treatment of RNA with TURBO DNase (Ambion), cDNA was synthesized using the RevertAid First Strand cDNA Synthesis Kit for reverse transcriptase PCR according to the manufacturer’s protocols (Thermo Fisher Scientific). In parallel, RNA samples were subjected to agarose gel electrophoresis and NanoDrop (Thermo Fisher Scientific) to verify quality and yield. qPCR with primers specific to the rpoA gene (the α subunit of the RNA polymerase core enzyme) (S7 Table) [63] was performed on cDNA samples prepared with and without reverse transcriptase to confirm no genomic DNA contamination of the RNA preparations following DNase treatment. The relative distance between oriC and transposon insertions sites were calculated as described previously [19]. Flow cytometry was performed as described previously [65] using an Apogee A10 instrument. For each sample, a minimum of 30. 000 cells were analyzed. Numbers of origins per cell and relative cell mass were determined as described previously [65]. Asynchrony was calculated as described by Løbner-Olesen et al. [65]. Initiations were considered asynchronous when A>0. 1. The fitness of DARS2 Clone rrpH, DARS1 Clone fimE #3, and DARS1 Clone ydeS were compared to the wild-type and either ΔDARS1 or ΔDARS2 (indicated in the text) during direct competition in LB medium. The competing strains were inoculated pairwise at an approximate concentration of (107 CFU/mL) each. The populations were propagated by continuously transfers in LB medium. Samples from each population were taken at 10-generation intervals. Each sample was diluted in 0. 9% NaCl and plated on LB plates with appropriate antibiotics. All plates were incubated for 18–24h at 37°C prior to counting.

Replication of the E. coli chromosome is the central event in the cell cycle, with the control of replication enforced at the level of initiation. DnaA is the key protein responsible for initiation at the origin of replication (oriC), and is active in this when bound to ATP and inactive when bound to ADP. The activity of DnaA is in part controlled by three non-coding DNA regions; datA, DARS1, and DARS2. Here we show that the chromosomal position of datA, DARS1, and especially DARS2 relative to oriC, is important for cell cycle control and bacterial fitness. Based on this and previous work, we propose that the functionally redundant DARS1 and DARS2 regions play different roles in replication control. DARS1 is mainly involved in maintaining the origin concentration, while DARS2 is also involved in maintaining single cell synchrony. Both regions are needed for proper replication control, and this also provides an explanation for the conservation observed in all sequenced E. coli strains. Further, the present literature indicates that these observations can be applied to other related bacteria.

Abstract Introduction Results Discussion Materials and Methods

bacteriology sequencing techniques cell cycle and cell division cell processes microbiology cloning genome sequencing genetic elements bacterial genetics molecular biology techniques microbial genetics microbial genomics research and analysis methods bacterial genomics artificial gene amplification and extension chromosome biology molecular biology genetic loci cell biology genetics transposable elements biology and life sciences genomics mobile genetic elements polymerase chain reaction chromosomes

2016

DNA Replication Control Is Linked to Genomic Positioning of Control Regions in Escherichia coli

Mlh1-Mlh3 (MutLγ) is a mismatch repair factor with a central role in formation of meiotic crossovers, presumably through resolution of double Holliday junctions. MutLγ has DNA-binding, nuclease, and ATPase activities, but how these relate to one another and to in vivo functions are unclear. Here, we combine biochemical and genetic analyses to characterize Saccharomyces cerevisiae MutLγ. Limited proteolysis and atomic force microscopy showed that purified recombinant MutLγ undergoes ATP-driven conformational changes. In vitro, MutLγ displayed separable DNA-binding activities toward Holliday junctions (HJ) and, surprisingly, single-stranded DNA (ssDNA), which was not predicted from current models. MutLγ bound DNA cooperatively, could bind multiple substrates simultaneously, and formed higher-order complexes. FeBABE hydroxyl radical footprinting indicated that the DNA-binding interfaces of MutLγ for ssDNA and HJ substrates only partially overlap. Most contacts with HJ substrates were located in the linker regions of MutLγ, whereas ssDNA contacts mapped within linker regions as well as the N-terminal ATPase domains. Using yeast genetic assays for mismatch repair and meiotic recombination, we found that mutations within different DNA-binding surfaces exert separable effects in vivo. For example, mutations within the Mlh1 linker conferred little or no meiotic phenotype but led to mismatch repair deficiency. Interestingly, mutations in the N-terminal domain of Mlh1 caused a stronger meiotic defect than mlh1Δ, suggesting that the mutant proteins retain an activity that interferes with alternative recombination pathways. Furthermore, mlh3Δ caused more chromosome missegregation than mlh1Δ, whereas mlh1Δ but not mlh3Δ partially alleviated meiotic defects of msh5Δ mutants. These findings illustrate functional differences between Mlh1 and Mlh3 during meiosis and suggest that their absence impinges on chromosome segregation not only via reduced formation of crossovers. Taken together, our results offer insights into the structure-function relationships of the MutLγ complex and reveal unanticipated genetic relationships between components of the meiotic recombination machinery. During meiosis, cells undergo DNA recombination to form crossovers between homologous pairs of chromosomes (homologs). Crossovers promote accurate segregation of homologs at the first meiotic division and increase genetic diversity by breaking up linkage groups [1]. Recombination is initiated by DNA double-strand breaks made by Spo11 [2–4], which remains covalently attached to the DNA and is released by endonucleolytic cleavage [5]. Double-strand breaks are then resected to form 3′ single-stranded tails, which serve as a substrate for strand exchange proteins to invade a homologous template [6,7]. Subsets of these initial invasions further mature, after DNA synthesis and capture of the second end, into double Holliday junction (dHJ) intermediates, which are finally resolved into crossovers [1,8, 9]. Because crossovers are crucial to meiosis, the cell tightly controls their number and distribution [10–13]. MutLγ is important for crossover formation in many organisms, including yeast and mammals [14–18]. MutLγ is believed to catalyze the nuclease reaction that resolves the dHJ intermediate into a crossover [17,19]. Other proteins implicated in regulated crossing over include the ZMMs (Zip1-Zip2-Zip3-Zip4-Spo16, Msh4-Msh5, Mer3), a biochemically and functionally diverse group of proteins that channel recombination intermediates toward a crossover fate [20,21]. In addition to the major MutLγ- and ZMM-dependent pathway, another crossover pathway in S. cerevisiae depends on the structure-specific nuclease Mus81-Mms4 [22,23]. Mus81-Mms4 is thought to be responsible for ~15% of crossovers in wild-type yeast but can partially substitute when MutLγ is compromised [17,22]. Several additional systems can also take apart dHJ intermediates. The structure-specific nucleases Yen1 and Slx1/Slx4 are largely cryptic in wild-type cells and presumably contribute primarily as failsafe mechanisms that can scavenge recombination intermediates that escape the normal resolution pathways [17,24–27]. In addition, another process referred to as dHJ dissolution uses a single-strand decatenase formed by the combined activity of Sgs1 helicase plus Top3 topoisomerase in complex with Rmi1 protein [28–30]. However, the fraction of dHJ intermediates that is acted upon by this system in normal meiosis is currently unclear. Mlh1 and Mlh3 are also involved in post-replication mismatch repair (MMR) [31]. Mlh1 and Pms1 form the central MLH complex in yeast (MutLα) that is targeted to DNA mismatches by an MSH complex (Msh2-Msh6 or Msh2-Msh3) and that introduces DNA nicks to initiate degradation and repair of a mismatch-containing strand [32,33]. Mlh3 also participates in MMR along with Mlh1, but in a minor role [34,35]. MutLα also functions in repair of mismatches formed within the heteroduplex DNA intermediates of meiotic recombination [14,36]. Furthermore, Mlh1 along with Mlh2 forms a third heterodimeric complex, MutLβ, which has as yet poorly understood functions in controlling meiotic gene conversion patterns [36,37]. Importantly, however, MutLγ is the only MLH complex critical for meiotic crossing over per se, as MutLα and MutLβ are fully dispensable for formation of normal crossover numbers [14,36,37]. The ZMM proteins Msh4 and Msh5 (MutSγ) are also related to MMR factors but play no role in MMR [38,39]. By analogy with MMR, it has been proposed that MutSγ binds to recombination intermediates and recruits MutLγ to catalyze crossover formation [17]. Experimental evidence consistent with this idea includes cytological studies in mice and other organisms that show that the appearance of MutSγ precedes the appearance of MutLγ foci and that the number, timing and distribution of MutLγ foci correlate with chiasmata, indicating that MutLγ marks crossover sites [40–44]. In vitro, human MutSγ binds to Holliday junctions (HJ) and related branched structures [45]. However, MutLγ alone also binds specifically to HJs, independently of MutSγ [46]. In addition, nuclease activity of MutLγ has been demonstrated using plasmid DNA substrates, but HJ resolution activity has not yet been reconstituted [46,47]. Thus, key steps in meiotic crossing over remain poorly understood. In addition to DNA-binding and cleavage activities, MutLγ possesses ATP-binding and hydrolysis activities that appear to be essential for its MMR and meiotic functions, although controversy remains as to whether ATP hydrolysis is required in meiosis [18,47–50]. As in other MLH proteins, MutLγ ATPase activity is carried within an N-terminal GHKL domain, which is connected to the C-terminal domain by a flexible linker [51]. The nuclease active site is in the C-terminal domain of Mlh3, which dimerizes with the Mlh1 C-terminal domain [19,46,47]. In bacterial MutL and eukaryotic MutLα, cycles of ATP binding, hydrolysis, and nucleotide release modulate the conformational state of the complexes through dimerization of the N-terminal domains. These structural changes are proposed to act as a molecular switch to transduce signals between mismatch recognition factors and repair [52–54]. Nevertheless, how the DNA-binding, nuclease, and ATPase activities of eukaryotic MLH complexes relate to one another and to their MMR and/or crossover-promoting properties are unclear. We set out in this study to address this lack by combining genetic approaches with detailed biochemical characterization of the DNA-binding properties of S. cerevisiae MutLγ. To study the biochemical properties of S. cerevisiae MutLγ, we purified N-terminally-tagged Mlh1-Mlh3 heterodimers from baculovirus-infected insect cells (Materials and methods) (Fig 1A). We verified that the tagged proteins are functional in yeast, using strains that express identically tagged versions of Mlh1 (HisFlagmlh1) and Mlh3 (HisFlagmlh3) from their endogenous loci. Using genetic assays (described below), we found that the HisFlagmlh1 strain displayed a possible mild increase in chromosome missegregation that was not statistically significant (0. 6%, p = 0. 057, Fisher’s exact test (two tailed P value) ) and retained wild-type levels of crossing over and MMR (Fig 1B and 1C). The HisFlagmlh3 strain had a crossover defect (80% of wild-type levels, p < 0. 05, G test) but essentially wild-type levels of chromosome missegregation (0. 3%, p = 0. 29 Fisher’s exact test) and wild-type levels of MMR (Fig 1B and 1C). The biochemical experiments presented below used MutLγ complexes tagged on Mlh1, unless stated otherwise. Purified MutLγ displayed the expected nuclease activity on a supercoiled plasmid substrate. Under these conditions, wild-type MutLγ converted ~15% of the supercoiled substrate to nicked product, but nicking was undetectable when Mlh3 carried the nuclease domain mutation D523N (MutLγ-nd) (Fig 1D). This is in agreement with published results [46,47]. To assay for ATPase activity, MutLγ was incubated with [α-32P]-ATP, then ATP-hydrolysis products were separated by thin layer chromatography. MutLγ exhibited a low ATPase activity (kcat = 0. 1 min-1; Fig 1E), similar to other MLH complexes [47,52,55]. ATPase activity was not significantly stimulated by DNA (Fig 1F), as reported previously [47]. To address whether MutLγ undergoes ATP-driven conformational changes, we performed partial trypsin digestions of MutLγ in the presence and absence of ATP (Fig 1G). Gel electrophoresis of proteolysis reactions revealed that the presence of ATP results in specific changes in the pattern of trypsin cleavage fragments (see asterisks in Fig 1G). We imaged the protein particles by atomic force microscopy (AFM) to gain insights into the molecular organization of the complex (Fig 1H). A volume analysis of the particles revealed that Mlh1 and Mlh3 exist as an equilibrium between monomers and dimers, with about one third of dimers at the concentration of this experiment (10 nM) (Fig 1J). Dimers exhibited different configurations that could be classified as extended, one-arm folded, semi-condensed, and condensed (Fig 1I), as previously reported for yeast and human MutLα [54]. In the presence of 1 mM ATP, the population of semi-condensed and condensed particles increased at the expense of extended and one-arm folded particles (Fig 1K). This is consistent with the idea of a molecular switch modulated by ATPase cycles [52,54]. We assembled protein-DNA complexes of MutLγ using DNA substrates immobilized on streptavidin-coated beads (Fig 2A). MutLγ was efficiently pulled-down on beads coated with ssDNA (80 nt poly-dT) or HJs (40-bp arms), but not double-stranded DNA (dsDNA, 80 bp), as revealed by SDS-PAGE of the bound fractions detected by silver staining and anti-Flag western blotting (Fig 2A). To measure the affinity of MutLγ for ssDNA and HJ substrates, we titrated MutLγ in the presence of 1 nM 5′ end-labeled ssDNA or HJ substrates and separated the protein-DNA complexes by gel electrophoresis (Fig 2B). Substrate binding reached over 90% completion at ~100–200 nM MutLγ. The affinity of MutLγ for the poly-dT ssDNA substrate was higher than for the HJ substrate, with a KD of 46 nM and 73 nM, respectively. The titrations fit Hill equations with slopes of 2. 9 and 2. 0 for ssDNA and HJ substrate, respectively, indicating that MutLγ binds cooperatively to both DNA substrates (Fig 2B, right). In addition, most of the nucleoprotein complexes remained stuck in the wells even when unbound substrate remained, consistent with cooperative formation of higher-order multimers. Poorly resolved complexes that migrate within the gel could be observed under some sub-saturating conditions (e. g. , see lanes with 100 nM MutLγ). This may reflect a MutLγ dimer bound to the substrate or a multimer of lower order than the well-shifted complexes. To address the specificity of MutLγ for the ssDNA and HJ substrates, we performed competition experiments in an electrophoretic-mobility shift assay (EMSA) (Fig 2C). Nucleoprotein complexes were assembled with 100 nM MutLγ, 0. 5 nM 5′ end-labeled HJ or poly-dT ssDNA substrate, and 0–200-fold molar excess of cold competitor. In the absence of competitor DNA, MutLγ bound most of the DNA substrates and generated well-shift complexes (Fig 2C, lanes 2 and 13). With a 5′ end-labeled HJ substrate, binding was efficiently competed with a cold HJ competitor (lanes 9–11). In contrast, dsDNA and poly-dT ssDNA competed less efficiently (lanes 3–8), indicating that the binding activity is specific to the HJ substrate. Conversely, with a 5′ end-labeled poly-dT ssDNA substrate, binding was efficiently competed with cold ssDNA (lanes 14–16), but dsDNA and HJ competed less efficiently (lanes 17–22), indicating that binding to the ssDNA substrate is also specific. This does not contradict earlier reports that observed MutLγ binding on dsDNA [46,47]. It indicates instead that, compared to dsDNA, MutLγ binds with higher affinity to both HJ and poly-dT ssDNA substrates and that these two DNA-binding specificities are biochemically separable. Although the presence of poly-dT ssDNA or dsDNA competitors did not efficiently disrupt the binding of MutLγ to the HJ substrate per se, both competitors qualitatively changed the EMSA pattern by allowing complexes to migrate within the gel (Fig 2C, lanes 3–8). Thus, presence of these competitors appears to preferentially disrupt protein-DNA (and/or protein-protein) interactions needed to form higher order (well-shifted) complexes, without titrating higher affinity interactions of MutLγ with the HJ. This effect was not seen when the labeled substrate was ssDNA (Fig 2C, lanes 17–22), which suggests that the presence of competitor dsDNA or HJ cannot occlude the ssDNA-binding sites responsible for the formation of the higher order complexes. To further investigate these two distinct DNA-binding activities, we depleted the protein sample of the ssDNA-binding activity by preincubating MutLγ with poly-dT ssDNA-coated streptavidin beads and tested the supernatant for DNA binding (Fig 2D). We found that the HJ-binding activity of MutLγ had also been depleted (compare lanes 6 and 9), which suggests that both activities are present in the same pool of MutLγ molecules rather than representing distinct subpopulations of the protein. If so, a further implication is that MutLγ can bind multiple substrates simultaneously. To address this, we incubated 50 nM MutLγ in the presence of 25 nM of a radiolabeled substrate and 25 nM of a biotin-labeled substrate. After complex assembly, biotin-labeled DNA was pulled down with streptavidin-coated beads. The beads were washed extensively, then radiolabeled DNA was deproteinized, separated by gel electrophoresis, and detected by autoradiography (Fig 2E, left). In the presence of MutLγ, radiolabeled substrates were pulled down with streptavidin beads, indicating that biotin-labeled and radio-labeled substrates were bound simultaneously by MutLγ (lanes 3,5, 8 and 10). Pulldown of radiolabeled substrate was dependent on the presence of the biotin label, thus cannot be ascribed to nonspecific interactions between MutLγ and the streptavidin beads (Fig 2E, right, compare lanes 13 and 18 to lanes 15 and 20). These findings are consistent with individual MutLγ heterodimers binding more than one substrate simultaneously, or binding of multiple substrates through multiple proteins bound to one another. To further investigate formation of higher-order nucleoprotein complexes, we used AFM to image MutLγ particles bound to large ssDNA or plasmid substrates (Fig 2F). Reactions were assembled with 2 nM DNA and 10–40 nM MutLγ, then plated on mica slides and dried. Large protein-DNA structures were visible with both substrates (compare scales in Fig 2F with protein alone images in Fig 1H). These higher order structures appeared similar to those previously reported with MutLα [56]. A preference of MutLα for large DNA substrates has been observed and suggested to take part in the cooperative assembly of higher-order nucleoprotein complexes [56]. To test whether this was also the case for MutLγ, we performed substrate competition reactions where binding of MutLγ to an 80 nt labeled poly-dT ssDNA substrate was competed with 500-fold excess (in nucleotides) of cold ssDNA substrates ranging from 10 to 80 nt (Fig 2G). No competition was observed with substrates smaller than 30–40 nt. Competition increased with larger substrate size and reached ~50% with 60–70 nt substrates. A previous study also reported binding of MutLγ to HJs but did not detect binding to ssDNA [46]. Two factors are responsible for this apparent discrepancy: the size and the sequence of the ssDNA substrate used. When 80 nt or 50 nt substrates that were either poly-dT or another (non-poly-dT) sequence were compared as competitors in an EMSA with radiolabeled ssDNA substrate, the best competitor was the longest poly-dT substrate, whereas the shorter non-poly-dT sequence was almost completely ineffective under these conditions (Fig 2H). It is possible that, in addition to an affinity for larger ssDNA substrates, MutLγ prefers ssDNA that is less able to form secondary structures. We mapped the ssDNA and HJ-binding interfaces of MutLγ by FeBABE footprinting (Fig 3). The technique takes advantage of an Fe3+ ion chelated by an EDTA moiety that can be chemically conjugated to a sulfhydryl group present on phosphorothioate-modified DNA substrates [57]. Upon activation with hydrogen peroxide, FeBABE generates hydroxyl radicals that cleave peptide or DNA chains within ~15–20 Å of the FeBABE binding site. Using terminally tagged proteins, the sizes of the protein fragments can be estimated by western blotting and provide an estimate of positions in the protein that were in proximity to the DNA. The FeBABE positions along the ssDNA and HJ substrates are illustrated in Fig 3A and 3B. We assembled nucleoprotein complexes on immobilized FeBABE-conjugated DNA substrates using MutLγ complexes that carried an N-terminal tag either on Mlh1 (HisFlagMlh1-Mlh3, Fig 3C) or Mlh3 (Mlh1-HisFlagMlh3, Fig 3D). After the hydroxyl radical cleavage reaction, the samples were separated by SDS-PAGE and the cleavage fragments were detected by anti-Flag western blotting. The cleavage fragments were dependent on the presence of the FeBABE modification and activation with H2O2 (e. g. , compare lane 2 with lanes 1,3 and 4 in Fig 3C). The cleavage pattern was also specific to which protein was tagged and to the DNA substrate used (compare Fig 3C and 3D, left and right panels). We mapped positions of the cleavages using molecular weight markers (see Materials and methods). With the ssDNA substrate, the predicted cleavage positions map within the linker and the N-terminal domains of both Mlh1 and Mlh3 (Fig 3C, 3D and 3E). With the HJ substrate, the pattern was similar but the fragments corresponding to cleavage in the linkers were enriched relative to cleavages mapping to the N-terminus, as compared to the ssDNA substrate (Fig 3F). These findings suggest that the specificity for the HJ substrate relies more on the linker regions of Mlh1 and Mlh3 than does the specificity for the ssDNA substrate (Fig 3E). We further asked whether the pattern of the hydroxyl radical cleavage fragments was affected by the position of the FeBABE moieties along the HJ substrate (Fig 3G and 3H). With Mlh1, the intensity of the cleavage fragments decreased as the FeBABE probes were moved more than ~25 bp away from the center of the HJ substrate, indicating that the main protein-DNA contacts are located within approximately two helical turns from the center of the HJ (Fig 3B and 3G). However, the sizes of the cleavage fragments changed little if at all as the FeBABE positions of the probes were shifted. In the case of Mlh3, a minor change in the cleavage pattern was observed when the probes were moved (Fig 3H). However, this effect was difficult to reproduce and was not investigated further. Globally, the cleavage patterns remained relatively constant, which may suggest that the structure and disposition of the Mlh1 and Mlh3 linker regions around the HJ core is not well defined and that the cleavage pattern results from a heterogenous mixture of protein/DNA complexes. Alternatively, it may reflect the limit of resolution of the footprinting strategy, e. g. , because of the effective radius of hydroxyl radical damage and/or locations of protein segments that are particularly susceptible to cleavage. We next sought to identify amino acid residues in the vicinity of the hydroxyl radical cleavage sites that might be directly involved in protein-DNA contacts. An alignment of fungal Mlh1 and Mlh3 proteins revealed conserved basic residues near each predicted cleavage site (S1 Fig and Table 1). When mapped to a homology-based model of the N- and C-terminal domains of the Mlh1-Mlh3 heterodimer, these candidate residues clustered to positively charged surfaces within the N-terminal domain, suggesting that they may indeed be part of a DNA-binding surface (compare position of the amino acids highlighted in yellow in Fig 3I to positive (blue) surfaces in Fig 3J). Predicted DNA-binding residues located within the linker regions of Mlh1 and Mlh3 could not be modeled because no structural information is available, but sequence alignments revealed that mapped cleavage sites are located within the most conserved regions of the linkers (S1 Fig). For each subunit, we mutated eight sets of amino acids that corresponded to eight FeBABE cleavage sites mapped either in the N-terminal domain (N-terminal mutants) or in the flexible linker region (linker mutants) and tested the effects of these mutations on meiotic recombination and MMR in vivo (Table 1). To address the roles of the predicted DNA-binding residues of Mlh1 and Mlh3 in meiotic recombination we introduced targeted mutations by gene replacement in yeast cells that harbor spore-autonomous fluorescent reporters. The mutant proteins were expressed as untagged versions from their endogenous promoters. The diploid cells had transgenes encoding red and green fluorescent proteins located next to the centromere and ARG4 loci, respectively, of one chromosome VIII homolog, and blue fluorescent protein next to the THR1 locus of the other homolog (Fig 4A). The transgenes are under the control of the YKL050c promoter, which is expressed late in sporulation, after meiosis, and therefore allows fluorescent proteins to be detected only in spores that inherit a copy of the reporter gene [58]. Upon sporulation, the frequency of crossing over in the two test intervals (CEN8-ARG4 and ARG4-THR1) and the frequency of missegregation of chromosome VIII (MI nondisjunction) can be scored by the diagnostic fluorescence patterns of tetrads (Fig 4A). Crossover patterns can be scored unambiguously. However, the MI nondisjunction pattern is ambiguous because it can also arise from a four-chromatid double crossover in the ARG4-THR1 interval. In the wild type, this segregation pattern is seen at a frequency of about ~0. 1% (1 out of 1268 tetrads). Double crossovers in the CEN8-ARG4 interval, which can be scored unambiguously, were very rare in all strains (<0. 05 to 0. 2%). Because the genetic interval of ARG4-THR1 (4. 2 ± 0. 39 cM) is smaller than CEN8-ARG4 (11. 63 ± 0. 63 cM), double crossovers in this interval should be even less frequent. Using a separate assay that does not suffer from this ambiguity, Thacker et al. found that the MI nondisjunction frequency of chromosome VIII in wild type is ~0. 1% [58]. Since this is the frequency at which we detected the ambiguous pattern in wild type, we treated all of these events as MI nondisjunction. We may have therefore slightly underestimated the crossover frequency in the ARG4-THR1 interval in wild type, but this is unlikely to be a significant bias in crossover-deficient strains for which MI nondisjunction dominates. We scored over 600 tetrads per strain (Fig 4 and see S1 Table for genetic distances and statistical analyses). The crossover frequency in mlh1Δ and mlh3Δ strains was reduced to about 60–70% of the wild-type values in the two intervals, and the residual crossover levels were not significantly different between these strains (p > 0. 9, G test), in agreement with prior studies [14,36,50] (Fig 4B and 4C). MI nondisjunction frequency increased from ~0. 1% in wild type to 1. 15% ± 0. 54% in mlh1Δ and 2. 73% ± 1. 02% in mlh3Δ (mean ± 95% CI) (Fig 4D and 4E). This 2. 4-fold difference was statistically significant (p = 0. 005, Fisher’s exact test) and to our knowledge has not been reported previously. Most of the mlh1 mutations affecting the N-terminal domain resulted in a significant decrease in crossing over and an increase in MI nondisjunction (green bars, Fig 4B and 4D). Interestingly, at least three of the mlh1 N-terminal domain mutants (mlh1-K253E/K254E, mlh1-R273E/R274E and mlh1-K286E/R289E) showed a stronger phenotype than the mlh1Δ strain, particularly for MI nondisjunction (green bars, Fig 4B and 4D). In contrast, the mlh1 linker mutants (mlh1-R367E/R369E/K370E/R373E, mlh1-K393E/R394E and mlh1-K398E/R401E) conferred little or no meiotic defect (orange bars, Fig 4B and 4D). The mlh3 mutants were more variable: some had weak defects and there was no clear distinction between linker mutants and N-terminal mutants, unlike for mlh1. For example, in contrast to mlh1, three of the mlh3 linker mutants showed a significant meiotic defect (yellow bars, Fig 4C and 4E) and none of the mlh3 mutants had a significantly stronger phenotype than the mlh3Δ strain. When MI nondisjunction frequency was plotted as a function of crossover level in either test interval, MI nondisjunction was increased only in those mutants where the crossover frequency had dropped below ~70–80% of wild-type levels (Fig 4F and 4G). This agrees with a previous observation that spore viability only starts to decrease below a certain threshold of crossovers [59]. We also found that mlh3 linker mutants exhibited a proportionally stronger defect in crossing over than in chromosome segregation compared to the mlh3Δ strain (compare yellow points to the red point in Fig 4G). We used a Lys+ reversion assay to quantify the effects of the predicted DNA-binding mutations in MMR (Fig 5 and S2 Table). The assay takes advantage of a mutant lys2 gene containing a stretch of 14 A residues that creates a hotspot for DNA polymerase slippage [60]. The allele has a +1 bp frameshift, so changes of -1 bp or +2 bp can give Lys+ revertants. MMR-deficient strains such as mlh1Δ frequently fail to repair mutations introduced by DNA polymerase, causing elevated Lys+ reversion frequencies. With this sensitive assay, the Lys+ reversion frequency of an mlh1Δ strain was three orders of magnitude higher than wild-type levels (Fig 5A and [61]). Out of the eight mlh1 point mutants we examined, two were nearly indistinguishable from a wild-type stain (mlh1-R214E and mlh1-R367E/R369E/K370E/R373E), whereas the other six mutants reached Lys+ reversion levels close to mlh1Δ. Because Mlh3 plays only a minor role in MMR, the Lys+ reversion frequency of an mlh3Δ strain was only ~5-fold greater than in wild type (Fig 5B and [34,50]). All eight mlh3 mutants conferred a significant MMR defect, with the N-terminal mutants almost indistinguishable from mlh3Δ (Fig 5B and S2 Table). When the MMR defects were plotted versus crossover defects or chromosome missegregation frequencies, two patterns emerged. First, the mutants clustered by class on the basis of mutation location in either the N-terminal domain or in the linker region, indicating that different contributions in MMR or meiotic recombination can be attributed to specific actions of particular domains. Second, several mutants substantially deviated from the diagonals formed between wild type and the deletion strains; these are therefore separation-of-function mutants (Fig 5C and 5D). Specifically, mlh1 linker mutants (particularly mlh1-K398E/R401E) were above the diagonal, consistent with the possibility that DNA binding by the Mlh1 linker may be more important for MMR than meiotic recombination (orange dots, Fig 5C left and right panels). The mlh1 N-terminal mutants clustered right of the diagonal, reflecting their stronger defect in meiotic recombination than mlh1Δ (green dots, Fig 5C left and right panels). The mlh3 mutants also tended to cluster according to mutation position (Fig 5D). When the MMR defect was plotted against crossovers, three of the mlh3 linker mutants (yellow points; mlh3-R401E/K406E/R407E, mlh3-K414E/K416E and mlh3-K443E/K445E/R448E) clustered below the diagonal, whereas the N-terminal mutants (blue points) fell close to the diagonal. When MMR defect was plotted against MI nondisjunction, however, the linker mutants fell on the diagonal while N-terminal mutants lay above the diagonal. Thus, the mlh3 linker mutants confer a disproportionately stronger defect in crossing over than in either meiotic chromosome segregation or MMR. The 14-A insertion Lys+ reversion assay was chosen because it allowed us to score mutations in either mlh1 or mlh3, but it also yields high basal rates in wild type [50,60]. Therefore, we also measured the MMR deficiency of the mlh1 mutants in an independent Thr+ reversion assay that uses a +1 T insertion in a stretch of 6 T’s in the HOM3 gene (hom3-10) [62]. mlh3 mutants were not scored because the Thr+ reversion rate of mlh3Δ was not detectably higher than background. Levels of Thr+ reversion in the mlh1 mutants, when normalized to mlh1Δ, were similar to Lys+ reversion (Fig 5E, compare with Fig 5A). A comparison of MMR defects in the Thr+ reversion assay with meiotic recombination defects revealed similar clustering of the mlh1 mutant classes as with the Lys+ reversion assay, thus our conclusions are robust to differences between these experimental systems (Fig 5F). We went on to characterize a selection of MutLγ mutants in vitro. For both Mlh1 and Mlh3, we chose one N-terminal mutant (mlh1-K286E/R289E and mlh3-K316E/K320E/R323E) and one linker mutant (mlh1-K393E/R394E and mlh3-K414E/K416E) and purified single and double mutant complexes (Fig 6A). By EMSA with ssDNA or HJ substrates, all of the mutants displayed binding defects although all retained some DNA-binding activity (Fig 6B–6E). In addition, the double mutants were significantly more affected than either single mutant. The degree to which the ssDNA or HJ binding activities of each mutant were compromised varied. We asked whether these could be predicted from the intensity of the corresponding FeBABE-induced cleavage signal, but this was not straightforward. For example, the Mlh1 linker mutant had a stronger defect in HJ binding than the N-terminal mutant (Fig 6E). This is consistent with the hydroxyl radical footprinting result because the cleavage fragments with the HJ substrate mapped to the linker and not to the N-terminal domain (see Fig 3F). In the case of Mlh3, the difference in HJ binding defect between the N-terminal and linker mutants was not obvious (Fig 6E). This is explained by the fact that, in contrast to the mlh1 N-terminal mutant, the HJ substrate had significant hydroxyl radical cleavage signal that mapped to the region of the mlh3 N-terminal domain that we mutated (mlh3-K316E/K320E/R323E) (see Fig 3F). Hence, a difference in HJ binding affinity between the purified Mlh3 linker and N-terminal mutant complexes was not necessarily expected. In the case of the ssDNA-binding activity, the mapped DNA-binding interface is extensive (Fig 3) and no clear predictions can be made as to the effect of a specific DNA-binding mutation. Indeed, the FeBABE footprinting assay establishes proximity, not direct binding, so there is no a priori way of predicting which residues are the most critical for a specific binding activity based on hydroxyl radical cleavage intensity alone. Taken together, the DNA-binding defects of the MutLγ mutants show that the linker and N-terminal domains exert different contributions to the specificity of the DNA substrate. The binding defects globally agreed with the predictions of the hydroxyl radical footprinting assay, as far as it was possible to make clear predictions. To better understand the function of MutLγ in meiotic recombination and the effects of the DNA-binding mutants, we tested the effect of the mlh1 and mlh3 mutations in the context of mms4Δ or msh5Δ backgrounds (Fig 7 and S3 Table). Mms4 is a component, together with Mus81, of a structure-selective nuclease; it contributes to a quantitatively minor crossover pathway in wild type but is believed to resolve the excess of joint molecules that accumulate when the MutLγ pathway is compromised [1,17,22,27]. Msh5 is part of the MutSγ complex, which channels recombination intermediates toward the MutLγ-dependent crossover pathway [1,12,18]. MutLγ recognition of HJs seems easy to rationalize given that the enzyme is proposed to resolve dHJs into crossovers [1,18]. However, binding to ssDNA is more surprising because canonical models of meiotic recombination do not predict that ssDNA will be present at the stage of the dHJ intermediate. The primary ssDNA substrate used here was an 80 nt poly-dT sequence, to which MutLγ binds with particularly high affinity. However, we see clear evidence of binding to non-poly-dT ssDNA substrates if they are long enough (see the competition experiment with 80 nt competitor in Fig 2H, and the AFM results with >10,000 nt ssDNA substrates in Fig 2F). It remains unknown why poly-dT is an especially good binding substrate, but taken together, our data clearly point to a generalized ssDNA-binding activity as an intrinsic feature of MutLγ. In vivo, ssDNA binding would presumably require MutLγ to compete against ssDNA-binding proteins, like RPA. Whether MutLγ would be able to overcome this inhibition is currently unknown. Nevertheless, assuming that the ssDNA-binding activity observed is physiologically relevant, we suggest that MutLγ may play a role earlier in recombination than typically envisioned, for example, by binding to ssDNA exposed at nascent strand-exchange events. Another surprising observation was that the HJ and ssDNA-binding activities are separable—i. e. , largely non-competing—implying that the binding sites must be different and raising the possibility that the complex can bind two substrates simultaneously. FeBABE footprinting helped explain this observation by revealing that the DNA-binding interfaces for the two substrates only partially overlap. The interface for HJ binding appears to lie primarily within the linker regions of Mlh1 and Mlh3, while the ssDNA-binding interface appears more extensive and includes both the N-terminal domains and the linker regions. What might be the role of these overlapping binding sites? Perhaps affinity for ssDNA and simultaneous binding to different substrates assist in recruiting MutLγ to recombination sites. During meiotic recombination, there are thought to be two main stages where ssDNA is generated: during exonucleolytic resection of the double-strand breaks and during invasion of the homologous template, which exposes ssDNA from the template in the form of D-loops. We propose that MutLγ may be recruited to these early recombination intermediates, prior to formation of a dHJ. In this model, MutLγ may bind initially to ssDNA and then probe for nearby branched intermediates. Yeast MutLα binds DNA cooperatively and prefers long substrates, and AFM experiments have shown that MutLα forms long tracks along DNA and that several substrates can be bound within the same higher order complex [56]. We found that MutLγ also binds DNA cooperatively, forms higher order structures, has a preference for longer DNA, and can bind multiple substrates simultaneously. It is currently unclear which, if any, of these higher order complexes are physiologically relevant. Interestingly, however, MLH1 and MLH3 form bright immunostaining foci that mark presumed crossover sites in mammalian meiotic cells [40,41,68,69], suggesting that formation of higher order complexes is part of the crossover-promoting function of MutLγ. It is tempting to speculate that the oligomeric MutLγ-DNA complexes observed in vitro may be related to these foci. To better understand how DNA binding relates to MutLγ function in vivo, we mutated residues that appear to be in proximity to DNA, as judged from hydroxyl radical footprinting. We consider it likely that at least some of these residues lie within protein surfaces that contact DNA directly. In support of this inference, the subset of mutants tested in vitro had diminished DNA binding. Most mlh1 N-terminal mutants showed decreased crossing over and increased missegregation of chromosome VIII, and were also defective in MMR. The MMR defects were comparable to mlh1Δ, but the meiotic defects were unexpectedly more severe than in the null (discussed further below). In contrast, the mlh1 linker mutants showed little or no meiotic defect but were severely defective in MMR. For Mlh3, the N-terminal mutants had variable degrees of meiotic defects but were more uniformly defective in MMR. In contrast, most of the mlh3 linker mutants were defective in meiosis, but tended to show greater defects in crossing over than in chromosome segregation. Most mlh3 linker mutants also conferred a partial MMR defect, but interestingly, the allele that was normal for crossing over (mlh3-R419E/K426E) was not the same as the allele that was nearly normal in MMR (mlh3-K443E/K445E/R448E). Together, these findings indicate that these putative DNA-binding surfaces contribute to functions of both Mlh1 and Mlh3 in both MMR and meiotic recombination. The observed defects are consistent with DNA binding per se being important, but it is also possible that the mutations cause defects in vivo because of biochemical defects that are different from (or in addition to) DNA-binding alterations. For example, the mutations might impinge on protein-protein interactions with crossover- or MMR-promoting factors such as Exo1 [64]. Importantly, identification of separation-of-function mutations in both proteins indicates that the molecular roles of Mlh1 and Mlh3 can be distinguished in MMR vs. meiotic recombination. Specifically, the Mlh1 linker region appears to be more important in MMR, while the Mlh3 linker region appears to be more important for meiotic crossover formation. Several of the mutants were also analyzed in prior studies [48,62,70]. Mlh1-R273E/R274E had compromised DNA-binding activity for the Mlh1 N-terminal domain [70]. In addition, mlh1 linker mutations conferred MMR defects (K393A/R394A and R401A/D403A) and N-terminal domain mutations conferred MMR and meiotic recombination defects (K253A/K254A, R273A/R274A) [62]. The Alani group recently characterized a collection of mlh3 mutants and also identified alleles with separable functions in MMR and crossing over [71]. Several of their mutants were similar to the ones reported here, including for example a mlh3 linker mutant (mlh3-K414A/K416A, referred to as mlh3-32), which conferred a stronger defect in crossing over than in MMR. Our observations agree well with these independent studies. In vitro experiments with truncations of the Mlh1 and Pms1 linker domains previously implicated the linker regions of MutLα in DNA binding [72]. Here, we identified specific substitutions within the most conserved regions of the linkers of Mlh1 and Mlh3 that compromise the DNA-binding activity (residues ~393–418 in Mlh1 and ~401–448 in Mlh3). Unlike MutLγ, the Mlh1-containing MutLα and MutLβ heterodimeric complexes are dispensable for meiotic crossing over, although they do participate in mismatch correction and other aspects of gene conversion during recombination [1,18,36]. Moreover, crossing over plays a critical role in promoting chromosome segregation, but gene conversion does not. Therefore, when we began this work, we envisioned that absence of either Mlh1 or Mlh3 would cause indistinguishable defects in crossover formation and therefore in chromosome segregation. However, we found that the meiotic phenotypes of the mlh1Δ and mlh3Δ strains differed: the mutants had similarly decreased crossover levels, but MI nondisjunction of chromosome VIII was 2- to 3-fold more frequent in mlh3Δ. This difference was not observed in previous measures of spore viability, which have been reported to be reduced to similar extents in the two mutants (to ~70–80%, e. g. [14,19]). One possible reason for the apparent discrepancy with spore viability measurements could be chromosome-specific effects, since we assayed missegregation of only chromosome VIII, whereas spore viability reflects segregation of all chromosomes and is probably driven principally by behaviors of the smallest chromosomes. Another possibility is that the greater MMR defect in mlh1Δ results in a higher burden of haplolethal mutations accumulating during premeiotic growth, in turn yielding more inviable spores than accounted for by chromosome missegregation alone [14]. The spore-autonomous fluorescent reporter assay is independent of spore viability [58], so it should not suffer from this complication. We further found that mlh1 N-terminal mutants were worse than mlh1Δ for both crossing over and chromosome missegregation. We also identified instances where decreases in crossing over did not correlate closely with increases in MI nondisjunction. For example, mlh3 linker mutants (mlh3-R401E/K406E/R407E, mlh3-K414E/K416E and mlh3-K443E/K445E/R448E) had crossover defects as strong as in mlh3Δ but showed less MI nondisjunction than mlh3Δ. In summary, mlh3 linker mutants behaved similarly to mlh1Δ (strongly reduced crossovers but only moderate MI nondisjunction), whereas mlh1 N-terminal mutants were more similar to mlh3Δ (strongly reduced crossovers and high MI nondisjunction). The mlh1 and mlh3 N-terminal mutants were also similar to mlh3Δ in that they were epistatic to mms4Δ for MI nondisjunction but were non-epistatic with mms4Δ for crossing over. Taken together, these results further illuminate the ways in which Mlh1 and Mlh3 contribute differently to meiotic recombination. For the null mlh1Δ and mlh3Δ mutations, differences may reflect fundamental distinctions between the proteins in terms of the complexes they inhabit: Mlh1 is part of MutLα, MutLβ, and MutLγ, whereas Mlh3 is only known to be a component of MutLγ. Thus, the mlh1Δ phenotype may reflect absence of Mlh1 itself; absence of all three heterodimeric complexes (MutLα, MutLβ, and MutLγ); aberrant behaviors of Mlh2, Mlh3, and Pms1 in the absence of their binding partner; or some combination of these defects. In contrast, the mlh3Δ phenotype could arise from absence of a specific function of Mlh3 itself; absence of MutLγ; and/or aberrant behavior of MutLα or MutLβ caused by there no longer being competition between Mlh3 and other proteins for binding to Mlh1. Interestingly, MutLβ has recently been implicated in regulating meiotic gene conversion tract lengths [37], raising the possibility that differences between mlh1Δ and mlh3Δ could trace in part to differences in activity of MutLβ. Two main scenarios can be envisioned to explain why the mutants have different nondisjunction levels despite roughly similar crossover defects: MI chromosome segregation may be more accurate than expected given the crossover defect in mlh1Δ and the mlh3 linker mutants, or conversely, MI nondisjunction may be more frequent than should be expected from the crossover defect in mlh3Δ and the mlh1 N-terminal mutants. What could account for such behaviors? DNA joint molecules (dHJ and/or other recombination intermediates) persist longer than normal in the absence of the obligate MutLγ-accessory factor Exo1 [64]. Perhaps in mlh1Δ and the mlh3 linker mutants these joint molecules persist just long enough to facilitate chromosome biorientation at metaphase I, but then at anaphase I are resolved as a mix of crossovers and noncrossovers (by Yen1, Mus81/Mms4, or Slx1/Slx4) [17,25] and/or dissolved to form noncrossovers by Sgs1/Top3/Rmi1 [28–30]. In mlh3Δ and the mlh1 N-terminal mutants in contrast, the joint molecules might be taken apart too early to facilitate chromosome biorientation, or might persist too long and thereby interfere with chromosome separation at anaphase. The possibility that alternative chromosome segregation mechanisms are at play in some MutLγ mutants, not strictly dependent on crossovers, has been noted previously based on the relatively high spore viability of an mlh1Δ mms4Δ strain [50]. Furthermore, the possibility that MutLγ might interfere with alternative HJ resolution systems has also been suggested previously [17]. Epistasis experiments with mms4Δ provide possible support for a role of joint molecules and alternative resolution pathways in explaining differences between mlh1Δ and mlh3Δ mutants. In the absence of MutLγ, Mus81/Mms4 is thought to substitute in resolving most of the joint molecules [17,22,23]. We speculate that Mlh1 (in the absence of Mlh3) can partially inhibit action of Mus81/Mms4, whereas Mlh3 (in the absence of Mlh1) cannot. In this model, action of Mus81/Mms4 should be more readily detected in an mlh1Δ mutant (where Mus81/Mms4 is not inhibited) than in an mlh3Δ mutant. Our observation that mlh3Δ is epistatic to mms4Δ for MI nondisjunction is thus consistent with this hypothesis. We further note that the N-terminal mutants mlh1-K286E/R289E and mlh3-K316E/K320E/R323E mutants behaved similarly to mlh3Δ, suggesting that these mutants may also prevent the resolution of recombination intermediates by Mus81/Mms4. We further found that mlh1Δ (but not mlh3Δ) partially alleviated the phenotype of msh5Δ, in terms of both crossovers and MI nondisjunction. A similar effect was seen but not remarked on in a prior study [23]. This result can also be interpreted in terms of competition between MLH complexes and alternative resolution systems. Similar to the reasoning above, presence of MutLα and/or MutLβ might interfere with access of alternative crossover-promoting activities in the absence of Msh5. Absence of all MLH complexes in the mlh1Δ mutant is thus envisioned to allow Mus81/Mms4 and/or other activities to form more crossovers. The spore viability defect of msh5Δ can also be partially suppressed by mlh2Δ [37,73]. Since absence of MutLβ increases gene conversion tract length [37], an interesting possibility might be that longer heteroduplex DNA tracts in the absence of MutLβ could stabilize early recombination intermediates, which would normally have been stabilized instead by MutSγ. Taken together, these observations draw a more complex picture of the role of MutLγ in meiotic recombination, and highlight the dynamic interplay between enzymatic activities that can act on recombination intermediates. Oligonucleotides (oligos) used in this study were purchased from Integrated DNA Technologies. The sequence of the oligos is listed in S6 Table. The plasmids used in this study are listed in S5 Table. Vectors for expression of untagged S. cerevisiae Mlh1 and Mlh3 proteins were generated by cloning the corresponding gene in pFastBac1 (Invitrogen). Vectors for expression of N-terminally 6xHistidine and 2xFlag-tagged (HisFlag) Mlh1 and Mlh3 were generated by cloning the corresponding gene in pFastBac-HTbFlag, which has a sequence coding for two Flag epitopes inserted into the BamHI site of pFastBac-HTb (a gift from V. Bermudez, MSKCC). The MLH1 sequence was amplified from an S288c strain (SKY4158) using primers cb111 and cb112. The PCR product was digested with BamHI and SalI restriction endonucleases and cloned into the BamHI and SalI sites of pFastBac1 to create pCCB313, or pFastBac-HTbFlag to create pCCB317. Similarly, the MLH3 sequence was amplified from strain SKY4158 using primers cb113 and cb114. The PCR product was digested with BamHI and EcoRI and cloned into the BamHI and EcoRI sites of pFastBac1 to create pCCB318 or pFastBac-HTbFlag to create pCCB314. The cloned PCR products were verified by Sanger sequencing. The plasmid constructs coding for mutant proteins were generated by QuickChange mutagenesis using the wild-type plasmids as templates. The primers were the same ones used to create the corresponding mutant yeast strains as listed in section: Yeast strains and targeting vectors. The viruses were produced by a Bac-to-Bac Baculovirus Expression System (Invitrogen) following the manufacturer’s instructions. Expression of HisFlagMlh1-Mlh3 heterodimer used viruses produced from vectors pCCB313 and pCCB318. Mlh1-HisFlagMlh3 heterodimer used viruses produced from vectors pCCB314 and pCCB317. Spodoptera frugiperda Sf9 cells were co-infected with both viruses at a multiplicity of infection of 2. 5 each. Cells were harvested 62 hours after infection, washed with phosphate buffer saline (PBS), frozen in dry ice and kept at -80°C until use. A typical purification was performed with cell pellets from 500 ml culture. All the purification steps were carried out at 0–4°C. Cell pellets were resuspended in 4 volumes of lysis buffer (25 mM HEPES-NaOH pH 7. 5,500 mM NaCl, 0. 1 mM DTT, 25 mM imidazole, 1× Complete protease inhibitor tablet (Roche) and 0. 1 mM phenylmethanesulfonyl fluoride (PMSF) ). Cells were lysed by sonication and centrifuged at 43,000 g for 30 min. The cleared extract was re-sonicated and filtered through 0. 2 μm filters before being loaded on a pre-equilibrated 1 ml HisTrap column (GE Healthcare) using an AKTA purifier (GE Healthcare). The column was washed extensively with wash buffer (25 mM HEPES-NaOH pH 7. 5,500 mM NaCl, 10% glycerol, 0. 1 mM DTT, 25 mM imidazole, 0. 1 mM PMSF). The tagged MutLγ complexes were then eluted with a 25–500 mM gradient of imidazole. Fractions containing MutLγ were pooled and diluted in 3 volumes of Flag buffer (25 mM HEPES-NaOH pH 7. 5,500 mM NaCl, 10% glycerol, 1 mM DTT, 2 mM EDTA). Next, the complexes were bound to 1 ml pre-equilibrated Anti-Flag M2 affinity resin (Sigma) in a poly-prep chromatography column (Bio-Rad). The resin was washed extensively with Flag buffer and the bound proteins were eluted with Flag Buffer containing 250 μg/ml 3xFlag peptide (Sigma). Fractions containing MutLγ were pooled and 200–300 μl each were loaded on 5 ml 15–40% glycerol gradients in 25 mM HEPES-NaOH pH 7. 5,400 mM NaCl, 1 mM DTT, 2 mM EDTA, 0. 01% NP40. Gradients were centrifuged at 250,000 g for 20 hours and separated in about 30 fractions of 7 drops each (~175 μl). Fractions containing MutLγ were then dialyzed in 25 mM HEPES-NaOH pH 7. 5,300 mM NaCl, 10% glycerol, 1 mM DTT, 2 mM EDTA using 50 kDa cut-off Slide-a-lyzer cassettes (Thermo Scientific) and concentrated in 50 kDa cut-off Amicon centrifugal filters (Millipore). Aliquots were frozen in dry ice and stored at -80°C. Mutant MutLγ complexes were prepared using the same procedure as the wild-type proteins. DNA substrates were generated by annealing complementary oligos (sequences listed in S6 Table). Oligos over 40 nt were first purified on 10% polyacrylamide-UREA gels. They were subsequently mixed in equimolar concentrations (typically 10 μM) in STE (100 mM NaCl, 10 mM Tris-HCl pH 8,1 mM EDTA), heated and slowly cooled on a PCR thermocycler (98°C for 3 min, 75°C for 1 h, 65°C for 1 h, 37°C for 30 min, 25°C for 10 min). Substrates for electrophoretic mobility shift assays (EMSA) were assembled with the following primers: Holliday Junction (HJ80): cb095, cb096, cb097 and cb098; single-strand DNA (ssDNA80): cb099; double-strand DNA (dsDNA80): cb095 and cb100. Substrates were 5′ end-labeled with [γ-32P]-ATP (Perkin Elmer) and T4 polynucleotide kinase (New England Biolabs) and labeled substrates were purified by native polyacrylamide gel electrophoresis. Substrates for DNA pulldown assays used the same combination of oligos as EMSA substrates except that cb095 and cb099 were replaced by 3’-biotinylated versions, cb256 and cb316, respectively. After annealing and gel purification, the substrates were bound to M280 streptavidin-coated Dynabeads (Invitrogen) in TENT buffer (10 mM Tris-HCl pH 7. 5,1 mM EDTA, 1 M NaCl and 0. 1% Triton X-100) for 4 hours at 4°C. Beads were washed in TENT buffer and were stored in binding buffer (25 mM Tris-HCl pH 7. 5,10% glycerol, 100 mM NaCl, 200 μg/ml BSA, 5 mM EDTA, 2 mM DTT, 0. 1% Triton X-100). Final concentration was estimated at 500 pmol DNA/μl beads (final substrate concentration of 500 nM). Phosphorotioate-containing DNA substrates were identical to the DNA pulldown assays except that the appropriate oligos were replaced by modified versions that carried a phosphorothioate modification at the positions indicated with an asterisk in S6 Table. The substrate used in the FeBABE assays were (with oligos between brackets): ssDNA- (cb316); ssDNA+ (cb361); HJ- (cb256, cb96, cb97, cb98); HJ+ (cb256, cb351, cb97, cb98); HJ+1 (cb256, cb351, cb369, cb98); HJ+2 (cb256, cb349, cb367, cb98); HJ+3 (cb256, cb347, cb365, cb98); HJ+4 (cb256, cb345, cb363, cb98). Oligos were annealed in 40 μl STE at a concentration of 2. 5 μM biotinylated oligos and a 1. 2× excess of non-biotinylated oligos. Substrates were purified on a 5% Tris-acetate-EDTA-polyacrylamide gel, eluted in STE, ethanol precipitated, and resuspended in 20 mM MOPS pH 7. 9. FeBABE conjugation to phosphorothioate-containing DNA substrates were in 20 μl reactions in 20 mM MOPS pH 7. 9 containing 4 μM DNA and 3. 5 mM FeBABE (Dojindo). After 16 hours incubation at 50°C, the substrates were immobilized to 800 ng M280 streptavidin-coated Dynabeads (Invitrogen) in 400 μl 20 mM MOPS (pH 7. 9) for 4 hours at 4°C. Excess FeBABE was removed by washing three times with 500 μl 20 mM MOPS pH 7. 9 and one time with 200 μl storage buffer (25 mM Tris-HCl pH 7. 5,10% glycerol, 100 mM NaCl, 200 μg/ml BSA). Substrates were resuspended in storage buffer at an estimated concentration 1000 pmol/μl beads (final substrate concentration of about 1 μM) and stored at 4°C. Binding reactions (20 μl) were carried out in 25 mM Tris-HCl pH 7. 5,15% glycerol, 100 mM NaCl, 2 mM DTT, 5 mM EDTA and 1 mg/ml BSA with 1 nM DNA substrate and the indicated concentrations of MutLγ. Complexes were assembled for 30 minutes at 30°C and separated on a 5% Tris-acetate-EDTA-polyacrylamide/bis (80/1) gel. Gels were dried and radioactivity was detected by phosphorimaging (Fuji). Binding reactions (20 μl) were carried out in 25 mM Tris-HCl pH 7. 5,10% glycerol, 100 mM NaCl, 2 mM DTT, 5 mM EDTA, 200 μg/ml BSA, 0. 1% Triton X-100 with 25 nM immobilized DNA and the indicated concentrations of MutLγ. Complexes were assembled for 1 hour at 4°C on a rotating wheel. The beads were washed three times with 200 μl binding buffer without BSA, resuspended in 1× Laemmli sample buffer, boiled for 5 minutes and loaded on 4–12% Bis-Tris NuPAGE gels in MOPS running buffer. Proteins were detected by silver staining or anti-Flag western blotting. Binding reactions (20 μl) were carried out in 25 mM Tris-HCl pH 7. 5,10% glycerol, 100 mM NaCl, 2 mM DTT, 5 mM MgCl2,1 mg/ml BSA, with 25 nM biotin-labeled DNA, 25 nM 32P-labeled DNA and 50 nM MutLγ. Complexes were assembled for 30 minutes at 4°C. Reactions were then supplemented with 1 μl of M280 streptavidin-coated dynabeads (10 mg/ml) preequilibrated in binding buffer and 500 nM competitor dsDNA (80 bp substrate). Nucleoprotein complexes were pulled down for 1 hour at 4°C on a rotating wheel, washed three times with 200 μl of binding buffer with 0. 01% NP-40 without BSA. Beads were then resuspended in loading buffer containing 0. 5 mg/ml proteinase K. After 30 minutes at 30°C samples were separated on a 5% TAE-polyacrylamide gel, dried, and developed by autoradiography. Binding reactions (20 μl) were carried out in 25 mM Tris-HCl pH 7. 5,10% glycerol, 100 mM NaCl, 1 mg/ml BSA, 5 mM MgCl2 with 100 nM immobilized substrates and 100 nM MutLγ. Complexes were assembled for 10 minutes at room temperature, washed twice with 200 μl FeBABE buffer (25 mM Tris-HCl pH 7. 5,10% glycerol, 100 mM NaCl, 5 mM EDTA, 0. 01% NP-40) with 2 mM DTT and once with 200 μl buffer without DTT. Reactions were resuspended in 20 μl FeBABE buffer and separated in two. One half (10 μl) was treated with 1. 25 μl of 50 mM sodium ascorbate followed rapidly with 1. 25 μl of 50 mM H2O2,10mM EDTA and the cleavage reaction performed at 30°C for 5 minutes (for ssDNA substrates) or 10 minutes (for HJ substrates). The other half was left untreated as a negative control. Reactions were quenched with 6 μl 4× LDS sample buffer and 1 μl 1 M DTT, boiled for 5 minutes and loaded on 4–12% NuPAGE Bis-Tris gels in MOPS running buffer. After electrophoresis, proteins were transferred onto Immobilon-FL PVDF membranes (Millipore) and membranes were blotted with anti-Flag M2 antibody (Sigma) followed by a IRDye 680RD goat anti-mouse IgG (Li-COR). Western blots were revealed using the Li-COR Bioscience Odyssey infrared imaging system. The molecular sizes of Flag-tagged fragments of Mlh1 and Mlh3 generated by the hydroxyl radical cleavage assay were determined by comparing their migration with MagicMark XP Western protein standards (Invitrogen). Profiles of each lane were quantified using ImageGauge. The peaks were identified for each molecular weight standard and used to deduce a fourth order polynomial equation of the migration distance as a function of molecular weight. The equation was used to deduce the molecular weight of the protein fragments, which in turn provide an estimated position of the hydroxyl radical cleavage sites. Based on multiple experiments we estimate that the cleavage sites are mapped within 5–10 residues. For the Michaelis-Menten kinetics experiment, ATP-hydrolysis reactions were performed with 200 nM MutLγ complexes in 25 mM Tris-HCl pH 7. 5,5 mM MgCl2,1 mg/ml BSA and 33 nM [α-32P]-ATP and the indicated amounts of cold ATP. Reactions were incubated at 30°C for 30 minutes and stopped with a final concentration of 50 mM EDTA. ATP hydrolysis was measured by thin layer chromatography. Briefly, 1 μl of the reactions were spotted on a polyethyleneimine-cellulose TLC plate (EMD Chemicals) and developed in 0. 5 M LiCl, 1 M formic acid. Plates were exposed to a phosphorimager screen and [α-32P]-ADP levels quantified in ImageGauge. ATPase assays in the presence of DNA had 280 nM MutLγ with 2 μM ssDNA, 1 μM dsDNA and 0. 5 μM HJ substrates, which correspond to equal amounts (52 ng/μl) of total DNA. For nuclease assays, 20 μl reactions contained 100 nM MutLγ in the presence of 5. 7 nM supercoiled substrate (200 ng of pUC19 plasmid) in 25 mM Tris-HCl pH 7. 5,5 mM MnCl2,20 mM NaCl and 100 μg/μl BSA. After 1-hour incubation at 30°C, the reactions were stopped with 30 mM EDTA, 0. 1% SDS and 10% glycerol. Reactions were deproteinized by treatment with 0. 5 mg/ml of proteinase K for 15 min at 37°C and the DNA was separated on a 0. 8% TBE-agarose gel. The gel was stained with 0. 3 μg/ml ethidium bromide and imaged using a ChemiDoc MP imaging system (Bio-Rad). Trypsin digestions were performed in 30 μl reactions with 2 μM MutLγ in 25 mM HEPES-NaOH pH 7. 5,15 mM CaCl2,100 mM NaCl, 3% glycerol, with or without 5 mM ATP (sodium salt, pH 7) with 25 ng trypsin (Worthington). After 5 minutes digestion at room temperature, reactions were stopped in Laemmli sample buffer. Proteins were separated on 10% NuPAGE Bis: Tris gels in MOPS SDS running buffer and visualized by Coomassie staining. The structure of the C-terminal domain of Mlh1 is from PDB accession number 4e4w [74]. Homology-based models for the N-terminal domain of Mlh1 and the N and C-terminal domains Mlh3 were generated by Phyre2 [75]. The model for the MutLγ heterodimer was generated using Pymol. The PDB file is available upon request. AFM images were captured on an Asylum Research MFP-3D-BIO (Oxford Instruments) microscope in tapping mode at room temperature. For studies of protein alone, MutLγ was diluted to a final concentration of 10 nM in 25 mM HEPES-NaOH pH 7. 5,5 mM MgCl2,100 mM NaCl, 10% glycerol with or without 1 mM ATP at room temperature. A volume of 20 μl of the diluted protein was deposited onto freshly cleaved mica for 10 seconds. Samples were rinsed with 10 ml of ultrapure H2O and the surface was dried using a stream of nitrogen. A Bruker FMV-A AFM probe with resonance frequencies of approximately 75 kHz and spring constant of approximately 2. 5 N/m was used for imaging. Images were collected at a speed of 0. 8 Hz with an image size of 1 μm at 512 × 512 pixel resolution. Volume analyses of MutLγ particles were performed using a custom-written code in Metamorph. Particles with a volume between 170 and 350 nm3 were classified as dimers and scored as extended, one-arm folded, semi-condensed or condensed dimers based on their shape. For protein-DNA interaction studies, reactions contained 2 nM ssDNA (M13-based ~10,400 nt circle, a gift from A. Chatterjee, MSKCC) or plasmid DNA (pUC19) with 10 or 40 nM MutLγ, respectively, in 25 mM HEPES-NaOH pH 6. 8,5 mM MgCl2,50 mM NaCl and 10% glycerol. A volume of 40 μl of the reaction was deposited onto freshly cleaved mica (SPI) for 2 minutes. The sample was rinsed with 10 ml of ultrapure H2O and the surface was dried using a stream of nitrogen. An Olympus AC240TS-R3 AFM probe with resonance frequencies of approximately 70 kHz and spring constant of approximately 1. 7 N/m was used. Images were collected at a speed of 0. 6 Hz with an image size of 2 μm at 512 × 512 pixel resolution. Yeast strains were from the SK1 background. All the strains used in this study are listed in S4 Table. mlh1Δ and mlh3Δ strains were constructed by replacing the coding sequence by the kanMX4 cassette, which was amplified from plasmid pFA6a-kanMX4 using primers cb412 and cb413 for mlh1Δ and cb414 and cb415 for mlh3Δ. The vector used to generate the mlh1 and mlh3 DNA-binding mutant strains was constructed as follows. The hphMX4 cassette was PCR amplified from pSK742 using primers cb416 and cb417 for MLH1 and primers cb428 and cb429 for MLH3. The PCR products, which target MLH1 and MLH3 about 50 bp downstream of the corresponding gene, were transformed into fluorescent reporter strains SKY3576 and SKY3579 to achieve SKY5087 and SKY5088 for MLH1-hphMX4 and SKY5089 and SKY5090 for MLH3-hphMX4. The genomic regions of MLH1 and MLH3 together with the downstream hphMX4 cassettes were PCR amplified using primers cb424 and cb425 (for MLH1) and cb426 and cb427 (for MLH3) and cloned into a Topo vector using the zero Blunt Topo cloning kit (Invitrogen), creating plasmids pCCB419 (MLH1) and pCCB420 (MLH3). The vectors used to generate tagged MLH1 and MLH3 strains were constructed by inverse PCR followed by blunt-end ligation using template pCCB419 and primers cb696 and cb697 (for MLH1) and template pCCB420 and primers cb698 and cb699 (for MLH3), to create vectors pCCB573 and pCCB574, respectively. Mutant plasmids were generated by QuickChange mutagenesis. Plasmid number and mutagenesis primers for each mutant are as follows: mlh1-R214E (plasmid pCCB423, primers cb430 and cb431); mlh1-K253E/K254E (plasmid pCCB424, primers cb432 and cb433); mlh1-R273E/R274E (plasmid pCCB425, primers cb434 and cb435); mlh1-K286E/R289E (plasmid pCCB426, primers cb436 and cb437); mlh1-R341E/K344E (plasmid pCCB427, cb438 and cb439); mlh1-R367E/R369E/K370E/R373E (plasmid pCCB428, primers cb470 and cb471); mlh1-K393E/R394E (plasmid pCCB429, primers cb442 and cb443); mlh1-K398E/R401E (plasmid pCCB430, primers cb444 and cb445); mlh3-R171E/R172E/R173E (plasmid pCCB433, primers cb454 and cb455); mlh3-R220E/K222E (plasmid pCCB434, primers cb456 and cb457); mlh3-K316E/K320E/R323E (plasmid pCCB435, primers cb458 and cb459); mlh3-K347E/K351E (plasmid pCCB436, primers cb460 and cb461); mlh3-R401E/K406E/R407E (plasmid pCCB437, primers cb462 and cb463); mlh3-K414E/K416E (plasmid pCCB438, primers cb464 and cb465); mlh3-R419E/K426E (plasmid pCCB439, primers cb466 and cb467); mlh3-K443E/K445E/R448E (plasmid pCCB440, primers cb468 and cb469). All mutant constructs were verified by sequencing. Mutant and tagged strains were constructed by transformation of fluorescent reporter strains SKY3576 and SKY3579 or mismatch repair strain SKY5139 with a NheI and EcoRV (for MLH1) or PvuII and BglI (for MLH3) restriction fragment of the mutant or tagged plasmids. Selection for hygromycin resistant clones gave strains with the endogenous locus replaced with the mutant locus together with a downstream hphMX4 cassette. Integrations at the MLH1 locus were confirmed by Southern blotting of ApaI and SacI digested genomic DNA using a probe amplified with primers cb450 and cb451. Integrations at the MLH3 locus were confirmed by Southern blotting of SalI and BamHI digested genomic DNA using a probe amplified with primers cb452 and cb453. The presence of the mutations or the tags was also confirmed by restriction digestion or sequencing of a PCR product amplified with primers cb111 and cb112 for mlh1 mutants or cb113 and cb114 for mlh3 mutants. All yeast cultures and sporulation were carried out at 30°C. The spore-autonomous fluorescence assay was performed as described previously [58]. Briefly, wild-type or mutant haploid strains carrying the fluorescence reporter cassettes were mated and streaked on YPD (1% yeast extract, 2% peptone, 2% dextrose, 2% agar) to isolate diploid colonies. At least three independent diploids were grown in liquid YPD overnight, transferred into YPA (1% yeast extract, 2% peptone, 2% potassium acetate) for 13. 5–14 hours and sporulated in 2% potassium acetate for two days. Tetrads were then scored by fluorescence microscopy for crossovers in two test intervals and for MI-nondisjunction events. Genetic distances (cM) were calculated using the Perkins equation: cM = (100 (6NPD + TT) ) / (2 (PD + NPD + TT) ), where PD is the number of parental ditypes, NPD is the number of nonparental ditypes and TT is the number of tetratypes. Standard errors of genetic distances were calculated using the Stahl laboratory online tools (http: //molbio. uoregon. edu/~fstahl/). The Lys+ reversion assay used the lys2: : insE-A14 mutation (from strain EAY1062 from E. Alani) and the threonine reversion assay used the hom3-10 mutation (from strain HTY1213 provided by E. Alani) [60,62]. Mutant strains were streaked on YPD. For each mutant, at least three colonies were grown in YPD until saturation and dilutions were plated on lysine or threonine dropout medium, as appropriate, and on YPD to measure the frequencies of Lys+ or Thr+ revertants per colony-forming unit.

Sexual reproduction involves the fusion of two gametes that each contain half of the DNA from each parent. These gametes are generated through a specialized cellular division called meiosis. During meiosis, the cell faces the challenge of identifying the appropriate pairs of chromosomes that need to be separated. This involves an elaborate mechanism whereby the parental chromosomes recombine and form crossovers, i. e. exchange DNA fragments. These crossovers are thus important for the accurate segregation of chromosomes and are also fundamental to evolution because they help shuffle linkage groups from one generation to another. Here, we have studied a complex of proteins called MutLγ that is important for the formation of crossovers, and is also involved in an unrelated mechanism that repairs mistakes that spontaneously arise in DNA when it is synthesized. We uncovered intriguing features of the interaction of this complex with DNA. In addition, by studying a collection of mutants of MutLγ, we identified mutants that affect one biological function but not another. For example, surprisingly, we found mutations that decrease the frequency of crossovers but did not affect chromosome segregation as much as expected. Taken together, our findings allow us to reconsider the ways in which we think about these processes.

Abstract Introduction Results Discussion Materials and methods

2017

Distinct DNA-binding surfaces in the ATPase and linker domains of MutLγ determine its substrate specificities and exert separable functions in meiotic recombination and mismatch repair

Increasing evidence of a role of chronic inflammation in type 2 diabetes progression has led to the development of therapies targeting the immune system. We develop a model of interleukin-1β dynamics in order to explain principles of disease onset. The parameters in the model are derived from in vitro experiments and patient data. In the framework of this model, an IL-1β switch is sufficient and necessary to account for type 2 diabetes onset. The model suggests that treatments targeting glucose bear the potential of stopping progression from pre-diabetes to overt type 2 diabetes. However, once in overt type 2 diabetes, these treatments have to be complemented by adjuvant anti-inflammatory therapies in order to stop or decelerate disease progression. Moreover, the model suggests that while glucose-lowering therapy needs to be continued all the way, dose and duration of the anti-inflammatory therapy needs to be specifically controlled. The model proposes a framework for the discussion of clinical trial outcomes. Despite more than 350 million patients worldwide and the concomitant expensive socioeconomic burden, the pathogenesis of type 2 diabetes (T2D) is not yet completely understood. T2D is a progressive disease. The most important physiological components of T2D are insulin resistance, which is characterized by impaired response to insulin in insulin-sensitive tissues, and β-cell failure, which is characterized by β-cell dysfunction and reduced β-cell mass. The progression of T2D is clearly divided into at least two phases, pre-diabetes and overt diabetes [1]–[5]. In the pre-diabetes phase, insulin resistance is compensated by increased single β-cell secretion capacity and/or β-cell number. If insulin resistance is not completely compensated, the blood glucose level would grow slowly, manifested as higher fasting glucose (impaired fasting glucose, IFG) and/or higher post-load glucose (impaired glucose tolerance, IGT) [6]. Overt T2D is characterized by compensation failure and continuous loss of functional β-cells [7], [8], hence accompanied by continuously aggravated hyperglycaemia. Although insulin resistance is usually present in the early phase of pre-diabetes, it is the pace of β-cell failure that determines the onset of overt T2D [9]. The mechanisms leading to the transition from pre-diabetes to overt T2D are unclear [10]. However, there is evidence that the transition from β-cell compensation to β-cell failure happens in a comparably short time span [2], [4], typically within 3 years [2]. This is further supported by a recent longitudinal study in a large population [5]. The trajectory of glycaemia before diagnosis of T2D was shown to be composed of a slow and stable adaptation, which lasts 12 years, followed by a rapid rise of glucose to overt T2D within 2 years [5]. Once overt T2D is started, hyperglycaemia continues to worsen regardless of treatments based on oral anti-diabetic agents [11]–[13]. The evidence of a role of inflammatory responses in the pathogenesis of T2D was increasing in recent years. Interleukin-1β (IL-1β) has been reported to contribute to β-cell failure [14]–[18]. β-cells themselves secrete IL-1β upon glucose stimulation [14]. Furthermore, IL-1β stimulates its own production in β-cells [17] and attracts macrophages [18] which can act as an extra source of IL-1β and other cytokines. Although it is currently unclear whether inflammatory responses are a primary cause or a secondary effect in T2D progression, therapies targeting IL-1β have shown encouraging progress albeit diverse results in different clinical trials [19]–[26]. These results motivated our working hypothesis that the pre-diabetic and overt T2D might be characterised by two qualitatively different states and that IL-1β is a potential candidate for promoting the transition between these two states. The hypothesis is consistent with results from clinic trials in which IL-1β blockade by interleukin-1-receptor antagonist (IL-1Ra), a naturally occurring competitive inhibitor of IL-1β, induced sustained improvements of β-cell function and the systemic inflammation state in patients with a mean disease duration of 11 years, even 39 weeks after cessation of treatment [19], [20]. IL-1Ra competes with IL-1β for IL-1 receptors but does not trigger any signalling event. It is an important regulator of the effect of IL-1β in many cell types, including human pancreatic β-cells which secrete IL-1Ra themselves [27], [28]. In several liver, autoimmune, and infectious diseases IL-1Ra is a better indicator of disease severity than IL-1β [29]. Results from a recent longitudinal study show that an accelerated increase in circulating IL-1Ra starts about 5 years before diagnosis of T2D [30], coinciding with the accelerated deterioration of insulin-sensitivity and the compensation of β-cell function [5]. In addition, long-term effects of temporary intensive insulin therapy on diabetes remission (or at least a prolonged normoglycaemia phase) in newly diagnosed diabetic patients may be partially attributed to the anti-inflammatory activity of insulin [31]. All these results highlight the role of inflammatory responses in the pre-diabetes phase and the possibility that the long-term effects induced by temporary anti-inflammatory therapy are mediated by a switch, i. e. a sudden transition, between the qualitatively different compensation and overt T2D states. Interestingly, in vitro experiments show that the effect of IL-1β on β-cell mass and insulin secretion is twofold [32]: low concentration of IL-1β stimulates β-cell proliferation, inhibits β-cell apoptosis and enhances glucose-stimulated insulin secretion while high concentration of IL-1β has opposite effects. In other words, IL-1β may contribute to both the β-cell compensation phase and the β-cell failure phase. Here, we present a mathematical model showing the possibility that T2D onset is induced by a sudden transition of IL-1β to a high level. The IL-1β switch results from the coexistence of two different states, which is often termed as bi-stability. Bi-stable switch models have been widely used in modeling developmental processes, such as cell cycle progression, cellular differentiation and apoptosis. The characteristics of bi-stable switches include sudden transition, and hysteresis. The latter means that the state of a system depends not only on the current environment but also on its past state. We combine the IL-1β bi-stable switch model with a previously published T2D progression model [33], by assuming extrapolated β-cell turnover rates caused by exogenous IL-1β [32]. Similar concepts describing diabetes progression by different steady states have been already developed for both type 1 and 2 diabetes [33]–[35]. Our model falls into the same class as those previously published by Topp et al. [33] and De Gaetano et al. [35], which considered the evolution of β-cell mass, insulinemia and glycaemia over a time-scale of years and rely on glucose toxicity. β-cell mass is controlled by an empirical parabolic function of glucose in [33]. Subsequently a physiologically amenable concept of pancreatic reserve was introduced, which controls the direction of β-cell mass change [35]. Both β-cell mass and the pancreatic reserve were modeled phenomenologically as functions of glucose, where some parameters bear the potential to act as bifurcation parameters. The model presented here is built on these results. The phenomenological implementation of glucose toxicity is replaced by a model of IL-1β, which is explicitly defined and relatively easy to measure. We focus on an IL-1β bi-stable switch giving rise to hysteresis which is not discussed in previous studies. IL-1β hysteresis turns out to be of utmost importance because it widely shapes the strategy of anti-inflammatory therapies. While the complexity of the disease will never be captured by a mathematical model, the structural insights suggested by the model will help exploiting treatment approaches more efficiently and elaborating the effect of different therapies on β-cell mass and disease progression. The IL-1β bi-stable switch model implies that T2D onset and later irreversible progression of β-cell failure arise from altered stability properties of the β-cell compensation state. Therefore, the strategy of glucose control would be more effective in the compensation phase than after the onset of overt T2D. This model result is consistent with recent data from the Diabetes Prevention Program Outcome Study (DPPOS) showing that IGT people who once regressed to normoglycaemia have a 56% lower relative risk of developing diabetes [36]. While glucose-lowering therapy bears the potential of stopping disease progression in the compensation phase, the IL-1β bi-stable switch model suggests that a combined glucose-lowering and anti-inflammatory therapy is necessary to stop the loss of functional β-cell mass as well as hyperglycaemia progression in the overt T2D phase, and eventually reestablish the compensation phase. The mathematical model (Fig. 1) is constructed in three steps. First, we build the core of the model which describes the competition between IL-1β and IL-1Ra, where glucose stimulates IL-1β production. Since the effect of IL-1β on β-cells is mediated by the binding to its receptor, which is also the target of IL-1Ra antagonism, the fraction of IL-1β bound receptor F is used as a measure for IL-1β stimulation. Second, the interaction between glucose (G) and insulin (I) is modeled with β-cell mass being a parameter. This subsystem is mainly adopted from [33], which is a single-compartment model justified by the slow dynamics of glucose and insulin over a time-scale of days to years. G and I are associated with long-term fasting glucose and insulin levels. As short-term (daily) fluctuations are averaged out, the model cannot describe disease onset derived from details of glucose and insulin dynamics. Third, the two subsystems which describe the inflammatory signals and glucose-dependent insulin secretion are connected. For that purpose the IL-1β-dependent β-cell turnover rates are adopted from human islets which were cultured with exogenous IL-1β. The data suggest a bimodal effect of IL-1β on the β-cell mass [32]: IL-1β stimulates β-cell proliferation and inhibits β-cell apoptosis when presented in low concentrations. Conversely, at high concentrations IL-1β enhances β-cell apoptosis and reduces β-cell proliferation. The unknown model parameters are determined from steady state conditions of the IL-1β/IL-1Ra subsystem and from the natural history of fasting glucose in the pre-diabetic phase [5]. There are three different time scales in the model (Fig. 1). The IL-1β/IL-1Ra subsystem evolves on the fastest scale, since ligands and receptors bind/unbind on a time scale of seconds and proteins are synthesised/secreted on a time scale half an hour. The glucose/insulin subsystem evolves on an intermediate time scale of days [33]. β-cell evolves on the slowest scale. The turnover rate of β-cell is rather slow: typically, one or two cells divide per month. Since our main interest lies on long-term evolution of fasting glucose, the IL-1β/IL-1Ra subsystem is assumed in steady-state equilibrium and β-cell mass is assumed to be constant on this relatively fast time scale of fasting glucose. The different time-scales in the model make it possible to separate the model into different subsystems and justify our three-step modelling approach. All effects of IL-1β (L) are mediated by the IL-1 receptor and are blocked by IL-1Ra (A). IL-1Ra competes with IL-1β for the receptor but does not trigger any signalling event. Therefore, ligand-receptor binding and unbinding are considered to describe the antagonist effect of IL-1Ra. The ligand-receptor binding/unbinding happens on a time scale of seconds and is fast compared to other time-scales in the model, such as IL-1β and IL-1Ra production and secretion which happens on a time-scale of tens of minutes. Therefore, it is reasonable to assume that ligand and receptor are always in equilibrium, which enables us to describe the fraction of IL-1β bound receptors (F) by (1) where KL and KA are the dissociation constants of IL-1β and IL-1Ra, respectively. We assume a constant number of IL-1 receptors on the β-cell membrane. Consequently, F can be considered as a measure of IL-1β stimulation. F is used in the following to fit certain functions in the model to corresponding experimental data. In these experiments, different amounts of exogenous IL-1β were added to cultured human islets and the corresponding effects were measured, while endogenous IL-1Ra and IL-1β were not measured. It is safe to neglect endogenous IL-1β and IL-1Ra when exogenous IL-1β is dominant, which is the case for most measured data points. However, it is necessary to estimate the endogenous IL-1β and IL-1Ra concentrations for data points with zero exogenous IL-1β. Secretion of IL-1Ra from cultured human islets was measured after 4 days which led to the endogenous IL-1Ra concentration of 114 pg/ml [32]. Endogenous IL-1β is estimated to be 0. 228 pg/ml using the ratio of IL-1Ra/IL-1β of 500 [27]. IL-1Ra (A) stimulation by IL-1β (L) is bimodal [32]. IL-1Ra is induced by IL-1β at exogenous concentrations below 20 pg/ml. At higher concentrations, IL-1β restores IL-1Ra to nearly the basal level. This nonlinear effect is phenomenologically captured by the nonlinear equation (2) where dA is the natural degradation rate of IL-1Ra. The parameters k1 to k3 are determined in the Section “steady states conditions” (see below). IL-1β dynamics are described by (3) Glucose induces β-cells to secrete IL-1β [17] (k5 g (G) ). IL-1β is also secreted from islets via auto-stimulation [17] (k6 l (F) ). High levels of IL-1β mRNA inhibit the stimulating effect of glucose and of auto-stimulation [17] (b (L) ). IL-1β degrades with a rate dL. In the following, the functions g (G), l (F), and b (L) are derived from experimental data. The parameters k5 and k6 are determined in the Section “steady state conditions”. G and I are quantities associated with long-term changes (days to years) of fasting levels of glucose and insulin, respectively. Therefore, a single-compartment model for glucose and insulin is used. The ordinary differential equations describing glucose and insulin are adapted from Topp et al. [33] with the modification that the dynamics of insulin resistance (R) is included as input to the model. The dynamics of the β-cell mass (12) is controlled by IL-1β induced β-cell apoptosis and proliferation. The apoptosis rate a (F) (13) and the proliferation rate p (F) (14) are fitted to data from human islets [32], which were cultured for 4 days with different IL-1β concentrations applied at day 0 (Fig. 2 D and E, respectively). We assume that the initiation of apoptosis and proliferation is faster than IL-1β-induced IL-1β and IL-1Ra secretion. Thus, the IL-1β and IL-1Ra dynamics are de-coupled from β-cell turnover and the initial F (at day 0 of the experiment) can be used as an approximation in fitting the data. Since the apoptosis and the proliferation rates depend on log (F), eventual modifications of F caused by newly secreted molecules would anyway remain small on the logarithmic scale. β-cell apoptosis and proliferation were measured as percentage of control (i. e. no IL-1β added) [32] such that the absolute rates remained undetermined. Scaling of the rates is represented by an extra parameter τ (Eq. (12) ), which is fitted to the natural history of fasting glucose in the pre-diabetic phase [5]. The relative importance of apoptosis and proliferation is controlled by the parameter λ in Eq. (12), which is determined in the Section “steady state conditions”. While a large number of model parameters could be determined by experimental constraints, the parameters k1–6 and λ remained undetermined and are derived from steady state conditions for Eqs. (2,3, 12). Eq. (12) is used to determine λ. Eq. (2) and (3) each exhibit two stable steady states and a bifurcation point, such that these 6 additional conditions can be used to determine k1–6. The last two undetermined parameters, m and τ defined in Eq. 11 and 12, are determined by fitting the model to the fasting glucose history of both incident diabetes cases and the non-diabetics controls [5]. The best fit is shown in Fig. 4A. Interpolated insulin sensitivity [5] (Fig. 4B) which is an input to the model during fitting, the behaviour of the β-cell mass (Fig. 4C) and of IL-1β (Fig. 4D) are also shown. The fitting is implemented in the SBTOOLBOX2 [40] using a differential evolution algorithm, which reports multiple possible parameter sets (Fig. 5A). A strong correlation between m and τ was identified (Fig. 5B), which gives rise to large variations of glucose and β-cell mass in the simulations of overt diabetes (Fig. S1). In other words, the pre-diabetes glucose trajectory does not allow for a precise determination of m and τ. The values from the best fit are reported in Table 1 and used throughout the paper. The above approach demonstrates that the bifurcation of the IL-1β/IL-1Ra subsystem is sufficient to account for the fitted data set. To investigate whether the bifurcation of the IL-1β/IL-1Ra subsystem is necessary, a complementary approach is employed: the model is fitted to the data without bifurcation as a presupposed condition. The steady state equations for Eq. (2) and (3), in the compensation phase and in the overt diabetes phase, allow reducing the degree of freedom from 8 (k1∼6, m and τ) to 4: k1, k2, k5, and k6 are explicitly expressed as functions of k3 and k4, and only k3, k4, m and τ are fitted. Strong correlations are found to exist between m and τ, which cause large variations of glucose and β-cell mass in the simulations of the overt diabetes (Fig. S2 A and B). Strong correlations are also found between m and k4, which cause large variations of the time point of the IL-1β/IL-1Ra switch (Fig. S2 C and D). Please note that the complementary approach found, although with low probability, cases allowing an extra transition (see Fig. S2 C and D). Although some parameters cannot be determined precisely, the key qualitative property, i. e. the switch of the IL-1β/IL-1Ra subsystem before T2D diagnosis, remains robust. Taking together the results of both complementary approaches, it is suggested that the bifurcation of the IL-1β/IL-1Ra subsystem is both sufficient and necessary to account for the data set, in the framework of the current model. Please note that despite the model being validated by quantitative experimental data, the results presented here remain of qualitative nature. In particular, in the overt T2D state we expect that additional physiological process as well as medication will alter the results on a quantitative level. An IL-1β switch is observed in a large range of parameter values including the physiological relevant range. Below the critical glucose level of 5. 84 mM the model gives rise to five steady states, three stable and two unstable ones (Fig. 6A). The stable steady state with lowest IL-1β (green in Fig. 6 A) is stable only if glucose is below the critical level. It is interpreted as corresponding to the compensation phase. The hyperglycaemia parts of the two other stable steady states may be associated with overt T2D. Nullcline analysis facilitates understanding the transition between the different inflammatory levels. Steady states appear as the crossing points of the IL-1Ra versus IL-1β nullclines where both quantities become static (Fig. 6B). The IL-1β nullcline (Fig. 6B red line) exhibits a nadir followed by a peak, which is the typical feature of an auto-stimulation species. The IL-1Ra nullcline (Fig. 6B black line) exhibits one peak. This implies that at large IL-1β its further increase is accompanied by decreased IL-1Ra. The measured non-linear dependence of IL-1Ra on IL-1β [32] is at the origin of the third stable steady state with highest IL-1β. If IL-1Ra is stimulated by IL-1β in a linear manner, only two stable steady states would have been found. Such a dysregulation of IL-1Ra may also happen in other diseases involving IL-1β, such as chronic myelogenous leukaemia and hairy cell leukaemia [29]. Hence, dysregulation of IL-1Ra at high IL-1β concentrations might be associated with a boost of disease progression. The stable steady state with lowest IL-1β corresponds to the pre-diabetic state. As glucose influences IL-1β production [14], [17] the nadir of the IL-1β nullcline moves up in response to increased glucose (Fig. 6B inset) while the IL-1Ra nullcline remains unchanged (Fig. 6B, black line). The two involved steady states approach each other and ultimately merge into one steady state (Fig. 6B inset, blue line). At even higher glucose this steady state disappears and the system makes a transition (Fig. 6B, black arrow) to the next available stable steady state, which is characterised by substantially higher IL-1β and moderately increased IL-1Ra. This transition is interpreted to correspond to the sudden transition from the compensation phase to disease progression. At even higher levels of IL-1β, the effect of glucose on IL-1β production is inhibited by IL-1β itself (represented in Eq. 3 by b (l) ), such that the IL-1β nullcline remains unaffected by glucose at high IL-1β concentrations. Therefore, the current model only reflects the transition from the compensation phase to the mild disease phase and not to the strong disease phase. An extended model including detailed IL-1β mRNA dynamics and other factors like free fatty acids, leptin or tissue resident macrophages might generate two transitions. Glucose-lowering therapy, simulated in silico by fixing the level of glucose at normal level (5. 47 mM), is represented by the IL-1β nullcline which now, as in the pre-diabetic state, intersects with the IL-1Ra nullcline (Fig. 7A) around its nadir at low IL-1β. The therapy reconstitutes the existence of the pre-diabetic stable steady state (Fig. 7A, red circle). However, the original pathological steady state (Fig. 7A, black square) is neither deleted nor does it lose stability. The disease state remains stable irrespective of the glucose levels reached by glucose therapy. Therefore, the glucose-lowering therapy does not induce a transition from overt T2D to pre-diabetes and the functional β-cell mass is further decreasing (Fig. 7D, black squares). IL-1Ra is a naturally occurring inhibitor of IL-1β, and it is upregulated in obesity [28] and reduced in poorly controlled T2D patients [27]. By injecting exogenous IL-1Ra, IL-1Ra therapy aims at inhibiting IL-1β signalling. This is simulated in silico by fixing the level of IL-1Ra at a higher concentration (Fig. 7B, red cross). The simulated therapy slows down the loss of β-cell mass (Fig. 7D, red crosses) because increased IL-1Ra drives IL-1β into a less pathological regime. The higher IL-1Ra is set, the more IL-1β is reduced. At a threshold level of IL-1Ra, which is determined by the peak of the IL-1β nullcline (Fig. 6B, red line), the response of IL-1β becomes qualitatively different and it is reset below its physiological normal level (Fig. 7B, green triangle). Though the speed of β-cell loss is significantly reduced, it is not stopped (Fig. 7D, green triangles). Releasing IL-1Ra fixation leads to a restoration of the T2D state associated with accelerated loss of β-cell mass (Fig. 7B, black square), because the stable pre-diabetic steady state associated with low IL-1β is not restored by this therapy (Fig. 7B). A combined IL-1Ra and glucose-lowering therapy (Fig. 7C) is required to induce functional β-cell mass reconstitution. Strong IL-1Ra therapy induces a transition from the overt T2D steady state to a transient state at low IL-1β (Fig. 7C, from black square to red dot) associated with reduced speed of β-cell loss (Fig. 7D, red dots). A concurrent glucose-lowering therapy restores a stable steady state at low IL-1β (Fig. 7C, blue circle). Upon withdrawal of the IL-1Ra fixation, the in silico patient makes a transition to the restored stable steady state (Fig. 7C, from red dot to blue circle) which is associated with β-cell growth (Fig. 7D, blue circles), thus, inducing a long-standing improvement of the disease state. Long-term combined IL-1Ra and glucose-lowering therapy prohibits β-cell mass regrowth in silico (Fig. 7D, red dots). The model predicts that IL-1Ra therapy, simulated in the model as the fixation of IL-1Ra, must be released in order to allow for functional β-cell mass reconstitution. β-cell mass regrowth is suppressed during IL-1Ra therapy as it reduces IL-1β to levels below the pre-diabetic state. An IL-1β switch model of T2D onset after long-term compensation was presented and a strategy for an anti-inflammatory therapy of T2D was proposed. The model is characterised by a pre-diabetic state during which increasing insulin resistance drives hyperglycaemia. This ultimately leads to a changed steady state configuration and progression into the overt T2D disease state. The model parameters are derived from in vitro islet measurements, from steady state conditions, and from glucose history of T2D patients, such that the model parameters are fully determined. In the following we first discuss its relevance to the empirical theory of five-stage β-cell dysfunction [3], then describe the predictions of the model, its relevance and potential contributions to clinical trials [23], [26], and its relation to previous modelling works [33], [35]. Last but not least, we discuss its limitations. Interestingly, the model shows three steady-states of IL-1β. This result is beyond our expectations when designing the model but is consistent with the five-stage diabetes development theory [3]. This theory is mainly based on clinical observations, where stage 2, “adaptation” (termed as compensation in this paper) corresponds to the model state of lowest IL-1β (green in Fig. 6A). Stage 3 “transient unstable early de-compensation” corresponds to the sudden switch of IL-1β/IL-1Ra in the model, where a 55 fold increase in IL-1β level is accompanied with only 2 fold more IL-1Ra, and the transient adaptation time thereafter. Stage 4 “stable de-compensation” corresponds to the state with medium IL-1β level (red in Fig. 6A) and stage 5 “severe de-compensation” corresponds to the state with highest IL-1β level (black in Fig. 6A). Stage 1 “normal” is not represented in the model because glucose level is already higher even 14 years before T2D diagnosis (see Fig. 4A and [5]). Thus, the IL-1β switch model qualitatively reflects all stages of T2D development, starting from the compensation phase and progressing to the mild disease and ultimately to the strong disease state. Stage 3 is observed to be transient, persisting a few weeks in different animal models that were subjected to partial pancreatectomies [41] or islets transplantation [42], [43]. Interestingly, islets transplantation [43] and pancreatectomies [41] in rodents gave rise to glucose levels in two discrete groups: either normal glucose or severe diabetes, with almost no glucose levels in between, a pattern consistent with the idea of two distinct states in the presented model. Stage 3, transient and unstable, lies between the clinically non-pathologic and pathologic states, and thus, may be of great interest in terms of providing a time window during which therapies are more efficient in reversing the pathological processes right after diagnosis. Results from newly diagnosed diabetic patients who received intensive insulin therapy supports this notion [31]. However, in the current model, the switch of IL-1β/IL-1Ra, which is relevant to stage 3, is instantaneous (see Fig. 6 and Fig. S2) because the model does not describe the dynamics of the IL-1β/IL-1Ra subsystem during the transition. The equations of the IL-1β/IL-1Ra are mainly based on assumed steady state relations, which were derived from in vitro data. In addition, the model does not describe the dynamics of insulin resistance, the improvement of which has been assumed to be responsible of preventing/postponing the transition during stage 3 [3]. Rather, insulin resistance is implemented as an input to the glucose/insulin subsystem. These limitations of the current modeling approach (see also later discussions about the meaning of G and I in the model) make information about the duration of stage 3 unavailable from the model. The IL-1β bi-stable switch model suggests four requirements for a successful anti-inflammatory treatment of T2D: Then, the functional β-cell mass is predicted to increase. But even if it reaches its healthy level, the in silico T2D patient would remain in the pre-diabetic state, since the release of glucose control would restart T2D progression due to the pathological state of insulin resistance. The power of this therapy is not to cure T2D but to keep the in silico patient in the pre-diabetic state. Dose and duration of anti-IL-1β therapy are critical. On the one hand, an above threshold IL-1Ra is required for inhibition of IL-1β auto-stimulation and the reverse transition from β-cell failure to compensation. On the other hand, IL-1β has to be kept above its healthy level for β-cell reconstitution. This infers a U-shaped dose-effect relationship of anti-IL-1β therapy in silico, which has been confirmed in a phase 1 clinical trial using a monoclonal antibody of human IL-1β [23]. Large scale clinical trials with anti-IL-1β agents revealed that glycaemia was less efficiently reduced in patient groups with short (6 years) compared to groups with long (11 years) disease duration [26]. We have performed simulations of IL-1Ra therapy on patients with short and long T2D duration, where short and long T2D are associated with mild and strong T2D steady states, respectively (Fig. 8). The strong T2D state is initiated by manually increasing IL-1β and decreasing IL-1Ra to the levels shown in Fig. 6B at 11 years. Then, anti-IL-1β therapy (ignoring other adjuvant medication) is applied by an IL-1Ra fixation for 90 days to both in silico patient groups. The IL-1Ra dose-effect relationship for the two groups, expressed by the glycaemia improvement versus control, is compared (Fig. 8). In agreement with the clinical-trials results, optimal IL-1Ra dosage induced more efficient glucose improvement for strong than for mild T2D. This happens because in strong T2D the aggravation of blood glucose levels is faster in the control group. The optimal dose of IL-1Ra is different in both groups in silico (Fig. 8). This calls for an individualised treatment according to the endogenous inflammatory state. These simulations do not aim at explaining the diverse outcomes from clinical trials because other adjuvant medications are neglected during the simulations. Furthermore, the capability of the current model is limited to the low to medium hyperglycaemia/IL-1β region (see later discussion). However, the model still provides a framework in which different outcomes can be discussed in respect to patients' internal inflammatory levels and strategies may be suggested for patient stratification with respect to their inflammatory levels at the beginning of a clinical trial. The model proposed by De Gaetano et al [35] is able to account for different arms of the Diabetes Prevention Program (DPP) results [44], by modeling the effects of different drugs on insulin resistance development. Some differences and similarities are worth noting here, besides those already discussed in the introduction. In both models glucose toxicity introduces a positive feedback in the disease progression. Therefore, both models emphasize the importance of an early treatment of hyperglycaemia. The unique property of the current model is that the inflammatory signal, once switched on by glucose, cannot be switched off by glucose-lowering therapies. It has to be treated separately. This point seems to be consistent with recently discovered anti-inflammatory effects of some long-standing anti-diabetic agents. In fact, insulin has long been known for its anti-inflammatory effect, which, early presumed to be due to its glucose-lowering ability, recently was established to be a complement to its metabolic role [for a recent review see 45]. Additionally, there is emerging evidence suggesting that metformin, the first-line diabetic drug, have anti-inflammatory effects that are independent of its hypoglycaemia effect [46], even though its anti-inflammatory effect in T2D being not as robust as in other diseases, such as hypertension [47]–[49]. The impact of currently employed T2D drugs onto the inflammatory state of the patients prohibits falsification or verification of the model prediction that the T2D disease state cannot be alleviated by controlling glycaemia alone. The presence of the IL-1β switch suggests that the system is operating in two distinct regimes, which defines a clear threshold for an anti-inflammatory therapy. A striking prediction of the model is that the anti-inflammatory therapy has to be released after a certain time. These inferences are derived from the IL-1β/IL-1Ra hysteresis and are consistent with some long-standing findings [11], [12], as well as recent progresses in the field, such as multistage development of the disease [3], [5], long-term effect of temporary IL-1Ra in overt-diabetes [20] and the efficiency of glucose control in pre-diabetes but not overt-diabetes [31], [36]. The present IL-1β switch model is built on in vitro data and several assumptions, both of which cast limitations on its applicability. The in vitro data show regulation of IL-1β and IL-1Ra by glucose and exogenously added IL-1β [14], [17], [27], [32]. Other factors involved in in vivo IL-1β control, such as free fatty acids [50], infiltrated macrophages and islet amyloid polypeptides [51] are not considered. These factors are associated with serious hyperglycaemia and overt T2D. However, they do not or only weakly affect the results of the core IL-1β model, especially concerning the onset bifurcation at glucose levels below 6 mM. IL-1β induced β-cell proliferation and apoptosis were extrapolated from data of cultured islets [32]. It is unclear to what degree the in vitro data can reflect relevant in vivo processes, but several weak points of this assumption could be delineated. There is evidence that β-cells can re-enter cell cycle in vivo and that β-cell mass can adapt to metabolic needs [52], both supporting adaptive β-cell proliferation in vivo. However, it is not the only possible source of functional β-cells. Dedifferentiation of other cell types with a common precursor [53], as well as reactivation of dysfunctional β-cells [8], contributes to β-cell homeostasis. Furthermore, data from autopsy indicate that the proliferation rate of existing β-cells is very low. Consequently, increased β-cell mass seen in the compensation phase may be mainly attributed to new islet formation from other cell types [54]. A more realistic description of β-cell mass dynamics may further improve the model and extend its validity to later stages of the disease. By representing the inflammatory stimulus via F instead of IL-1β, it is implicitly assumed that the amount of membrane IL-1β receptor per β-cell is constant. In support of this, there is evidence that the amount of IL-1β receptor in the β-cell membrane is much higher than in other cell types [55], suggesting that it is saturated. However, it is unlikely that the receptor is not down-regulated at high concentrations of IL-1β. Consequently, the model may overestimate the inflammatory signal at high IL-1β levels. This limitation has to be kept in mind when interpreting model results at high IL-1β. Another important assumption concerns the IL-1β mRNA quantity used in the model. The IL-1β mRNA quantity is assumed to be proportional to the IL-1β protein level, in order to establish the quantitative relationships in IL-1β production. However, IL-1β production is characterized by the dissociation of transcription and translation. As a result, high levels of IL-1β mRNA are always associated with low levels of protein [29]. Alternatively, a sigmoidal relationship may be considered instead of the linear one used here. While it is difficult to estimate the quantitative error induced by the assumed linear relationship, it is straightforward to predict that the alternative sigmoidal relationship would increase the nonlinearity of IL-1β auto-stimulation, and of the interaction between glucose and IL-1β. Therefore, the essential elements for the bi-stable switch remain unchanged (or even would be increased), such that the qualitative results of the IL-1β switch model are robust. It is difficult to define the exact physiological meaning of G and I in the model [35]. Since G and I are described by one-compartment dynamics in the model, only changes in long-term fasting levels are accessible to the model. Hence, we propose to interpret G as an abstract parameter associated with long-term fasting glucose on a time scale of days to weeks. Fasting levels represent more the internal property of the body' s glucose control system than the history of how it was perturbed by meal or exercise [56]. However, the driving stimulus of IL-1β production may depend on peaks of glycaemia rather than on fasting levels. It is possible that dysregulations of fasting glucose levels and postprandial peaks are independent, as evidenced by patients with impaired glucose tolerance but normal fasting glucose. Eventually, the modifications of the dynamics of insulin release (first-phase insulin) are more important for glycaemia control than the long-term fasting levels of glucose. We would like to emphasize that these details of glucose and insulin dynamics are beyond the resolution of the model presented here and have to be addressed with a model focusing on the short-term dynamics of these quantities. T2D is a very complex disease. The current model is simplified on purpose in order to generate structural insight. Other players of the glucose control system, for example glucagon, also play very important roles in glucose homeostasis and T2D development. Augmented glucagon secretion, together with impaired β-cell function, happens in the very early phase of pre-diabetes [57], [58]. The underlying islet defects were identified as reduced maximal insulin response and reduced glucose-sensitivity of β-/α-cell [58]. According to the physiological integral rein control theory [56], [59], glucose-sensitivity of β-/α-cell are important factors in setting steady state glucose level. In fact, recent studies support an independent role of glucose-sensitivity in hyperglycaemia development [60], [61]. Future improvement of the model should include the glucose-sensitivity of β-/α-cell to account for the rising of fasting glucose in the compensation phase. In particular, this might improve the fitting to the data in Fig. 4A during the compensation phase. In addition, post-load glucose of incident diabetes cases showed a rapid increase 5 years before diagnosis [5], which is followed by a rapid decrease in insulin sensitivity. This important change is not reflected in the history of fasting glucose while it is represented in the model output (Fig. 4A). On the one hand, the absence of a corresponding change in fasting glucose at ∼5 years before diagnosis suggests that fasting glucose level is dominated by factors other than insulin resistance. Again, this might point to a role of glucagon. On the other hand, the coincident post-load glucose increase and insulin sensitivity decrease at ∼5 years before diagnosis might suggest a cause-effect relationship between β-cell function and insulin resistance [62]. The complexity of the defects during the pre-diabetes phase highlights the necessity of the mathematical modeling approach in the field. Although results from both in vitro experiments and animal models were promising, the actual effect of anti-IL-1β agents in clinical trials of T2D was a matter of debate because the observed effect was very modest so far. Besides the possible role of polymorphisms of the IL1RN gene which encodes IL-1Ra [20], more research is needed to clarify for example, drug interaction with other IL-1 family members, such as soluble IL-1 receptors and soluble IL-1 receptor accessory proteins, in the pancreatic local environment as well as in blood and liver. On the structural level, the present analysis proposes that the transition from pre-diabetes to overt T2D is associated with changed stability properties of the endocrine-immune system. The transition is initiated at a threshold value of glucose and associated with qualitatively different inflammatory states. The notion of an IL-1β hysteresis is supported by data from different studies, which showed the effectiveness of glucose control in pre-diabetes [36] or newly diagnosed patients [31] and the ineffectiveness of glucose control in long-standing T2D patients [11]–[13]. It is also supported by the clinical trial based on a combined IL-1Ra and glucose-lowering therapy: the improvement of the disease state remained for a year after release of the IL-1Ra therapy [20]. The existence of the IL-1β hysteresis requires a combined glucose-lowering and anti-inflammatory therapy for overt T2D patients, where the anti-inflammatory therapy should be kept for only a limited time.

Insulin resistance and relative insulin deficiency are two hallmarks of type 2 diabetes. While insulin resistance is always present in the early phase, it is β-cell failure that determines the pace of the disease onset. Increasing evidence that the immune system is activated and plays an important role in type 2 diabetes has stimulated efforts of developing drugs targeting inflammatory cytokines. We built a model to describe the principles of type 2 diabetes onset under the influence of interleukin-1β. The disease onset is understood in terms of bifurcation. It is found that inflammatory cytokines are required to be suppressed for a limited time only, while glucose has to be controlled over the long term. These structural insights may serve as a guideline for future clinical trials.

Abstract Introduction Materials and Methods Results Discussion

diabetes mellitus type 2 diabetes metabolic disorders medicine and health sciences

2014

Possible Role of Interleukin-1β in Type 2 Diabetes Onset and Implications for Anti-inflammatory Therapy Strategies

Lignin is a polymer in secondary cell walls of plants that is known to have negative impacts on forage digestibility, pulping efficiency, and sugar release from cellulosic biomass. While targeted modifications of different lignin biosynthetic enzymes have permitted the generation of transgenic plants with desirable traits, such as improved digestibility or reduced recalcitrance to saccharification, some of the engineered plants exhibit monomer compositions that are clearly at odds with the expected outcomes when the biosynthetic pathway is perturbed. In Medicago, such discrepancies were partly reconciled by the recent finding that certain biosynthetic enzymes may be spatially organized into two independent channels for the synthesis of guaiacyl (G) and syringyl (S) lignin monomers. Nevertheless, the mechanistic details, as well as the biological function of these interactions, remain unclear. To decipher the working principles of this and similar control mechanisms, we propose and employ here a novel computational approach that permits an expedient and exhaustive assessment of hundreds of minimal designs that could arise in vivo. Interestingly, this comparative analysis not only helps distinguish two most parsimonious mechanisms of crosstalk between the two channels by formulating a targeted and readily testable hypothesis, but also suggests that the G lignin-specific channel is more important for proper functioning than the S lignin-specific channel. While the proposed strategy of analysis in this article is tightly focused on lignin synthesis, it is likely to be of similar utility in extracting unbiased information in a variety of situations, where the spatial organization of molecular components is critical for coordinating the flow of cellular information, and where initially various control designs seem equally valid. Lignin is a phenolic heteropolymer in the secondary cell walls of vascular plants. It is derived mainly from three hydroxycinnamyl alcohol monomers, namely p-coumaryl, coniferyl, and sinapyl alcohols, which, when incorporated into the lignin polymer, give rise to p-hydroxyphenyl (H), guaiacyl (G), and syringyl (S) subunits, respectively. The development of lignin biosynthesis is considered to be one of the key factors that allowed vascular plants to dominate the terrestrial ecosystem [1]. This evolutionary advantage is in part due to the fact that lignin, when deposited in the cell wall, contributes to the structural integrity of the cell, facilitates transport of water and minerals through the tracheary elements, and serves as a defensive barrier against pathogens and herbivores [2]. In addition to lignin, the secondary cell walls of vascular plants contain several polysaccharides, such as cellulose and hemicellulose. Extraction of these polymers from lignocellulosic biomass for biofuel production has attracted extensive interest partly because exploitation of this source of fermentable sugars could minimize the competition for food, which has been criticized in the case of corn- or sugarcane-based biofuel production. However, the natural resistance of lignocellulosic biomass to enzymatic or microbial deconstruction has rendered the task of generating sustainable and cost effective biofuels from lignocellulosic feedstocks very challenging. While impressive advances have been made toward the reduction of biomass recalcitrance [3], [4], it was also shown that the amount of fermentable sugars released through chemical and enzymatic treatments is inversely proportional to that of lignin present in biomass, and that some transgenic plants with reduced lignin content yield up to twice as much sugar from their stems as wild-type plants [5]. These observations suggest that lignin biosynthesis may be targeted for generating engineered crops with reduced recalcitrance. The challenge of this task derives from the fact that a rational design of less recalcitrant varieties would require a thorough, multi-level understanding of the lignin biosynthesis in wild- type plants, which we do not yet have. Such an understanding would include a grasp of the system of interactions between the enzyme-encoding genes, proteins and metabolites involved in the biosynthesis of lignin, as well as details of the regulation of this multi-tiered system. The metabolic scaffold for the biosynthesis of the three building blocks of lignin originally was seen as a grid-like structure [6], but this initial structure has been revised and refined and is now understood as an essentially linear pathway with only a few branch points (Figure 1). Although this generic pathway structure is now widely accepted, it has become clear that different lineages of vascular plants have evolved variants that engage distinct biosynthetic strategies. An interesting example is the model legume Medicago truncatula, where the characterization of two distinct cinnamoyl CoA reductases, CCR1 and CCR2, has suggested parallel routes from caffeoyl CoA to coniferyl aldehyde (Figure 1) [7]. A more unusual case is the lycophyte Selaginella moellendorffi. Functional analyses of the two enzymes recently discovered from this species, SmF5H and SmCOMT, support the notion that S. moellendorffi may have adopted a non-canonical pathway from that in angiosperms to synthesize coniferyl and sinapyl alcohol (Figure 1) [8], [9], [10]. Given such variations, it would appear reasonable to consider genus- or species-specific similarities and differences. However, such data are seldom available, and even if a customized pathway structure can be established, its regulation often remains obscure. This shortcoming tends to become evident with new, precise data. For instance, experiments using genetically modified M. truncatula lines with reduced CCR1 activity exhibited an unexplainable decrease in the ratio of S to G lignin over wild type [7]. Such discrepancies between expectation and observation suggest that the currently accepted pathway diagrams may require further revisions that include regulatory mechanisms affecting the physiological outcome when the pathway is perturbed. The focus of this article is an assessment of such a regulatory system associated with lignin biosynthesis in Medicago. This genus includes model species like M. truncatula, as well as alfalfa (Medicago sativa L.), an important forage legume. Medicago is particularly suited for these studies, because comparatively extensive information is available. For instance, a detailed dataset was established that characterized different lines in which seven lignin biosynthetic enzymes were independently down-regulated, and the resulting lignin content and monomer composition were determined in several stem segments [11]. In a recent study, we demonstrated that these types of data contain substantial, although hidden, information. In particular, we used these data to show that certain enzymes may co-localize and/or assemble into two independent channels for the synthesis of G and S lignin, and that salicylic acid acts as a potential regulatory molecule for the lignin biosynthetic pathway [12]. Although these earlier results provided significant insights into the mechanisms of regulation in this pathway, several critical questions, especially regarding the biological function as well as the operating mode of the channels, remain unanswered: For instance, are these channels always active in vivo? Are they sufficient to explain all available data in Medicago? Is there crosstalk between them, and if so, how is it organized? Exploring all pertinent scenarios associated with such questions would be experimentally intractable because they are simply too numerous. Instead, we present here a novel computational approach to investigate exhaustively all regulatory schemes involving the key reactions associated with G and S channels in the lignin biosynthetic pathway (Figure 2). The specific hypothesis is that the formerly postulated and validated channels may have two different modes of operation. Either they are permanent in a sense that the component enzymes are persistently assembled into a complex; such a complex could be realized through membrane co-localization, thereby ensuring that the corresponding alcohol is always synthesized. As an alternative, the channels could be facultative, thereby displaying a functionality that depends on the sub-cellular localization of the component enzymes and the metabolic milieu. This hypothesis, in turn, leads to 19 possible topological configurations (Figure 3A). For each of these topologies, we consider an additional level of regulation, involving individual or combined regulatory mechanisms that may serve as a means of “crosstalk” between the two channels (Figure 3B). The emphasis of this approach is on mechanisms at the metabolic level, but one must not forget that the transcriptional network governing the system could be involved in regulation of the pathway as well [13]. The goal is thus to assess and compare the functionality of all given combinations of topological configurations and crosstalk patterns, each of which we call a design. To obtain insights that are independent of parameter choices, we constructed for each design a library of 100,000 loosely constrained dynamic models and tested each of them against the observed ratios of S to G lignin in four lignin-modified Medicago lines. The resulting analysis of hundreds of designs and millions of models led to the intriguing hypothesis that either a single activation mechanism or a dual-inhibition mechanism lies at the core of all experimentally supported designs. The former mechanism was not supported by an in vitro enzyme assay, while the latter is consistent with several lines of evidence from Medicago and other species. As an added insight, the analysis suggested that functionality of the G lignin channel is more important than that of the S lignin channel. Overall, these findings not only enrich our current understanding of how lignin biosynthesis is regulated, but they also demonstrate the possible application of the proposed approach in entirely different biological scenarios where the task is to identify true regulatory circuit among many theoretically feasible designs that depend on the functionality and localization of interacting molecules. The base scaffold on which the different topological variants were built is shown in Figure 2. It consists of all relevant steps in the lignin biosynthetic pathway that possibly affect the relative amounts of G and S lignin. Prior work [12] provided evidence that CCR1 and cinnamyl alcohol dehydrogenase (CAD) may organize into a functional complex through which the substrate coniferyl aldehyde is transferred from CCR1 to CAD without much leakage, thereby acting as a channel leading specifically to the synthesis of G lignin. Similarly, caffeic acid O-methyltransferase (COMT) and ferulate 5-hydroxylase (F5H) were suggested to form an analogous complex contributing specifically to the synthesis of S lignin. These two complexes, which we called G and S lignin channels, are represented in Figure 2 as two directed edges, one linking feruloyl CoA and coniferyl alcohol (G channel) and the other linking caffeyl aldehyde and 5-hydroxyconiferyl aldehyde (S channel). The experimentally validated channeling hypothesis permits 19 different topological configurations (Figure 3A) that satisfy the following constraints. First, at least one edge must be leaving caffeyl aldehyde and feruloyl CoA, and at least one edge must be entering coniferyl alcohol and 5-hydroxyconiferyl aldehyde; otherwise mass would unduly accumulate in intermediate pools. Second, if coniferyl aldehyde can be produced by a free CCR1 and/or COMT, it must also be consumed by a free enzyme, thereby decreasing the metabolic burden that would otherwise be imposed on the cell. For reasons that will be explained below, we also consider for each topological configuration various crosstalk patterns between the CCR2/COMT and CCoAOMT/CCR1 pathways. Each pattern is composed of documented or postulated mechanisms of metabolic regulation (activation or inhibition) (Figure 3B). The specific combinations of topological configurations and crosstalk patterns lead to hundreds of different designs, which were analyzed and compared (see Figure S4). For each design, we first constructed 100,000 Generalized Mass Action (GMA) models by randomly sampling loosely-constrained parameter combinations from a parameter space that was deemed biologically realistic. A notable feature of this approach was that the parameter space was not only constrained at the level of individual parameters (e. g. kinetic orders), but also at the level of steady-state fluxes. For instance, the ratio of fluxes leading to S and G lignin was fixed at a value observed in the wild-type Medicago species (see Materials and Methods and Text S1 for details). Once all parameters for a given GMA model instantiation were specified, we determined steady-state fluxes under conditions that mimic CCoAOMT and COMT down-regulated alfalfa lines as well as ccr1 and ccr2 M. truncatula mutant lines and computed the S/G ratios for which we had experimental data. We declared a model as valid if it yielded quantitatively and qualitatively correct results for both transgenic alfalfa and M. truncatula plants (see Materials and Methods). To assess the robustness of a design to parametric perturbations, we defined Q as the total number of valid model instantiations. As a reasonable baseline, we first assumed the absence of crosstalk between the CCR2/COMT and CCoAOMT/CCR1 pathways (Figure 4). Of all possible topological configurations lacking crosstalk, only six had at least one parameter combination that yielded quantitatively correct predictions of S/G ratios for CCoAOMT and COMT down-regulated alfalfa plants. Supporting our previous findings [12], all six configurations include either one or both channels, suggesting that the channels are necessary. In other words, the pathway models are consistent with the observed changes in the S/G ratios of CCoAOMT and COMT down-regulated alfalfa plants only if at least one channel is present. To assess these initially feasible parameter combinations further, we used the models with these parameter values to predict the S/G ratios for ccr1 and ccr2 knockout mutants. The M. truncatula lines harboring transposon insertions in CCR1 and CCR2 show a corresponding reduction in CCR1 and CCR2 activity, and their S/G ratio is decreased or increased, respectively, compared to the wild-type level [7]. Moreover, the activities of CCR1 and CCoAOMT, as well as their mRNA transcripts and proteins, are increased in the ccr2 knockout mutant, indicating that part of their activation might be processed through a hierarchical control of gene expression [14]; see also Figure S4. Figure 4 shows simulation results for those topological configurations where at least one out of 100,000 randomly parameterized models yielded quantitatively correct predictions of S/G ratios for both CCoAOMT and COMT down-regulated alfalfa plants. In these plots, a model is valid only if its predicted S/G ratios for ccr1 and ccr2 knockout mutants fall within the northwest quadrant. In the case of no hierarchical regulation, i. e. , the ccr2 mutant exhibits only reduced CCR2 activity, some model instantiations from configuration A showed a decreased S/G ratio for the ccr1 knockout mutant, but not a single case exhibited an increased S/G ratio for the ccr2 knockout mutant. This outcome did not improve much when hierarchical regulation was considered: not one of the 1. 9 million model instantiations from the 19 possible configurations yielded qualitatively acceptable predictions for both ccr1 and ccr2 knockout mutants. These findings indicate that the S and G channels alone are not sufficient to explain all available transgenic data, and that some type of crosstalk is highly likely to occur between the CCR2/COMT and CCoAOMT/CCR1 pathways. One potential source of crosstalk between the CCR2/COMT and CCoAOMT/CCR1 pathways is substrate competition. CCR1/2 converts hydroxycinnamoyl CoA esters to their corresponding cinnamyl aldehydes, whereas CCoAOMT and COMT together complete the methylation of the aromatic C3 and C5 positions of the aldehydes and alcohols (Figure 1). All these enzymes are known to be multi-functional, acting upon multiple substrates with distinct catalytic efficiency. Because of their promiscuous nature, different substrates compete with each other if the supply of enzyme is limited. As a consequence, the enzymatic conversion of one substrate is effectively subjected to competitive inhibition by another substrate, and vice versa. This type of cross-inhibition is not necessarily equally strong in both directions because a promiscuous enzyme often displays preference for some substrates over others. In the case of lignin biosynthesis, two regulatory mechanisms could arise from substrate competition. First, recombinant Medicago CCR2 exhibits similar kcat/KM values for caffeoyl CoA (0. 49 µM−1•min−1) and feruloyl CoA (0. 40 µM−1•min−1) [7], suggesting that the CCR2-mediated conversion of caffeoyl CoA to caffeyl aldehyde in Medicago might be competitively inhibited by feruloyl CoA (Figure 3B; Mechanism 1). Furthermore, CCR2 is inhibited by feruloyl CoA at a concentration above 20 µM [7]. Conversely, it is highly unlikely that the CCR1-mediated conversion of feruloyl CoA to coniferyl aldehyde is significantly affected by caffeoyl CoA, because CCR1 has a kcat/KM value for caffeoyl CoA (0. 019 µM−1•min−1) that is 60 times lower than that for feruloyl CoA (1. 14 µM−1•min−1) [7]. Second, the methylation of caffeoyl CoA by the combined activity of COMT and CCoAOMT may be subject to weak competitive inhibition by caffeyl aldehyde (Figure 3B; Mechanism 2). This assumption is based on the following observation. Although the combined O-methyltransferase (OMT) activity against caffeoyl CoA in extracts from internodes 6 to 8 of CCoAOMT-down-regulated alfalfa was reduced by 4. 2-fold compared with the empty vector control line, about ∼25% of OMT activity remained [15]. This activity is presumably associated with COMT, for which caffeyl aldehyde is the preferred substrate. Notably, both mechanisms are independent of each other and may work individually or collaboratively to establish crosstalk between the two channels, thereby leading to three different crosstalk patterns and 57 different designs. In the case where only Mechanism 1 (Figure 3B) was incorporated in the design, we observed a substantial increase in the number of model instantiations showing a decreased S/G ratio for the ccr1 knockout mutant (Figure S1). Yet, even when we accounted for the effect of hierarchical regulation, none of the models was capable of delivering a qualitatively correct change in the S/G ratio for the ccr2 knockout mutant. This finding indicates that the experimentally inferred inhibition evidently exists but is not sufficient. Similarly, we found no valid models when Mechanism 2, either by itself or coupled with Mechanism 1, was employed (Figures S2 and S3). An explanation may be that, with caffeyl aldehyde inhibiting the 3-O-methylation of caffeoyl CoA, knocking down CCR2 activity will consistently lead to a deregulation of CCoAOMT by caffeyl aldehyde, thereby increasing the flux to G lignin and reducing the S/G ratio. One could surmise that the 3-O-methylation of caffeoyl CoA, for which CCoAOMT is the primary enzyme, is actually activated by caffeyl aldehyde. This conjecture is based on the following argument. When the production of S lignin is compromised due to a knockout of ccr2, the only way of raising the S/G ratio beyond its wild-type level is to further reduce the flux through the CCoAOMT/CCR1 pathway, which can be accomplished if CCoAOMT is activated by caffeyl aldehyde. The simulation results using this type of postulated mechanism, either by itself (Figure 5) or coupled with the documented inhibition of CCR2 by feruloyl CoA (Figure 6), are very intriguing: For each crosstalk pattern where millions of randomly parameterized models were generated, we found thousands of valid instantiations that yielded quantitatively and qualitatively correct predictions for both transgenic alfalfa and M. truncatula plants. Perhaps more surprisingly, only six topological configurations (A, B, E, F, I, O) had at least one valid model (Q>0; see Materials and Methods). To ensure that this result was not due to the use of overly restrictive thresholds, we relaxed the criteria and found more parameter combinations that qualified. Nevertheless, the same six topological configurations always passed the screening test by a wide margin (Table S1). Collectively, these findings suggested that this activation mechanism, acting alone or with the inhibition of CCR2 by feruloyl CoA, is necessary for consistency with the ccr1 and ccr2 knockout data. This conclusion immediately translated into a targeted hypothesis that was independent of specific parameter choices and readily testable by experiment. To examine whether caffeyl aldehyde indeed activates CCoAOMT, we expressed alfalfa CCoAOMT in Escherichia coli and assayed the purified recombinant enzyme with caffeoyl CoA as substrate and caffeyl aldehyde as the putative activator. As shown in Figure 7, the CCoAOMT activity increased by 16% at 2 µM of caffeyl aldehyde and 20 µM of caffeoyl CoA; at higher substrate concentrations (i. e. , 30 and 40 µM of caffeoyl CoA), the increase in mean CCoAOMT activity became less. Assays using lower concentrations of the substrate caffeoyl CoA (2,4, 5 and 10 µM) and the putative activator caffeyl aldehyde (0. 5,1, 2 and 4 µM) showed no increase in CCoAOMT activity compared to the reaction without caffeyl aldehyde (data no shown). The maximal activation achieved in vitro was only 16%, which may not be biologically significant. Since a direct activation of CCoAOMT by caffeyl aldehyde was not observed with recombinant enzymes, we tested other regulatory mechanisms by themselves and in combination with known mechanisms. According to one possible mechanism, based again on the concept of substrate competition, caffeoyl CoA could be a competitive inhibitor for the 3-O-methylation of caffeyl aldehyde (Figure 3B; Mechanism 4). This proposal agrees with the fact that CCoAOMT may contribute up to ∼10% of the methylation reaction in alfalfa [15]. In addition, evidence in ryegrass (Lolium perenne) points to the possibility of COMT being inhibited by different substrates, such as caffeyl aldehyde and 5-hydroxyconiferyl aldehyde [16]. Interestingly, substrate inhibition by caffeyl alcohol and 5-hydroxyconiferyl alcohol has also been observed in Selagniella moellendorffii COMT [9]. Thus, we hypothesized that COMT might be inhibited by caffeyl aldehyde (Figure 3B; Mechanism 5) in Medicago as well; direct evidence supporting this hypothesis in Medicago remains to be determined. In total, there are 24 = 16 different crosstalk patterns that can result from the combination of four independent regulatory mechanisms (Figure 3B; Mechanisms 1,2, 4 and 5). However, only four of them, when combined with the same six topological configurations (A, B, E, F, I and O) that were identified previously (cf. Figures 5 and 6), gave rise to designs with at least one valid model instantiation (Figure 8). Interestingly, all these crosstalk patterns require that caffeyl aldehyde is an inhibitor of the 3-O-methylation of both caffeoyl CoA and itself (Figure 3B; Mechanisms 2 and 5), providing computational evidence that this synergy between the two seemingly unrelated mechanisms is necessary for consistency with the ccr1 and ccr2 knockout data. Indeed, with respect to the ccr2 knockout, such a combination of two inhibition mechanisms appears to have a similar ultimate effect as a single activation mechanism (see Discussion section). Inspecting the crosstalk patterns giving rise to at least one design with valid model instantiations (rows colored in red in Figure 8), one might surmise that caffeyl aldehyde would accumulate to an unduly high level, because Mechanism 5, which is employed in all these patterns, reflects substrate inhibition of COMT by caffeyl aldehyde. To examine the validity of this inference, we checked, for all designs with valid model instantiations, the predicted changes in caffeyl aldehyde under conditions that mimic the down-regulation of four lignin biosynthetic enzymes. As shown in Figure 9, it appears that down-regulation of CCoAOMT or COMT is consistently associated with a lower caffeyl aldehyde level compared with wild type, regardless of the crosstalk pattern being considered. Similarly, knocking out ccr2 consistently raises the caffeyl aldehyde level in all crosstalk patterns examined. However, in the case of the ccr1 knockout mutant, the results are mixed in a sense that some crosstalk patterns are associated with significantly higher caffeyl aldehyde levels, whereas others are associated with only modest changes. Interestingly, both crosstalk patterns suffering from an undue accumulation of caffeyl aldehyde contain Mechanism 1. By contrast, this mechanism is absent from other patterns, which maintain a relatively stable caffeyl aldehyde level. This finding suggests that the control pattern in Mechanism 1 may disrupt the metabolic homeostasis via accumulation of caffeyl aldehyde when CCR1 drops below its normal level. As any cellular system is constantly afflicted by a variety of intrinsic and extrinsic noises, this type of fluctuation must be expected to occur frequently and spontaneously, suggesting that Mechanism 1 is disadvantageous. Investigation of the six robust topological configurations, which contain at least one valid model instantiation, revealed interesting structural features of the pathway. In particular, the G lignin channel is common to all robust designs and thus may be considered critical for the proper functioning of the pathway, at least for the cases studied. The evolutionary conservation of such a feature, one may argue further, is not due to the fact that it cannot possibly be altered, but that this particular design can sustain maximally tolerable changes and variability in other features [17]. These arguments lead to an interesting follow-up question, namely: Are the robust topological configurations related in an evolutionary sense? To address this question, we constructed a “topology graph” where each node corresponds to a topological configuration. Two nodes are connected if the corresponding topological configurations differ only by one edge. For instance, configurations A and B are directly linked to each other because the only difference between them is whether caffeyl aldehyde can be converted, via free COMT, to coniferyl aldehyde. In other words, moving from a node to its neighbor may be considered a singular evolutionary event where an enzyme' s preferred mode of action is changed. Two outcomes are possible for the structure of such a topology graph. First, the graph may be disconnected, that is, there exist pairs of topological configurations such that no evolutionary path (defined as a series of evolutionary events) connects one to the other. In the most extreme case, the graph would consist exclusively of isolated nodes. Second, the graph is fully connected, so that any pair of topological configurations is connected by at least one evolutionary path. As shown in Figure 10, the actual topology graph of the six robust configurations of lignin biosynthesis is indeed connected, and so is the graph of all configurations, except for design S. This interconnectedness can be interpreted as facilitating the evolvability of the system [17], because the gain or loss of specific features that are needed to produce phenotypically novel traits will be tolerated and survive during evolution if robustness is preserved. Of course, this evolutionary aspect, which was derived purely with computational means, will require additional analysis. The spatial organization of cooperating enzymes, known as metabolic channeling, has long been recognized as an effective means of regulation in primary and secondary plant metabolism [18], [19], [20]. This channeling phenomenon involves the organization of enzymes into complexes and/or the co-localization of enzymes at the plasma membrane or on the surfaces of organelles, as was demonstrated for the two initial enzymes, L-phenylalanine ammonia-lyase (PAL) and cinnamate 4-hydroxylase (C4H), in the phenylpropanoid pathway [21], [22]. Interestingly, some complexes or interactions are persistent, while others are temporary. In fact, many of the component enzymes such as PAL may be operationally soluble and are therefore only facultatively channeled. Such short-lived or dynamic complexes, while being readily responsive to the metabolic status of the cell, are inherently difficult to study with existing or emerging experimental models. Using the lignin biosynthetic pathway as a model system, we propose here a novel strategy for studying metabolic channeling in unprecedented detail. Specifically, we consider all possible modes of action for both the G lignin and S lignin channels, and these can be mapped into 19 different topological configurations (Figure 3A). Metabolic channeling is clearly not the only process that affects the functionality of this system, and it is therefore necessary to study control processes affecting a channeled system. In the present case, this control is potentially exerted by individual or combined mechanisms of crosstalk between the CCR2/COMT and CCoAOMT/CCR1 pathways (Figure 3B). Some of these were documented in the literature, while others were hypothesized. Taken together, a topological configuration and a specific crosstalk pattern constitute a design. We evaluated each design with or without consideration of non-allosteric or hierarchical regulation which could involve transcription, as well as a variety of non-transcriptional processes such as phosphorylation, methylation, and targeted degradation of proteins and mRNA. Ideally, the comparative assessment of design features would be entirely symbolic and independent of specific parameter values. However, systems of a realistic size are rarely analyzable in such fashion. As a reasonable alternative, we analyzed the possible design space comprehensively with widely varying parameter values, which resulted in a computational analysis of millions of models from hundreds of designs. This analysis yielded several interesting findings. Importantly, it predicted that CCoAOMT is directly or indirectly activated by caffeyl aldehyde. This piece of information by itself is essentially unbiased, but insufficient to explain the exact mechanism of regulation. Nevertheless, it offered a specifically targeted hypothesis and was therefore experimentally testable. However, the hypothesis of a direct activation was refuted by subsequent experiments using the recombinant Medicago CCoAOMT, which failed to provide evidence confirming the putative role of caffeyl aldehyde as an allosteric activator. It might still be possible that activation exists in vivo, but it seems more likely that the activation is indirect rather than direct. As a possible mechanism, the design analysis suggested that caffeyl aldehyde inhibits the 3-O-methylation of both caffeoyl CoA and itself. Several lines of evidence, although not exclusively from Medicago, support this computational prediction. Most importantly, the same six topological configurations were identified in the indirect design analysis and in the initial analysis of a putative activation mechanism. However, the two most parsimonious mechanisms differ in their proposed control strategies. The original analysis suggested just one activation mechanism, while the second analysis proposed two inhibition mechanisms. To some degree, these two mechanisms have the same ultimate effect. If ccr2 is knocked out, the flux entering the CCR2/COMT pathway and the subsequent synthesis of S lignin decline. The only possibility to increase the S/G ratio is to reduce the flux entering the CCoAOMT/CCR1 pathway. This task can be accomplished either through a diminished activation, as suggested for the single activation mechanism, or through an enhanced inhibition, as suggested for the dual-inhibition mechanism. The latter mechanism seems sufficient to restore consistency with the data, but it is of course possible that more complicated control patterns are present. The computational analysis suggests that the G lignin channel is necessary for the system to respond correctly and robustly to certain genetic perturbations. By contrast, the S lignin channel appears to be dispensable. This theoretical deduction is indirectly in line with the fact that S lignin has arisen much later in the evolution of higher plants than G lignin [1]. It is also consistent with the observation that its formation, which in many plant species is dictated by F5H expression [23], [24], [25], is directly regulated by a secondary cell wall master switch NST1/SND1 and not by MYB58, a SND1-regulated transcription factor that can activate other lignin biosynthetic genes [26]. It could also be possible that S lignin, which is specifically involved in the pathogen defense of some plants [27], was relatively recently recruited for lignin biosynthesis and thus may not be essential for plant growth. Evidence supporting this postulate includes an Arabidopsis NST1/SND1 double knockout mutant that shows a complete suppression of secondary cell wall thickening in woody tissues, including interfascicular fibers and secondary xylem, but otherwise grows quite well as compared to the wild-type plants [28]. Within an evolutionary context, the multiplicity of robust solutions can be represented with a graph representation that connects any two (robust) topological configurations differing by a single edge. This graph is reminiscent of the “neutral network” concept that was initially proposed in genotype-phenotype models for RNA secondary structures [29] and protein folds [30], but also more recently extended to Boolean models for gene regulatory networks [31]. In the case of proteins, neutral networks are defined as sets of amino acid sequences that are connected by single-mutation neighbors and that map into the same tertiary structure. Such degeneracy of the mapping from genotype to phenotype allows a neutral drift in genotypic space, which is critical for accessing adjacent neutral networks with novel phenotypes that may confer higher fitness to the cells. As of yet, it is unclear whether individual plants within a Medicago population use the same or different designs, or whether the response to selected perturbations is an adequate phenotypic feature. Further investigation of the protein-protein interactions between lignin biosynthetic enzymes is thus necessary to confirm that a G lignin channel is indeed necessary for optimal functioning. The work in this article describes a novel computational approach that shows promise in deciphering the principles of channel assembly in a biosynthetic pathway when relevant information is limited. It also provides a clear direction in which to proceed with more targeted experiments. Beyond the application described here, the proposed strategy might be beneficial in entirely different biological contexts, such as gene regulatory and signaling networks, where the task is to analyze how information flow is controlled by the spatial organization of molecules in the cell. Since the two metabolic channels of interest are assumed to affect only the relative amounts of G and S lignin, the analysis is restricted to those critical steps within the lignin biosynthetic pathway system that govern the flow of material either toward G or S (Figure 2). For each possible design, we first formulate the corresponding generalized mass action (GMA) model [32], [33] in a symbolic format, where each intermediate is represented by a dependent variable and each enzymatic process by a product of power-law functions. An important reason for this choice of a modeling format is that it is mathematically sound and minimally biased, because it does not require the a priori specification of a biological mechanism [34]. The model contains either six or seven dependent variables, depending on whether coniferyl aldehyde is explicitly included, and 10 to 16 distinct power-law terms, depending on the topology in a specific design. Also, there are six independent variables, each of them representing the extractable activity of an enzyme. Each power-law term contains two different types of parameters: a non-negative rate constant γi that represents the turnover of reaction i, and kinetic orders fi, j, each of which characterizes the effect of a variable Xj on a reaction i. A kinetic order can take any real value, and the sign of its value has a directly interpretable meaning: a positive value indicates an activating effect, a negative value an inhibiting effect, and zero no effect. As a typical example, the differential equation for caffeoyl CoA, defined as X1, has two representations (cf. Figure S4B and [33] for further details): (1) The variables Vi in the first equation represent the reaction rates, or fluxes, in a generic format. In the corresponding GMA model, these fluxes are specifically modeled as products of power-law functions of dependent and independent variables Xi. Aside from X1, the fluxes contain two other dependent variables, X2 (caffeyl aldehyde) and X3 (feruloyl CoA), as well as two independent variables, Xn+1 (CCR2) and Xn+2 (CCoAOMT). In GMA representations, n typically denotes the number of dependent variables, so that n+1 and n+2 refer to the first two independent variables. At the steady state, the derivative on the left-hand side becomes zero, thereby turning the differential equation into a linear equation of fluxes (ViS; where S denotes the steady state) (2) with the following constraints (3) Notably, X3 and X2 are included in the power-low representations of V2 and V3, respectively, because they potentially modulate the two consuming fluxes of X1. Applying the rules for kinetic orders described above, we can immediately impose bounds on the values of f2,3 and f3,2 for different regulatory mechanisms (Figure 3B). For instance, modeling Mechanism 1 requires the following constraints, (4) because X3 is considered an inhibitor, so that f2,3<0, while X2 has no influence on the degradation of X1 through reaction 3 in this design, so that f3,2 = 0, which in effect eliminates the factor from the term on the far right. By convention, all independent variables have a kinetic order of 1. Determination of all parameters in a GMA model, including kinetic orders and rate constants, is required prior to most simulation tasks. While numerous methods have been developed over the years, parameter estimation is seldom straightforward as each pathway and each dataset has its own challenges [35]. For the lignin pathway in Medicago, very little information is available on exact concentrations of intermediates or fluxes through the pathway; in fact, many metabolites in vivo are below detection level with standard HPLC [36]. To address this issue of insufficient data, we sample parameter values from relatively wide, biologically realistic ranges. The procedure involves the following steps. First, we sample the steady-state fluxes ViS from a set P in an m-dimensional vector space (Rm), where m equals the number of reactions. The sampled steady-state fluxes can be thought of as possible representatives of a wild-type Medicago species, while P defines the boundaries within which the system is able to operate. Biologically, these boundaries are given by many linear equality or inequality constraints with physiological meaning, such as the reaction stoichiometry (e. g. Eq. (2) ), the ratio of S to G lignin in a mature stem internode of wild-type alfalfa, and the degree of reversibility of individual reactions (see Text S1 for further information). Mathematically, P is a bounded polyhedron (or polytope) and therefore has a concise parametric description (5) where the m-dimensional vectors ui can be identified using first principles [37]; in a different context, the vectors ui have been called “extreme pathways” [38]. Once a set of steady-state reaction rates is randomly generated, we sample kinetic orders (fi, j) from their respective ranges (Table S2), which are chosen based on the unique role of each kinetic order. Even with this information, the lack of concentration data from a wild-type Medicago species remains an issue that needs to be solved. To this end, we perform two transformations. First, we define a normalization of variables by replacing Xi with Yi≡Xi/XiS, where XiS are the unknown steady-state levels of Xi in wild type. As an example, the differential equation for caffeoyl CoA assumes the form (6) where ViS are the steady-state reaction rates sampled from P. This representation is well suited for the current analysis because the exact values of XiS become irrelevant once all the equations are set to zero, that is, at a wild-type or perturbed steady state (cf. Figure S4C). Second, after all parameters for a given GMA model instantiation are specified, we derive the corresponding S-system equations with straightforward mathematical manipulations that do not require any additional biological information [33; Chapter 3]. At the steady state, GMA and S-system models are equivalent, but they offer different advantages for further analyses. In particular, S-system differential equations, despite being intrinsically nonlinear, become linear at the steady state after a logarithmic transformation, thereby facilitating the computation of secondary steady-state features and bypassing the time-consuming numerical integration that is otherwise required for assessing nonlinear models. Given this convenient feature, we are able to obtain, in a very efficient manner, estimates of steady-state fluxes under conditions that mimic the two transgenic alfalfa lines and two M. truncatula mutant lines; we can also easily compute the S/G ratios for which we had experimental data. Down-regulation of specific lignin biosynthetic enzymes is simulated by setting the corresponding normalized independent variables Yi to values between 0 and 1 that represent the degree of down-regulation, and solving the steady-state equations. In cases where hierarchical regulation might be effective, such as in ccr2 knockout mutants, all affected Yi are given values that mirror the specific changes in activities (cf. Figure S4D). The Parallel Computing Toolbox™ in MATLAB (version R2009b, The MathWorks, Natick, MA) was used to divide the simulation job among multiple cores for speedup. Not all models behaved properly during simulation, and some ill-behaved models were excluded from further analysis. These were defined, arbitrarily, as models that showed a more than 1000-fold increase or decrease in any dependent variable during any simulation. Further, a properly behaved parameter set was deemed valid if the following criteria were met: These criteria for success of a model instantiation were initially determined in an ad hoc fashion: We applied more lenient, qualitative criteria to the predictions of S/G ratios for ccr1 and ccr2 knockout mutants because the experimental data were only available in Medicago truncatula, but not in our model organism alfalfa (the S/G ratio of every model instantiation was set to the experimentally determined value for the sixth internode of a wild-type alfalfa plant). However, when we relaxed the criteria for screening the predictions of S/G ratios for CCoAOMT and COMT down-regulated lines to allow a percentage error as large as 25%, we obtained the exact same set of pathway designs that are consistent with the data, suggesting that the main conclusions are quite robust to the choice of thresholds. The cloning of the alfalfa CCoAOMT cDNA into the expression vector pET15b was as described previously [15]. E. coli Rosetta strains containing the constructed plasmid were cultured at 37°C until OD600 reached 0. 6–0. 7, and protein expression was then induced by adding isopropyl 1-thio β-galactopyranoside (IPTG) at a final concentration of 0. 5 mM, followed by 3 h incubation at the same temperature. Cell pellets from 25 ml induced medium were harvested and frozen at −80°C for further use. Induced cell pellets were thawed at room temperature, resuspended in 1. 2 ml of extraction-washing buffer (10 mM imidazole, 50 mM Tris-HCl pH 8. 0,500 mM NaCl, 10% glycerol and 10 mM β-mercaptoethanol), and sonicated three times for 20 s. Supernatants were recovered after centrifugation (16,000×g), and incubated at 4°C for 30 min with equilibrated Ni-NTA beads (Qiagen, Germantown, MD) under constant inversion to allow the His-tag protein to bind to the beads. The beads were washed three times with 1 ml of extraction-washing buffer, and the target protein was eluted with 250 µl of elution solution (250 mM imidazole, 50 mM Tris-HCl buffer pH 8. 0,500 mM NaCl, 10% glycerol and 10 mM β-mercaptoethanol). The concentration of the eluted target protein was determined using the BioRad protein assay (BioRad, Hercules, CA) and its purity was verified by SDS-PAGE. Caffeoyl CoA for the enzyme assays, and feruloyl CoA for the calibration curve, were synthesized as described previously [39]. Caffeyl aldehyde was synthesized as described by Chen et al. [40]. Pure recombinant CCoAOMT enzyme (100 ng) was incubated at 30°C for 20 min with 60 mM sodium phosphate buffer pH 7. 5,200 µM S-adenosyl methionine, 600 µM MgCl2 and 2 mM dithiothreitol. The substrate (caffeoyl CoA) concentration was 20,30 or 40 µM and the putative activator (caffeyl aldehyde) concentration was 0,2, 5 or 10 µM. Since caffeyl aldehyde was in dimethyl sulfoxide solution, the final concentration of dimethyl sulfoxide in the reaction was 4% and the final volume of the reaction was 50 µl. The reactions were stopped by adding 10 µl of 24% w/v trichloroacetic acid. Reaction products were analyzed by reverse-phase HPLC on a C18 column (Spherisorb 5 µ ODS2, Waters, Milford, MA) in a step gradient using 1% phosphoric acid in water as solvent A and acetonitrile as solvent B. Calibration curves were constructed with authentic standard of the product feruloyl CoA. Activity assays using lower concentrations of the substrate caffeoyl CoA (2,4, 5 and 10 µM) and the putative activator caffeyl aldehyde (0. 5,1, 2 and 4 µM) were performed using a sensitive radioactive assay method as described previously [15].

The organization of cooperating enzymes into complexes is a pervasive feature of metabolism. In particular, this phenomenon has been shown to participate in the regulation of flux through the networks of both primary and secondary metabolism in plants. It remains a challenging task to unravel the organizing principles of such “metabolic channels, ” which can be temporary or persistent, and to understand their biological function. In this article, we analyze metabolic channels in the biosynthetic pathway of lignin, a complex polymer that stiffens and fortifies secondary cell walls within woody tissues. This system is well suited because the present analysis can be based on a computational, experimentally validated model demonstrating that several enzymes are spatially associated into channels specific for the production of two lignin monomers. To characterize the functioning of these channels, we develop a novel computational approach that is capable of identifying interesting structural and regulatory features of metabolic channeling and permits the formulation of targeted and readily testable hypotheses. Since the spontaneous or controlled assembly of molecules into functional units is known to occur in many biological contexts where information flow is tightly coordinated, the proposed approach might have broad applications in the field of computational systems biology.

Abstract Introduction Results Discussion Materials and Methods

systems biology biochemistry plant science mathematics plant biology applied mathematics biology computational biology genetics and genomics

2012

Functional Analysis of Metabolic Channeling and Regulation in Lignin Biosynthesis: A Computational Approach

Even across genomes of the same species, prokaryotes exhibit remarkable flexibility in gene content. We do not know whether this flexible or “accessory” content is mostly neutral or adaptive, largely due to the lack of explicit analyses of accessory gene function. Here, across 96 diverse prokaryotic species, I show that a considerable fraction (~40%) of accessory genomes harbours beneficial metabolic functions. These functions take two forms: (1) they significantly expand the biosynthetic potential of individual strains, and (2) they help reduce strain-specific metabolic auxotrophies via intra-species metabolic exchanges. I find that the potential of both these functions increases with increasing genome flexibility. Together, these results are consistent with a significant adaptive role for prokaryotic pangenomes. Prokaryotes exhibit remarkable genome flexibility, with strains from the same species often containing dramatically different gene content [1–4]. Intraspecific differences in gene content are often characterized by a “core” genome (genes common to all strains) and “accessory” genome (genes found in a fraction of strains) [5]. While the core genome might represent a set of species-specific indispensable genes, we do not yet understand whether the accessory genome of a species is the result of neutral or adaptive evolution. Indeed, this is the subject of an ongoing debate: do the majority of prokaryotic accessory genes have negligible or positive fitness effects, i. e. are they neutral or adaptive? Recent population genetics arguments support roles for both neutral and adaptive evolution as possible factors driving accessory genome evolution [6–9]. For example, microbial species with more accessory genes also tend to have larger effective population sizes, as expected of genetic variation in a population under neutral evolution [9]. On the other hand, models in which microbial genomes evolve in large, migrating populations, suggest that acquired genes can often be beneficial, as expected under adaptive evolution [7]. However, these studies have only addressed broad aspects of microbial populations such as effective population size, migration, and the fitness effects of gene loss and gain. In response, subsequent criticisms of these studies have strongly expressed the need for more functional, gene-explicit and ecological analyses [10–11]. Here I present the first such systematic analysis of 96 phylogenetically diverse prokaryotic species, which suggests that prokaryotic accessory genomes often provide significant metabolic benefits. I chose to study metabolism as a possible explanatory factor for three reasons. (1) Metabolic genes dominate the functional content of accessory genomes [12] (S8A Fig). (2) Metabolic interactions between microbes—especially interdependencies—can often be adaptive [13–14]. For instance, microbes that obligately cross-feed, i. e. that critically depend on exchanging metabolites with each other, can grow faster than their wild-type counterparts [13]. Such a fitness benefit can also drive genomes, in many cases, to lose genes and become metabolically dependent [14,15]. If different genes are lost between different conspecific strains, this can lead to both metabolic interdependence, as well as accessory genomes (since different strains will have different metabolic repertoires) [16]. (3) Databases such as KEGG contain already-curated genomes for several fully-sequenced strains. KEGG contains high-quality gene and reaction annotations, allowing us to accurately predict the biosynthetic capabilities of each strain under different conditions [17]. In this study, I ask to what degree accessory genes can metabolically benefit conspecific strains. For this, I have used genome-scale metabolic network reconstructions of 1,339 prokaryotic strains (corresponding to 92 bacterial and 4 archaeal species) from the KEGG database over 59 distinct nutrient environments. In general, my analyses reveal two beneficial roles for the accessory metabolic content of prokaryotes. First, I find that the accessory genome of most species harbours extensive biosynthetic potential, with several accessory genes providing strains with additional nutrient utilization abilities. Second, I find that pairs of strains from the same species often display a remarkable potential for metabolic interdependence, which scales with the amount of accessory genome content. These interdependencies have the ability to alleviate strain-specific auxotrophies in a particular niche through the exchange of secreted metabolites. My results are, from a metabolic standpoint, consistent with a possible adaptive evolution of accessory genomes. To obtain a large set of species pangenomes, I first collected a list of all prokaryotic species in the KEGG GENOME database, and filtered those that had complete genomes for 5 or more conspecific strains. This gave me 1,339 genomes (96 species), which I used in all my subsequent analyses (S1 Table). To account for potential biases due to uneven phylogenetic sampling, I verified that a more restrictive choice of one species per genus did not significantly impact my results (55 species; S1 Fig). For each strain, I then extracted all annotated genes and metabolic reactions from the KEGG GENES and REACTION databases, respectively. To quantify accessory genome content for every species, I used the well-studied genome fluidity measure, φ [18]. For this, I calculated, across each pair of conspecific strains, the fraction of all genes in the pair that were unique to each strain. The average of this fraction over each species gave me its genome fluidity φ (see Methods). I constructed metabolic networks for every individual strain, where each network contained the set of reactions corresponding to the strain’s genome in KEGG. I included gap-filled reactions when curated models were available [19], though I verified that their addition did not impact my results (S2 Fig). I used these reaction networks to infer the biosynthetic capabilities of each strain under several different conditions. To define these conditions, I selected 59 different carbon sources, previously shown and commonly used to sustain the growth of diverse microbial metabolisms in laboratory experiments [20–22] (S2 Table). I associated with each carbon source a different nutrient environment or condition. In each condition, I included exactly one of the 59 carbon sources, say glucose, along with a set of 30 commonly available metabolites, which I assumed were always available (for example, water and ATP; S3 Table). To assess biosynthetic capability, I curated a list of 137 crucial biomass precursor molecules, often essential for growth (hereafter, “precursors”) from 70 experimentally verified high-quality metabolic models [23] (S4 Table). Finally, to calculate what each strain could synthesize in a particular environment, I used a popular network expansion algorithm: called scope expansion [24–25]. This algorithm determines which metabolites each strain can produce—its “scope”—given an initial seed set of already available metabolites. To start with, only those reactions whose substrates are available in the environment can be performed, and their products constitute the initial set of metabolites that can be produced. These metabolites can then be used as substrates for new reactions that can then be performed, and step by step, more metabolites can be produced. When no new reactions can be performed, the algorithm stops, giving the full set of metabolites that could be synthesized in the given environment. Such a calculation sidesteps the need for arbitrary assumptions of binary (yes/no) growth and optimality typically used in more complex metabolic modeling approaches such as flux balance analysis [26] and is well-known for its ability to infer what metabolic networks can synthesize in diverse conditions [27–28]. I first investigated the capabilities of individual strains. Specifically, I was interested in the extent to which the accessory genes in each strain expanded the set of precursors that could be synthesized. For each species, I calculated, via network expansion, the list of precursors that could be produced per strain per condition. I then counted how many unique precursors each strain could synthesize across all conditions, i. e. by the accessory genes alone. From this, I computed, for every species, its accessory metabolic capacity α, defined as the average number of precursors (per strain per condition) produced exclusively due to the accessory genome. I found that while for 18 species this quantity was zero, for the majority of prokaryotic species (81%), this number lay between 0. 1 and 15. 0 (mean 3. 1; median 2. 0). Further, α scaled positively with genome fluidity φ (Spearman' s rho = 0. 44; P value = 7 x 10−6; Fig 1). Since I observed that the accessory genome of different strains typically imparted different biosynthetic capabilities to different strains, I wondered if, in the same conditions, metabolic interactions between conspecific strains could further expand these capabilities. This could, for instance, indicate a potential dependence of an auxotrophic strain on another strain, i. e. a strain that cannot produce a crucial precursor in an environment. Such auxotrophies have been previously shown for example, in different strains of Escherichia coli co-inhabiting the human gut [29]. For each pair of conspecific strains in each condition, I calculated a metabolic dependency potential (MDP), defined as the average number of new precursors each strain has the potential to synthesize when grown as a pair versus alone. Here I assessed, in every pair, which metabolites that could be produced and secreted by one strain could subsequently allow the production of a new precursor in the other strain that it would not otherwise be able to make (i. e. was auxotrophic for). Note that this method does not count those metabolic interactions that can provide extra (functionally redundant) pathways to produce a precursor and supplement growth, and is thus more likely to represent actual or obligate dependencies. I verified that my approach can successfully predict such obligate dependencies by comparing with some well-documented intra- and inter-species pairs [13–14,30–32] (see Methods) (S3 Fig). I found that while I could not detect any dependency potential for 17 species, surprisingly, the majority of species (82%) showed an MDP per strain per condition between 0. 1 (for Bacillus thuringiensis) and 3. 3 (for Ralstonia solanacearum), with a mean 1. 7 and median 1. 4. Interestingly, the 17 species for which I could not detect any MDP matched those with zero α (the leftover Legionella pneumophila showed low MDP = 0. 5). Over all tested pairs with detected dependency potential (48%), commensal interactions were more common than mutualisms (29% versus 19%; Fig 2B). This is because, in species with detected dependencies, not all pairs show dependency potential (on average 46% conspecific pairs do). The auxotrophies relieved by these dependencies varied from those for amino acids, vitamins, carbohydrates, and organic acids, among others (Fig 2C). Strikingly, like α, MDP also scaled positively with genome fluidity φ, suggesting that greater amounts of accessory content can potentially sustain more conspecific metabolic dependencies (Spearman' s rho = 0. 56, P value = 4 x 10−9; Fig 2A). For both α and MDP, considering medians instead of means did not impact my results (for α: Spearman' s rho = 0. 38, P value = 10−4; for MDP: Spearman' s rho = 0. 51, P value = 10−7; S4 and S5 Figs). To verify that such potential conspecific dependencies are indeed ecologically realizable, I repeated my analysis, this time restricting it to those genomes, which were known to co-occur in microbial communities (29 strains across 14 species; see Methods). I found that my observed trend was still valid, namely MDP still scaled with φ, suggesting that several auxotrophies may indeed be reduced through within-species metabolic exchanges in nature (Spearman’s rho = 0. 56, P value = 0. 03; S6 Fig). Given the extent to which I detected the potential for obligate metabolic interactions between conspecific strains, I wondered whether such interactions are possibly common among prokaryotes. For this, I extended my study to analyze metabolic dependency potential between inter-specific strains (see Methods). I found that, indeed, strains from all species can metabolically depend on strains from at least one other species to alleviate potential auxotrophies across many different environments (with MDP ranging from 1. 6 to 3. 2, with mean 2. 2; S7 Fig). Interestingly, I found that these interspecific metabolic interactions often involve accessory metabolic genes as well. Taken together, my results suggest that a considerable fraction of prokaryotic accessory genomes contains potentially beneficial metabolic functions (upto 70% of accessory genes per strain across this study, with median 40%; see S8D Fig). Specifically, I found that the accessory “metabolome”: (1) expands a genome’s biosynthetic potential, possibly allowing for niche-specific adaptations [29,33–34]; and (2) reduces potential auxotrophies via obligate metabolic interactions, also explaining how conspecific strains can coexist despite high competitive potential [35–38]. The accessory genes that impart these functions are often different (median overlap 10%), suggesting that these are indeed distinct, non-redundant benefits. Moreover, apart from these additional biosynthetic abilities, the y-intercept for genome fluidity (at φ = 0. 03 for both α and MDP) provides an estimate of metabolic redundancies (such as extra pathways). My findings may additionally help explain the following observations: (1) metabolic functions are enriched in accessory genomes (median 50% in accessory versus 38% in core; S8C Fig); and (2) the variation in accessory metabolic genes exceeds the variation in genes of many other functions (metabolic variation being dominant in 81% of examined cases; S8B Fig). Previous studies have suggested that the evolution of metabolic dependencies likely occurs via adaptive gene loss [15,39] (e. g. the Black Queen hypothesis). Such a mechanism suggests that metabolic dependency evolution can often lead to reduced genome sizes, but makes no comment on genome flexibility (i. e. gene content variability). My results also indicate that metabolic dependency evolution can impact genome flexibility as well. Specifically, more flexible genomes (with more variable gene content) are more likely to display a potential for metabolic interactions. Can stochastic accessory gene turnover explain these results? To test this, I repeated my study with randomly assembled pangenomes. Within each species, I retained the core genes in every strain and shuffled the accessory genes between strains (see Methods). During this randomization, I preserved the observed within-species genome size distribution, strain number distribution, and ensured that any change in species’ genome fluidity was insignificant. I found that not only did this significantly diminish the metabolic benefits observed in each species, both measurements of α and MDP yielded non-significant correlations, suggesting that the mere presence of additional accessory genes is unlikely to explain my observed trends (for α: Spearman' s rho = 0. 01, P value = 0. 9; for MDP: Spearman' s rho = 0. 17, P value = 0. 1; S9A and S10A Figs). The measured benefits remained lower than observed, even when I shuffled known operons of genes together instead of shuffling genes one by one (S9B and S10B Figs; see Methods). I believe this is because often, prokaryotic operons do not contain complete metabolic pathways, but instead parts of them (in my data set, each metabolic operon encoded 1. 5 reactions on average, while pathways typically had 4 to 5 steps). Collectively, this suggests that accessory gene acquisition is consistent with the coordinated gain of functional and beneficial pathways, which I believe provides further support for the accessory genes being maintained for adaptive reasons. To summarize, here I addressed the debate on whether the accessory genomes of prokaryotes are beneficial. I found that, indeed, large fractions (about 40%) of the prokaryotic accessory gene pool can contribute to metabolic benefits. Specifically, such genes can allow microbes to produce a larger repertoire of crucial molecules, and facilitate the exchange of others. Since these functions can improve growth in many habitats, my results suggest that adaptation may explain accessory gene maintenance. Note, however that my analyses are only capable of detecting obligate metabolic dependencies and biosynthetic potential, and do not consider signaling, regulation, metabolic redundancy, etc. that could also play important functional roles and might indicate potential benefits due to additional accessory genes. Further work might also explain accessory genomes in those species, where I could not detect additional metabolic functions, if such roles are indeed there. Moreover, even in the context of metabolism, more detailed metabolic models, when available, may be used to probe even more precise fitness effects of intraspecies metabolic variability, including the effect of higher-order interactions. However, these studies would require knowledge of a large number of parameters such as reaction kinetics, thermodynamics, and exact strain biomass compositions before they are feasible. Finally, systematic measurements of the fitness effects of all accessory genes, metabolic and otherwise, are needed for more complete estimates of the fraction of accessory genomes consistent with adaptive versus neutral evolution. I used the KEGG GENOMES database [17] to extract a list of all prokaryotic species with complete genomes for 5 or more strains. This yielded a list of 1,339 strains or genomes corresponding to 96 species (92 bacteria, 4 archaea), which I used for all subsequent analyses (see S1 Table for the full list of species and strains, along with their taxonomic classification). For each strain with a unique genome abbreviation, I extracted the full set of annotated genes under the KEGG GENES database and reactions under the KEGG REACTION database using an in-house Python script. I also extracted the full list of reactions with their stoichiometries and participating metabolites in the database. The metabolic reaction network for each strain was considered to be the complete set of annotated reactions detected in that strain’s genome in KEGG. Note that my analyses systematically ignore genes without known functions. For strains for which genome-scale metabolic reconstructions were available in the Model SEED database [19], I also included gap-filled reactions. Specifically, I extracted the list of all gap-filled reactions for 130 genomes from table S3 in ref. 19. I mapped all genomes from this table to KEGG genomes by matching strain names, and all reaction IDs to KEGG reactions by searching the Model SEED database online (https: //modelseed. org/biochem/reactions) using a custom Python script. This resulted in a total of 562 gap-filled reactions, spread across 22 genomes (20 out of 96 species; S6 Table). I then added these reactions to the metabolic networks already constructed via KEGG. Separately, I verified that adding these gap-filled reactions did not impact my results (S2 Fig). For nutrient environments or conditions, I selected a set of 59 diverse carbon sources known to sustain microbial biomass and energy synthesis from previous genome-scale metabolic studies of phylogenetically broad species [20–22] (S2 Table). Every condition was assumed to contain one of these carbon sources (such as glucose and maltose), along with a set of 30 commonly available metabolites (assumed to be present in all conditions, such as water, oxygen and ATP), similar to the aforementioned studies (S3 Table). To infer biosynthetic potential, I separately collected a set of all prokaryotic species biomass compositions and their constituent metabolites from high-quality experimentally verified metabolic models in the BiGG database [23]. I curated from this a list a union of 137 biomass precursors across diverse microbial metabolisms (S4 Table). To infer what each strain could synthesize in each nutrient environment or condition, I used a well-documented network expansion algorithm—scope expansion [24–25]. Briefly, this algorithm is given a reaction network (one from each genome) and an initial “seed” set of available metabolites (each nutrient environment). It first determines which reactions can be performed by the network using only the nutrients in the environment. I assume that metabolites that are products in this initial set of reactions can be synthesized by the network, and can be subsequently used as reactants in new reactions. Again, I consider that the products of such new reactions can be synthesized by the network, and may allow additional new reactions to be performed. This continues step by step, till no no new reactions can be performed. All metabolites that can be produced over all such steps are defined as the “scope” of the metabolic network, i. e. I assume that these metabolites can be synthesized by the reaction network from the initial nutrients in the environment. I calculated genome fluidity φ as prescribed in a previous study [17], using a custom Python script. For every genome, I considered each constituent gene’s KO number as its unique identifier. Then, to estimate φ for every species, I calculated, for all conspecific pairs, the ratio of the number of genes unique to a strain in the pair to the total number of genes in their sum. The average over all pairs for a species was considered its genome fluidity φ. Note that though using KEGG orthologous groups underestimates the exact values of φ, my estimates still scale well with previously reported values [9] (Spearman' s rho = 0. 60, P value = 7 x 10−5; S11 Fig). I calculated an accessory metabolic capacity α for every species. For each conspecific strain, I first calculated, using the network expansion algorithm described, the scope of each of the 1,339 reaction networks across all 59 conditions. Then, species by species, for every condition, I calculated a “core” metabolome, i. e. metabolites that were present in the scope of every conspecific strain. I then explicitly removed these metabolites within every species from the scope of each strain and counted how many precursors remained in the corresponding “accessory” metabolome of every strain across all conditions. This gave me a number of additional precursors that could be synthesized per strain per condition for each species, and was defined as the species’ accessory metabolic capacity α (S5 Table). I calculated a metabolic dependency potential (MDP) for every species. For this, I considered within each species, all conspecific pairs across all 59 conditions. For each pair, I calculated the scope for both strains first in “monoculture”, i. e. when grown alone. I then calculated, for every metabolite that could be produced by one strain but not the partner strain, whether or not its secretion could alleviate an auxotrophy in the partner. I specifically considered auxotrophies only for the 137 key precursors I had selected. I then counted each alleviated auxotrophy as a potential metabolic dependency, and the average number of dependencies (per strain per pair per condition) for each species was defined as its metabolic dependency potential, or MDP (S5 Table). To quantify the extent of metabolic interactions between inter-specific strains, I calculated a separate metabolic dependency potential for every species. For each species, I paired each conspecific strain with 25 randomly chosen strains from other species, also picked at random. For all inter-specific pairs generated this way, I calculated metabolic dependency potential using the same method as described above, for conspecific pairs. In this way, the average number of dependencies identified per strain per condition was defined as the inter-specific metabolic dependency potential, or MDP (S5 Table). To test whether the conspecific metabolic interactions detected in my MDP analysis could be realized in natural microbial communities, I analyzed genome co-occurrence data from Chaffron et al [40]. These data list all 16S rRNA sequences co-detected across several microbial community samples. Here, sets of sequences are clustered into operational taxonomic units (OTUs) corresponding to different sequence similarity thresholds. To map these OTUs to the genomes in my study, I first obtained 16S rRNA sequences for all the 1,339 genomes I analyzed from KEGG. When multiple sequences were available for a given genome, I used the longest sequence and maped that as the unique 16S identifier for that strain. Then, using BLAST, I mapped OTUs in the co-occurrence data to the genomes in my study (where OTUs were binned with a sequence similarity threshold of 99%). Here, I used the BLAST bit score as my assignment criterion. I used the 689 genomes that could be mapped this way for further analysis. Here, across all microbial community samples, I asked which conspecific genomes co-occurred in at least one sample—from which I found 29 genomes corresponding to 14 species (S7 Table). I then repeated my metabolic dependency potential analysis for these conspecific strains, as described above. To test if my observed correlations between φ and α as well as φ and MDP could be explained by random accessory gene turnover, I repeated my study with a “randomly assembled” pangenome dataset. I randomized genomes species by species. I first collected all available genomes for a species, and picked a random pair of these. I then shuffled the accessory genes in this pair in two ways: (1) gene by gene, and (2) operon by operon. When shuffling gene by gene, for each strain pair, I randomly picked two genes, one from each strain in the pair, and swapped them. I repeated this several times before picking another conspecific pair from the same species. The number of swaps per pair was chosen such that each accessory gene was swapped once on average. I verified that the exact number of swaps does not affect my results. By the end of this process, I had a new set of genomes which had undergone “stochastic accessory gene turnover”. Note that in order to avoid any potential biases, this process preserves the observed genome sizes and strain numbers while only slightly affecting genome fluidity. I then repeated my α and MDP calculations for these “shuffled” genomes. This would test if the mere acquisition of extra genes from a species’ accessory genome could allow the expanded biosynthetic potential and metabolic dependencies observed in the data. To identify operons, I used the ProOpDB database, which lists operon compositions for more than 1200 prokaryotic genomes [41]. I found that this database had operon compositions available for 795 strains across 64 of the species in my study, which I used for the operon shuffling analysis (S8 Table). Here, when shuffling operons, I used a similar method as when shuffling genes, but instead of swapping merely randomly chosen genes from a pair of strains, I identified which operon they belonged to in their respective strain’s genome, and swapped all genes in those operons across the pair. I repeated these operon swaps several times for each strain pair, and for several pairs, at the end of which, I had another new set of randomly shuffled genomes. To test if my metabolic dependency potential (MDP) measure could accurately predict metabolic dependencies between different pairs, I compared its performance on genome-scale metabolic networks corresponding to some well-studied experimentally verified metabolically dependent pairs. Specifically, I considered 2 conspecific and 4 inter-specific pairs. For conspecific dependencies, I used 2 Escherichia coli cross-feeding pairs [13] and for inter-species, I used (1) a Desulfovibrio vulgaris and Methanococcus maripaludis pair [30]; (2) an E. coli and Acinetobacter baylyi pair [14]; (3) an Lactobacillus bulgaris and Streptococcus thermophilus pair [31]; and (4) a Bifidobacterium longum and Eubacterium rectale pair [32]. In all cases, I obtained the metabolic models for the closest available strains from KEGG and, when needed, modified the genes present to best match those described in the respective studies. I then used my MDP approach as described to infer which potential dependencies were detected in each pair for the specific conditions mentioned in each study. I found that my method could accurately identify the extent of dependencies (number of auxotrophies) and interaction directionality (commensal or mutualistic interactions; S3 Fig). To calculate correlation coefficients throughout the study, I used Spearman’s nonparametric rho, and for P values, I used a one-way asymptotic permutation test for positive correlation. All statistical tests were performed using standard implementations in the SciPy (version 0. 18. 1) and NumPy (version 1. 13. 1) libraries in the Python programming language (version 3. 5. 2). Linear regression and prediction interval calculations were performed using the Seaborn library function regplot (version 0. 7. 1). All computer code and extracted data files used in this study are available at the following URL: https: //github. com/eltanin4/pangenome_dep.

Recent and rapid advancements in genome sequencing technologies have revealed key insights into the world of bacteria and archaea. One puzzling aspect uncovered by these studies is the following: genomes of the same species can often look very different. Specifically, some “core” genes are maintained across all intraspecies genomes, but many “accessory” genes differ between strains. A major ongoing debate thus asks: do most of these accessory genes provide a benefit to different strains, and if so, in what form? In this study, I suggest that the answer is “yes, through metabolic interactions”. I show that many accessory genes provide significant metabolic advantages to different strains in different conditions. I achieve this by explicitly conducting a large-scale systematic analysis of 1,339 genomes across 96 diverse species of bacteria and archaea. A surprising prediction of this study that in many ecological niches, co-occurring strains of the same species may help each other survive by exchanging metabolites exclusively produced by these different accessory genes. More pronounced gene differences lead to more underlying metabolic advantages.

Abstract Introduction Results and discussion Methods

functional genomics genome evolution statistics metabolic networks operons regression analysis genomic databases mathematics metabolites network analysis genome analysis dna research and analysis methods computer and information sciences prokaryotic cells mathematical and statistical techniques biological databases molecular evolution statistical methods biochemistry cell biology nucleic acids database and informatics methods linear regression analysis genetics biology and life sciences cellular types physical sciences genomics evolutionary biology metabolism computational biology

2018

Metabolic adaptations underlying genome flexibility in prokaryotes

Ectotherms rely for their body heat on surrounding temperatures. A key question in biology is why most ectotherms mature at a larger size at lower temperatures, a phenomenon known as the temperature–size rule. Since temperature affects virtually all processes in a living organism, current theories to explain this phenomenon are diverse and complex and assert often from opposing assumptions. Although widely studied, the molecular genetic control of the temperature–size rule is unknown. We found that the Caenorhabditis elegans wild-type N2 complied with the temperature–size rule, whereas wild-type CB4856 defied it. Using a candidate gene approach based on an N2 × CB4856 recombinant inbred panel in combination with mutant analysis, complementation, and transgenic studies, we show that a single nucleotide polymorphism in tra-3 leads to mutation F96L in the encoded calpain-like protease. This mutation attenuates the ability of CB4856 to grow larger at low temperature. Homology modelling predicts that F96L reduces TRA-3 activity by destabilizing the DII-A domain. The data show that size adaptation of ectotherms to temperature changes may be less complex than previously thought because a subtle wild-type polymorphism modulates the temperature responsiveness of body size. These findings provide a novel step toward the molecular understanding of the temperature–size rule, which has puzzled biologists for decades. For many decades biologists have been intrigued by the relation between body size and temperature. It was discovered that ectotherms—animals that maintain their body temperature by absorbing heat from the surrounding environment such as fish and all invertebrates—reproduce later at a larger size when reared at lower temperatures [1–3]. This phenomenon is known as the temperature–size rule, and nearly 90% of ectothermic species studied so far follow this rule [4]. The magnitude of this phenomenon is illustrated by Azevedo et al. [5] who found a 12% increase in wing and thorax size in Drosophila melanogaster when grown at relatively low temperatures. In the case of the nematode C. elegans (strain Bristol N2), an environmental temperature of 10 °C resulted in adults that were ~33% larger than those grown at 25 °C [6]. About 99. 9% of all species are ectothermic, and the temperature–size rule is observed in bacteria, protists, plants, and animals, making it one of the most widespread phenomena in ecology. From the perspective of life-history evolution it is not well understood why growing bigger at lower temperatures is beneficial for organisms. Because this thermal plasticity of body size is taxonomically widespread, the reasons are probably diverse and may vary among groups of organisms. It has been suggested that a large body size is advantageous, because it compensates for delayed reproduction by yielding more offspring [7]. Other explanations may be that a larger body size at maturity enables individuals to produce larger offspring or to provide better parental care [2]. Since body size and temperature are the two most important variables affecting fitness [8,9], many experimental and theoretical attempts have been made to explain the mechanism underlying the temperature–size rule. Essentially, an increase in body size can be achieved by increasing cell number, cell size, or by both. Various studies point at the second (cell size) and the third option (cell size and number) as being the most likely explanation for the observed increase in body size at lower temperatures (Drosophila spp. [10–12], yellow dung fly [13], and the nematode C. elegans [6]). Next to these empirical observations, various models have been proposed that are based on a combination of changes in cell size and number. Biophysical models show that the temperature–size rule is the result of unequal effects of temperature on cell growth and cell division [14]. When the effect of temperature on the rate of division is greater than its effect on the rate of cell growth, the model predicts that a low temperature should lead to a larger body size. Recently, a physiological model was proposed by Atkinson et al. [15], which assumes that temperature induced changes in cell size and number depend on the optimisation of oxygen supply at different temperatures. Yet, these empirical and theoretical findings give little insight into the molecular genetic control of body size at lower temperatures. Unravelling the molecular mechanism underlying the temperature–size rule is hampered by the fact that temperature affects nearly all biochemical processes in a cell, and in theory growing bigger at lower temperatures may have numerous causes. However, low temperatures also have been shown to induce a number of specific physiological and genetic responses in ectotherms [16]. In D. melanogaster gene expression analysis revealed a senescence marker smp-30 to be induced by low temperature [17]. Van ‘t Land et al. [18] reported the association of the gene Hsr-omega with low temperatures in D. melanogaster. Next to these specific gene responses, an early indicator of low temperature is a transient elevation of the cytosolic calcium concentration [Ca2+]i. Higher cytosolic calcium levels occur not only in response to a rapid cooling but also to more gradual reductions in temperature, and it is a widespread phenomenon observed in plants [19,20] and ectothermic animals [21–24]. Here we aimed to identify and characterize genes underlying the temperature–size rule in a model ectotherm, the nematode C. elegans. C. elegans is a suitable model for studying the molecular control of temperature–body size responses because of its completely sequenced genome, isomorphic growth, and cell constancy, and because nematode life-history traits are easy to observe [25]. We found that wild-type Bristol N2 (designated as N2) grew bigger at lower temperatures and thus complied with the temperature–size rule, whereas wild-type CB4856 (designated as CB) defied the rule. The natural variation in body size response to temperature between CB and N2 was caused by a single mutation F96L in a calpain-like protease TRA-3 encoded by tra-3. Homology modelling predicts that F96L is likely to reduce the ability of TRA-3 to bind calcium. We studied the thermal reaction norm for body size (TRB), which is a plot of body size at maturity versus temperature, and defined compliance with the temperature–size rule if body size is significantly and negatively related to temperature. To assess differences in the TRB between the two wild-type strains we measured body size at 12 °C and 24 °C. Body-size measurements were taken from Gutteling et al. [26]. We found a marked difference in TRB between the two wild types. The body size of wild-type N2 exhibited a significant negative relationship with temperature, i. e. , N2 grew larger at low temperature (F = 3. 49; p = 0. 02). In contrast, CB defied the temperature–size rule because body size was not significantly affected by temperature (F = 0. 8; p = 0. 47) (Figure 1). The results for N2 are in agreement with previous findings where increased body size was found in C. elegans N2 hermaphrodites as well as males at lower temperatures [27,6]. To further study the genetic control of the TRB, we first developed an N2 × CB recombinant inbred panel and performed a quantitative trait locus (QTL) analysis for detecting genomic regions associated with the TRB. By selfing the CB × N2 F1 offspring for 20 generations, we obtained a segregating population of recombinant inbred lines (RILs), which were also exposed to 12 °C and 24 °C. We found large differences in TRB slopes among the RILs (Figure 1). As generally observed in recombinant inbred crosses between divergent strains, the mean trait values for many of the RILs exceeded the mean value for either parental strain. Apparently the differences between the N2 and CB phenotypes (the slope of the TRB) capture a great deal of genetic variation. This was evident in the variation exhibited in the RILs for the TRB slope. Such transgressive segregation has been reported for many organisms and indicates that alleles at different loci act in the same direction, and when combined these alleles will result in phenotypes more extreme than either parent [28]. In general RILs matured at 12 °C at a bigger size than at 24 °C, which is in accordance with the temperature–size rule (see Atkinson [4] for an overview). We found strong genetic variation among RILs for body size across the two temperatures (F = 40. 1; p < 0. 001). We then sought to determine which loci were associated with the TRB by genotyping the RILs and performing a QTL mapping study using the recombinant inbred panel. For the QTL analysis we used a dense single nucleotide polymorphism (SNP) map. A full description of the genetic architecture of the RILs can be found in [29]. In summary, the overall average distance between two SNP markers was 835 kb or 2. 38 cM. The overall average chromosomal coverage was 96% if measured in bp or 95% if measured in cM. Compared to the Wormbase F2-derived genetic maps (http: //www. wormbase. org, release WS106), the genetic maps showed on average an ample 2-fold expansion. This is common for RILs bred by self-fertilization or sib-mating and can be explained by the multiple rounds of meiosis undergone [30]. Figure 2 shows the detected QTLs associated with the slope of the TRB. Two QTLs on Chromosome IV were associated with a negative effect on the slope of the TRB and were linked to CB alleles. The distal QTL at Chromosome IV showed pleiotropy or linkage for body size at 24 °C (additive effect of 4%). We aimed to identify the gene (s) controlling the QTL at Chromosome IV with a peak at marker pkP4095 at 12 cM, because this QTL was uniquely associated with TRB (hence we named it the TRB-locus) and not with body size itself at 12 °C or 24 °C. This locus had a relatively large additive effect of 34% of the total standard deviation and explained 11% of the among-RIL variance. Introgression of a CB segment spanning the TRB-locus into an N2 background confirmed the QTL analysis. Phenotyping of NIL WN17–9 carrying an ~6-cM region of the TRB locus revealed no significant body-size difference between low and high temperature (Figure 3). Three other QTLs on Chromosome III increased the slope and each of these QTLs was linked to N2 alleles and showed a pleiotropic or close linkage effect for body size at 12 °C [26]. The 2. 5-cM genome segment covered by the confidence interval (CI) of the TRB locus harbours a number of mutationally mapped genes of which only one (dpy-4) [31] is known to affect body size. To identify promising candidate genes, we reasoned as follows. Previous studies have shown that body size in C. elegans is controlled by genes that affect cell size and not cell number [32,33]. It is also known that this is one of the main mechanisms, next to cell number, by which ectotherms grow bigger in colder environments [7]. Furthermore, we sought to identify and characterize genes that encode a calcium-activated protein because [Ca2+]i is a key signal of low temperature. Lower temperatures lead to an increase of [Ca2+]i [21–24]. Given these two facts (increased [Ca2+]i and cell size) we searched for genes that are activated by [Ca2+]i and that play a role in increased cell size. Among the few genes with known function in the TRB locus, the most likely candidate gene was tra-3. TRA-3 has a high homology with mammalian calpains [34], which are known to be activated by [Ca2+]i and have been reported to regulate cell size during oncosis (cell swelling) [35]. dpy-4 is not known to be activated by [Ca2+]i [31]. We therefore selected tra-3 as a candidate gene that might explain the difference in temperature responsiveness between N2 and CB. The gene tra-3 seems to be important for the TRB slope because a significant difference was found between tra-3 allelic variants (using the linked marker pkP4095) and the TRB slope (t-test, p = 0. 03). RILs with the N2 allele had a larger slope than RILs with a CB allele. To investigate the hypothesis that tra-3 controlled the TRB, we first sequenced this region in CB. One SNP was found within the coding region where phenylanaline-96 in N2 was mutated into leucine-96 in CB. To see whether other tra-3 mutants displayed the same phenotype as observed in CB, we selected two homozygous artificial allelic mutants in an N2 background, tra-3 (e1107) carrying a nonsense mutation [34] and tra-3 (e2333). We also sequenced tra-3 (e2333) in the ORF ± 1 kb and found a nonsense mutation at nucleotide position 1,779 (G to A) of the spliced tra-3 transcript. This resulted in a premature stop (W to stop) at position 593 of the TRA-3 protein. Both mutants were phenotyped for body size at 12 °C and 24 °C and compared to the wild-type N2. Like CB, body size was not affected by temperature in both mutants (Figure 4). The N2 phenotype was rescued by the fully suppressed mutant tra-3 (e1107) sup-24 (st354) IV, which promotes translational readthrough of the tra-3 (e1107) mutation (Figure 4). We then tested whether a larger body size could also be obtained by mimicking a low temperature environment through an artificial increase of [Ca2+]i at 24 °C. Although TRA-3 does not have a specific EF calcium-binding site in C. elegans, a well-conserved region has been shown to bind calcium [36,37]. We used thapsigargin (TG) to increase [Ca2+]i [38,39]. We found a clear dose–response relationship between TG and body size, showing that N2 grew larger at 24 °C at increased levels of TG (Figure 4). A significant increase in size was found at 0. 015 μM TG compared to a positive control that included the solvent dimethyl sulfoxide (DMSO). Calpain activity was required for the TG-induced body-size enlargement because treatment with 0. 015 μM TG did not result in a larger body size in homozygous tra-3 (1107) mutants (Figure 4). These results indicate that calcium activation of TRA-3 may be controlling body size at different temperatures. In addition to the F96L mutation, the observed phenotypic differences could be due to differential expression of tra-3. Therefore, we performed quantitative RT-PCR experiments on cDNA obtained from N2 and CB at 12 °C and 24 °C. It was found that expression was slightly enhanced at 24 °C in both wild types. There was no significant difference in tra-3 expression across temperatures between N2 and CB (results not shown). Based on these findings we hypothesised that observed TRB differences between N2 and CB were the result of a polymorphism in tra-3. To further investigate the role of tra-3 in the wild-type CB, we performed a complementation analysis by crossing the near-isogenic line (NIL WN17–9) with tra-3 (e1107). Heterozygous F1 from a cross between NIL WN17–9 and N2 revealed the recessive nature of the CB–TRB allele (Figure 3). The body size for the e (1107) /+ F1 offspring exhibited increased size at 12 °C indicating that tra-3 (e1107) was recessive (Figure 3). Complementation analysis in which NIL WN17–9 was crossed with tra-3 (1107) showed no differences in body size of F1 between high and low temperature (Figure 3). These results show that tra-3 is required for regulating body size in response to changing environmental temperatures and that an SNP in tra-3 is able to reduce this ability. We did not attempt to perform a complementation test between NIL WN17–9 and tra-3 (e2333) because of the dominant nature of tra-3 (e2333) over other tra-3 mutants. Homozygous tra-3 (e1107) worms show partial masculinisation whereas homozygous tra-3 (e2333) animals are wild-type hermaphrodites. Heteroalleles of these two mutants are also wild-type hermaphrodites indicating a dominance effect of e2333 over e1107 [40]. We next asked whether the N2 version of the tra-3 gene could transform CB to have a larger body size at low temperature. Therefore we carried out a transgenic assay in which tra-3 from N2 was transferred to the CB background. We exposed independently derived strains of CB (gfp) (control strains) and CB (gfp and tra-3 (+) ) to 12 °C and 24 °C. Figure 5 shows that the N2 phenotype was rescued in CB (gfp and tra-3 (+) ) because it grew 24% larger at the low temperature. CB (gfp) retained the CB phenotype because it did not grow larger at low temperature. We next sought to determine whether F96L could lead to a diminished activity of TRA-3 in CB by conducting homology modelling of the 3D structure of TRA-3. The TRA-3 protein consists of four domains (I–III and T), where domain II is the protease catalytic site, and domain T does not have a critical calcium-binding function [34,41] but may be important for protein folding. Although TRA-3 does not have a specific EF calcium-binding site in C. elegans, a well-conserved region spanning the boundaries of domain II and III has been shown to bind calcium [36]. In addition, Moldoveanu et al. [37] reported on non-EF calcium-binding sites in domain II between position 62–74 (Ca-1). Homology modelling shows that F96 is located at the beginning of a short helix, H6, contiguous in space to the loop hosting Ca-1. In “open” configuration, corresponding to the absence of calcium, the distance between the α-carbons of F96 and E68, G69, and A70 reduces to 8–10 Å, as compared to ~10–14 Å corresponding to the “closed” configuration. In addition, the side chain of F96 is oriented toward the Ca-1 loop making their atoms to come frequently in van der Waals contact (<3. 0 Å) (Figure 6). As the length of a leucine side chain is ~1. 5 Å smaller than that of a phenylalanine, F96L will introduce a void in this region. Therefore, F96L can make a small but important difference by increasing the conformational space that the “opened” Ca-1 loop can sample during its dynamics. As the number of configurations increases this might reduce the probability to find the loop in its “closed” configuration and consequently reduce the ability for calcium binding. The genetic control of the C. elegans body size has been intensively studied. Mutants such as sma-2,3, 4, and daf-4 have a small body size and are defective in the TGF-β signalling pathway, which underlies body growth and development [42]. The lon mutants have been found to grow longer but not larger in volume [32,43]. It was shown that egl-4 mutants, defective in a gene encoding a cGMP-dependent protein kinase, have a much larger body size than N2 [32]. Here it is shown that TRA-3 has a prominent role in regulating the thermal plasticity of body size in C. elegans. Homology modelling shows that the F96L mutation in CB4856 attenuates the ability to grow bigger at lower temperatures by destabilizing the calcium-binding site in TRA-3. These data indicate that calcium signalling in response to temperature changes may lead to the activation of TRA-3. This mechanism to control the temperature–size rule is supported by various reports on the elevation of the free cytosolic calcium concentration in response to lower temperatures. Increase of cytosolic calcium levels in response to a gradual reduction of temperature is widely observed in plants [19,20] and ectothermic animals. Many studies in other organisms have shown the importance of calpains in oncosis showing calpain-mediated cell swelling and disruption of plasma membrane permeability followed by cell death [35]. In C. elegans calcium-activated TRA-3 is known to be involved in the sex determination pathway by activating TRA-2A, a membrane protein that indirectly activates the zinc finger transcriptional regulator TRA-1A by binding and inhibiting a masculinising protein FEM-3 [44]. Current insights are insufficient to link aforementioned findings and to infer a putative pathway by which calcium activation of TRA-3 results in larger cell sizes in C. elegans. Many different theories have been proposed to unravel the underlying mechanism of the temperature–size rule [2,7, 45,46]. Our results partly fit the theory by Van der Have et al. [14] who suggested that the temperature–size rule is regulated by two distinct processes underlying temperature effects on body size: growth rate (which is the biomass increase per time unit) and differentiation rate (which is the reciprocal of development time). Their model presupposes that the temperature–size rule depends on a wide range of alleles differing in sensitivity to temperature. Our results show that a polymorphism in a single gene may attenuate the TRB in C. elegans. CB was originally isolated in Hawaii while N2 originates from the UK. Whether the F96L mutation in CB reflects adaptive change or a fortuitous event is unknown. Both parental strains have been isolated decades ago and kept in the laboratory ever since, and additional field research is needed to establish whether this polymorphism and/or others in tra-3 are typical for strains isolated from tropical regions. Our results do not provide insight into how natural selection modifies the temperature–size rule, yet they provide the basis for a more mechanistic understanding of the evolutionary outcomes. Like C. elegans the increased body size at lower temperatures in flatworms, Drosophila spp. , and protists [47–49] is caused primarily by increased cell size. Because tra-3 shows a high homology with other ectothermic calpains [34,37], our findings may imply a possible role of calpain in the control of the temperature–size rule in other organisms as well. We have presented genetic and structural evidence that an SNP in the gene tra-3 encoding a calpain-like protease is required for the regulation of the temperature–size rule in wild-type C. elegans. First, we found that the wild-type N2 complied with the temperature–size rule, whereas wild-type CB4856 defied it, and demonstrated that the genetic variation in the temperature–size response mapped to a single QTL on Chromosome IV harbouring tra-3. Second, we showed similar expression levels in tra-3 between the two wild types. Third, transgenic CB carrying an N2 allele of tra-3 complied with the temperature–size rule. Fourth, we found that F96L in TRA-3 attenuates the ability of wild-type CB4856 to grow larger at low temperatures. Finally, we showed that, based on homology modelling, the CB4856 mutation decreased the calcium-binding activity of TRA-3 rendering it less active. Because TRA-3 shows a high homology with other ectothermic calpains, our findings imply a possible role of tra-3 in the control of the temperature–size rule in other organisms as well. Together our data show that the response of a quantitative trait to temperature changes can be simple and far less complex than previously thought. Both N2 and CB parental strains were homozygous. Strains were grown in 9-cm petri dishes at 15 °C or 20 °C on standard nematode growth medium with Escherichia coli strain OP50 as food source [50] and transferred to new dishes by a chunk of agar once a week. RILs were constructed according to [29]. NIL WN17–9 was constructed by crossing a single L4 hermaphrodite of RIL WN17 with five males generated from N2 on a 6-cm petri dish. The proportion of males in the offspring was approximately 0. 5 indicating a successful cross. Subsequently, 12 crosses were set up, each with the use of single L4 hermaphroditic offspring of the former cross and five males derived from N2. Backcrossing procedure was continued with two L4 hermaphroditic offspring per successful cross. After described three generations backcrossing, six L4 hermaphrodites were picked from each successful cross and placed individually on a 3-cm petri dish to self. Selfing was continued for ten generations for each of the lines. All derived lines were subsequently genotyped at seven marker positions (including marker pkP4095) distributed equally over the fragments that were identified in RIL WN17 to be of CB origin. A total of five lines that appeared to have N2 alleles in all genotyped positions except for the marker pkP4095 were used for detailed genotyping. These lines were genotyped at all remaining marker positions. The results for one of the genotyped lines (NIL WN17–9) showed at all genotyped markers N2 alleles except for CB allele at marker pkP4095, and three neighbouring marker positions at Chromosome IV indicating a single ~6-cM DNA fragment of CB origin introgressed into N2 background. Genotyping was according to Li et al. [29]. Prior to an experiment, all lines (80 RILs and two parental) were synchronised at room temperature by transferring five adult nematodes to fresh 6-cm petri dishes and allowing them to lay eggs for 3–4 h, after which the nematodes were removed. Eggs were allowed to develop at 20 °C, and three days later synchronisation was repeated to get double-synchronised lines. Measurements for parental and RIL body size at maturation were taken from Gutteling et al. [26]. Maturation was defined as the moment that the first few eggs are laid and can be easily observed. Because of this, body size at maturity can be precisely measured. For each RIL, three replicate experiments were performed using double-synchronised lines as a start. In each replicate, four adult nematodes per RIL were transferred to a fresh 6-cm dish, allowed to reproduce at room temperature for 2–4 h (average 2. 5 h), and removed. Dishes were then stored at 12 °C and 24 °C climate chambers (Elbanton, http: //www. elbanton. nl). Temperature was recorded with Tinytag Transit temperature loggers (Gemini Data Loggers, http: //www. geminidataloggers. com). After 1 d (24 °C) or 4 d (12 °C), 12 juvenile nematodes were transferred at room temperature to individual dishes (3 cm diameter). Dishes were randomised and put back at the appropriate temperature. After 38 h (24 °C) or 145 h (12 °C) dishes were scanned at room temperature at regular intervals (1. 5 h for 24 °C and 4 h for 12 °C) for the presence of eggs. If one or more eggs were observed, time and number of eggs were noted and the dish was put at −20 °C to prevent further development; a pilot study (unpublished data) showed that freezing did not affect body size. Dishes were defrosted and nematodes were transferred to new dishes with NGM-agar. Digital pictures were taken with a CoolSnap camera (Roper Scientific Photometrics, http: //www. photomet. com). Each nematode was measured automatically with Image Pro Express 4. 0 (Media Cybernetics, http: //www. mediacy. com). Using a measurement ocular we calibrated 10. 000 pixels3 as 753. 516 μm3. We assumed a rod-like shape of a worm where volume Vszm = π· (D/2) 2·L = (1/4) ·π·A2/L where D is diameter, L is length, and A = L·D. Because perimeter P = 2L + 2D ~ 2L we get: Area (A, pixels2) and perimeter (P, pixels) of each worm were measured digitally. In subsequent analyses Vszm was used as input value for body size [26]. We studied the TRB, which is a plot of body size versus temperature, and used the slope of the reaction norm as a mapping trait. For mutant phenotyping the following strains were used for body-size measurements at 24 °C and 12 °C: wild-type Bristol N2 and CB4856 isolate, tra-3 (e2333), tra-3 (e1107) /dpy-4 (e116) IV, and tra-3 (e1107) sup-24 (st354) IV. The tra-3 (e1107) /dpy-4 (e116) IV heterozygotes segregate dpy-4 homozygotes, heterozygotes, and tra-3 (e1107) homozygote hermaphrodites, which due to maternal effects are phenotypically wild type and segregate pseudomales [51]. We measured body size in these homozygote pseudomales, as well as the homozygote and heterozygote hermaphrodites. Body size was larger only in the hermaphrodites at 12 °C. Body size in the pseudomales was measured after the characteristic male tail [52] was completely formed. Experiments were performed on agar dishes (3 cm diameter) as described above. Samples were not frozen, but body size was measured directly when worms started laying eggs. Crosses with the mutants and NIL WN17–9 were conducted by transferring J2 stage worms on small agar dishes (3 cm diameter) with three to five males. The worms were allowed to mate at 24 °C after which the females were transferred to new plates thus allowing them to lay eggs for 3–4 h. Mating was considered to be successful if the ratio of males: hermaphrodites was approximately 1: 1 in the F1 hybrids. After this period females were removed and eggs allowed to develop at subsequent high or low temperature. When reproduction started body size was measured as described above. TG (Sigma, http: //www. sigmaaldrich. com) was applied to agar plates dissolved in DMSO. Different concentrations were added in a volume of 200 μl to petri dishes (3 cm diameter) each containing 2 ml of agar (end concentration in the agar: 0. 004,0. 0075, and 0. 015 μM) and seeded with E. coli. A positive control was included containing 200 μl of DMSO. After 24 h eggs were transferred to each dish and allowed to hatch. The size at maturity was recorded as described above. The number of replicate worms measured for their body size were at 24 °C (italics) and 12 °C (bold): tra-3 (e1107) 24,16; tra-3 (e1107) 24 °C DMSO 10; tra-3 (e1107) 24 °C TG, 10; tra-3 (e1107) sup-24 (st354) 11,13; tra-3 (e2333) 19,10; +/+ DMSO control 5; +/+ 0. 004 μM TG 6; +/+ 0. 0075 μM TG 6; +/+ 0. 015 μM TG 6; NIL/+ 11,15 16,11; e1107/+ 8,9, 8,8, 8,9; e1107/NIL 10,16,10,18,6, 4; NIL/NIL 22,31. Populations of N2 and CB were bleached (0. 5 M NaOH, 1% hypochlorite) to collect synchronized eggs, which were then inoculated into fresh dishes. For each wild-type strain, four replicate dishes of synchronized eggs were kept in each of the two temperatures until maturity was reached. The nematodes were then collected and frozen in liquid nitrogen. Three independent samples were used for each strain and temperature. For each sample, individuals were synchronized and RNA was extracted using the Trizol method. RNA was subsequently purified (with genomic DNA digestion step) with the RNeasy Micro kit from Qiagen (http: //www. qiagen. com). RNA concentration and quality were measured with Nano Drop (http: //www. nanodrop. com). From each sample 2 μl of RNA were used to obtain cDNA using Superscript II reverse transcriptase from Invitrogen (http: //www. invitrogen. com) and oligo d (t) primers from Genisphere (http: //www. genisphere. com). cDNA was diluted 20× and used for RT-PCR with iQ Sybr Green Supermix from Bio-Rad in 20 μl reactions (http: //www. bio-rad. com). Standard curves for each sample were generated by serial dilutions of the cDNA to select for primer efficiencies of 90%–110% and correlation coefficients greater than 0. 99. We selected two reference genes (rps-20 and rpl-3) using geNorm on the basis of Vandesompele et al. [53]. All primers were designed with Beacon Designer avoiding secondary structures and cross homology. RT-PCR runs were done with MyIQ from Bio-Rad, and expression levels were calculated with the Bio-Rad Gene Expression Macro version 1. 1 using the selected reference genes for normalization. Expression levels are presented relative to the lowest expression of the gene. At least two independent experiments were carried out for each gene. Transgenic worm strains containing tra-3 (+) from the Bristol N2 wild-type strain in the CB background were obtained from the Umeå Worm/Fly Transgenic Facility (http: //www3. umu. se/utcf/index_eng. html). Standard microinjection methods were used [54]. A DNA fragment spanning the entire tra-3 locus and containing the endogenous tra-3 promoter was injected at a concentration of 25 μg/ml. The coinjection marker was pCC [55], a plasmid containing gfp under the control of the unc-122 promoter, which is active in coelomocytes. pCC was injected at a concentration of 50 μg/ml. Body size was measured as described above for five independently derived strains of CB (gfp) (control strains) and CB (gfp and tra-3[+]). For RIL analysis a randomised block design was used (three blocks per RIL). Statistical analyses were performed in SAS. All data were found to be normally distributed according to the Box-Cox method. Comparison between treatments was tested with one-way ANOVA using PROC MIXED. In case of crossing experiments, replicate crossings were performed, and the data were analysed with a nested design where each cross was nested within temperature (cross[temperature]). In PROC MIXED we defined cross (temperature) as a random factor. The number of replicates was optimal to obtain the mean to be within the 95% CI. ANOVA was performed to study the effect of temperature, RIL, block, and interactions on body size. QTL mapping was used to identify the genomic regions (Wormbase release WS100) controlling various life-history traits. Composite interval mapping was used to identify responsible QTL because it is statistically a well-established and powerful tool; it has a better resolution of QTL peaks compared to interval mapping and is able to control for a number of background markers [56]. QTL analyses were performed with the software package QTL Cartographer version 2. 0 [57] using forward regression, a maximum of five background parameters, and the default window size of 10 cM. The experiment-wise likelihood-ratio threshold significance level was determined by computing 1,000 permutations of each trait [58] as implemented by QTL Cartographer. These permutations can account for non-normality in marker distributions and trait values. A peak in the likelihood ratio LR was taken to indicate a significant QTL if LR > 10. Composite interval mapping is sensitive to the number of background markers included in the analysis. The relatively low number of five background markers was used because too many background markers can over-parameterise the model. However, in order to assess whether detected loci that were close to one another also suggested one QTL, we examined the inclusion of ten background markers. The results show the significant QTL based on this ten-marker correction. CIs for QTL were taken based on a 1-LOD support interval corresponding to 95% CI [59]. Template identification was performed with 3D-PSSM [60]. Results show that the first part of TRA-3 sequence corresponding to domains I–III best matches rat calpains M and μ with E-values of 0. 0142 and 0. 0332 respectively, corresponding to over 95% fold recognition confidence. By contrast, the last part corresponding to the T domain matches the C2 domain fold with a best hit to protein kinase C alpha C2, E-value: 0. 583, corresponding to 90% confidence. Restricting to only the TRA-3 DII domain where the F96L mutation occurs (Figure S1), best templates were searched for the “open” and “closed” configurations—corresponding to calcium free and bound states respectively. For the “open” configuration the best match is with human M-calpain (pdb code—1kfu) with an E-value of 0. 000686,36. 5% identity, and 67. 5% similarity. However this calpain was not crystallized in its “closed” configuration as well, and further search for templates was needed to model this state. Structural analysis of the existing M and μ calpains crystallized in “closed” state (1tl9,1mdw, 1kxr, 1tlo, 1zcm, 2ary) showed that all of these are practically identical from a structural point of view, with main chain rms deviations of only 0. 731–1. 124 Å. Consequently all are equally good templates for TRA-3 DII and the closest sequence homologue can be used. This was found to be the rat μ-calpain (pdb code – 1tl9, identity: 38. 6%; similarity: 67. 5%, but an E-value of only 0. 0553). In building the models, target and template sequences were first aligned using MULTALIN [61]. This alignment was further optimised manually in several steps by incorporating information on secondary structure, accessibility, contacts, and functionality of important residues. Secondary structure profile of the target was raised by a consensus based on the top five prediction methods according to CASP6 (Critical Assessment of Structure Prediction Methods 6): JPRED [62], HNN [63], SSPRO [64], PROF [65], NNPREDICT [66]. The alignment was further refined by moving the gaps to correct for unfavourable exposures in the 3D model. The 3D models were then built by coordinate transfer in the sequence conserved regions. Loops with insertions or deletions were generated ab initio, then subjected to multiple rounds of conformational search by simulated annealing and local minimization. Packing of long insertions was investigated using Modeller [67] by generating large number of loop conformers and subjecting them to statistical analysis. Simulated annealing was then used to optimise the top contenders followed by extensive rounds of energy minimisation. In the end, the entire model was subjected to repeated rounds of minimization to relieve unfavourable contacts. Model building, refinement, and analysis were performed using the Accelrys programs: Insight II, Discover, Homology, Modeller, Charmm, Cdiscover, and the free-ware 8v1 version of Modeller on an Silicon Graphics, Octane 2 station.

Biologists are fascinated by variation in body size, which is hardly surprising, considering that the range of body sizes spans orders of magnitude from bacteria to blue whales. Even within species, body sizes can vary dramatically. This intraspecies variation is intriguing because it suggests strong associations between body size and environment. Already in 1847, Bergmann noticed that mammals tend to be larger in colder environments. More recently similar relationships were found for ectotherms, which rely for their body heat on the temperature of their surroundings, where more than 85% of the species studied grew larger at lower temperatures. This phenomenon, dubbed the temperature–size rule, has caused a renewed interest to understand how temperature affects body size. Yet the control of the temperature–size rule remains enigmatic, and the hypotheses proposed have been inconclusive. In this paper the authors show that a single nucleic acid change in one gene is required for regulation of the temperature–size rule in the nematode C. elegans. Using protein modelling they also show that this subtle change in DNA decreases the function of the encoded protein. The data suggest that temperature adaptation can be simple and far less complex than previously thought.

Abstract Introduction Results/Discussion Materials and Methods

developmental biology ecology caenorhabditis evolutionary biology genetics and genomics

2007

A Caenorhabditis elegans Wild Type Defies the Temperature–Size Rule Owing to a Single Nucleotide Polymorphism in tra-3

The dual specificity protein/lipid kinase, phosphoinositide 3-kinase (PI3K), promotes growth factor-mediated cell survival and is frequently deregulated in cancer. However, in contrast to canonical lipid-kinase functions, the role of PI3K protein kinase activity in regulating cell survival is unknown. We have employed a novel approach to purify and pharmacologically profile protein kinases from primary human acute myeloid leukemia (AML) cells that phosphorylate serine residues in the cytoplasmic portion of cytokine receptors to promote hemopoietic cell survival. We have isolated a kinase activity that is able to directly phosphorylate Ser585 in the cytoplasmic domain of the interleukin 3 (IL-3) and granulocyte macrophage colony stimulating factor (GM-CSF) receptors and shown it to be PI3K. Physiological concentrations of cytokine in the picomolar range were sufficient for activating the protein kinase activity of PI3K leading to Ser585 phosphorylation and hemopoietic cell survival but did not activate PI3K lipid kinase signaling or promote proliferation. Blockade of PI3K lipid signaling by expression of the pleckstrin homology of Akt1 had no significant impact on the ability of picomolar concentrations of cytokine to promote hemopoietic cell survival. Furthermore, inducible expression of a mutant form of PI3K that is defective in lipid kinase activity but retains protein kinase activity was able to promote Ser585 phosphorylation and hemopoietic cell survival in the absence of cytokine. Blockade of p110α by RNA interference or multiple independent PI3K inhibitors not only blocked Ser585 phosphorylation in cytokine-dependent cells and primary human AML blasts, but also resulted in a block in survival signaling and cell death. Our findings demonstrate a new role for the protein kinase activity of PI3K in phosphorylating the cytoplasmic tail of the GM-CSF and IL-3 receptors to selectively regulate cell survival highlighting the importance of targeting such pathways in cancer. A key mechanism by which growth factors and cytokines promote cell survival is via the phosphoinositide 3-kinase (PI3K) pathway and constitutive PI3K signaling is known to promote autonomous cell survival and transformation [1]. The recruitment and activation of class 1A isoforms of PI3K (p110α, p110β, p110δ) by cytokine and growth factor receptors leads to the phosphorylation of phosphatidyl inositol phosphates (PIPs) and the subsequent docking of pleckstrin homology (PH) domain proteins such as Akt that activate downstream signaling cascades and biological responses [1]. However, in addition to their lipid kinase activity, all members of the class 1 PI3K family also possess intrinsic protein kinase activity [2]–[4]. While much is known regarding the targets and biological functions of PI3K lipid signaling, little is known of the substrates and functional roles of its protein kinase activity. We and others have shown that the phosphorylation of specific serine residues in the cytoplasmic tails of growth factor and cytokine receptors is critical for initiating intracellular signaling pathways that selectively control cell survival [5]–[9]. In non-transformed cells, physiological picomolar (pM) concentrations of GM-CSF and IL-3 are able to promote Ser585 phosphorylation in the cytoplasmic domain of the βc receptor subunit to regulate cell survival in the absence of other biological responses such as proliferation (the “survival-only” response) [7]. Importantly, this “survival-only” pathway is deregulated in leukemia with constitutive Ser585 phosphorylation clearly detectable in >85% of primary AML samples [10]. Such findings suggest that the kinase responsible for cytokine receptor serine phosphorylation and cell survival becomes constitutively activated in leukemia and may therefore represent a potential therapeutic target. We therefore sought to identify the kinases that promote cellular transformation through their ability to constitutively phosphorylate serine residues in cytokine receptors. Using primary human AML patient samples, we have isolated a kinase that phosphorylates Ser585 in the cytoplasmic tail of the GM-CSF/IL-3 βc receptor. We have identified this Ser585 kinase as the p110α catalytic subunit of PI3K and show that physiological picomolar concentrations of cytokine activate the protein kinase activity of PI3K leading to Ser585 phosphorylation and cell survival. Inhibition of p110α using pharmacological and RNA interference approaches reduced Ser585 phosphorylation in multiple cell types including primary human AML blasts whereas expression of a mutant form of p110α that was lipid kinase-defective but protein kinase-active restored Ser585 phosphorylation. Our findings identify a new role for the protein kinase activity of PI3K in promoting cytokine-mediated cell survival and provide a novel functional link between the deregulated PI3K protein kinase activity and phosphotyrosine-independent survival programs in leukemia. GM-CSF and IL-3 receptor signaling regulate both proliferation and survival of normal myeloid cells and play an important role in myeloid leukemia [11]. However, while GM-CSF promotes cell proliferation in both AML blasts and K562 chronic myeloid leukemia (CML) cells in a tyrosine kinase-dependent manner, we observed that cell survival was autonomous, growth factor-independent, and resistant to tyrosine kinase inhibition (Figure 1A and 1B). Consistent with our previous findings [10], Ser585 phosphorylation of the GM-CSF/IL-3 βc receptor was constitutive in primary AML blasts (Figure 1C) and K562 CML cells (Figure 1D) and was not affected by tyrosine kinase inhibitors (TKIs). Furthermore examination of a panel of primary AML patient samples demonstrated that neither Ser585 phosphorylation nor cell survival was affected by JAK (JAKI) or FLT3 (AG1296, CEP-701) TKIs (Figure S1A–S1E). These results indicate that cell survival pathways in leukemia, such as those regulated by Ser585, are constitutively activated and are largely resistant to tyrosine kinase inhibition. In order to identify the kinases responsible for phosphorylating Ser585 and promoting cell survival we performed chromatographic fractionation of an AML patient sample exhibiting constitutive Ser585 phosphorylation (Figure 2A). Eluted fractions were tested for Ser585-kinase activity in vitro using a βc peptide encompassing Ser585 and a single peak of activity was observed (Figure 2A). Pharmacological profiling of the eluted Ser585-kinase activity (peak activity, fraction 8) revealed that only the PI3K inhibitor, LY294002, significantly reduced Ser585 phosphorylation (Figure 2B). Western blotting of eluted fractions confirmed that the p85 regulatory subunit of PI3K co-eluted with the peak of Ser585 kinase activity (Figure 2A, immunoblots). Further analysis using a panel of four independent PI3K inhibitors indicated that each was able to inhibit the Ser585-kinase activity in a dose-dependent manner (Figure 2C). Although little is known of the protein substrates of PI3K, our results suggested the possibility that the serine kinase activity of PI3K could phosphorylate Ser585. We next immuno-purified PI3K from the TF-1 cytokine-dependent hemopoietic cell line and examined its ability to phosphorylate Ser585 in vitro. PI3K immuno-purified using an anti-p85 antibody was able to phosphorylate a Ser585 peptide but not a control peptide in an LY294002-dependent manner (Figure 3A). Consistent with the known divalent cation and redox requirements for PI3K [2], robust Ser585 phosphorylation only occurred under conditions where both Mn++ and DTT were present (Figure 3B). Using isoform-specific antibodies, we immuno-purified individual class 1A p110 catalytic subunits (p110α, p110β and p110δ) from TF-1 cells and examined their ability to phosphorylate Ser585 in vitro. Immunoblotting precipitates with anti-p85 antibodies indicated that TF-1 cells express predominantly p110α (Figure S2A), which was confirmed by PI3K lipid kinase activity assays (Figure 3C). Consistent with this activity profile, our results show that immuno-purified p110α was able to phosphorylate Ser585 in an LY294002-dependent manner (Figure 3D). To determine whether immuno-purified PI3K could phosphorylate βc within the context of a full-length protein, we performed in vitro kinase assays using the purified recombinant intra-cytoplasmic portion of βc (βic) [5]. PI3K was able to phosphorylate the p85 subunit (as has been previously described [2]) as well as purified recombinant βic (Figure S2B). We then examined whether a mutant form of p110α in which 4 lysine residues (K941-944) in the lipid-binding pocket were substituted for alanine (p110α-4KA) that has been previously described as being defective in its lipid kinase activity but retains full protein kinase activity was able to phosphorylate βic [12]. Although the p110α-4KA mutant was defective in phosphorylating PIPs (Figure S2C), it was not only able to phosphorylate p85, but also βic (Figure 3E). Furthermore, purified recombinant p110α and p110β were able to phosphorylate βic in an LY294002-sensitive manner (Figure S2D). Importantly, we also showed that purified recombinant p110α can directly phosphorylate Ser585 in the context of the full-length purified recombinant βic protein by immunoblot analysis using a phospho-specific anti-phospho-Ser585 pAb (Figure 3F). While it remains possible that PI3K can phosphorylate serine residues in addition to Ser585, purified recombinant p110α was able to directly phosphorylate a Ser585 peptide and this phosphorylation was blocked by the PIK-75 p110α-selective inhibitor (Figure 3G) [13]. Taken together, these results indicate that the protein kinase activity of p110 can directly phosphorylate Ser585 of the GM-CSF and IL-3 receptors. While the ability of PI3K to promote cell survival has almost exclusively been attributed to its lipid kinase activity, the potential biological roles of PI3K protein kinase activity remain unknown. Our previous studies have shown that very low cytokine concentrations in the picomolar range can promote the phosphorylation of Ser585 within the GM-CSF and IL-3 βc receptor to promote cell survival in the absence of both phosphotyrosine pathways and proliferation [7]. Our current studies indicate that PI3K protein kinase activity can phosphorylate Ser585. Thus, if PI3K was able to phosphorylate Ser585 in cells, then picomolar concentrations of cytokine that induce Ser585 phosphorylation should also activate PI3K protein kinase activity. We therefore examined the regulation of both the protein kinase and lipid kinase activities of PI3K in response to increasing concentrations of cytokine. In order to examine the regulation of PI3K protein kinase activity, we analysed p85-Ser608 phosphorylation which has been shown to be a direct substrate of p110 [2], [14]. Low picomolar concentrations of GM-CSF that were able to promote Ser585 phosphorylation (0. 1–1 pM) were also able to activate the protein kinase activity of PI3K as evidenced by increased p85-Ser608 phosphorylation (Figure 4A). However, such low picomolar concentrations did not detectably activate PI3K lipid signaling as evidenced by the lack of both Akt and GSK-3 phosphorylation (Figure 4B), p85 tyrosine phosphorylation (Figure 4A), or activation of PI3K lipid kinase activity (Figure 4C). Thus, PI3K demonstrates two distinct modes of signaling with PI3K protein kinase signaling being regulated by low picomolar cytokine concentrations and PI3K lipid kinase signaling being regulated by higher nanomolar concentrations. We then examined whether PI3K lipid kinase activity was essential for regulating cell survival by examining the role of the key downstream lipid signaling target of PI3K, Akt. Our results show that there was no significant defect in the ability of 1 pM murine GM-CSF to promote the survival of primary mouse bone marrow (BM) progenitor cells derived from either Akt1−/− mice (Figure 4D, left) or in BM progenitor cells derived from wild-type (wt) mice and treated with an Akt1 inhibitor (AKTI-1) (Figure 4D, right). Further, inducible-expression of a constitutively active form of Akt (myr-Akt-1) (Figure S3A) was not sufficient to support the long-term viability of factor-dependent myeloid (FDM) cells in the absence of cytokine (Figure 4E). Together, these results indicate that the PI3K lipid signaling target, Akt, was not required for promoting the survival-only response in the presence of low picomolar cytokine concentrations. We next examined the ability of low picomolar cytokine concentrations to promote cell survival under conditions where the protein kinase activity of PI3K was blocked (using the YM024 or PIK-75 p110α-selective PI3K inhibitors) while downstream PI3K lipid signaling was enforced by expression of myr-Akt-1. Blockade of PI3K protein kinase activity induced by 1 pM GM-CSF using either YM024 or PIK-75 resulted in cell death despite constitutive signaling by myr-Akt-1 suggesting that the protein kinase activity of PI3K was required for cell survival and could not be rescued by enforced Akt1 signaling (Figures 4F and S3B). We also performed an inverse experiment and examined the effect of selectively blocking PI3K lipid signaling while allowing PI3K protein kinase signaling. For these experiments we over-expressed the PH domain of Akt1 fused to GFP (Akt1-PH-GFP) in order to block the binding of endogenous PH-domain proteins (such as Akt) to PIPs in the plasma membrane thereby abrogating PI3K lipid signaling but permitting PI3K protein kinase signaling. Using this approach, we examined the regulation of cell survival in response to either 1 pM cytokine (that was able to promote PI3K protein kinase activity) or 1,000 pM cytokine (that was able to promote PI3K lipid kinase activity) (Figure 4A–4C). While expression of Akt1-PH-GFP was able to block lipid signaling as evidenced by the lack of detectable Akt phosphorylation (Figure S3C) and reduce cell viability under conditions where PI3K lipid kinase signaling is activated by high cytokine concentrations (Figure 4G, 1,000 pM), it had no significant effect on cell survival under conditions where only the protein kinase activity of PI3K is induced by low cytokine concentrations (Figure 4G, 1 pM). We then examined whether constitutive activation of the protein kinase activity of PI3K was able to promote Ser585 phosphorylation of the endogenous βc subunit of the GM-CSF receptor and cytokine-independent cell survival. For these experiments we utilized a doxycycline-inducible system for the expression of a cytokine-independent membrane-localized form of p110α with both lipid and protein kinase activity (p110α-CAAX) or only protein kinase activity (p110α-4KA-CAAX). Induction of p110α-CAAX using doxycycline in the absence of cytokine resulted in increased Ser585 phosphorylation as well as downstream Akt phosphorylation, both of which were blocked by the YM024 PI3K inhibitor (Figure 4H). Importantly, induction of p110α-4KA-CAAX, which is lipid kinase defective (Figure S2C) and was unable to promote Akt phosphorylation (Figure 4H, lower panels), also resulted in increased Ser585 phosphorylation in a cytokine-independent manner (Figure 4H, upper panels). In line with its ability to promote increased Ser585 phosphorylation, p110α-4KA-CAAX was able to significantly increase cell survival to levels approaching that observed in the presence of 1 pM GM-CSF (Figure 4I). Together, these findings demonstrate that the protein kinase activity of PI3K can phosphorylate Ser585 of the GM-CSF receptor to regulate cell survival. We next examined the impact of inhibiting PI3K on Ser585 phosphorylation and cell survival. As shown in Figure 5A, TF-1 cytokine-dependent cells rapidly lose viability in the absence of GM-CSF (0 pM) and are able to proliferate in response to higher concentrations of cytokine (1,000 pM). Importantly, lower concentrations of cytokine (1 pM) that were able to promote PI3K protein kinase signaling but not lipid signaling (Figure 4A–4C) were also able to maintain the viability of TF-1 cells for up to 2 wk in the absence of detectable proliferation (“survival-only” response) (Figure 5A and 5B). To test whether Ser585 of βc was a substrate for p110 under these “survival-only” conditions, cells were pretreated with both pan-specific (LY294002 and Wortmannin) and isoform-selective PI3K inhibitors (Table S1) and then stimulated with 1 pM cytokine. Our results show that both LY294002 and Wortmannin inhibited Ser585 phosphorylation induced by 1 pM GM-CSF (Figure 5C and 5G). Furthermore, two different p110α-selective inhibitors (YM024 and PIK-75) and a p110α-selective and mTOR dual inhibitor (PI-103) were able to down-regulate Ser585 phosphorylation (Figure 5D–5G). In contrast, p110β-selective (TGX-221) and p110γ-selective (AS252424) inhibitors had no detectable effect on Ser585 phosphorylation while p110δ-selective (IC87114) and the protein kinase A inhibitor, H89, had modest effects (Figure 5G). Furthermore, inhibition of DNA-PK and the related PI3K family member ATM had no effect on either Ser585 phosphorylation or the survival of AML blasts (Figure S4A and S4B). Consistent with their ability to block Ser585 phosphorylation, both YM024 and PIK-75 were also able to significantly block the survival-only response in both TF-1 cells (Figure 5H) and lineage-negative primary mouse BM progenitors in the presence of 1 pM cytokine (Figure 5I). Thus, our results show that under survival-only conditions in which low picomolar cytokine concentrations activate the protein kinase activity of PI3K but not its lipid kinase activity, inhibition of p110α not only blocks Ser585 phosphorylation of endogenous βc but also cell survival. We then screened a panel of siRNAs for their ability to knockdown p110α in HEK 293T cells and examined the impact on Ser585 phosphorylation. As shown in Figure 6A, siRNA-p110α-1 resulted in decreased p110α protein levels and an almost complete loss of Ser585 phosphorylation. We then examined the ability of siRNA-p110α-1 to reduce constitutive Ser585 phosphorylation in a panel of primary human AML samples. We observed a significant decrease in Ser585 phosphorylation following transfection of the siRNA-p110α-1 in 6/6 AML samples (Figures 6B, 6C, and S5A; p = 0. 001, Mann-Whitney U). We then tested the ability of YM024 and PIK-75 to induce apoptosis in AML blasts derived from patient samples that were sensitive to down-regulation of Ser585 phosphorylation following PI3K inhibition. Our results show that YM024 and PIK-75 were able to induce cell death in primary AML blasts whereas inhibition of p110β (TGX-221), p110δ (IC87114), and p110γ (AS252424) were less effective (Figures 6D and S5B). Furthermore, siRNA-p110α-1 also significantly reduced the survival of primary human AML blasts (Figure 6E). Thus, both pharmacological and siRNA-mediated targeting of p110α results in a significant decrease in the phosphorylation of Ser585 in the GM-CSF and IL-3 βc receptor in primary human AML cells and the induction of cell death. While many cytokines and growth factors are able to regulate PI3K lipid signaling, little is known of their ability to regulate PI3K protein kinase signaling or whether the protein kinase activity of PI3K is also important in promoting cellular responses in certain contexts. Previously, we and others have identified key serine residues in the cytoplasmic tails of cytokine and growth factor receptors that selectively control cell survival [5]–[9]. In the case of the GM-CSF and IL-3 βc receptor, constitutive Ser585 phosphorylation is associated with deregulated cell survival programs in AML [10]. Importantly, constitutive Ser585 phosphorylation in leukemic cells is refractory to tyrosine kinase inhibition (Figure 1) and thus may provide a receptor-dependent mechanism by which transformed cells are able to survive in the presence of TKIs. We have now isolated a kinase activity from primary AML samples that is able to phosphorylate Ser585 in vitro and shown that this activity is uniquely sensitive to PI 3-kinase inhibitors (Figure 2). We have further shown that purified recombinant p110α can directly phosphorylate Ser585 in vitro (Figure 3) and that inhibition of p110α using either RNA interference or p110α-selective inhibitors down-regulated Ser585 phosphorylation of endogenous βc (Figures 5 and 6). Furthermore, inducible expression of a p110α-4KA-CAAX mutant of PI3K that is defective in lipid-kinase activity but retains protein kinase activity not only promotes Ser585 phosphorylation but also cell survival in the absence of cytokine (Figure 4). These results reveal Ser585 in βc as a direct substrate of the protein kinase activity of PI3K and show that p110α (rather than p110β, p110δ, or p110γ) is the predominant isoform responsible for this activity at least in the myeloid context (Figure 7, model) While the lipid kinase activity of PI3K is clearly pivotal in regulating a wide array of cellular responses including cell survival [1], little is known regarding the protein substrates of PI3K and their functional significance [2]–[4]. Initial reports identified several autophosphorylation sites in either the p85 regulatory subunit, or the p110 catalytic subunits (Table 1) [2], [15], [16]. Additionally, a number of other protein substrates of PI3K have been reported including p101, insulin receptor substrate 1, PDE3B phosphodiesterase, eukaryotic initiation factor 4E-binding protein 1, mitogen-activated protein kinase kinase, and H-Ras [17]–[21]. However, these earlier studies did not determine the specific residues phosphorylated by PI3K nor their functional significance. To our knowledge, only one other specific phosphorylation site has been identified for the protein kinase activity of PI3K for which a functional role has been ascribed. Prasad et al. have shown that p110γ can phosphorylate Ser61 of non-muscle tropomyosin, which is required for agonist-dependent β-adrenergic receptor internalization (Table 1) [12]. From the limited protein substrates so far identified for PI3K, no clear consensus motif is apparent (Table 1); however, the known auto-phosphorylation sites are located within disordered flexible regions either at the C-terminus of p110 isoforms or between the inter-SH2 and C-terminal SH2 domains of p85 suggesting that primary and/or secondary structures may be more important for substrate recognition than tertiary structures. The activation of canonical Type 1A PI3K lipid signaling requires the recruitment of p85 SH2 domains to pYXXM (where pY is phosphotyrosine) phosphotyrosine docking sites, either in the cytoplasmic tails of cell surface receptors or their associated signaling proteins [22]. This mode of signaling is triggered by higher concentrations of ligand in the nanomolar range that induce receptor dimerization/oligomerization and the trans-activation of tyrosine kinases [23]. However, several lines of evidence indicate that low picomolar concentrations of ligand promote Ser585 signaling and cellular survival in the absence of phosphotyrosine pathways, PI3K lipid signaling, and proliferation. Firstly, while high concentrations of cytokine clearly activate the lipid kinase activity of PI3K, we were unable to observe any detectable activation of lipid kinase activity in response to 1 pM cytokine (Figure 4) despite the ability of these concentrations of cytokine to promote long-term cell survival (Figure 5). Secondly, genetic or pharmacological blockade of the key downstream target of PI3K lipid signaling, Akt, had no effect on hemopoietic cell survival in response to 1 pM cytokine (Figure 4). Thirdly, while key downstream targets of PI3K lipid signaling such as Akt or GSK were clearly phosphorylated in response to high nanomolar doses of cytokine, phosphorylation was not detected in response to 1 pM cytokine (Figure 4). Fourthly, although we found no evidence of PI3K lipid signaling in response to 1 pM cytokine, we were clearly able to detect cytokine-regulated PI3K protein kinase activity as evidenced by the induction of p85-Ser608 and βc-Ser585 phosphorylation (Figure 4). Fifthly, enforcing downstream lipid kinase signaling by targeting Akt1 to the plasma membrane while blocking the protein kinase activity of PI3K in response to 1 pM cytokine with YM024 was able to block cell survival (Figure 4). Sixthly, selectively blocking the lipid kinase activity of PI3K by over-expression of an Akt1 PH domain that dominant-negatively blocks PIP docking sites in the plasma membrane but allowing PI3K protein kinase activity in the presence of 1 pM cytokine permitted cell survival (Figure 4). Finally, inducible expression of a p110α mutant that is defective in lipid kinase activity but retains protein kinase activity (p110α-4KA-CAAX) was able to restore Ser585 phosphorylation and promote cell survival in myeloid cells in the absence of cytokine and detectable Akt activation. Thus, our results highlight an important distinction between the regulation of PI3K lipid kinase and protein kinase signaling. On the one hand, higher concentrations of cytokine can regulate phosphotyrosine pathways, PI3K lipid signaling, and the phosphorylation of downstream lipid signaling targets to promote both cell proliferation and survival. On the other, lower concentrations of cytokine promote the activation of PI3K protein kinase activity, Ser585 phosphorylation, and cell survival in the absence of other biological responses such as proliferation (Figure 7, model). Others have also suggested that PI3K can provide multiple independent signaling outputs with p110γ regulating Akt signaling via its lipid kinase activity and regulating ERK signaling via its protein kinase activity [24]. While the functional significance of this signal bifurcation remains unclear, it is intriguing that the insulin and IFNα receptors have been reported to activate the protein kinase activity of PI3K in a phosphotyrosine-independent manner [20], [25]. In the case of the βc subunit, the mechanism by which PI3K is recruited and activated leading to Ser585 phosphorylation is not clear. It is possible that in addition to recruitment to phosphotyrosine docking sites, class 1A PI3Ks such as p110α can also be recruited via phosphotyrosine-independent mechanisms similar to those employed for the recruitment of p110γ to G-protein coupled receptors or p110β and p110δ to ErbB3 [26], [27]. Consistent with this notion, our previous studies have shown that a βc receptor mutant in which all eight cytoplasmic tyrosine residues were substituted for phenylalanine (βcF8) is not only phosphorylated in Ser585 in response to cytokine but is also able to promote cell survival in the absence of proliferation [6] indicating that βc tyrosine phosphorylation is not required for regulating the Ser585-survival pathway. One possible mechanism by which PI3K is recruited may involve the binding of the p85 SH3-domain to a conserved PXXP motif in the cytoplasmic tail of the α-subunit of the GM-CSF and IL-3 receptors as proposed by Perugini et al. [28]. While the mechanisms by which PI3K is recruited to protein targets to phosphorylate substrates such as Ser585 in the GM-CSF/IL-3 receptors (identified in these studies) or Ser61 in tropomyosin (identified by others [12]) requires further study, it is interesting that significant levels of PI3K can be found at the plasma membrane under basal conditions in at least some transformed cell types and that this translocation may be enhanced by the 14-3-3 proteins [29]. It is important to note that siRNA-mediated knockdown of p110α or pharmacological inhibition inhibits both the protein and lipid kinase activity of PI3K. Thus, the induction of apoptosis following PI3K inhibition may not only result from inhibition of PI3K protein kinase targets (such as Ser585), but also lipid kinase targets (such as Akt) (Figure 7, model). Deregulated PI3K lipid signaling has been widely observed in many cancers and activating mutations in p110α are frequently observed in solid tumors. However, p110α mutations are rare in AML [30]. Nevertheless, constitutive PI3K lipid signaling is prevalent in AML with elevated Akt phosphorylation being observed in most patient samples [30]. Our previous studies suggest that the protein kinase activity of PI3K is also deregulated with high prevalence in AML with constitutive Ser585 phosphorylation observed in >85% of primary AML patient samples [10]. While kinases other than PI3K may be responsible for constitutive Ser585 phosphorylation in at least some AMLs, siRNA targeting of p110α significantly reduced Ser585 phosphorylation in 6/6 primary AML samples (Figure 6). Additionally, siRNA-mediated knockdown of p110α or inhibition of p110α using YM024 in two AML samples analysed resulted in increased apoptosis (Figure 6) consistent with a role for p110α in regulating AML cell survival. Most importantly, this pathway appears refractory to FLT3 and JAK kinase inhibition (Figure 1). Others have shown that targeting p110δ with IC87114 prevents the proliferation of AML blasts, but the effect on the cell survival has not been determined [31]. In our studies, IC87114 as well as p110β-selective (TGX-221) and p110γ-selective (AS252424) inhibitors were not effective in down-regulating either Ser585 phosphorylation (Figure 5) or promoting apoptosis in AML blasts (Figure 6) suggesting that p110α is likely to be the primary isoform promoting Ser585-survival signaling in AML. Thus, our results identify a new role for PI3K in which its protein kinase activity phosphorylates cytokine receptors to initiate downstream signaling leading to cell survival. The ability of PI3K to switch between protein kinase and lipid kinase activities would thus allow two independent modes of signaling each functionally linked to a distinct cellular outcome. How these two distinct arms of enzymatic activity are perturbed and hijacked in cancer remains to be elucidated. Discovery of other protein kinase substrates of PI3K that are constitutively phosphorylated in cancer may reveal useful biomarkers and therapeutic targets for PI3K-pathway drug development. Bisindolylmaleimide I, rapamycin, LY294002, GSK-3 inhibitor IX, JAK inhibitor 1 (JAKI), U0126, quercetin, PI3Kγ-1, genistein, PI-103, TGX-221, and AS252424 were from Calbiochem; SB203580 from Promega; H89, PP1, staurosporine, and kemptide from Biomol; imatinib and dasatinib were from Selleck Chemicals; Akt inhibitor 1 was from MBL; 5,6-dichlorobenzimidazole riboside (DRB), AG1296, and Wortmannin were from Sigma; CEP-701 was from Tocris Biosciences; IC87114 and YM024 were generously provided by Shaun Jackson (ACBD). PIK-75 and A66 were synthesized as previously described [13]. Peptide sequences encompassing Ser585 of βc (Mimotopes) were 579LGPPHSRSLPDILG591 and 579LGPPHSRpSLPDILG591 (where pS is phospho-Ser585 which was used as a non-phosphorylatable control). Recombinant purified p110α was from Meredith Layton (Monash University). Murine GM-CSF and IL-3 were from Prospect. BM from Akt1−/− knockout mice were from Rick Pearson (Peter MacCallum Cancer Centre). Akt1-PH domain plasmid obtained from Christina Mitchell (Monash University). A CAAX box was engineered into the C-terminus of p110α by PCR amplification of a 3′ fragment from pcDNA3. 1-myc-p110α using GCGGCCATCGATTTGTTTACAC and TTTCGCGCGGCCGCTCAAGAGAGCACACACTTACAGTTCAAAGCATGCTGCTTAA and cloned into the Cla1/Not1 sites of pcDNA3. 1-myc-p110α and pcDNA3. 1-myc-p110α-4KA (gifts of Lazaros Foukas, University College London), which expresses a mutant form of p110α in which lysines 941–944 within the lipid binding pocket are mutated to alanine, which results in defective lipid kinase activity while protein kinase activity is unaffected. The full length myc-p110α-CAAX cDNAs were then PCR amplified using GAGGAGGACCTGCTGCCTCCAAGACCATCATCAGGTGAACTG and GAACTGTAAGTGTGTGCTCTCTTGAAGCGCTCCGAAA followed by PCR using AAACGGACCGGTGCCACCATGGAGCAGAAGCTGATCTCCGAGGAGGACCTGCTGCCTC and TTTCGGAGCGCTTCAAGAGAGCACACACTTACAGTTC and the products cloned into pTripz using Age1 and Afe1 to give pTripz-myc-p110α-CAAX and pTripz-myc-p110α-4KA-CAAX. HEK 293T cells were transfected with using lipofectamine (Invitrogen) in 0. 5% fetal calf serum (FCS; JRH Laboratories) and DMEM for 4 h. TF-1 factor-dependent cell line was cultured in 10% FCS/RPMI with 2 ng/ml human GM-CSF and transfected by electroporation (1,000 µF at 250 V). FDM cell lines were generated by HoxB8 transformation as described in Figure S3A and cultured in DMEM/10% FCS with 0. 25 ng/ml murine IL-3 as previously described [32]. Primary murine hemopoietic progenitor cells were isolated from the BM of SV129 or BL6 mice as previously described and lineage negative (Lin−) cells were isolated by negative selection using a Lineage Cell Depletion Kit (Miltenyi Biotec) [7]. Apheresis product, BM, or peripheral blood samples were obtained from patients with AML and one patient with CML. Patient samples were collected after informed consent according to institutional guidelines and studies were approved by the Royal Adelaide Hospital and Alfred Hospital Human Ethics Committees. Diagnosis was made using cytomorphology, cytogenetics and leukocyte antigen expression and evaluated according to the French-American-British classification. For patient characteristics see Table S2. Mononuclear cells (MNCs) were isolated by Ficoll-Hypaque density-gradient centrifugation and resuspended in PBS containing 0. 1% human albumin (CSL) [10]. Morphological analysis revealed >70% blasts after Ficoll-Hypaque density-gradient centrifugation. Primary AML MNCs (3×108) from patients were lysed in a hypotonic buffer (20 mM Tris-Hcl [pH 7. 4], 0. 5 mM EDTA, 0. 5 mM EGTA, 10 mM βME, 5% glycerol) containing 2 mM NaF and Complete Mini EDTA-free protease inhibitor cocktail (Roche). Hypotonic lysis in the absence of detergents was used to ensure that the activity of multi-subunit kinases was preserved during the purification. The lysate was then subjected to centrifugation at 16,000 g for 10 min followed by ultracentrifugation of the supernatant at 186,000 g for 1 h. The clarified lysate was then subjected to fast protein liquid chromatography (FPLC) on a Superdex 200PC 3. 2/30 column (Amersham Biosciences). Chromatography was performed using a running buffer (Tris-Cl [pH 7. 5], 200 mM NaCl, 0. 1 mM EDTA, and 10 mM βME) and a flow-rate of 40 µl/min and 40 µl fractions were collected. Protein kinase activity was examined in (i) aliquots of eluted fractions following chromatography of primary AML samples, (ii) p85 and p110 immunoprecipitates, or (iii) purified recombinant p110 catalytic subunits of PI3K as described in detail in Figure S2. Reaction mixtures comprised of 50 µM Ser585 peptide, 50 µM Kemptide or 0. 5 µg of recombinant beta subunit cytoplasmic domain (βic) in kinase buffer (50 mM Hepes [pH 7. 4], 5 mM EDTA, 10 mM MnCl2,0. 250 mM dithiothreitol [DTT], 0. 02% Tween 20) with 0. 25 µCi[γ-32P]ATP, 1 µM cold ATP. Production and purification of the histidine-tagged recombinant βic protein encompassing amino acids 445–881 of the intracellular domain of βc has been previously described [11]. Reactions were incubated at 30°C for 30 min and aliquots examined for 32P-labelled peptide on phosphocellulose filters (Whatmann, P81) and liquid scintillation counting [5]. For βic kinase assays, reactions were stopped by adding 2× SDS load buffer followed by SDS-PAGE and autoradiography. For PI3K lipid kinase assays, cells were lysed in NP-40 lysis buffer (137 mM NaCl, 1. 0% NP-40,10% glycerol, 50 mM Tris-HCl [pH 7. 4]) containing 10 mM β-glycerol phosphate, 1 mM phenylmethylsufonylfluoride, 10 mM NaF, 10 mM Na orthovanadate, 4. 5 U/ml aprotinin (Sigma), and 1 mg/ml leupeptin (Sigma) and immunoprecipitated proteins were examined for PI3K lipid kinase activity using PIP and 0. 25 µCi[γ-32P]ATP as substrates as described in Figure S2C and previously reported [5]. HEK 293T cells or primary AML blasts were transfected for 48 h with 100 nM of siRNAs to p110α or a scrambled control using lipofectamine RNAiMAX (1∶300) in OptiMEM medium (Invitrogen) and 0. 5% FCS. siRNA sequences for p110α knockdown were Silencer control (Ctl siRNA-1, Ambion), Stealth control (Ctl siRNA-2, Invitrogen), GCAUUGACUAAUCAAAGGATT (siRNA-p110α-1, Ambion), AAUAGUGUGAGAAUUUCGCACCACC (siRNA-p110α-2, Invitrogen), and UUACCCAGAUCACCACUAUUAUUUG (siRNA-p110α-3, Invitrogen). Transfection efficiency was monitored using a BLOCK-iT Alexa Fluor red fluorescent oligonucleotide (Invitrogen) and we routinely obtain >85% transfection efficiency using siRNAs [10]. TF-1 cells were factor-deprived in RPMI containing 0. 5% FCS for 12 h and then stimulated with different GM-CSF concentrations before lysis in NP-40 lysis buffer [5]. The βc subunit was immunoprecipitated using 1 µg of 1C1 or 8E4 anti-βc mAbs; p85 and various isoforms of p110 were immunoprecipitated with anti-p85 pAb (Upstate) at 1∶1,000, anti-p110α (Cell Signalling), anti-p110β pAb (Santa Cruz), anti-p110δ mAb A-8 (Santa Cruz). Anti-myc (9E10) and anti-α-tubulin antibody (Abcam) was used at 1∶1,000; anti-Flag and anti-HA mAb HA7 (Sigma) was used at 1∶10,000; Anti-phospho-Ser473Akt, anti-phospho-Ser21/9GSKα/β (Cell Signalling), anti-phosphotyrosine 4G10 (Upstate), anti-Ckl, anti-phospho-STAT5 (Tyr694) (Cell Signalling), and anti-phospho-Ser608 [14] were used at 1∶500. Affinity-purified phospho-Ser585 of βc pAb was used at 1∶500 [5]; affinity-purified phospho-Tyr577 of βc pAb was used at a dilution of 1∶1,000 [5]. Cell survival was determined by either trypan blue exclusion, annexin V-FITC (Roche) staining, propidium iodide staining, or counting viable cell number in reference to Flow Count Fluorospheres (BD Biosciences) essentially as described previously [10]. Cell proliferation was determined by BrdU incorporation as described previously [10], using the in situ cell proliferation kit (Roche).

The ability of cells to survive in the absence of proliferation (cell division), differentiation (cell maturation) or activation allows tissues to maintain cell populations that are poised for rapid responses to damage, infections, or other physiological demands. While this “survival-only” response is fundamental to all physiological processes, the underlying mechanisms are not understood. Many growth factors are potent regulators of cell survival through their ability to bind specific cell surface receptors, which in turn activate specialized enzymes called kinases. Phosphoinositide 3-kinase (PI3K) is a dual specificity kinase that is known to be involved in cell survival and malignant transformation, and it is able to phosphorylate both lipid and protein substrates. While the PI3K lipid kinase activity has been extensively studied, the functional significance of its protein kinase activity remains unclear. Here we show that PI3K protein kinase activity can directly phosphorylate growth factor receptors on human hematopoietic (blood) cells to promote a “survival-only” response. We further show that the protein kinase activity of PI3K can be hijacked to result in uncontrolled growth factor receptor phosphorylation and the deregulated survival of leukemic cells. Our studies provide the first evidence that the protein kinase activity of PI3K can control cell survival and that this activity may be deregulated in cancer.

Abstract Introduction Results Discussion Materials and Methods

protein kinase signaling cascade signal transduction molecular cell biology immune physiology cytokines physiology biology anatomy and physiology signaling cascades

2013

Protein Kinase Activity of Phosphoinositide 3-Kinase Regulates Cytokine-Dependent Cell Survival

Translation of mRNA into a polypeptide chain is a highly accurate process. Many prokaryotic and eukaryotic viruses, however, use leaky termination of translation to optimize their coding capacity. Although growing evidence indicates the occurrence of ribosomal readthrough also in higher organisms, a biological function for the resulting extended proteins has been elucidated only in very few cases. Here, we report that in human cells programmed stop codon readthrough is used to generate peroxisomal isoforms of cytosolic enzymes. We could show for NAD-dependent lactate dehydrogenase B (LDHB) and NAD-dependent malate dehydrogenase 1 (MDH1) that translational readthrough results in C-terminally extended protein variants containing a peroxisomal targeting signal 1 (PTS1). Efficient readthrough occurs at a short sequence motif consisting of a UGA termination codon followed by the dinucleotide CU. Leaky termination at this stop codon context was observed in fungi and mammals. Comparative genome analysis allowed us to identify further readthrough-derived peroxisomal isoforms of metabolic enzymes in diverse model organisms. Overall, our study highlights that a defined stop codon context can trigger efficient ribosomal readthrough to generate dually targeted protein isoforms. We speculate that beyond peroxisomal targeting stop codon readthrough may have also other important biological functions, which remain to be elucidated. Although translation of mRNA usually occurs with high fidelity, different recoding mechanisms exist that alter the amino acid sequence of the resulting polypeptide [1], [2]. Programmed ribosomal frameshifting and stop codon readthrough are commonly used to maximize genomic coding capacity in viruses [3]. Bacteria containing suppressor tRNAs and Saccharomyces cerevisiae strains expressing a non-functional prion form of the eukaryotic release factor 3 display significantly enhanced bypass of stop codons [4], [5]. This results in C-terminal extension of many proteins and increases phenotypic variability. Stop codon readthrough of individual genes was first described for viral replicases in bacteriophage Qβ and tobacco mosaic virus [6]–[9]. In their pro- and eukaryotic hosts, however, only a few cellular genes have been identified, where a biological function could be attributed to readthrough derived extended polypeptides [2], [10], [11], [12]. Although in mammals readthrough derived isoforms have been detected for beta-hemoglobin [13] and for myelin P0 [14], it is unknown whether the derived C-terminal extensions are functionally important. Recently, bioinformatic analysis and ribosome profiling revealed evidence for abundant and regulated bypass of termination in different developmental stages of Drosophila and other metazoa, suggesting a conserved function for translational readthrough in animals [15], [16]. We have recently reported that translational readthrough and alternative splicing are used in fungi to generate peroxisomal isoforms of several glycolytic enzymes [17]. Peroxisomes are single-membrane compartments with a major role in fatty acid degradation [18], [19]. In humans, peroxisomes are essential and peroxisome disorders cause severe syndromes [20]–[22]. The majority of proteins destined for the peroxisomal matrix contain a type 1 peroxisomal targeting signal (PTS1), a short C-terminal motif derived from the prototypical sequence Ser-Lys-Leu (SKL) [23], [24]. Here, we report that translational readthrough at a short conserved stop codon context is used in animals and fungi to generate peroxisomal isoforms of important metabolic enzymes. We have previously shown that in the fungus Ustilago maydis, a PTS1 containing isoform of triosephosphate isomerase (Tpi1) is generated by stop codon readthrough (Fig. 1A) [17]. To determine the minimal sequence requirements for efficient readthrough, we fused the reporter gene gfp (green fluorescent protein) at different positions downstream of the tpi1 stop codon (Fig. 1B). Western analysis revealed that three nucleotides (CUA) following the UGA stop codon were sufficient to trigger efficient stop codon readthrough (Fig. 1B). Readthrough was also observed if this sequence motif was inserted between two reporter genes (Fig. 1C). Stop codon identity was found to be important as replacement of UGA by UAG or UAA reduced translational readthrough (Fig. 1C). Mutational analysis revealed that the identity of the first and the second nucleotide following the stop codon mainly determine readthrough efficiency (Fig. 1D). Thus, a UGA stop codon followed by the dinucleotide CU is sufficient to trigger efficient translational readthrough in U. maydis. Readthrough cannot result from incorporation of the non-canonical amino acid selenocysteine at the UGA termination codon since fungi do not contain the required specialized translation machinery [2]. It remains to be elucidated which near-cognate tRNA is required for recoding of the UGA stop codon within the UGA CU context. It has already been shown that tryptophan, arginine, cysteine or serine can be incorporated at UGA stop codons [25], [26]. Aminoglycoside antibiotics inhibit protein biosynthesis by reducing the accuracy of ribosomal translation. They also cause misreading of termination codons and thus enhance the level of ribosomal readthrough [27], [28], [29]. If the aminoglycoside G418 was added to growing U. maydis cells we observed a concentration dependent increase of C-terminally extended fusion proteins, for both Tpi1 and the readthrough reporter construct (Fig. 1E and F). To identify additional U. maydis genes coding for proteins with readthrough derived PTS1 motifs we screened all open reading frames ending on TGA that are followed by CT (Tab. S1). In addition to the previously described glycolytic enzymes with cryptic PTS1 motifs [17], we identified D-ribulose-5-phosphate-3-epimerase and an NADH-dependent aldehyde reductase (Fig. 2A). We could demonstrate that the readthrough derived extensions of both proteins trigger peroxisomal targeting in U. maydis (Fig. 2B). While the biological function of the aldehyde reductase is yet unknown, D-ribulose-5-phosphate-3-epimerase is an enzyme of the pentose phosphate pathway. Other components of this pathway have already been described to be located in peroxisomes [30]–[32]. Genome comparison revealed that peroxisomal targeting of both enzymes via ribosomal readthrough at UGA CU occurs also in other fungal species (Fig. 2C). It has been described previously that nucleotides downstream of the termination codon influence termination efficiency in pro- and eukaryotes [2], [33], [34]. In the mammalian Sindbis virus the motif UGA C is important to trigger stop codon readthrough in vitro [35]. Remarkably, a similar but extended stop codon context was found to suppress the lethal phenotype of a nonsense mutation in a human patient [36]. Indeed, we were able to show that the UGA CU motif identified in U. maydis also triggered efficient readthrough in HeLa cells (Fig. 3A). We observed a similar dependence on the identity of the stop codon and the two downstream nucleotides (Fig. 3A and B). An adenine residue at position +3 further stimulated readthrough efficiency (Fig. 3B). Taken together this suggests that the short sequence motif UGA CU (A) promotes stop codon readthrough in a wider range of species. Next, we screened all human protein coding genes ending on TGA CT for potential peroxisomal isoforms generated by stop codon readthrough. To exclude false positives arising from chance occurrence of putative PTS1 motifs, all candidate genes were analyzed for phylogenetic conservation of both stop codon context and readthrough dependent PTS1 (Tab. S2). This analysis revealed two striking candidates, which were characterized in greater detail: cytosolic NAD-dependent malate dehydrogenase 1 (MDH1) (Fig. 4A) and NAD-dependent lactate dehydrogenase B (LDHB) (Fig. 4B). Both enzymes are critical for intracellular redox homeostasis and participate in redox-shuttling across organellar membranes [17], [37]–[39]. Accordingly, these enzymes are generally found in different cellular compartments. Lactate dehydrogenase activity has been detected in rat liver peroxisomes and was found to be required for regeneration of NAD+ [40]. MDH1 was recently identified as a peroxisomal protein in human liver cells [41]. However, the import mechanism of both enzymes is not obvious. Our data suggest that peroxisomal isoforms of MDH1 and LDHB containing a PTS1 motif are generated by stop codon readthrough. We confirmed targeting of the predicted peroxisomal isoforms of MDH1 and LDHB by fluorescence microscopy of GFP fusion proteins in HeLa cells (Fig. S1). Phylogenetic analysis revealed that readthrough dependent peroxisomal localization of LDHB is conserved in eutherian mammals (Fig. 4C). Remarkably, the cryptic PTS1 motif of MDH1 is highly conserved in the animal kingdom and occurs already in the cnidarian Hydra vulgaris (Fig. 4C). Ribosome profiling experiments in human foreskin fibroblasts revealed evidence for stop codon readthrough for MDH1 but not for LDHB [16]. We used HeLa cells to address ribosomal readthrough of full-length LDHB harboring both an N-terminal HA-tag and a C-terminal Myc-tag (Fig. 5A and B). The latter epitope is only generated upon ribosomal readthrough. Leaky termination again depended on the identity of both the UGA stop codon and the nucleotides immediately downstream (Fig. 5A and B). Quantification of readthrough revealed only a small fraction (approx. 1%) of C-terminally extended proteins (Fig. 5A). However, this amount is sufficient to yield a similar protein concentration in peroxisomes and the cytosol as peroxisomes comprise only a small fraction (1–2%) of the cellular volume [42]. To further confirm that the C-terminal extensions are generated by stop codon readthrough we analyzed the quantity of the extended LDHB isoform in the presence of the aminoglycoside gentamicin. We detected an increased amount of the extended LDHB isoform corroborating that it is generated by ribosomal readthrough (Fig. 5C). To determine the intracellular localization of MDH1 and LDHB we monitored full-length N-terminal GFP-fusion proteins in HeLa cells by live cell imaging. Strong cytosolic fluorescence prevented direct observation of peroxisomal LDHB and MDH1 isoforms (Fig. 5D and S2). However, depletion of cytosolic GFP-LDHB and GFP-MDH1 by repeated photobleaching [43] unveiled peroxisomal targeting of both enzymes (Fig. 5D and S2). Deletion of the PTS1 containing extension abolished peroxisomal localization demonstrating that intracellular sorting of both enzymes depends on ribosomal readthrough (Fig. 5D and S2). To confirm these data we inserted GFP between the TGA CTA motif and the PTS1 encoding sequence of either LDHB or MDH1. Microscopic analysis confirmed the presence of readthrough derived GFP fusion proteins that predominantly co-localize with the peroxisomal marker mCherry-SKL (Fig. 5E and F). We identified additional interesting candidate genes also in other model organisms. D. melanogaster NADP-dependent isocitrate dehydrogenase and Caenorhabditis elegans inorganic pyrophosphatase contain both the stop codon context UGA CU and a predicted PTS1 in the readthrough derived extension. (Fig. 6A, S3A and S3B; Tab. S2). Interestingly, mammalian and fungal orthologs of NADP-dependent isocitrate dehydrogenase contain a PTS1 motif at their regular C-termini (Fig. S3B and S3C) [44]. Peroxisomal localization of NADP-dependent isocitrate dehydrogenase has already been demonstrated in the fungi Aspergillus nidulans and S. cerevisiae [45], [46], [47]. The predicted peroxisomal targeting of inorganic pyrophosphatase was found to depend on a variety of molecular mechanisms in different species (Fig. 6B). While alternative splicing generates a C-terminal PTS1 in D. melanogaster Nurf-38, the U. maydis Ipp1 enzyme contains an N-terminal PTS2 signal (Fig. 6B). Mammals contain two genes coding for cytosolic and mitochondrial versions of inorganic pyrophosphatase (PPA1 and PPA2). Putative readthrough derived PTS1 motifs could be identified in either of these genes depending on the organism (Fig. 6B and C). Inorganic pyrophosphatase is an essential enzyme that drives the direction of all biochemical reactions in which pyrophosphate is generated, e. g. DNA and RNA polymerization [48]. Activity of this enzyme has already been detected in isolated rat liver peroxisomes, where it might be important for ATP dependent activation of fatty acids [49]. In this study, we have shown that the short readthrough core motif UGA CU (A) is widely used for generation of C-terminally extended proteins. Recently, readthrough derived isoforms of four other mammalian proteins were reported [50]. All corresponding genes contain a TGA CTA motif but readthrough efficiency was found to be further increased by downstream sequences. The presence of a guanine nucleotide at position +4 was of particular importance. We did not test the effect of this residue systematically but noticed that many of the genes identified in our study contain a G at this position (see Fig. 1,2 and 4). This suggests that the extended motif UGA CUA (G) represents a weak context for termination of translation. While in some cases the amount of readthough product generated by this context alone might be sufficient, higher levels of readthrough require downstream mRNA sequences, which often form secondary structures [51]. Such structural elements might also be involved in efficient readthrough of cellular and viral genes that lack the TGA CT context [1], [3], [16], [17]. A genome wide study in D. melanogaster has shown that in genes subject to readthrough secondary structures occur more often if the corresponding stop codons are TAA or TAG [15]. Both stop codons are known to be less leaky than TGA [52]. Another property of the UGA CU (A) element might be relevant for its function in generating peroxisomal isoforms. Recently, it was shown that hydroxylation of a proline residue in the decoding center of ribosomes regulates translational fidelity in response to oxygen availability [53], [54]. Mutational analysis revealed that hydroxylation decreased termination efficiency if the stop codon was followed by a cytosine residue [37]. Thus, intracellular sorting of enzymes targeted to peroxisomes via translational readthrough at the conserved UGA CU motif is likely to be regulated by oxygen levels, which are critical for peroxisomal β-oxidation [30], [55]. Therefore, we speculate that the use of specific readthrough signals allows differential regulation to modulate protein function in response to physiological conditions. Escherichia coli strain TOP10 (Invitrogen) was used for all cloning purposes and amplification of plasmid DNA. All U. maydis strains used and created in this study are derivatives of strain Bub8 [56] and listed in table S3. Liquid cultures of U. maydis were grown either in YEPSlight medium [57] or yeast nitrogen base medium (YNB; Difco) supplemented with 2% glucose at 28°C. E. coli and U. maydis cells were transformed as described previously [56], [58]. To analyze the effect of G418 on readthrough levels cells of the U. maydis strains Bub8 mCherry-TGACTA-GFP and Bub8 TPI+3-GFP were grown overnight and diluted to reach OD600 0,5 after 2 h of growth. At this point 0 µg/ml, 10 µg/ml, 20 µg/ml or 40 µg/ml G418 (Roth) were added to the cultures. After 6 hours the cells were harvested and protein isolation and Western blotting was performed as described below. HeLa cells were kindly provided by Lucie Sauerhering (Institute for Virology, Marburg). The cells were maintained in Dulbecco' s modified Eagle' s medium (DMEM, Life Technologies) supplemented with 10% fetal bovine serum (FBS, Life Technologies) and 1% antibiotics (10. 000 units/ml Penicillin, 10. 000 µg/ml Streptomycin, Life Technologies) at 37°C, 5% CO2. For transfection, the cells were seeded on 6-well plates and transfected the next day using Lipofectamine2000 (Life Technologies) according to the manufacturer' s protocol. 24 h after transfection protein extracts were prepared with RIPA-buffer (50 mM Tris-HCl, pH 7,5; 1% NP-40; 0. 25% sodium deoxycholate; 150 mM NaCl; 1 mM EDTA) supplemented with protease inhibitors (Complete, Roche) and phosphatase inhibitors (Phosphatase inhibitor cocktail 2, Sigma-Aldrich). Post nuclear supernatants were stored at −80°C and used for Western blot analysis. To analyze the effect of aminoglycosides on readthrough levels HeLa cells were transfected as described above. 24 h after transfection the medium was replaced with DMEM supplemented with 10% FBS, 1% Pen/Strep and 800 µg/ml Gentamicin. 48 h after transfection protein isolation and Western blotting was performed as described. Standard procedures were followed for all DNA manipulations [59]. Plasmids used for expression analysis in U. maydis are derivatives of the plasmid otef-Ala6-MMXN [60]. Oligonucleotides and resulting plasmids are listed in table S4. Genomic DNA from U. maydis was prepared as described previously [61]. Plasmids used for human cell culture are derivatives of plasmids pEGFP-C1 (Clontech), pEGFP-N1 (Clontech), pmCherry-C1 (Clontech) and pcDNA3. 1+ (Invitrogen). All DNA-fragments were inserted between the EcoRI and BamHI sites. cDNA clones encoding human LDHB and MDH1 are from Origene. All plasmids were verified by sequencing. For protein isolation, U. maydis strains were grown to mid-log phase in YEPSlight. 10 ml of cells were centrifuged for 5 min at 4000 rpm. The pellet was resuspended in 200 µl TBS containing 0. 1% (v/v) Triton-×100 and 1% protease inhibitor cocktail for use with fungal extracts (Sigma) and supplemented with glass beads. Suspensions were frozen to −80°C and disrupted by 30 min incubation on a vibrax shaker (IKA) at 4°C. Suspensions were centrifuged for 15 min at 12000 rpm at 4°C. Supernatants were quantified by a Bradford assay (Biorad). Preparation of proteins from human cell culture was performed as described (see Human cell culture). For protein detection Western blotting analysis was performed. Proteins were separated by SDS-PAGE and blotted on polyvinylidene difluoride membranes or nitrocellulose membranes. Antibodies used in this study: anti-GFP mouse monoclonal (Santa Cruz), anti-mCherry mouse monoclonal (Abcam), anti-α-Tubulin mouse monoclonal (Calbiochem), anti-HA mouse monoclonal (Sigma Aldrich) and anti-Myc mouse monoclonal (New England Biolabs) and secondary antibody (HRP-conjugated goat anti-mouse, Santa Cruz). Detection was performed with Supersignal West Femto and Supersignal West Pico Chemiluminescent Substrate (Thermo Fisher Scientific) and a ChemoCam Imager (INTAS Science Imaging). Western blots were quantified using ImageJ [62]. U. maydis cells from logarithmically growing cultures (YNB-medium) were placed on agarose cushions and visualized by phase contrast (PC) and epifluorescence microscopy using a Zeiss Axiovert 200 microscope. Images were taken using a cooled CCD camera (Hamamatsu Orca-ER) with an exposure time of 30–300 ms. Image acquisition was performed using Improvision Volocity software and processing was carried out with ImageJ. For microscopic analysis, HeLa cells were grown on Glass Bottom dishes (Mattek) in DMEM medium supplemented with 10% FCS. Cells were transfected with equal amounts (1 µg of total DNA) of the tested construct and the peroxisomal marker mCherry-SKL using Fugene6 reagent (Promega). Prior to imaging, cells were shifted to Life Cell Imaging buffer (Life Technologies) and kept at 37°C. Life cell imaging and photobleaching experiments were performed using a LSM710 confocal microscope (Zeiss) equipped with a 63×1. 4 NA objective and 405,488 and 561 nm laser lines. For both co-localization and photo-depletion experiments, single image planes with an optical section thickness of 0. 7 um were acquired. In order to exclude spurious co-localization by cross-excitation or bleed-through, only a narrow band of emission fluorescence was detected (GFP: Excitation 488 nm, Emission 493–532 nm, mCherry: Excitation 561 nm, Emission 599–708 nm). Using those settings, we did not observe any bleed-through or cross-excitation, even when observing extremely bright samples. To deplete the freely diffusible cytoplasmic pool of the respective GFP fusions, a small region of interest (ROI) was repeatedly bleached using 405 nm and 488 nm laser lines. Loss of fluorescence was monitored using a second ROI at the opposite end of the observed cell. During bleaching, only these two areas of the cell were recorded to minimize photobleaching of the rest of the cell. After complete depletion of the cytoplasmic GFP-fusion pool, as estimated by the lack of further photobleaching in the second ROI, an image of the whole cell was acquired. Image processing and analysis was performed using ImageJ. Only cropping and linear adjustments of brightness and contrasts were performed. In order to visualize the weak signals after photobleaching, a different scaling of the display range in comparison to the non-bleached images was used; also this adjustment was linear. For the identification of C-terminal protein extensions containing a peroxisomal targeting signal the TransTerm [63] database was used to retrieve 3′ flanking sequences from several organisms. The sequences were processed using regular expressions to retain only sequences between the regular stop codon to the next in-frame stop codon. Sequences were translated using Virtual Ribosome [64]. C-terminal tripeptides of the translated sequences were scanned for putative PTS1 motifs using the following regular expression: (∧\*[∧*]*? ([ASTPCE]RL|[SATPCVNG]KL|S[SNH]L|ARI|S[KR]M|[AG]NL|SN[IM]|[SA][KRQ]Y|HHL|[QS][KRQ]F) \*). PTS2 motifs were predicted with TargetSignalPredictor (http: //www. peroxisomedb. org). Data analysis was performed with NCBI (National Center for Biotechnology Information) Basic local alignment search tool (Blast) [65], other NCBI resources and GenRE - MIPS Ustilago maydis DataBase.

Eukaryotic organisms use various strategies to generate protein isoforms with different function or intracellular localization from a single gene. While differential splicing of mRNA is the most common mechanism to expand the number of encoded proteins, translational readthrough of stop codons is an alternative strategy to create protein variants with C-terminally extended proteins. Recently, it has been shown that fungi use both alternative splicing and translational readthrough to specify peroxisomal isoforms of glycolytic enzymes. Here we show that stop codon readthrough is also used in the animal kingdom to target important metabolic enzymes to peroxisomes. Interestingly, several of these enzymes have a function in peroxisomal redox homeostasis and energy metabolism. It has been described that termination fidelity is modulated by oxidation of specific ribosomal proteins. This suggests that dual targeting via translational readthrough allows adaptation of peroxisomal metabolism to the oxidative status of the cell.

Abstract Introduction Results/Discussion Materials and Methods

cell biology protein translation fungal genetics gene expression genetics translation termination biology and life sciences peroxisomes cellular structures and organelles

2014

Ribosomal Readthrough at a Short UGA Stop Codon Context Triggers Dual Localization of Metabolic Enzymes in Fungi and Animals

In Hedgehog (Hh) signaling, binding of Hh to the Patched-Interference Hh (Ptc-Ihog) receptor complex relieves Ptc inhibition on Smoothened (Smo). A longstanding question is how Ptc inhibits Smo and how such inhibition is relieved by Hh stimulation. In this study, we found that Hh elevates production of phosphatidylinositol 4-phosphate (PI (4) P). Increased levels of PI (4) P promote, whereas decreased levels of PI (4) P inhibit, Hh signaling activity. We further found that PI (4) P directly binds Smo through an arginine motif, which then triggers Smo phosphorylation and activation. Moreover, we identified the pleckstrin homology (PH) domain of G protein-coupled receptor kinase 2 (Gprk2) as an essential component for enriching PI (4) P and facilitating Smo activation. PI (4) P also binds mouse Smo (mSmo) and promotes its phosphorylation and ciliary accumulation. Finally, Hh treatment increases the interaction between Smo and PI (4) P but decreases the interaction between Ptc and PI (4) P, indicating that, in addition to promoting PI (4) P production, Hh regulates the pool of PI (4) P associated with Ptc and Smo. The Hedgehog (Hh) signaling pathway plays important roles in both embryonic development and adult tissue homeostasis [1–3]. In Drosophila, the Hh signal is transduced through a receptor system at the plasma membrane, which includes the receptor complex Patched-Interference Hh (Ptc-Ihog) and the signal transducer Smo [4–6]. Binding of Hh to Ptc-Ihog relieves the Ptc-mediated inhibition of Smo, which allows Smo to activate the cubitus interruptus (Ci) /Gli family of zinc finger transcription factors and thereby induce the expression of Hh target genes such as decapentaplegic (dpp), ptc, and engrailed (en) [7,8]. Over the last 30 years, many Hh pathway components have been identified, including those that control transmission, propagation, receipt, and transduction of the Hh signal. However, it is still unclear how Ptc inhibits Smo to block the activation of the Hh pathway and how Ptc inhibition of Smo is relieved by Hh stimulation. It is unlikely that Ptc inhibits Smo by direct association [9,10], as the inhibition occurs even when Smo is present in 50-fold molar excess of Ptc, and substochiometric levels of Ptc can repress Smo activation [10,11]. These findings suggest that the inhibition process is catalytic [10]. The involvement of small molecules, rather than a protein ligand, has been proposed: Ptc may inhibit the production of positive regulators or promote the synthesis of inhibitory molecules [10]. Smo, an atypical G protein-coupled receptor (GPCR), is essential in both insects and mammals for transduction of the Hh signal [8,12,13]. The activation of Smo appears to be one of the most important events in Hh signaling. Hh induces cell surface accumulation and phosphorylation of Smo [9] by multiple kinases, including protein kinase A (PKA), casein kinase 1 (CK1) [14–16], casein kinase 2 (CK2) [17], G protein-coupled receptor kinase 2 (Gprk2) [18], and atypical PKC (aPKC) [19]. These phosphorylation events activate Smo by inducing a conformational change [20] to promote Smo interaction with the Costal2-Fused (Cos2-Fu) protein complex [21–23]. It is believed that Hh-induced phosphorylation counteracts the autoinhibition imposed by arginine clusters in the Smo C-terminal tail (C-tail), which induces an open conformation that promotes the dimerization of Smo proteins [1,20]. Similar to other GPCRs, Smo cell surface accumulation is controlled by endocytic trafficking mediated by ubiquitination [24,25]. In mammals, Hh signal transduction depends on the primary cilium, and Smo accumulation in the cilium is required for Smo activation [26–28]. Therefore, the cilium represents a signaling center for the Hh pathway in mammals [29]. Phosphorylation by multiple kinases promotes the ciliary localization of mammalian Smo [30], but it remains unclear how Smo cell surface or ciliary accumulation and intracellular trafficking are controlled. A previous study has shown that mutation in INPP5E, a lipid 5-phosphotase, results in signaling defects in primary cilium [31], indicating a role for phospholipids to regulate the function of cilium. In the preparation of this manuscript, two studies found that phospholipids regulate ciliary protein trafficking [32,33]; however, it is unknown whether and how phospholipids directly regulate Smo. A very well characterized system for studying Hh signaling is the Drosophila wing disc. Hh proteins expressed and secreted from the posterior (P) compartment cells act on neighboring anterior (A) compartment cells located adjacent to the A/P boundary to induce the expression of Dpp [34,35]. As a morphogen, Dpp diffuses bidirectionally into both the A and P compartments to control the growth and patterning of cells in the entire wing [36–38]. Other genes, including en and ptc, are also induced by Hh to specify cell patterning at the A/P boundary [39,40]. Expression of dpp monitors the low levels of Hh activity, and ptc expression indicates higher levels of Hh activity, whereas en induction appears to be an indicator of the highest doses of Hh signaling activity [39]. The transcription factor Ci is only expressed in A compartment cells that receive the Hh signal. In this study, we found that Hh stimulation increases the levels of phosphatidylinositol 4-phosphate (PI (4) P) in both wing discs and cultured cells. We further found that PI (4) P activates Smo by promoting Smo phosphorylation. Mechanistically, we identified an arginine motif in the Smo C-tail that is responsible for the interaction of Smo with PI (4) P. Arginine to alanine mutation abolishes, whereas arginine to glutamic acid mutation elevates, Smo activity. We also found that, in addition to the kinase activity of Gprk2, its pleckstrin homology (PH) domain increases PI (4) P in wing discs and is required for Gprk2 to fully function in Hh signaling. The findings that Ptc interacts with PI (4) P and that Ptc inactivation increases the levels of PI (4) P indicate that PI (4) P acts downstream of Ptc to activate Smo in the Hh signaling cascade. Finally, we show that PI (4) P promotes phosphorylation and ciliary accumulation of mouse Smo (mSmo) in mammalian cells, and that PI (4) P prevents the ciliary accumulation of mouse Ptc1 and Ptc2. Taken together, our findings suggest that PI (4) P acts as a special small molecule shuttling between Ptc and Smo to modulate Hh responses. In an effort to identify novel regulators in Hh signaling, we collected RNAi lines from the Vienna Drosophila Resource Center (VDRC) when the library was available and screened for kinases, phosphatases, and E3 ubiquitin-protein ligases using the wing-specific MS1096-Gal4. We tested selected RNAi lines for the ability to modify the phenotype of SmoPKA12, a weak dominant negative form of Smo that results in a reproducible wing phenotype with partial fusion between Vein 3 and Vein 4 when combined with the C765-Gal4 (S1D Fig), which represents a very sensitive genetic background for screening Smo regulators [19,25,41]. One of the “hits” was Stt4 kinase, the yeast homolog of PI4KIIIalpha required for the generation of PI (4) P. We found that, although knockdown of Stt4 alone did not produce any change in the wild-type wing (S1B Fig), Stt4 RNAi combined with SmoPKA12 expression enhanced the fusion of Vein 3 and Vein 4 (S1E Fig). In contrast, inactivation of Sac1 phosphatase, which dephosphorylates PI (4) P to phosphatidylinositol (PI), attenuated the fusion phenotype (S1F Fig). These results suggest that Stt4 and Sac1 regulate the activity of Smo in the wing. Consistently, Sac1 RNAi partially rescued the abdominal cuticle loss caused by Hh RNAi, although Sac1 RNAi alone did not show any cuticle phenotype (S1G–S1J Fig). A recent study established a genetic link between Smo and Stt4-Sac1 [42]; however, the molecular mechanisms are unclear. Smo accumulates in P compartment cells as well as A compartment cells near the A/P boundary (Fig 1A) [9,15]. We found that the level of PI (4) P is elevated in the A compartment cells that abut the A/P border (Fig 1B), suggesting that Hh induces the accumulation of both Smo and PI (4) P in these cells. Consistently, the accumulation of Smo in Drosophila embryo correlated with the accumulation of PI (4) P (S2A–S2C Fig). We further found that PI (4) P levels were increased by the expression of Ci-3P and SmoSD123, the constitutively active forms of Ci and Smo, respectively (S2D and S2E Fig). To accurately measure and quantify the absolute concentration of PIP in cells, we established a mass spectrometry-based multiple reaction monitor (MRM) assay and examined whether Hh indeed induces the production of PIP. Based on a detailed method published recently for quantifying PIP2 (PI (3,4) P2, PI (3,5) P2, PI (4,5) P2), and PIP3 (PI (3,4, 5) P3) [43], we optimized the conditions to examine PIP lipids. Trimethylsilyl diazomethane was used to protect the phosphate groups, which allowed for more efficient ionization of the methylated PIP species and a marked improvement in the sensitivity of the assay. We found that treatment of S2 cells with 60% Hh-conditioned medium [44] induced the formation of PIP in a timely manner (Fig 1C, left panel). Consistent with this, treatment of NIH3T3 mouse fibroblasts with mouse sonic Hh N-terminus (ShhNp) [30] stimulated the production of PIP (Fig 1C, right panel). Total PIP was quantified, since this assay was unable to distinguish PI (4) P from PI (3) P and PI (5) P. To further characterize the regulation of PI (4) P by Hh, we used an enzyme-linked immunosorbent assay (ELISA) and found that Hh stimulated the production of PI (4) P in S2 cells in a concentration-dependent manner (Fig 1D, left panel). In addition, knockdown of Stt4 downregulated, whereas knockdown of Sac1 upregulated, the production of PI (4) P (Fig 1D, left panel), suggesting that the Hh-regulated formation of PI (4) P was mediated by Stt4 and Sac1. We further found that overexpression of Ptc prevented the production of PI (4) P, whereas RNAi-mediated knockdown of Ptc elevated production (Fig 1D, right panel), suggesting that Ptc regulates the levels of PI (4) P. To delineate the involvement of PI (4) P in Hh signaling, we used a ptc-luciferase (ptc-luc) reporter assay to monitor the activity of Hh signaling [44] and found that Hh-induced ptc-luc activity was suppressed by RNAi of Stt4 but elevated by RNAi of Sac1 (Fig 1E). Furthermore, treatment with PI (4) P and the expression of Inp54p (a PI (4,5) P2-specific phosphatase to produce PI (4) P) enhanced, whereas IpgD (converts PI (4,5) P2 into PI (5) P) suppressed, the basal and Hh-induced ptc-luc activity (Figs 1F and S3A). As a control, Inp54pD281A phosphatase-dead mutant had no effect on ptc-luc activity (Fig 1F). These data suggest that PI (4) P is a specific phospholipid that regulates Hh signaling in cultured S2 cells. The PH domain is a known phosphoinositide-binding module that is important for signal transduction by sensing alterations in the membrane lipid composition. To visualize PI (4) P pools in wing discs, we used an RFP-PHOSBP reporter that contains two copies of the PH domain from the oxysterol binding protein (OSBP), which is known to specifically bind PI (4) P [45]. In wing discs, expression of RFP-PHOSBP accumulated PI (4) P (Fig 2B) and Smo (Fig 2C), compared to the expression of RFP alone (Fig 2A). In cultured S2 cells, treatment with PI (4) P enhanced Smo activity, indicated by an elevated ptc-luc reporter activity (S3B Fig), thus prompting the question of whether PI (4) P regulates Smo phosphorylation, since phosphorylation promotes Smo activation. Indeed, we found that PI (4) P, but not other phospholipid forms, increased the levels of basal and Hh-induced Smo phosphorylation detected by a phospho-specific antibody (SmoP) [44] recognizing phosphorylation within the second PKA/CK1 cluster (Fig 2D). In addition, PI (4) P treatment induced Smo phosphorylation to a lesser extent compared to Hh treatment, and the combination of Hh and PI (4) P induced hyperphosphorylation of Smo (Fig 2E). Consistently, treatment with PI (4) P induced mSmo phosphorylation in cultured NIH3T3 cells, which was detected by a phospho-specific antibody (PS1) [30] recognizing mSmo phosphorylation at the first CK1/GRK cluster (Fig 2F). In an in vitro kinase assay using glutathione S-transferase (GST) -Smo fusion protein containing Smo amino acids 656–755, we found that Smo phosphorylation by PKA and CK1 kinases was enhanced by the addition of PI (4) P, but not PI (4,5) P2 or PIP3 (Fig 2G), suggesting that PI (4) P directly regulates the phosphorylation of Smo. In support of this notion, we found that Smo interacted with the PH domain from OSBP, and that this interaction was enhanced by the treatment with either PI (4) P or Hh (Fig 2H and 2I). It is possible that PI (4) P directly interacts with Smo and facilitates Smo interaction with the PH domain of OSBP. To test this, we used a solid-phase lipid-binding assay and found that purified full-length Myc-SmoWT strongly associated with PI (4) P, and weaker binding to PI (5) P was detected as well (Fig 3A, left column). Because the level of PI (5) P is much lower than that of PI (4) P in cells [46], PI (4) P is likely the primary lipid that binds Smo. We further found that Myc-SmoΔC (Smo lacking the C-tail) did not interact with PI (4) P (Fig 3A, right column), suggesting that the C-tail of Smo is required for binding to PI (4) P. To identify the residues in the Smo C-tail interacting with PI (4) P, we used an in vivo approach to examine whether the activity of Smo is regulated by the PH domain expression. Membrane-tethered Smo C-terminal truncations by the myristoylation signal (Myr-SmoCT) possess the activity to induce ectopic dpp-lacZ expression [21]. We found that the membrane-tethered PH domain of OSBP (Myr-PHOSBP) increased the ectopic dpp-lacZ expression induced by Myr-Smo730-1035 but did not change dpp-lacZ expression induced by Myr-Smo764-1035 (compare Fig 3E to 3D and 3G to 3F), although Myr-PHOSBP itself had no effect on dpp-lacZ expression in the wing. This suggests that PHOSBP likely regulates Smo activity through aa 730–764. This domain of Smo contains one of the four positively charged arginine clusters, which are known to negatively regulate Smo activity by counteracting phosphorylation (Fig 3B) [20]. Surprisingly, using GST-Smo656-755 in the solid-phase lipid-binding assay, we found that SmoWT strongly interacted with PI (4) P; however, an Arg to Ala mutation (GST-SmoRA4) abolished this interaction (Fig 3H), suggesting PI (4) P interacts with only the fourth arginine cluster. In support of this finding, mutation in the fourth arginine cluster (R4) was sufficient to block PI (4) P binding in a PI (4) P beads pull-down assay (Fig 3I). To further characterize R4, we generated the Arg to Ala mutation in Smo full-length (Myc-SmoRA4) and found that Myc-SmoRA4 lost both the ability to bind PI (4) P (Fig 3J) and the interaction with Cos2-Fu complex (Fig 3K). In addition, phosphorylation of SmoRA4 was no longer regulated by Hh and PI (4) P (Fig 3K). SmoRA4 also had no responsiveness to PI (4) P stimulation in the ptc-luc assay (S4E Fig). Using the fluorescence resonance energy transfer (FRET) assay to test C-terminal CFP/YFP dimerization [20,44], we found that SmoRA4 had a much lower FRET signal and was much less responsive to both Hh and PI (4) P stimulation compared to SmoWT (S4F Fig). RA4 mutation did not cause protein misfolding because there was no difference between SmoRA4 and SmoWT regarding the expression and subcellular distribution in S2 cells. Taken together, our findings suggest that SmoRA4 is inactive. Indeed, driven by the tubulinα promoter that expresses Smo at a level close to endogenous gene expression [17], the expression of SmoRA4 did not rescue ptc and en expression in smo mutant cells (Figs 3L, 3M and S4G). Furthermore, we found that overexpression of SmoRA4 by MS1096-Gal4 did not induce ectopic expression of dpp-lacZ, as noted with SmoWT (Fig 3N and 3O), and that an R>A mutation in the constitutively active form of Smo (SmoSD123RA4) had lower ptc-luc activity compared to SmoSD123 (S4E Fig) and was unable to induce en expression in ventral A compartment cells (Fig 3R, compared to Fig 3Q). These data suggest that Smo activity is compromised by R>A mutation in the fourth arginine motif. In contrast, we found that R>E mutation (SmoRE4), which mimics negative charges caused by PI (4) P binding, elevated Smo activity to induce higher levels of dpp-lacZ expression in the wing disc (Fig 3P) and higher levels of ptc-luc activity, but had no responsiveness to PI (4) P stimulation (S4E Fig). Taken together, our findings suggest that the fourth arginine motif is required for Smo activation. The binding position of PI (4) P in Smo is very critical, because fusion of the PH domain from OSBP to either the third intracellular loop (SmoL3PH) or the C-tail (SmoPH) retained PI (4) P with Smo (S4A Fig) but compromised Smo activity in the wing (S4C and S4D Fig, compared to S4B Fig) and resulted in loss of responsiveness to PI (4) P stimulation (S4E Fig). In comparison, fusion of CFP and YPF to the third intracellular loop and C-tail, respectively, did not change the activity of Smo [20]. Our findings suggest that PI (4) P binds Smo in a position-dependent manner. Considering that the PH domain of OSBP interacts with and activates Smo, we wondered whether a PI (4) P transport protein (PITP) facilitates the interaction between PI (4) P and Smo, since Smo itself does not contain a PH domain. We used RNAi lines from the VDRC to screen a total of 15 typical PH domain-containing PITPs in the fly genome for their ability to modulate Hh phenotypes; inactivation of these proteins by RNAi did not affect Smo accumulation in wing discs, although RNAi of some candidate PITPs modified the wing phenotype of C765-SmoPKA12 (S1 Table). Interestingly, all Gprks contain a PH domain in their C-terminus, and this domain contributes to agonist-dependent translocation by facilitating interaction with lipids and other membrane proteins [47,48]. We next investigated whether the C-terminus PH domain of Gprk2 was important for its role in Hh signal transduction. Wild-type Gprk2 fully rescued en expression in gprk2 mutant cells (Fig 4A). However, deletion of the PH domain in the Gprk2 C-tail (Grpk2ΔC) abolished its ability to rescue en expression (Fig 4B), whereas replacing the PH domain in Gprk2 with the PH domain from OSBP (Gprk2-PHOSBP) restored this ability (Fig 4C). This is consistent with our previous finding that Gprk2KM, a kinase-dead form of Gprk2, has a kinase activity-independent role in regulating Smo [18]. These findings suggest that the PH domain is required for Gprk2 to fully function in transducing the Hh signal. In support of these results, Gprk2, Gprk2ΔC, and Gprk2-PHOSBP, but not Gprk2KM, were able to phosphorylate mSmo in vitro (Fig 5A), indicating that the removal or replacement of the PH domain does not affect the kinase activity. Thus, the function of the Gprk2 PH domain likely accounts for the kinase-independent role of Gprk2 in Smo regulation. Because both Gprk2 transcription and Gprk2 protein expression are upregulated by Hh signaling, and Gprk2 is enriched at the A/P boundary [18,49], we hypothesize that, in addition to promoting the production of PI (4) P, Hh may regulate PI (4) P accumulation by enhancing the expression of Gprk2 as the endogenous carrier for PI (4) P. To examine the ability of Gprk2 to enrich PI (4) P in vivo, we knocked down Gprk2 in the wing disc and found that the levels of PI (4) P were decreased (S5A Fig). We also overexpressed Gprk2 or Gprk2ΔC and found that the expression of Gprk2 elevated the levels of both Smo and PI (4) P (Fig 4D and 4E), whereas the expression of Gprk2ΔC had no effect (Fig 4G and 4H). Similar to RFP-PHOSBP (Fig 2B and 2C), overexpression of PHGprk2 resulted in increased PI (4) P and Smo accumulation (S5B and S5C Fig). In addition, Gprk2 and PI (4) P were largely localized at the cell surface (Fig 4F and 4F′), whereas Gprk2ΔC was cytosolic (Fig 4I). These results suggest that the PH domain of Gprk2 is required for the enrichment of PI (4) P in vivo by localizing Gprk2 at the cell surface. To further characterize the kinase activity-independent role of Gprk2 in regulating Smo, we examined Gprk2-regulated Smo phosphorylation in cultured S2 cells. We found that RNAi targeting the coding region of Gprk2, but not OSBP, attenuated PI (4) P-induced Smo phosphorylation detected by the anti-SmoP antibody (Fig 5C). RNAi targeting the 3′-UTR region of Gprk2 consistently inhibited Smo phosphorylation (Fig 5D, lane 4, top panel). We found that the expression of HA-Gprk2 or HA-Gprk2-PHOSBP rescued Smo phosphorylation inhibited by RNAi of Gprk2 3′-UTR but the expression of HA-Gprk2ΔC did not (Fig 5D), suggesting that the PH domain is responsible for Gprk2 to promote Smo phosphorylation increased by PI (4) P. We also found that deletion of the PH domain decreased the Gprk2-PI (4) P interaction in the PI (4) P beads pull-down assay (Fig 5E). Moreover, the PH domain of Gprk2 (PHGprk2) interacted directly with PI (4) P; mutation of arginine (PHGprk2RA) or phenylalanine (PHGprk2FA) abolished this interaction (Fig 5F). Finally, similar to the PHOSBP interaction with Smo (Fig 2H and 2I), the PHGprk2 interaction with Myc-SmoWT was increased by Hh and PI (4) P treatments in cultured S2 cells (Fig 5G and 5H). Taken together with the observation that deletion of the PH domain does not alter the kinase activity of Gprk2 in vitro (Fig 5A) and in cultured NIH3T3 cells (Fig 5B), our findings suggest that the Gprk2 PH domain plays a positive role in mediating Smo regulation by PI (4) P. The finding that expression of the PH domain from OSBP accumulated PI (4) P (Fig 2B) prompted the notion that an endogenous protein may attract PI (4) P away from Smo in the absence of Hh. Ptc contains a sterol-sensing domain (SSD) and has structural similarity to the resistance, nodulation, division (RND) family of bacterial proton gradient-driven transmembrane molecular transporter [50]. SSD was first identified in proteins implicated in cholesterol metabolism but is now more broadly associated with vesicle trafficking. The Ptc SSD is essential for suppression of Smo activity [51], and mutations of SSD abrogate the Ptc-mediated repression of Smo, although these mutations do not compromise either binding or internalization of Hh [10,52]. It is possibe that the Ptc SSD controls the influx or the efflux of PI (4) P or attracts PI (4) P away from Smo. To test this hypothesis, we generated three vectors: HA-tagged wild-type full-length Ptc (HA-PtcWT), HA-tagged Ptc lacking its SSD domain (HA-PtcΔSSD), and HA-tagged SSD domain (HA-SSD). We transfected S2 cells with these constructs and evaluated the ability of each to interact with PI (4) P. When expressed in S2 cells, all proteins were expressed at low levels, detected only after immunoprecipitation (Fig 6A, top panel). We found that HA-PtcWT and HA-SSD strongly bound PI (4) P, whereas HA-PtcΔSSD did not bind (Fig 6A, lower panel). To further determine whether the SSD domain from Ptc directly interacts with PI (4) P, we used the solid phase lipid-binding assay similar to that used for detecting Smo binding. We found that the SSD fragment protein purified from bacteria strongly associated with PI (4) P, but not with PIP2 or PIP3 phospholipids (Fig 6B), suggesting that the interaction between SSD and PI (4) P in the lipid beads protein pull-down assay is direct. The SSD association with PI (3) P or PI (5) P (Fig 6B) suggests that the expression of a single SSD domain may lose specificity for interaction, or, alternatively, that such interaction may also promote Ptc regulation of PI (3) P and PI (5) P. We also found a very strong interaction between SSD and phosphatidic acid (PA) or phosphatidylserine (PS) (Fig 6B); these may be nonspecific, as PA and PS binding to short protein fragments has often been considered questionable [53]. Our findings in cultured cells led us to examine the correlation of Ptc and PI (4) P in the wing. We found that mutation of ptc or knockdown of Ptc by RNAi increased PI (4) P levels in the wing disc (Fig 6C and 6D), similar to the observation that Ptc inactivation elevates PI (4) P in the salivary gland [42]. These indicate that the activation of Hh signaling by the inactivation of Ptc elevated the production of PI (4) P. Moreover, we found that the overexpression of PtcWT also increased the level of PI (4) P (Fig 6E), which is likely due to the ability of Ptc to accumulate PI (4) P. In support of these findings, PtcΔSSD overexpression had no effect on regulating the accumulation of Smo, Ci, and PI (4) P in wing discs. In addition to promoting the production of PI (4) P, Hh may also regulate the pools of PI (4) P between Smo and Ptc. To test this hypothesis, we purified Smo and Ptc proteins from S2 cells treated with Hh-conditioned medium or control medium and accessed the protein interaction with PI (4) P. As shown in Fig 6E, the level of PI (4) P-bound Smo was increased by the treatment of Hh (Fig 6F, left panel). In contrast, the level of PI (4) P-bound Ptc was decreased by Hh treatment (Fig 6F, right panel). These data indicate that Hh treatment releases PI (4) P from Ptc, suggesting an additional layer of regulation beyond Hh promoting PI (4) P production. It would be interesting to understand how Hh regulates PI (4) P. However, we found that Hh treatment did not significantly change the mRNA levels of Stt4 and Sac1 (S6A Fig). In addition, Hh did not change the protein levels of the overexpressed Stt4 and Sac1 in P compartment cells of the wing disc (S6B and S6C Fig). Hh also did not regulate the accessibility of the Stt4/Sac1 to Smo or Ptc in an immunoprecipitation assay with S2 cells. To examine whether Hh regulates the activity of Stt4 or Sac1, or both, we carried out in vitro kinase/phosphatase assays using purified Stt4 and Sac1 combined with PI substrate. We found that the phosphorylation of PI was enhanced when using Stt4 from cells treated with Hh (Fig 6G, lane 3, compared to lane 2, left panel). In addition, Sac1 from cells treated with Hh had less activity to dephosphorylate the constitutive PI phosphorylation (Fig 6G, lane 3, compared to lane 2, right panel). These data suggest that Hh elevates the activity of Stt4 and inhibits the activity of Sac1. Next, we wondered whether PI (4) P plays a role in regulating mSmo, because PI (4) P induces mSmo phosphorylation (Fig 2F). We first tested different phospholipids for their effects in activating mSmo and found that, similar to Shh, PI (4) P treatment elevated mSmo activity as monitored by a Gli-luc reporter (Fig 7B). In contrast, PIP2 and PIP3 treatment had no effect on mSmo activity. In addition, the activity of the constitutively active form of mSmo (mSmoSD), which mimics mSmo phosphorylation by GRK2 and CK1 [30], was further increased by PI (4) P (Fig 7B). Consistently, Drosophila SmoSD123 activity was increased by PI (4) P (S4E Fig). Similar to Drosophila Smo, mSmo contains arginine clusters in its C-tail (Fig 7A) [20]. We next examined whether the arginine motif (s) were responsible for regulation of mSmo by PI (4) P. As shown in Fig 7C, PI (4) P treatment increased Gli-luc reporter activity, although to a lesser extent compared to Shh treatment in the control group of NIH3T3 cells. PI (4) P and Shh treatment consistently increased Gli-luc activity when cells were transfected with mSmoWT (Fig 7C). However, R>A mutations in R3 and R4 arginine clusters (mSmoRA3 and mSmoRA4, respectively) attenuated the increased activity noted with PI (4) P, and mutations in both R3 and R4 (mSmoRA34) completely blocked the effect of PI (4) P on mSmo activation (Fig 7C). These data suggest that R3 and R4 are responsible for the regulation of mSmo by PI (4) P. Phosphorylation promotes ciliary accumulation of mSmo, which correlates with pathway activation [30], but the molecular mechanisms that control Smo ciliary accumulation are poorly understood. mSmoWT was found in about 5% of cilia, and Shh treatment increased SmoWT accumulation in 75% of cilia (Fig 7D and 7E) [30]. Treatment with PI (4) P induced mSmoWT accumulation in 47% of cilia (Fig 7D and 7E), which was correlated with changes in mSmoWT phosphorylation (Fig 2F) and activity (Fig 7B) induced by PI (4) P. In contrast, mSmoRA34 had no response to PI (4) P treatment (4% of ciliary accumulation by PI (4) P treatment) (Fig 7D and 7E), although it had a low response to Shh stimulation (from 3% to 21% of ciliary localization). Consistent with these findings, mSmoRA34 had much lower activity and much less responsiveness to Shh stimulation in a previous study [20]. Our findings suggest that R3 and R4 clusters are responsible for PI (4) P-associated binding and activation of mSmo. To determine whether Hh regulates the production of PI (4) P in vertebrate systems, using the ELISA assay combined with the anti-PI (4) P antibody, we examined the levels of PI (4) P in ptc1 mutant mouse embryonic fibroblasts (MEFs) and found that, compared to control MEFs, ptc1 MEFs had significantly increased PI (4) P (Fig 7F). Consistently, Shh treatment increased, whereas Smo inhibitor decreased, the levels of PI (4) P (S7A Fig), indicating that Hh signaling activity promotes PI (4) P production in cultured cells. We further investigated the ciliary localization of Ptc1 and Ptc2 and found that the ciliary localization of both Ptc1 and Ptc2 was decreased by Shh treatment or PI (4) P treatment (Fig 7G), which was consistent with the previous study that Hh inhibits the ciliary localization of Ptc1 [54]. Similar to Drosophila Ptc interaction with PI (4) P, we found that both Ptc1 and Ptc2 interacted with PI (4) P in the lipid beads protein pull-down assay (Fig 7H). These observations indicate a consistent regulation of Smo and Ptc by PI (4) P in Drosophila and mammalian systems. To incorporate the findings in this study and the findings published recently [32,33,42], we proposed a model in which Smo phosphorylation and ciliary accumulation is regulated by PI (4) P (Fig 8). To explore the possible involvement of a PITP protein to facilitate the interaction between Smo and PI (4) P, we unexpectedly found that the PH domain of Gprk2 is responsible for the accumulation of PI (4) P that activates Smo. It has been shown that Gprk2 is positively involved in Hh signaling by directly phosphorylating Smo C-tail [18]. In addition, Gprk2 forms a dimer/oligomer and binds Smo C-tail in a kinase activity-independent manner to promote Smo dimerization and activation [18]. However, how Gprk2 promotes Smo dimerization and activation is unclear. In this study, we found that the function for Grpk2 PH domain to activate Smo is independent of its kinase activity (Fig 5A and 5B) and that the PH domain of Gprk2 is responsible for enriching PI (4) P that promotes Smo phosphorylation and dimerization. There are instances in which the binding of lipids to the PH domain promotes dimerization of the protein [55,56], raising the possibility that PI (4) P interaction with the PH domain of Gprk2 promotes its dimerization. Taken together, our findings suggest that the function of PH domain in Gprk2 accounts, at least in part, for the kinase activity-independent role of Gprk2 in Hh signaling, a deeper mechanism for Smo activation by Gprk2. In the absence of Hh, Ptc inhibits Smo activity by promoting Smo endocytosis and turnover in intracellular compartments [9]. Ptc likely inhibits Smo catalytically [10], because substochiometric levels of Ptc are able to repress Smo activation [10,11]. Here, we found that Hh promotes the activity of Stt4 and inhibits the activity of Sac1 (Fig 6G), which, at least in part, explains the catalytic regulation. A previous study proposed a model in which Ptc represses Smo by regulating lipid trafficking; Ptc recruits lipoproteins to endosomes, changing their lipid composition, in order to regulate Smo degradation [57], but the class of lipids remains unidentified. In the presence of Hh, Smo is phosphorylated and accumulates at the cell surface, resulting in protein activation [1,8]. However, it is unknown how Ptc inhibition on Smo is relieved by Hh stimulation. The ability of the Ptc SSD domain to interact with PI (4) P (Fig 6A and 6B) raises the possibility that Ptc may control the pool of phospholipids regulating the accessibility of Smo to PI (4) P. Our finding that Hh treatment decreased the interaction between Ptc and PI (4) P (Fig 6F) suggests the possibility that binding of Hh to Ptc results in a conformational change in Ptc and releases phospholipids. Thus, this study uncovers an additional layer of regulation by indicating the release of PI (4) P from Ptc, which may account for the optimal regulation of Smo. The structure of the Smo N-terminal, including the extracellular cysteine rich domain (CRD), has been characterized [58,59]. Unlike other GPCRs, no ligand-binding function has been identified. It has been shown that Smo-mediated signal transduction is sensitive to sterols and oxysterol derivatives of cholesterol [60–62]. However, unlike vertebrate Smo, Drosophila Smo CRD does not interact with oxysterols [63]. In this study, we found that phospholipids activate both vertebrate and Drosophila Smo through binding to the arginine motif in the Smo C-terminus, although the C-tails have sequence divergence among species. Using the protein: lipid overlay assay, we found that PI (4) P directly binds Smo (Fig 3A) and that mutation in the R4 arginine motif abolishes this direct interaction (Fig 3H–3J). Importantly, R>A mutation abolished the activity of Smo (Figs 3L, 3M and S4G). It is likely that binding of PI (4) P to the arginine motif changes Smo conformation, thus allowing kinases to phosphorylate and activate Smo. In support of this notion, PI (4) P binding to specific arginine residues in specific locations is critical for Smo conformational change, because fusion of the PH domain to either the third intracellular loop or the C-terminus attracts PI (4) P to different locations in Smo, thus blocking Smo activation by PI (4) P (S4 Fig). In this study, we focused on the regulation of Smo by PI (4) P and found that Hh regulates the accessibility of Smo to PI (4) P, evidenced by the Hh-promoted interaction between Smo and PH domain from either OSBP or Gprk2 (Figs 2I and 5E) and by the Hh-enhanced interaction between Smo and PI (4) P (Fig 6F). Binding of PI (4) P likely changes the conformation of Smo, leading to Smo phosphorylation by kinases. It should be noted that phosphomimetic Smo mutations (SmoSD123 and mSmoSD) are still regulated by PI (4) P (Figs 7B and S4E), suggesting that PI (4) P either promotes phosphorylation at other residues or has additional role (s) in activating Smo. Hh signaling activity promoted the production of PI (4) P that was detected by the mass-spec assay (Fig 1C), which was a very sensitive approach. However, in a previous study, the overexpression of full-length wild-type Ci did not elevate the accumulation of PI (4) P in wing discs [42]. It is possible that low levels of Hh signaling activity induced by the expression of wild-type Ci are unable to induce detectable changes in PI (4) P accumulation in wing discs. The disc immunostaining with the anti-PI (4) P antibody might not be as sensitive as the mass-spec method. In support, by expressing SmoSD123 or the constitutively active Ci-3P, in which three PKA sites in the phosphorylation clusters were mutated to block Ci processing [64], we found that PI (4) P was accumulated in wing discs (S2A Fig), likely due to the high levels of Hh signaling activity induced by the active forms of Smo and Ci. Generation of the Myc-SmoWT, Myc-SmoΔC, Myr-Smo730-1035, and Myr-Smo764-1035 constructs and transgenes was previously described [21]. Myc-SmoPH and Myc-SmoL3PH were constructed by fusing one copy of the OSBP PH domain (aa 17–123) to the Smo C-terminus and two copies of the OSBP PH domain after aa 451 of Smo intracellular loop 3, respectively. Generation of the Myc-SmoSD123 and Myc-SmoRA1234 was previously described [15,20]. Myc-SmoSD123RA4 was generated by the combination of SD123 and RA4. Myc-tagged SmoRA1, SmoRA2, SmoRA3, SmoRA4, SmoRA12, and SmoRA34 were generated by PCR. To construct tub-SmoRA4, the previously described tubulinα promoter [17] was inserted upstream to the SmoRA4 sequence. Transgenic lines were generated using the VK5 attP locus to ensure Smo protein expression at the same levels without positional effects. Genotypes for examining the activity of Smo transgene in smo clones include: yw hsp-flp/+ or Y, smo3 FRT40/hs-GFP FRT40, and tub-SmoRA4/+. RFP-PHOSBP and RFP-PHGprk2 were constructed by an in-frame fusion of two copies of the PH domain from OSBP or Gprk2 at the RFP C-terminus in the attB-UAST backbone [41]. RFP-PHOSBP and RFP-PHGprk2 transgenes were generated by insertion at the VK5 attP locus. As a control, the UAS-RFP transgenic line was generated with the same approach. Flag-Gprk2WT and Flag-Gprk2KM constructs and transgenes have been described [18]. Flag-tagged Gprk2ΔC (containing Gprk2 aa 1–666) and Gprk2-PHOSBP (aa 667–714 replaced by the PH domain of OSBP) were inserted in-frame in the Flag-UAST vector, and their transgenic lines were generated using standard P-element-mediated transformation. Multiple independent transgenes were generated, and those on the second chromosomes were used for rescue experiments. Gprk2 mutant clones were generated with yw 122 and FRT82 Gprk2/FRT82 hs-Myc-GFP. ptc mutant clones were generated with yw hsp-flp/+ or Y and ptc[wII] FRT42D/hs-GFP FRT42D. Inp54p (from Saccharomyces cerevisiae, Addgene 20155) and IpgD (from Shigella flexneri, a gift from Dr. Frederique Gaits-Iacovoni) were subcloned into HA-UAST backbones. A phosphatase-dead version of Inp54p with a D281A mutation (Inp54pD281A) was generated by site-directed mutagenesis. GST-SmoWT containing aa 656–755 has been described [15]. GST-tagged SmoRA12, SmoRA34, and SmoRA4 were generated by PCR on the backbone of GST-SmoWT. GST-PHGprk2, GST-PHGprk2RA, and GST-PHGprk2FA were constructed by the same approach. HA-PtcWT, HA-PtcΔSSD, and HA-SSD were generated by PCR with Ptc or its fragments inserted into HA-UAST backbones, and transgenic lines were generated using the VK5 attP locus. His-SSD was generated by in-frame fusion of Ptc SSD domain (aa421-589) to pET30a backbone. Myc-tagged mSmoWT, mSmoRA mutants, and mSmoSD (previous mSmoSD0-5) have been described [20,30]. GFP-tagged mSmoWT and mSmoRA34 were generated by subcloning each cDNA into pEGFP-N1 backbone. Flag-tagged Stt4 (Flag-Stt4) was generated by combining seven RT-PCR fragments with unique cloning sites into the SK+ backbone and, finally, subcloning into Flag-UAST backbone. HA-tagged Sac1 (HA-Sac1) was constructed by inserting the RT-PCR fragment into the HA-UAST backbone. pGE-mPtch1-Myc and pGE-mPtch2-Myc were generated by subclone of the open reading frame from pEGFP-mPtch1 and pEGFP-mPtch2 (Gift from Dr. Chi-Chung Hui), respectively. MS1096-Gal4, ap-Gal4, C765-Gal4, and Gprk2 deletion mutants have been described [17,18]. Stt4 RNAi lines (v15993 and v105614), Sac1 RNAi lines (v44376, v37216, and #56013), Ptc RNAi line (#28795), Hh RNAi line (v1402), GFP-Sac1 (#57356 and #57357), and prd-Gal4 (#1947) were obtained from VDRC or Bloomington Stock Center (BSC). Generation of Gprk2 RNAi lines have been described [18]. The RNAi lines that targeted each PITP were obtained from either VDRC or BSC (S1 Table). PtdIns lipid extraction and quantification by mass spectrometry were carried out as previously described [43]. Briefly, 3 × 107 S2 cells (or 4 × 106 NIH3T3 cells) from a 100 mm dish were harvested and washed once with ice-cold PBS and suspended in 340 μL H2O and 1500 μL quench mix with 25 ng of 17: 0–20: 4 PI (4) P (Avanti Polar Lipids, Inc.) in 10 μL methanol as the internal standard. Lipids were extracted with 1450 μL CHCl3 and 340 μL 2 M HCl and subsequently derivatized with trimethylsilyl diazomethane (Sigma), as previously described [43]. After washing and drying under a stream of nitrogen at room temperature, samples were dissolved in 200 μL methanol. We applied 10 μl of each sample to LC-MS/MS analysis using a Shimadzu LC-20 HPLC and TSQ Vantage triple quadrupole mass spectrometer (ThermoFisher). A Jupiter 5μ C4 300A (50 × 1. 0 mm) column (Phenomenex) was used with the multiple reaction monitoring (MRM) transitions described [43]. PIP concentrations were calculated from MRM peak areas and the internal standard and were subsequently normalized to cell number. For PI (4) P quantification by ELISA assay, similar PtdIns lipid extraction procedures were used without adding the PI (4) P internal control. After washing and drying with nitrogen stream, lipid extracts were dissolved in ethanol and loaded into a microplate, dried under a vacuum, and incubated with 2% BSA in PBS at room temperature for 30 min. Mouse anti-PI (4) P or anti-PI (4,5) P2 monoclonal antibodies (Echelon Biosciences) were added for 1 h, followed by goat anti-Mouse IgG-HRP (Jackson ImmunoResearch) for 30 min, with 3 PBS washes after each inculation. Finally, chemiluminescence substrate (SuperSignal West Pico, Pierce) was added to the microplate, and luminescence intensity was determined by a luminometer. GST-Smo fusion proteins expressed in bacteria were pooled by GST beads (GE Healthcare), then eluted with elution buffer (10mM Glutathione pH 8. 0 in 50 mM Tris) at 4°C overnight. Myc-tagged Smo proteins expressed in S2 cells were immunoprecipitated with anti-Myc antibody combined with beads of protein A ultralink resin, followed by two sequential elutions with Myc peptide (Roche, 100mM KCl, 20% glycerol, 20 mM HEPES KOH, pH 7. 9,0. 2 mM EDTA, 0. 1% NP-40,5 mM DTT, and 0. 5 mM PMSF). The eluted purified GST-fusion proteins and Myc-tagged Smo proteins were concentrated by the Centrifugal filter units (Millipore) and incubated with lipid-dotted strips according to manufacturer’s instruction (Echelon Biosciences), followed by western blot with the anti-GST (Santa Cruz), anti-Myc (Santa Cruz), or anti-His (Millipore) antibodies. For the PI (4) P beads pull-down experiments, HA-tagged Gprk2 proteins were expressed in S2 cells, immunoprecipitated with mouse anti-HA antibody (F7, Santa Cruz), eluted with HA peptide (Sigma, in 500 mM NaCl), and concentrated by the Centrifugal filter units (Millipore). GST-PHGprk2, GST-PHGprk2RA, and GST-PHGprk2FA proteins were expressed in bacteria and purified using the protocol employed for GST-Smo purification. The purified and concentrated GST-fusion proteins or epitope-tagged proteins were incubated with PI (4) P beads (Echelon Biosciences) with wash/binding buffer (10 mM HEPES, pH7. 4; 0. 25% NP-40; 150 mM NaCl), and subjected to western blot to detect PI (4) P bound proteins. Wing discs from third instar larvae were dissected in PBS and then fixed with 4% formaldehyde in PBS for 20 min. After permeabilization with 1% PBST, discs were incubated with primary antibodies for 3 h and appropriate secondary antibodies for 1 h, and washed three times with PBST after each incubation. Affinity-purified secondary antibodies (Jackson ImmunoResearch) for multiple labeling were used. It was a challenge for disc staining with the mouse anti-PI (4) P antibody. We have adopted/modified a critical method for PI (4) P immunostaining from previous publications [46,65]. Discs were fixed in 4% formaldehyde in PBS and permeabilized in 1 M sucrose by freezing at -80°C for 1 h followed by thawing at room temperature. Then discs were washed with PBS and incubated with 50 mM NH4Cl for 15 min, followed by incubation with the anti-PI (4) P antibody at 4°C overnight. For Drosophila embryo primary antibody staining, stage 11 fly embryos with specific genotypes were dechorionated, fixed with Heptane solution, and immunostained with similar procedures. To examine mSmo ciliary localization, NIH3T3 cells were transfected with mSmo-GFP variants, treated with Shh or PI (4) P, and immunostained for mSmo localization in the cilium. Primary antibodies in this study were: mouse anti-Myc (9E10, Santa Cruz), anti-Flag (M2, Sigma), anti-SmoN (DSHB), anti-En (DSHB), and anti-PI (4) P (Z-P004, Echelon Biosciences); rabbit anti-β-Gal (Cappel), anti-GFP (Clontech), anti-Acetylated tubulin (Sigma), anti-PS1 [30], and rat anti-Ci (2A1, DSHB). Affinity-purified secondary antibodies (Jackson ImmunoResearch) for multiple labeling were used. Fluorescence signals were acquired on an Olympus confocal microscope and images processed with Olympus Fluoview Ver. 1. 7c. About 15 imaginal discs were screened and three to five disc images were taken for each genotype.

The Hedgehog (Hh) signaling pathway plays important roles in both embryonic development and adult tissue homeostasis. A critical step in Hh signal transduction is the inhibition of Smoothened (Smo), an atypical G protein-coupled receptor (GPCR), by the Hh receptor Patched (Ptc). It is a longstanding question how Ptc inhibits Smo and how Hh promotes Smo phosphorylation and activation. It is unlikely that Ptc inhibits Smo by direct interaction. Here, we uncover that phosphatidylinositol 4-phosphate (PI (4) P), a specific phospholipid, directly interacts with Smo through an arginine motif in the Smo C-terminal tail and promotes Smo phosphorylation, activation, and ciliary localization. Ptc also interacts with PI (4) P, which is inhibited by Hh stimulation, indicating that Hh triggers the release of PI (4) P from Ptc. We further uncover that Hh stimulation induces the overall production of PI (4) P, likely by regulating PI (4) P kinase and phosphatase. Finally, in addition to the direct role in regulating Smo phosphorylation, G protein-coupled receptor kinase 2 (Gprk2) facilitates PI (4) P interaction with Smo. This study suggest that PI (4) P acts as a special small molecule shuttling between Ptc and Smo to modulate Hh responses.

Abstract Introduction Results Discussion Experimental Procedures

phosphorylation rna interference enzyme-linked immunoassays chemical compounds organic compounds immunoprecipitation basic amino acids amino acids epigenetics immunologic techniques cellular structures and organelles research and analysis methods lipids genetic interference proteins gene expression immunoassays chemistry precipitation techniques hedgehog signaling biochemistry signal transduction rna arginine cell biology post-translational modification nucleic acids organic chemistry cilia genetics biology and life sciences physical sciences cell signaling

2016

PI(4)P Promotes Phosphorylation and Conformational Change of Smoothened through Interaction with Its C-terminal Tail

The HER/ErbB family of receptor tyrosine kinases drives critical responses in normal physiology and cancer, and the expression levels of the various HER receptors are critical determinants of clinical outcomes. HER activation is driven by the formation of various dimer complexes between members of this receptor family. The HER dimer types can have differential effects on downstream signaling and phenotypic outcomes. We constructed an integrated mathematical model of HER activation, and trafficking to quantitatively link receptor expression levels to dimerization and activation. We parameterized the model with a comprehensive set of HER phosphorylation and abundance data collected in a panel of human mammary epithelial cells expressing varying levels of EGFR/HER1, HER2 and HER3. Although parameter estimation yielded multiple solutions, predictions for dimer phosphorylation were in agreement with each other. We validated the model using experiments where pertuzumab was used to block HER2 dimerization. We used the model to predict HER dimerization and activation patterns in a panel of human mammary epithelial cells lines with known HER expression levels in response to stimulations with ligands EGF and HRG. Simulations over the range of expression levels seen in various cell lines indicate that: i) EGFR phosphorylation is driven by HER1-HER1 and HER1-HER2 dimers, and not HER1-HER3 dimers, ii) HER1-HER2 and HER2-HER3 dimers both contribute significantly to HER2 activation with the EGFR expression level determining the relative importance of these species, and iii) the HER2-HER3 dimer is largely responsible for HER3 activation. The model can be used to predict phosphorylated dimer levels for any given HER expression profile. This information in turn can be used to quantify the potencies of the various HER dimers, and can potentially inform personalized therapeutic approaches. The HER family (Human Epidermal growth factor Receptor, also known as the ErbB family) of cell surface receptors plays critical roles in normal cell physiology, development, and cancer pathophysiology [1], [2], [3], [4]. The family consists of the four closely related transmembrane receptor tyrosine kinases HER1 (EGFR), HER2 (NEU), HER3 and HER4, which when activated initiate downstream signaling, and affect a range of cellular decisions including proliferation, survival and motility [4], [5]. The HER receptor expression profile is a critical determinant of cell behavior [6], [7], and outcomes in cancer pathology. Overexpression of EGFR, HER2 and HER3 is associated with decreased survival in cancer, while HER4 overexpression is correlated with increased survival [8], [9]. HER2 is overexpressed in 25–30% of all breast cancers, as well as in other solid tumors [10], [11] and is associated with poor prognosis [8], [12], [13], [14]. While this has led to the development of a range of therapeutics targeting the HER2 receptor [15], the use of these drugs can often lead to resistance through a diverse set of mechanisms [16]. The overexpression of HER family members and their ligands are key compensatory mechanisms responsible for the development of resistance to HER-targeted therapies [17], [18], [19], [20]. In particular, the importance of HER3 expression in driving tumorigenesis [21], [22], [23], [24], and in the development of drug resistance [17], [25] is being increasingly recognized leading to an increased focus on HER3-targeted therapies [3], [15], [26], [27], [28]. While the importance of HER expression levels has been established for clinical prognosis and drug resistance, the mechanistic link between receptor expression, HER activation and downstream consequences is not as clear yet. HER activation is a complex process involving multiple sequential steps, which in general are as follows: the specific binding of ligands (growth factors) to HER receptors leads to conformational changes promoting dimerization between members of the family [29], [30], [31]; dimerization leads to the trans-phosphorlyation of receptor cytoplasmic tails via the kinase activities of the partners in the dimer leading to downstream signaling [31]. Although the HER receptors are homologous, there are key differences in their behavior. EGFR [32], HER3 [33], and HER4 [34] undergo ligand-induced conformational changes promoting dimerization. In contrast, HER2, which has no known ligand, has a structure that enables constitutive dimerization [35], [36]. HER3, on the other hand has impaired kinase activity, but can allosterically facilitate a partner' s kinase activity following dimerization [37]. Further, HER receptors have different trafficking properties with EGFR showing increased ligand-induced internalization and degradation compared to the other members of the family [38]. All of these aspects have a bearing on the number and types of dimers that are formed between the HER receptors following ligand addition. Since the HER dimerization pattern is an important determinant of the consequences of HER activation [39], [40] it is important to quantitatively predict this as a function of the receptor expression profile. Mathematical models have been extensively applied to understand HER activation dynamics [1], [41], [42]. Recent efforts have focused on a quantitative understanding of the interactions between multiple members of the HER family [43], [44], [45], [46]. Birtwistle et al. constructed a mathematical model for the early events (0 to 30 min) in HER activation and downstream signaling in cells coexpressing all four HER receptors [45]. They parameterized their model using HER1, HER2, Erk and Akt activation data in response to EGF and HRG stimulation in MCF-7 cells. Chen et al. constructed a more expanded model for receptor activation and signaling over longer time frames (0 to 120 min) and parameterized it using HER1, Erk and Akt activation data in response to EGF and HRG stimulation in three different cell lines – A431, H1666 and H3255 cells [43]. In each of these two manuscripts, the authors note problems with regards to parameter identifiability given the size of the models [43], [45]. Hendriks et al. focused on HER activation alone in cells expressing HER1, HER2 and HER3 [44]. They assumed parameter values based on the literature and compared simulations with receptor activation data collected in the H292 lung carcinoma cell line [44]. We have recently developed a panel of Human Mammary Epithelial (HME) cells that co-express EGFR with varying levels of HER2 and HER3 [47]. HME cells, like many epithelium-derived cell types require EGFR activation for proliferation and migration [48], and are an excellent system for developing physiologically relevant models of HER signaling. Importantly, our cell line library enables us to study the effects of varying HER expression levels in a common cellular background. We have published data on HER1-3, Erk and Akt activation in these cell lines for single doses of EGF and HRG [47]. Here, we focus further on the quantitative aspects of HER activation. We collect an expanded dataset for receptor activation that includes measurements of total and internal HER phosphorylation and HER receptor mass in four distinct HME cell lines in response to a range of EGF and HRG doses. We have identified the appropriate modeling approach (choices for model scope, granularity, etc.) for analyzing such datasets through a comprehensive model-based analysis of EGFR activation in cells that predominantly express this receptor alone [49]. Here, we expand this model by considering the co-expression of EGFR, HER2 and HER3, and parameterize it using receptor activation data collected in our cell line library. We explicitly consider model identifiability and show that the model can predict the dimer phosphorylation levels given the receptor expression level of a cell line. We note that the fourth member of the HER family is also very important in cancer [50] and should be included for completeness. However, our gene expression and proteomics studies (unpublished data) indicated that used HME cells do not express HER4. Therefore, it was not considered in this study. Our objectives here are to quantitatively link HER (specifically HER1-3) expression levels to receptor activation, and to understand how differential interactions between the members of the HER family drive the process. Towards this end we constructed a parsimonious mathematical model for HER dimerization and receptor activation (Figure 1, see Methods for details), and parameterized it using the appropriate experimental datasets. The model includes the ligands EGF and HRG, ligand-bound and unbound receptor monomers, as well as the feasible combinations of receptor homo- and hetero-dimers (Figure 1A). The reversible biochemical reactions of receptor-ligand binding and dimer formation are represented explicitly via mass action kinetics (Figure 1B). As in our recent manuscript [49], we expressed the level of phosphorylated HER1-3 as a linear combination of the contributions from various dimer species with dimer-specific phosphorylation factors (pfs in Figure 1B) accounting for the relative contribution of each species. The pf can be thought of as a lumped phosphorylation efficiency factor that combines the characteristics of all possible tyrosine sites in a dimer [49]. It enables us to calculate the HER1-3 phosphorylation signals emanating from the various dimer types. The model includes 3 compartments: the cell surface, early endosomes and late endosomes. Biochemical reactions are allowed to occur in the first 2 compartments, while the late endosome is assumed to be a site for the accumulation of dephosphorylated receptors prior to degradation (Methods, and also [49]). In all, the model consists of 51 species and 140 parameters. Given these parameter values, and a specified HER1-3 expression level, the model can be used to predict activated levels of HER1-3 at the cell surface and interior as a function of time in response to various concentrations of EGF and HRG. Values for several model parameters including receptor-ligand binding rates, receptor internalization, recycling and degradation rates are available in the literature (Methods; Tables S1, S2, S3 in Text S1). With these in place, there are 47 unknown model parameters (highlighted in red in Figure 1) that include: the compartment-specific dissociation rates for various receptor dimers, the compartment-specific phosphorylation factors that define the contribution of various dimers to the HER1-3 phosphorylation levels, and a parameter that defines how species are sorted (distributed) between the early and late endosomes (Methods). In order to determine these unknown model parameters, we measured HER1-3 activation dynamics in a panel of HME cell lines with relatively constant levels of HER1, and different levels of HER2 and HER3 [47]. The complete set of data collected for one of the four cell lines used in our study that expresses all 3 HER receptors at significant levels (HER2+3+; designated with clone tag D20) is presented in Figures 2–4. The data includes measurements of the total levels of phosphorylated HER1-3 in response to various doses of EGF (Figure 2A–C, markers), and various doses of HRG (Figure 2D–F). We also obtained detailed time course measurements of total levels of phosphorylated HER1-3 (Figure 3A–C), levels of phosphorylated HER1-3 in the cell interior (Figure 3D–F), and HER1-3 total protein levels (Figure 4) in response to a single specific dose each of EGF and HRG, either added separately or in combination. Corresponding datasets for the parental (HER2−3−) cell line that expresses very low levels of HER2 and HER3; a cell line that expresses HER2 but not HER3 (HER2+3−, designated 24H); and a cell line that expresses HER3 but not HER2 (HER2−3+, designated B5) are presented in Figures S3, S4, S5, S6, S7, S8 of Text S1. Note that consistent y-axis scales are used in Figures 2–4 and Figures S3, S4, S5, S6, S7, S8 of Text S1 for each measurement type to enable comparison of total receptor phosphorylation, internal phosphorylation levels and receptor mass across the four cell lines. In all, the experimental data consisted of 999 distinct measurements, with N> = 2 for each measurement. We estimated the 47 unknown model parameters by simultaneously fitting the model to all of the data described above. Distinct measurement types (receptor phosphorylation, receptor mass) were scaled appropriately to ensure that they contributed comparable amounts to the residual vector (Methods). We found that ∼188 of the optimization runs converged with an RMSE relatively close to the best overall RMSE (Figure S1 in Text S1). In order to better assess the location of the solutions in the 47-dimensional parameter space, we used k-means clustering to identify the existence of distinct solution clusters (Figure S2 in Text S1). Our analysis indicated that the solutions can be split into 7 distinct clusters (Figure S2A in Text S1). In order to examine the similarities and differences in the model' s behavior for these solutions, for each cluster we selected the parameter set that yielded the best fit (smallest RMSE), and used these 7 solutions (Figure S2B–C in Text S1) for additional analysis. In Figures 2–4 and Figures S3, S4, S5, S6, S7, S8 of Text S1 we compare the experimental data (markers) to model predictions (lines) generated using each of the 7 representative parameter sets. Predictions using the parameter set with the best overall RMSE are depicted using darker lines in these plots. As seen, the model predictions are in good agreement with the experimental data (compare lines vs. markers of same color). Further, predictions using the 7 distinct parameter sets are in good agreement with each other (compare lines of the same color). This can also be seen in Figure S9 of Text S1 where the mean and standard deviations of the 7 distinct model predictions are plotted against the experimental data for the various types of measurements. There is a strong linear relationship between model predictions and the experimental data with a slope between 0. 83 and 1. 1 for most measurement types (Figure S9 in Text S1). The exceptions are for the levels of phosphorylated HER2 in the cell interior (slope = 0. 64), and the EGFR receptor mass (slope = 0. 77). Overall, model predictions are in good agreement with the experimental data with each of the 7 solutions yielding comparable results. Parameter values for the 7 representative solutions are presented in Table S4 of Text S1. Although these solutions result in comparable fits to the various measurements (see Figures 2–4) and have similar overall RMSEs (Table S4 in Text S1), they involve substantial differences in the values of several parameters, with 20 parameters displaying a greater than 2-orders of magnitude variability. Reliability of the estimated parameters can depend on the information content of the training data, which in turn depends on the experimental design because the choice of which conditions are changed in the experiments can favor the identifiability of certain parameters. In other words, certain parameters would be more sensitive to the changed experimental conditions and this would allow for their better determination. For this reason, while some of the parameters can be extracted from the datasets reliably, we can only determine broad ranges for the rest of the parameters. The results indicate that both the dimerization affinities and the pf values are estimated with reasonable confidence for the R11 homodimer and the R12 heterodimer (in our notation Rij refers to the HERi-HERj dimer). However, there is more than two orders of magnitude variability in the dimerization and phosphorylation parameters related to the R13, R23, R22, R33 dimers (Table S4 in Text S1). Model predictions for the abundances of various receptor dimers generated using the 7 representative solutions reinforce these findings: predictions for R11 and R12 abundances fall within a narrow range, while there is considerable uncertainty in the estimates for the other dimers (Figure S10 in Text S1). Visualization of the correlation between the parameters calculated based on the 188 solutions with good RMSE values (Figure S11 in Text S1) revealed that the dimer dissociation rates and phosphorylation efficiencies (pf values) were strongly correlated with each other for the various dimer types (Figure S11 in Text S1). This is because the extent of receptor phosphorylation for each of the three HER receptors is determined by both the absolute number of dimers of each type, as well as the pf values for the dimers (e. g. , see equation in Figure 1B). Thus, estimating both the dimerization affinities and pf values by fitting the model to receptor phosphorylation data is expected to be challenging. The ability to overcome this limitation in the case of the R11 and R12 dimers is likely to be related to the availability of dose response datasets with strong HER1 and HER2 receptor phosphorylation in cell lines that express EGFR/HER1 alone, or HER1 and HER2, but not HER3. In order to independently validate the model, we collected additional data for receptor phosphorylation in various cell lines in the absence and presence of 2C4 (Pertuzumab) (Figure 5). This monoclonal antibody is considered to be a general inhibitor of HER2 dimerization due to its ability to bind the HER dimerization surface [51], [52], [53]. Model simulations for the antibody blocking experiments were performed by assuming that the addition of 2C4 renders 95% of the cellular HER2 unavailable for receptor dimerization. Note that the concentration of 10 µg/ml 2C4 used in our experiments is much higher than the Kd value for 2C4-HER2 binding [54]. Although none of the data in Figure 5 was used in “training” the mathematical model, the model does an excellent job of predicting these results. When model predictions for EGFR, HER2 and HER3 phosphorylation are plotted against experimental data from the validation experiments we obtain linear relationships with slopes of 0. 92,0. 96 and 0. 97, respectively (Figure S12 in Text S1). As before, model predictions of receptor phosphorylation based on the 7 distinct parameter sets are in excellent agreement with each other (see standard deviations of model predictions in Figure 5 and Figure S12 in Text S1). HER receptors display unique patterns of site-specific phosphorylation, adaptor protein recruitment, and downstream signaling depending upon their dimerization partners [39], [40]. Therefore, it is of interest to quantify the relative contributions of various dimer types to HER phosphorylation. We calculated the phosphorylated levels of HER1-3 emanating from various dimers by multiplying the dimer abundances with appropriate phosphorylation efficiencies (see Methods). We found that although dimer abundances could be uniquely determined for only a subset of dimers (Figure S10 in Text S1), the contributions of various dimers to HER phosphorylation could be determined with much higher confidence (Figure S13 in Text S1). Predictions for the time-dependent phosphorylation signal from the various dimer types in the HER2+3+ cells using the 7 distinct parameter sets were in good agreement with each other (Figure S13 in Text S1). Since, the phosphorylation levels were found to be relatively stable beyond 1 hour of ligand addition (Figure S13 in Text S1), we chose the t = 60 min time point for all subsequent analysis. Model predictions for dimer contributions to HER1, HER2, and HER3 phosphorylation in the HER2+3+ cells at 60 min following the addition of saturating levels of EGF and HRG are presented in Figures S14, S15, and S16, respectively in Text S1. Predictions from the 7 different solutions, in general, were in good agreement with each other. The exception was for HER2 phosphorylation where the R12 dimer was found to contribute between 48–69% of the HER2 phosphorylation signal, with a 4–5% contribution from the R22 homodimer and the rest from the R23 dimer (Figure S15 in Text S1). Since, the predictions overall were in reasonable agreement we picked the best fit parameter set – the one with the lowest RMSE (Table S4 in Text S1) – and used it to generate predictions for the other cell lines in our panel. To understand the effect of HER expression levels on the phosphorylation pattern, we calculated the relative contributions of the dimers to HER activation in each of the 4 cell lines used in our study (Figure 6). In the figure, relative dimer contributions are shown as pie charts where the size of the circles is proportional to the total phosphorylation level (Figure 6). HER1 phosphorylation was found to be consistently high in all four cell lines with the highest level in the HER2+3− cells (Figure 6A–D). As expected, ∼90% of HER1 phosphorylation in the parental (HER2−3−) cells was found to be due to the R11 homodimer, with most of this contribution coming from the species where both dimer partners were ligand-bound (R11EE; Figure 6A). In the HER2+3− (24H clone) cells, >60% of EGFR phosphorylation was from the R12 dimer (Figure 6C). Dimer contributions to EGFR phosphorylation in the HER2−3+ (B5 clone) and HER2+3+ (D20 clone) cells were similar to that in the parental and 24H cells, respectively. In other words, HER1 and HER2 expression levels were found to dictate the HER1 phosphorylation pattern, with HER3 expression having little to no effect. As expected, HER2 phosphorylation levels were predicted to be much higher in the HER2+3− and HER2+3+ cell lines compared to the HER2− cell lines (Figure 6E–H). The HER2+3+ cells were found to have the highest HER2 phosphorylation. Whereas in the HER2+3− cells 90% of HER2 phosphorylation was due to the R12 dimer (Figure 6F), there were substantial contributions from both the R12 and R23 dimers in the HER2+3+ cells (Figure 6H). Since the EGFR expression level in the HER2+3+ cells is an order of magnitude higher than the HER3 level (Table S5 in Text S1), this suggests a stronger propensity to form activated R23 dimers compared to R12 dimers. In all cases, we found that the R22 homodimer contributed less than 5% to HER2 activation. Overall, HER2 phosphorylation can be driven via interactions with either EGFR or HER3 with the latter appearing to be the preferred dimer partner. High levels of HER3 phosphorylation were found only in the HER2+3+ cell line (Figure 6I–L). The R13 dimers (EGF or HRG-bound) were found to contribute significantly to HER3 activation when HER2 levels were low (Figure 6I, 6K). In cell lines that expressed both EGFR and HER2, HER3 activation was found to be dominated by the R23 interaction, which points to the significance of this interaction for HER3 activation. Our model can be used to predict the extent of HER phosphorylation and the receptor dimerization pattern for any combination of HER1-3 expression levels. We generated model predictions over a wide range of receptor expression levels (Figure 7). To ensure the relevance of this analysis, we obtained information on the HER expression levels of various cell lines from the literature (Table S5 in Text S1). Since, most HER3-expressing cells typically display a receptor expression level of ∼40,000 molecules/cell (Figure S17 in Text S1), we fixed HER3 at this level. We varied EGFR from 103 to 106 and HER2 from 103 to 3×106 to encompass the receptor expression levels observed in various cell lines (Table S5 and Figure S18 in Text S1). HER1-3 phosphorylation levels (Figure 7A–C) and the percentage contribution from the R11 homodimer to EGFR phosphorylation (Fig 7E), the R12 dimer to HER2 phosphorylation (Figure 7F) and the R23 dimer to HER3 phosphorylation (Figure 7G) are presented in Figure 7 as a function of EGFR and HER2 expression levels. The contribution of the other dimer types to HER1-3 phosphorylation is presented in Figure S19 in Text S1. Simulation results indicate that EGFR phosphorylation increases with both EGFR and HER2 expression, with HER2 expression having a stronger effect at low to moderate EGFR expression (Figure 7A). The EGFR homodimer contributes anywhere from 0–100% of the EGFR phosphorylation signal with the actual contribution increasing with EGFR expression and decreasing with HER2 expression (Figure 7B). The contribution of the R12 dimer to HER1 phosphorylation displays the opposite pattern (Figure S19A in Text S1), while the R13 dimer contributes <10% to HER1 phosphorylation in all cases (Figure S19B in Text S1). HER2 phosphorylation increases with HER2 expression, with EGFR expression having only a minor effect (Figure 7B). The contribution of the R12 dimer to HER2 phosphorylation shows a broad range, increasing with EGFR expression (Figure 7G). The contribution of the R23 dimer shows the opposite pattern (Figure S19C in Text S1). The HER2 homodimer is predicted to contribute <15% of the HER2 phosphorylation signal in all cases (Figure S19D in Text S1). Interestingly neither the R12 contribution nor the R23 contribution is a strong function of HER2 expression (Figure 7G, Figure S19C in Text S1). Thus, the EGFR expression level is the strongest predictor of which dimer type dominates HER2 signaling. HER3 phosphorylation is found to be strongly dependent on HER2, but not EGFR expression levels (Figure 7C). Over the range of expression levels seen in actual cells (see dots in Figure 7D), >80% of the HER3 signal is predicted to be due to the R23 dimer (Figure 7D) with the remaining from the R13 dimer (Figure S19E in Text S1). In order to validate these simulation results, we compared model predictions for HER3 activation in two distinct cell lines – ADRr and ADRrE2 – with similar levels of EGFR and HER3, but distinct HER2 levels with previously published experimental data [27]. As seen, model predictions for the relative change in HER3 activation due to an increase in HER2 expression are in good agreement with the experimental data (Figure 7). Thus, our results indicated that the model predictions may be applicable to other cell lines as well. To enable cell type-specific comparisons, we have computed the phosphorylation levels of HER receptors and HER dimerization patterns for 52 distinct cell lines using their receptor expression levels compiled from the literature (Figures S20, S21, S22 in Text S1). One key aspect of receptor signaling is the prediction of how the changes in receptor phosphorylation levels would alter the activation patterns of the downstream elements of the involved signaling pathways. Here, we briefly illustrate how the constructed receptor activation model can be used to quantitatively predict the relative contributions of the HER receptor types and their dimers to the activations of Erk and Akt kinases. Erk and Akt are important regulators of the cell proliferation and mobility processes, and their activation kinetics in HME cells were subject of our earlier investigations [47], [55]. We pursued multilinear regression analysis to determine the relationship between Erk and Akt activation and HER phosphorylation by fitting the coefficients of the regression model to the data collected in our HME cell lines. This analysis was pursued in two different ways by assuming a) that the total receptor phosphorylations are the predictors, i. e. , pT (t) = b0+Σ bi * pRi (t), where pRi (t) is the contribution of receptor type i (= HER1,2, or 3) to the activation of the target protein T (= Erk or Akt) and the sum is over the receptor types, and b) that the receptor dimer contributions are the predictors [55], i. e. , pT (t) = b0+Σ bi_ji * pRi_ji (t), where pRi_ji (t) is the contribution of the dimer Rij to the phosphorylation of receptor type i and the sum is over the receptor dimer types. Comparison of the regression model predictions with the experimental data have shown that the EGFR/HER1 contribution to Erk phosphorylation (pERK) is the dominant predictor and that various receptor dimers could make comparable contributions to the prediction of pERK (Table S1; Figure S24 in Text S1). In contrast to Erk, Akt phosphorylation has a much stronger dependence to the activation through particular receptor dimers: regression analysis indicated that signaling through the HER1-HER3 receptor dimer was the dominant predictor of pAKT (Table S1; Figure S25 in Text S1). Results of this analysis were consistent with the results of other analysis methods such as clustering and targeted inhibition (Gong et al, in preparation). We have constructed a parsimonious mathematical model for HER1-3 activation that incorporates the important biochemical/biophysical steps involved in the process, and have parameterized it using the data collected in a panel of HME cells that express varying levels of HER1-3. Despite using rate constants from the literature where available, and considering an extensive dataset including total and internal receptor phosphorylation levels and receptor mass measurements as a function of ligand dose, our analysis indicates that not all aspects of the model are equally identifiable. Specifically, we find that while there is considerable uncertainty surrounding the absolute dimer abundances of all but the HER1 homo- and HER1-HER2 hetero-dimer types (Figure S10 in Text S1), the phosphorylation signal from all the dimer types can be predicted with good confidence using the model (Figure S13 in Text S1). Since the tyrosine phosphorylation levels in various dimers are the relevant quantities to consider in the context of signal transduction, the obtained results provide the needed information for probing dimer specific downstream responses. That said the lack of complete model identifiability still highlights the challenges encountered in the construction and parameterization of models for biomolecular networks. Additionally, as in almost all of the earlier studies, possible location-dependence of the HER receptor kinetics was not included in our study. Receptor placement in membrane ruffles or the corralling role of the cytoskeleton elements [56] could be important factors but such complexities cannot be captured with the design of our experiments and hence were omitted. Previous modeling studies of the co-expression of multiple HER receptors considered both receptor activation and downstream signaling (Erk and Akt activation) as part of an integrated analysis [43], [45]. They involved the use of a single cell type [45], or multiple distinct cell types [43]. These models are useful because they represent comprehensive quantitative frameworks for assembling information regarding HER-mediated signaling, and serve to document the various steps in the process. They have also been utilized in subsequent studies that have focused on therapeutic targets for HER3-mediated signaling [27], [28]. However, due to the large scope of these models, model identifiability is a challenge as noted by the authors themselves [27], [43]. Here, we adopted an alternate approach to establish the quantitative link between HER expression levels and downstream signaling: we constructed a relatively detailed mechanistic model for HER activation, since the available information and datasets allowed us to do so. As a next step, we have used the dimer phosphorylation levels predicted by the model (an aspect that is identifiable given the data), along with data on the activation of MAPK and Akt signaling pathways to quantify differential signaling by the various HER dimers (Supplementary Material). We and others have previously used this conceptual step-wise approach to analyze HER-mediated signaling in cells that co-express EGFR and HER2 [55], [57], [58], [59]. Interestingly, as briefly discussed in the Results section, our recent analysis also indicated that activation of the pro-survival Akt pathway correlated more with HER3 signaling from the smaller R13 dimer pool compared to the substantially larger signal from the R23 dimers (to be submitted). The current model can predict the abundance of R11 and R12 dimers with much higher confidence than that of the other dimer types (Figure S10 in Text S1). The predictions for dimer abundance, and the associated parameters for these two dimer types, can be compared with previous results. We recently used a simpler model for the activation of a single receptor type (EGFR) to analyze the data for the parental HME cell line, and showed that even under saturating concentrations of EGF, <40% of the receptors dimerize and are phosphorylated [49]. These findings were in agreement with previous experimental data from our laboratory where we quantified the fraction of phosphorylated EGFR [59]. While our previous model [49] neglected the presence of low levels of HER2 and HER3 in the parental cell line, our current analysis explicitly accounts for this aspect (Table S5 in Text S1). Further, our current analysis involves the simultaneous optimization of the model using data from four different HME cell lines. Despite these differences, the current model also predicts that for the parental cell line <40% of the EGFR form homodimers following ligand stimulation, with much lower abundances for the other dimer types (Figure S23 in Text S1). The value of the phosphorylation efficiency factor pf for the R11 dimer with two bound EGF molecules estimated here (see the pf11ees values in Table S4 of Text S1) ranges from 2. 8×10−2 to 3. 4×10−2, which is in excellent agreement with the mean value of 2. 97×10−2 estimated in our previous manuscript [49]. This suggests that the four HME cell lines behave in a consistent manner, since the simultaneous analysis of these cell lines yields findings that are consistent with the analysis of the parental cell line in isolation. We have previously used a much simpler model for HER activation that neglected explicit consideration of receptor-ligand binding, receptor recycling, and sorting to analyze receptor activation in cells co-expressing EGFR and HER2 alone [59]. In that analysis we assumed that the formation of active dimers occurred in a single lumped step, and for each dimer type we used a single lumped pf value applicable to all cellular compartments [59]. The analysis indicated decreased stability of the R12 dimer compared to the R11 dimer, which contradicted the assumptions used in other modeling papers [57], [60]. Further, we found that the pf11 and pf12 values were comparable indicating that EGFR phosphorylation occurred with equal efficiency in R11 and R12 dimers, and that pf22 was an order of magnitude smaller than pf21 indicating much lower HER2 phosphorylation efficiency in the R22 homodimer compared to the R12 dimer [59]. Comparison of the dimer dissociation constants obtained in our current analysis (see ku11ees and ku12es in Table S4 of Text S1) also indicates that the R11 dimer is far more stable than the R12 dimer. However, the model predicts that EGFR phosphorylation is 6–35 times more efficient at the cell surface and 1. 5–4 times more efficient in the early endosome vesicles when the EGFR is part of the R12 dimer as opposed to the R11 dimer (see pf11ees/pf12es and pf11eei/pf12ei ratios in Table S4 of Text S1). The model, in agreement with our previous findings [59], also predicts that the HER2 phosphorylation is far more efficient in the R12 dimer compared to that in the R22 homodimer (see pf21s/pf22s ratio in Table S4 of Text S1). One way of validating these results would be to measure the absolute abundances of R11 and R12 dimers in these cell lines. While it is possible to address these using FRET or co-IP experiments, these experiments would be challenging due to the difficulties in quantitative interpretation of the FRET signal, and possible differences in antibody pull down efficiencies, respectively. We used the model trained on data from HME cells to predict the HER activation levels and dimer contributions for a range of cell lines (Figure 7; Figures S20, S21, S22 in Text S1). For these predictions we first had to obtain information on receptor expression levels in various epithelial cell lines (see Table S5 in Text S1). Interestingly, while the data revealed wide variability in the expression levels of EGFR and HER2 among the cell lines, cells that expressed HER3 did so in a relatively narrow range of ∼30,000 to 60,000 receptors/cell (Table S5 in Text S1). Perhaps, this indicates that the in vivo quantitative regulation of HER3 signaling occurs via control of the expression level of its main partner HER2 and/or via the differential regulation of its ligands [61]. We partially validated the ability of our model to predict HER phosphorylation dynamics in other cell lines by comparing our results with experimental data [27] for HER3 phosphorylation in the ADRr and ADRrE2 cell lines (Figure 8). In such extrapolations we make the implicit assumption that the rate constants for HER activation processes (receptor-ligand binding, dimerization, phosphorylation, trafficking) are similar across various cell lines, and that knowledge of receptor expression levels alone is sufficient to predict dimerization and phosphorylation patterns. We caution that the validity of the assumption may be questionable, and that extrapolations to other cell lines should be specifically validated (for e. g. , by measuring the levels of HER2 phosphorylation relative to a benchmark cell line) when quantitative predictions of dimerization patterns are desired. That said our predictions appear to be in qualitative agreement with published results. For instance, we predict that the R12 dimer is an important component of EGFR and HER2 activation with the contribution of this dimer dependent on both EGFR and HER2 expression levels. This is in agreement with the finding of Defazio-Eli et al. who used the VeraTag™ (Monogram Biosciences, South San Francisco, CA) proximity-based ligation assay to quantify EGFR and HER2 expression levels, R12 dimer abundances and phospho-R12 levels in various cells [62]. Mukherjee et al. [63] also used VeraTag assays to quantify the phosphorylation levels of various HER receptors and relative dimer abundances in panel of breast tumors with particular focus on HER3 activation. They found that the level of HER3 phosphorylation correlated strongly with the level of the R23 dimer, and that the expression level of HER2 is a strong determinant of the level of HER3-PI3K signaling [63]. This is in agreement with our finding that the HER2 expression level is the strongest determinant of HER3phosphorylation (Figure 7C), and that it is the R23 dimer that contributes significantly to HER3 activation (Figure 7F). To summarize, we have constructed and parameterized a mathematical model that can be used to predict the levels of HER phosphorylation, and the levels of various phosphorylated HER dimers as a function of the HER expression profile. We present predictions of HER1-3 phosphorylation levels and their dimerization patterns for 52 distinct cell lines (Figures S20, S21, S22 in Text S1). These results can be used to determine the dominant dimer type that contributes to HER signaling in each cell line, and hence to device optimal strategies to disrupt HER signaling in a cell lines with known HER expression levels. Importantly, model predictions can be used to determine the relative potencies of the various HER dimers to activate distinct downstream cell signaling pathways, and drive specific cell decisions. In this regard, this manuscript represents a critical piece in the effort to mechanistically link HER expression levels to receptor dimerization, activation, and eventually to the cell phenotype. The parental human mammary epithelial (HME) cell line used in this study was originally provided by Martha Stampfer (Lawrence Berkeley National Laboratory, Berkeley, CA) as cell line 184A1-1. It expresses approximately 200,000 molecules of EGFR/HER1, and much lower levels of HER2 and HER3 [47], [57], and is designated here as the HER2−3− cell line. We used retroviral transduction to insert the HER2 gene and the HER3 gene into the parental cell line to obtain the 24H (HER2+3−) and B5 (HER2−3+) cell lines, respectively. The HER3 gene was then inserted into the 24H cell line to obtain the D20 cell line (HER2+3+) that expressed all 3 receptors. We have previously described the detailed protocols used for deriving these cell lines [47]. The parental cell were maintained in DFCI-1 medium supplemented with 12. 5 ng/ml EGF (PeproTech, Rocky Hill, NJ) as described previously [64]. Growth mediums for the 24H cell line, the B5 cell line, and the D20 cell line were the same as the parental cell line except for the addition of antibiotics G418 (250 µg/ml; Invitrogen, Carlsbad, CA), puromycin (2 µg/ml; Sigma, St. Louis, MO), and both, respectively to ensure selection [47]. When cells grew to near confluency, DFCI-1 medium was replaced with bicarbonate-free DFHB minimal medium lacking all supplements but 0. 1% bovine serum albumin. Cells were then brought to quiescence for 12–18 hours before treatment. Cells were activated through the HER receptors by the addition of known concentration of EGF and/or HRG (Peprotech, Rocky Hill, NJ) followed by incubation at 37°C for fixed amounts of time from 5 to 120 min. In the dimerization blocking experiments, cells were preincubated with 10 µg/ml of monoclonal antibody 2C4 (Pertuzumab; generous gift from Genentech, Inc, San Francisco, CA) for 4 hours prior to ligand stimulation. Following stimulation, cells were then solubilized with ice cold lysis buffer (1% NP-40,20 mM pH 8. 0 Tris buffer, 137 mM NaCl, 10% glycerol, 2 mM EDTA, supplemented with 1 mM heat activated sodium orthovanadate and 1% protease inhibitor cocktail III; Calbiochem, La Jolla, CA) for 20 min. Cell lysates were collected with a scraper. Lysates were centrifuged at 13,000 rpm for 10 min at 4°C, and the supernatants were transferred into fresh microtubes. Obtained cell lysates were either analyzed immediately or stored at a −80°C freezer until needed. Phosphorylated receptor levels in the internal compartments were determined using an acid-stripping protocol, which selectively dephosphorylates cell surface receptors without altering the phosphorylation of internalized receptors [65]. Following cell stimulation with ligands, and acid stripping, cells were washed 3X with ice cold PBS and incubated at room temperature for one minute to allow surface receptor dephosphorylation. After another round of cold PBS washing, cells were solubilized and lysates were prepared as described in the previous paragraph. ELISA assays to quantify the receptor mass and phosphorylation levels were performed using the R&D DuoSet IC ELISA kits (R&D Systems Inc. , Minneapolis, MN). Two types of ELISA data were collected as a function of time following ligand addition for each of the four cell lines used in our study: The ELISA results were normalized based on the total protein present in the cell lysate (measured using the Bicinchoninic Acid protein quantitation kit, Sigma, St. Louis, MO), and were expressed in units of picograms per microgram of total lysate protein. For each cell line and treatment condition at least two independent measurements were performed, with at least two biological replicates in each experiment. The mathematical model for HER activation (Figure 1) is an extension of our recently published multi-compartment model for cells expressing EGFR alone [49]. Here, we consider the interactions between multiple members of the HER family. There are 17 types of species in the mathematical model (Figure 1A) including the ligands EGF and HRG, free and ligand-bound HER monomers, and the various possible homo- and hetero-dimers that can be formed following the addition of EGF and/or HRG. These species are allowed to exist in 3 distinct compartments – the cell surface, early endosomes (EE) and late endosomes (LE) – resulting in a total of 51 model variables. The model combines the key biochemical reactions underlying HER activation (Figure 1B) with receptor trafficking between the cellular compartments (Figure 1C) to predict receptor mass, dimerization and phosphorylation dynamics following ligand stimulation. In the model biochemical reactions leading to receptor activation – receptor-ligand binding, dimerization and phosphorylation – are allowed to occur at the cell surface and in the EE. Following exit from the EE, receptors destined for degradation become part of multivesicular bodies (MVBs) where they undergo terminal dephosphorylation prior to degradation [66], [67]. To account for this process, we include an idealized LE compartment which is a site for the accumulation of dephosphorylated receptors prior to degradation [49]. We assume that receptors in the LE do not contribute to receptor phosphorylation measurements, but contribute to receptor mass measurements. Since it is unnecessary to track the receptor activation process in the LE, biochemical reactions for this compartment were excluded from the rate equations. Our general approach is to construct a parsimonious model to avoid over-fitting of the data. We use lumped parameters or scaling factors where detailed kinetic information is unavailable. To ground the model in reality, and to facilitate parameter estimation, we employ previously determined values for rate constants where available. Explicit consideration of the various HER homo- and hetero-dimer types in their different ligand-bound states demands the specification of a large number of model parameters due to the combinatorial complexity. However, we choose this approach because quantitative information is available in the literature regarding the relative affinities of various HER dimer types for EGF and HRG [68], [69], [70]. Further, we can use reasonable simplifying assumptions regarding the trafficking properties of the different species to reduce the number of unknown parameters in the model (see below). The different reaction types in the model are briefly discussed below along with their associated assumptions, known parameter values, and unknowns. The complete governing equations for the model are presented as part of the Supporting Information. Rate expressions and parameter values used for the biochemical reactions at the cell surface and the EE are in Tables S1 and S2, respectively of Text S1. The trafficking parameters are presented in Table S3 of Text S1. Estimates for unknown model parameters obtained here by fitting to the experimental data are tabulated in Table S4 of Text S1. There are 47 unknown model parameters (described above), which include the dimer dissociation rates, the pf values for the various dimer types and the trafficking parameter δ1. Given values for the 47 parameters and the HER1-3 expression levels (Table S5 in Text S1), the model can be simulated for any given concentration of EGF and/or HRG to predict the total (pRt) and internal (pRi) receptor phosphorylation levels as well as the receptor mass (mRt) for the three HER receptor types. In order to estimate the unknown parameters, we simultaneously considered these 3 distinct measurement types for each of the 4 HME cells. We constructed scaled residual vectors (residual = model prediction−experimental data) for the pR and mR predictions by dividing each of these residuals by the maximum values measured in the phosphorylation and receptor mass measurements. This ensures roughly equal importance to the distinct measurement types during parameter estimation. We then concatenated the scaled residual vectors and used lsqnonlin – the MATLAB (Natick, MA) nonlinear least squares regression function to determine optimal parameter values. During optimization, initial guesses for the unknown kinetic rate parameters were generated by sampling the parameters from broad uniform distributions: guesses for the dimer dissociation rates ranged from 10−3 to 103; the pf values from 10−6 to 1 and δ1 from 0. 1 to 10. To ensure convergence, we adopted a progressive optimization approach. HER1-related parameters were first estimated using the parental (HER2−3−) cell line. These values were then used as initial guesses in the estimation of parameters related to HER1-HER2 interactions using the parental and HER2+3− cell lines; and parameters related to HER1-HER3 interactions using the parental and HER2−3+ cell lines. The parameter sets obtained from these simpler optimizations were used as initial guesses in the final set of optimization runs where all 47 parameters were estimated by simultaneously considering the data from all four cell lines. These optimization runs were repeated 500 times with randomly generated initial guesses for parameters related to the HER2-HER3 interaction. We used the overall root-mean-squared error (RMSE) between the experimental data and model predictions to assess the goodness of the fit. Following parameter estimation, model predictions were generated for dimer abundances, the HER1-3 phosphorylation signal from various dimers (product of abundance and appropriate pf value), and the total HER1-3 phosphorylation levels (sum of the relevant dimer phosphorylation signals). These results were used to compute the fractional contribution of the various dimer types to the phosphorylation of the HER1-3 receptors. Predictions were generated both for HME cells as well as a panel of 48 cell lines for which HER expression levels were compiled from the literature (Table S5 in Text S1). Unless specified otherwise, all predictions represent the HER dimerization and activation pattern at t = 60 min following the addition of saturating doses of both ligands, specifically 30 ng/ml EGF and 100 ng/ml HRG.

A family of cell surface molecules called the HER receptor family plays important roles in normal physiology and cancer. This family has four members, HER1-4. These receptors convert signals received from the extracellular environment into cell decisions such as growth and survival – a process termed signal transduction. In particular, HER2 and HER3 are over-expressed in a number of tumors, and their expression levels are associated with abnormal growth and poor clinical prognosis. A key step in HER-mediated signal transduction is the formation of dimer complexes between members of this family. Different dimer types have different potencies for activating normal and aberrant responses. Prediction of the dimerization pattern for a given HER expression level may pave the way for personalized therapeutic approaches targeting specific dimers. Towards this end, we constructed a mathematical model for HER dimerization and activation. We determined unknown model parameters by analyzing HER activation data collected in a panel of human mammary epithelial cells that express different levels of the HER molecules. The model enables us to quantitatively link HER expression levels to receptor dimerization and activation. Further, the model can be used to support additional quantitative investigations into the basic biology of HER-mediated signal transduction.

Abstract Introduction Results Discussion Methods

mechanisms of signal transduction signaling networks membrane receptor signaling growth factors signaling pathways biophysics simulations biochemistry simulations proteins biology biophysics systems biology biochemistry biochemical simulations signal transduction transmembrane proteins biophysic al simulations computational biology molecular cell biology signaling cascades

2013

Model-Based Analysis of HER Activation in Cells Co-Expressing EGFR, HER2 and HER3

Little is known about immediate phases after viral infection and how an incoming viral genome complex counteracts host cell defenses, before the start of viral gene expression. Adenovirus (Ad) serves as an ideal model, since entry and onset of gene expression are rapid and highly efficient, and mechanisms used 24–48 hours post infection to counteract host antiviral and DNA repair factors (e. g. p53, Mre11, Daxx) are well studied. Here, we identify an even earlier host cell target for Ad, the chromatin-associated factor and epigenetic reader, SPOC1, recently found recruited to double strand breaks, and playing a role in DNA damage response. SPOC1 co-localized with viral replication centers in the host cell nucleus, interacted with Ad DNA, and repressed viral gene expression at the transcriptional level. We discovered that this SPOC1-mediated restriction imposed upon Ad growth is relieved by its functional association with the Ad major core protein pVII that enters with the viral genome, followed by E1B-55K/E4orf6-dependent proteasomal degradation of SPOC1. Mimicking removal of SPOC1 in the cell, knock down of this cellular restriction factor using RNAi techniques resulted in significantly increased Ad replication, including enhanced viral gene expression. However, depletion of SPOC1 also reduced the efficiency of E1B-55K transcriptional repression of cellular promoters, with possible implications for viral transformation. Intriguingly, not exclusive to Ad infection, other human pathogenic viruses (HSV-1, HSV-2, HIV-1, and HCV) also depleted SPOC1 in infected cells. Our findings provide a general model for how pathogenic human viruses antagonize intrinsic SPOC1-mediated antiviral responses in their host cells. A better understanding of viral entry and early restrictive functions in host cells should provide new perspectives for developing antiviral agents and therapies. Conversely, for Ad vectors used in gene therapy, counteracting mechanisms eradicating incoming viral DNA would increase Ad vector efficacy and safety for the patient. DNA viruses require nuclear transport of their genomes to productively infect the host cell and initiate efficient replication. Simultaneously, introduction of viral nucleic acids into the host cell nucleus triggers danger signals, and activates DDR (DNA damage response) prior to cell cycle arrest, and apoptosis. Many viruses counteract these regulatory measures in infected cells in order to ensure productive infection, which necessitates proper viral gene expression and adequate progeny synthesis [1]. In line with this, Ad (Adenoviruses) have gained functions to modulate DSBR (double-strand break repair), apoptosis, cellular gene expression, and host cell immune responses. The incoming viral genome is complexed with core factors and capsid protein VI after endosomal release. Our recent work showed that Ad DNA remains transcriptionally inactive until protein VI mediates activation of the viral E1A promoter by functionally inhibiting chromatin-associated transcription factor Daxx [2]. Ad E1A (early region 1A protein) is the first protein expressed after infection, playing an essential role in subsequent transcriptional activation, and induction of cell cycle progression [3]. Recently, E1A was seen to interact with the cellular PML-II isoform, and together this complex elevates transcription from Ad promoters [4]. E1B-55K (early region 1B 55 kDa protein) supports efficient viral replication by inhibiting anti-proliferative processes induced by the host cell [5]. However, additional functions of E1B-55K mainly require its interaction with E4orf6 (early region 4 open reading frame protein 6). So far, several reports have demonstrated that Ad E4orf6 connects Ad E1B-55K to an E3 ubiquitin ligase complex in the nucleus, containing cellular factors Rbx1/Roc1/Hrt1, Elongin B/C, and either Cullin 2 or Cullin 5 [6]. Recent work has shown that E1B-55K is the substrate recognition unit, while E4orf6 assembles the cellular components, and this functional complex sequesters cellular target proteins into a proteasomal degradation pathway C [6], [7], [8], [9], [10], [11], [12], [13]. Another hurdle to the viability and propagation of DNA viruses is imposed by host DDR and repair machinery. To counter this, several DNA viruses have acquired early viral genes that degrade or redistribute key cellular factors of the repair machinery to protect viral genome integrity [6]. Ad-mediated DDR inhibition by the Ad E1B-55K/E4orf6 E3 ubiquitin ligase complex and E4orf3 (early region 4 open reading frame protein 3) -dependent relocalization of the MRN complex into nuclear tracks and cytoplasmic inclusions together block concatemer formation and DNA damage signaling, therefore allowing productive infection and efficient virus growth [14], [15], [16], [17], [18]. The human SPOC1 (survival-time associated PHD protein in ovarian cancer 1/PHF13) protein has been identified as a novel regulator of DDR and chromatin structure [19], [20]. The spoc1 gene is located in chromosomal region 1p36. 23, a region with frequent heterozygous deletions implicated in tumor development and progression [21], [22]. Consistent with this, elevated SPOC1 RNA levels in primary and recurrent epithelial ovarian cancers have been associated with decreased survival rates in patients [23]. Moreover, SPOC1 RNA can be detected in most human tissues, with the highest levels in the testis, where it has been exclusively detected in spermatogonia [23], [24]. SPOC1 is a nuclear protein with a PHD (plant homeodomain), predicted to bind H3K4me2/3 and to regulate chromatin-specific interactions [20], [25]. In line with this, Kinkley and co-workers observed that SPOC1 is dynamically associated with chromatin and induces chromosome condensation to regulate proper cell division [20]. Particularly, SPOC1 plays a role in radiosensitivity and DNA repair by selective modulation of, and functional cooperation with chromatin modifiers and DDR regulators [19]. There is also evidence that SPOC1 is recruited to DSBs and regulates the kinetics of DSB repair and cellular radiosensitivity. It is proposed that H3K4me2/3-containing chromatin can be converted into more compact chromatin by SPOC1-mediated increase of H3K9 KMTs and H3K9me3. Hence, loss of SPOC1 promotes chromatin decondensation, and is associated with increased levels of DDR transducers and efficient DNA repair. The correlation between SPOC1 protein levels and H3K9me3, as well as expression of several H3K9 KMTs, implicates SPOC1 functions in both chromatin condensation and DDR [19]. In sum, SPOC1 is a multifunctional protein with a additional role in stem cell differentiation, oncogenesis, chromatin structure, and DNA repair processes. Here, we identified SPOC1 as a novel host restriction factor targeted during viral infection. SPOC1 protein levels decreased in Ad infected cells, which we could attribute to proteasomal degradation mediated by the E1B-55K/E4orf6 E3 ligase complex. Not only did SPOC1 interact with E1B-55K, but also Ad5 E2A-DBP, a marker for nuclear Ad replication sites, as seen by co-immunoprecipitation and immunofluorescence assays. When SPOC1 was overepressed, Ad virus yield, viral DNA synthesis, and viral protein synthesis decreased; reporter gene and chromatin immunoprecipitation assays showed that SPOC1 repressed gene expression at the level of transcription. Intriguingly, interaction with structural viral core protein pVII initially stabilized SPOC1 protein levels before expression of any viral early proteins. SPOC1 induces chromosome condensation [20], stimulates cellular gene silencing and influences DDR, which potentially contributes to its transformation potential [19]. Since host DDR prevents viability and propagation of DNA viruses, Ad efficiently targets a multitude of host cell DSB repair regulatory factors in order to promote productive infection [9], [10], [17], [26], [27]. Based on these data, we examined SPOC1 protein levels in infected human cells, and observed SPOC1 levels reproducibly reduced after 24 hours post infection (Fig. 1A). This decrease was even more pronounced at a higher multiplicity of infection (data not shown), and was comparable to the reduction of already known targets of the E1B-55K/E4orf6 E3-ubiquitin ligase complex [6]. To discover whether SPOC1 is a new host cell substrate of the Ad5 E3 ubiquitin ligase complex, we determined SPOC1 protein concentrations in wild type (H5pg4100), and mutant-virus infected cells lacking either E1B-55K (H5pm4149) or E4orf6 (H5pm4154) (Fig. 1B). As anticipated, SPOC1 was dramatically reduced in cells infected with the wild type virus H5pg4100 (Fig. 1B, lane 2), whereas the cellular protein accumulated to levels comparable to non-treated cells in infected cells lacking E1B-55K (H5pm4149; Fig. 1B, lane 3). Similarly, SPOC1 was not reduced in cells infected with the E4orf6-minus virus mutant H5pm4154 (Fig. 1B, lane 4). We further examined SPOC1 protein levels in H1299 cells infected with an E4orf6 virus mutant (H5pm4139) carrying point mutations in the BC Box that abrogate the formation of the Ad5 E3 ubiquitin ligase complex, by inhibition of E4orf6 binding to Elongins B and C. Again activity of the E1B-55K/E4orf6 E3 ligase complex apparently played a role in reducing SPOC1 protein concentrations (Fig. 1B, lane 5). Consistent with previous publications, Mre11 was not degraded during lytic infection with the BC Box virus mutant H5pm4139 (Fig. 1B, lane 5). Our results confirm that the formation and ligase activity of the E1B-55K/E4orf6 ubiquitin complex are essential to reduce SPOC1 protein levels. To confirm that we had identified SPOC1 as a novel target of virus-induced proteasomal degradation via E1B-55K/E4orf6-dependent E3 ligases, we treated infected cells with the proteasome inhibitor MG-132 (Fig. 2A). Functionally inhibiting host cell proteasomes abolished Mre11 and SPOC1 reduction seen in wild type infected cells, supporting the role of ubiquitin proteasome system in SPOC1 degradation. We finally validated our findings by analyzing the levels of SPOC1 protein in transfected cells (Fig. 2B). Expression of E1B-55K alone (Fig. 2B, lane 2) or E4orf6 alone (Fig. 2B, lane 3) had no effect on steady-state concentrations of endogenous SPOC1 protein, while expression of both E1B-55K and E4orf6 (Fig. 2B, lane 4) diminished SPOC1 steady-state concentrations to levels similar to virus-infected cells (Fig. 1 and 2A). Since E1B-55K is the substrate recognition unit of the SCF-like E3 ubiquitin ligase, we next tested whether E1B-55K interacts with the endogenous SPOC1 protein. As anticipated, in E1B-55K-transfected human H1299 cells E1B-55K co-immunoprecipitated with SPOC1-specific antibody, revealing an interaction between these factors (Fig. 3A, lane 2). No E1B-55K signal was observed in the corresponding negative controls (Fig. 3A, lane 1). The next question was whether E1B-55K interferes with the intracellular localization of SPOC1. Consistent with previous findings, immunofluorescence analysis in E1B-transfected human cells revealed that wild type E1B-55K protein localizes in the cytoplasm, mostly concentrated in perinuclear bodies Fig. 3B, panel F, H; [28], [29], [30], [31], [32], [33], [34]. In contrast to the diffuse nuclear localization in non-treated cells (Fig. 3B, panel C, D), SPOC1 was completely sequestered into the E1B-containing aggregates in the presence of the viral protein (Fig. 3B, panel G, H). To gain better understanding of SPOC1 role during Ad infection, we investigated the subcellular localization of SPOC1 in infected human DLD1 cells, which stably overexpress SPOC1 after doxycyclin treatment to overcome degradation of the cellular factor (Fig. S1). In mock-infected cells, SPOC1 is diffusely distributed inside the nucleus in approximately 90% of the cells investigated (n = 50; data not shown). When monitored together with Ad5 E2A-DBP, a marker for Ad replication sites in the nucleus, at 24 to 48 hours post infection SPOC1 colocalized with sites of viral replication in approximately 60% of cells (n = 50; Fig. 4A, panels D, H, L and P). We confirmed our observation by monitoring SPOC1/E2A-DBP association using immunoprecipitation analysis after wild type and mutant virus infection in DLD1 and U2OS cells stably overexpressing SPOC1 after doxycyclin treatment (Fig. 4B). As expected, E2A-DBP co-immunoprecipitated with SPOC1-specific antibody, revealing an interaction in all infected SPOC1-overexpressing human cells (Fig. 4B, lanes 2,3, 4,6, 7,8), whereas no signal was obtained in the corresponding negative controls (Fig. 4B, lanes 1 and 5). We also were able to confirm the SPOC1/E1B-55K interaction within the infected SPOC1-induced cells. To further test specificity of the SPOC1 antibody used in all of the coIPs, we tested SPOC1 binding after proteasomal degradation of SPOC1 (Fig. S2). Therefore, we infected H1299 cells with Ad5 wild type virus (H5pg4100) and virus mutants depleted for either E1B-55K (H5pm4149) or E4orf6 (H5pm4154) expression. After 24 and 48 hors post infection, total cell extracts were prepared. E1B-55K and E2A/DBP were immunoprecipitated using rabbit polyclonal SPOC1 antibody. Proteins were separated on 10% SDS-PAGE and visualized by immunoblotting. Input levels of total-cell lysates and co-precipitated proteins were detected using monoclonal antibody 2A6 (E1B-55K), B6-8 (E2A/DBP), SPOC-1-specific rat monoclonal antibody, and mouse monoclonal antibody AC-15 (β-actin) as a loading control. As expected, E2A-DBP co-immunoprecipitated with SPOC1-specific antibody (Fig. S2, lanes 3,5–8), only revealing binding in infected cells, where SPOC1 could not be degraded due to the time point (Fig. S2, lane 3) or the absence of either E1B-55K (Fig. S2, lanes 5 and 6) or E4orf6 (Fig. S2, lanes 7 and 8). We also were able to confirm the specific SPOC1/E1B-55K interaction within the infected cells expressing SPOC1 and E1B-55K (Fig. S2, lanes 3,7 and 8). Together these data improved our co-immunoprecpitation analysis and reassured specificity of the SPOC1 antibody usd in these experiments. Together, our results show that the cellular factor SPOC1 associates with E2A-DBP within Ad5 replication compartments in virus-infected human cells. This novel observation prompted us to investigate whether SPOC1 is linked to Ad transcriptional regulation during productive infection. Such recruitment to nuclear sites associated with viral genome replication suggests that SPOC1 is involved in regulating Ad5 gene expression. To better analyze the role of SPOC1 during Ad5 infection, we performed experiments in SPOC1-inducible human colon carcinoma (DLD1) and osteosarcoma (U2OS) cell lines, which express exogenous SPOC1 after doxycyclin (dox) treatment. Prior to our analysis, we investigated SPOC1 protein expression in the absence and presence of dox (Fig. 5A). Additionally, proliferation of the SPOC1 overexpressing cells was quantified, revealing no significant difference in comparison to the uninduced controls (Fig. S1). To assess the effect of SPOC1 on overall virus growth, we determined total virus yield after SPOC1 induction in DLD1 cells (Fig. 5B). SPOC1 overexpression reduced progeny production seven- (48 h p. i.) to eight-fold (72 h p. i.) compared to non-treated DLD1 control cells (Fig. 5B). When we analyzed Ad5 progeny production in SPOC1 inducible U2OS cells we observed similar effects (Fig. 5B). These results suggest that SPOC1 mediates repressive effects during Ad5 infectious cycle. To further validate this hypothesis, expression of viral early and late proteins was monitored at different time points after infection (Fig. 5C). Consistent with Ad5 progeny production, protein synthesis was inefficient in SPOC1 overexpressing DLD1 and U2OS cells (Fig. 5C). Particularly expression of E1A was affected, since it was substantially higher in non-treated DLD1 (Fig. 5C, upper panel lanes 1–6) and U2OS (Fig. 5C, lower panel lanes 1–6). Similar effects were observed when monitoring viral DNA synthesis in infected cells. As with Ad5 progeny production (Fig. 5D), DNA synthesis was less efficient in SPOC1 induced cells than in uninduced cells (Fig. 5D). Next we investigated whether Ad transcription is negatively regulated by SPOC1 expression. Therefore, we analyzed whether early and late viral mRNA levels are affected by SPOC1 overexpression in Ad wild type virus infected cells. We observed that viral early E1A and E1B mRNA production is lower in SPOC1-induced cells, compared to cells lacking treatment with doxycyclin (Fig. 5E). Similar results were obtained for hexon mRNA expression, suggesting either an impact of enhanced synthesis of early gene products or direct repression of the late promoter (Fig. 5E). Moreover, we tested cellular mRNA levels to exclude non-specific repression of gene expression by SPOC1 expression affecting cellular RNA Pol II transcription (Fig. 5E). Our results show similar GAPDH mRNA synthesis in the cells tested, indicating that SPOC1 expression does not reflect an overall reduction in cellular RNA pol II transcription. To validate these observations and to exclude non-specific impact of the overexpression studies, we depleted SPOC1 expression by transient knock-down with siRNA (Fig. 6A) and analyzed Ad5 replication. Virus growth was enhanced three-fold (24 h p. i.) to five-fold (48 h p. i.) in the absence of SPOC1, compared to control cells (Fig. 6B). Consistent with data observed in DLD1 and U2OS cells, early and late viral protein synthesis was more efficient in SPOC1-depleted cells (Fig. 6C). DNA synthesis was monitored and showed two-fold more efficient synthesis of viral DNA in SPOC1-depleted cells compared to the control cells (Fig. 6D). To further investigate a role of SPOC1 in Ad transcription and viral mRNA synthesis, we analyzed whether early and late viral mRNA levels are affected by the absence of SPOC1 expression in infected cells (Fig. 6E). We observed that viral early E1A and E1B mRNA production is two-fold reduced in SPOC1 expressing cells, compared to cells lacking this cellular transcription factor. Results obtained for hexon mRNA expression showed three-fold difference in viral mRNA expression (Fig. 6E). As shown above, we did not observe altered GAPDH mRNA levels, indicating that SPOC1 depletion does not affect cellular transcription in general (Fig. 6E). To exclude that SPOC1 expression impacts on virus entry, we analyzed whether capsid protein VI from incoming Ad capsids show altered accumulation in the host cell nuclear fraction. Using nucleo-cytoplasmic fractionation, we observed rapid protein VI accumulation in the nuclear fraction after infection of U2OS control cells and cells treated with doxycyclin to induce SPOC1 expression (Fig. S3). Together, these results indicate that cellular factor SPOC1 is a potent repressor of Ad5 growth and thus represents a novel host cell restriction factor during Ad productive infection. Due to the fact that SPOC1 is a chromatin bound protein, we next questioned whether SPOC1-dependent Ad restriction occurs on the transcriptional level. To assess this, we performed reporter gene assays with luciferase expression vectors under the control of specific Ad promoters in SPOC1-induced and non-induced DLD1 and U2OS cells. In the absence of other viral factors, SPOC1 was able to repress luciferase expression from all the viral promoters tested (Fig. 7A). To exclude non-specific repression of gene expression, we investigated whether we can detect SPOC1-dependent effects on the cellular E2F-dependent H2A promoter (Fig. S4). In contrast to the viral reporter constructs, co-expression of SPOC1 did not affect transcription from the cellular promoter (Fig. S4, lanes 3 and 4). The inhibitory effect of SPOC1 expression is specific at least for the promoter sequences tested so far. To further demonstrate direct association of SPOC1 with Ad promoter sequences, we performed chromatin immunoprecipitation assays from wild type virus infected cells, using SPOC1-specific monoclonal antibody, unrelated IgG control antibody and Ad promoter-specific oligonucleotides (Fig. 7B). The results show that SPOC1 is associated with E1A, E1B, E2early and MLP Ad promoters in infected cells (Fig. 7B). Conclusively, these results indicate that SPOC1 is a component of host cell antiviral mechanisms, playing an important role in the Ad gene expression program through transcriptional repression of viral promoters. SPOC1 has been previously reported to bind histone H3 and to be capable of promoting repressive epigenetic transitions. Since Ad major core protein VII shares homology with the N-terminal regulatory tail of histone H3 [35], we investigated if SPOC1 may also associate pVII. Whole extracts from H1299 cells that had been transfected with plasmid DNA expressing pVII products (Fig. 8A, lanes 1–3) as well as Ad wild type infected lysates (Fig. 8A, lanes 4 and 5) were immunoprecipitated with SPOC1 specific ab and analyzed by Western blotting using anti-HA mab (Fig. 8A). We detected co-immunoprecipitation of SPOC1 with pVII, when both factors were present (Fig. 8A, lanes 2,3 and 5). As core protein pVII remains associated with the viral genome during entry [36], we performed further pVII transfection studies mimicking immediate early phase of infection (Fig. 8B). First, we observed a significant increase in SPOC1 protein levels (Fig. 8B, lane 2); and repressive H3K9me3 histone marks as expected due to other reports [19]; Fig. 8B, lane 2. Next, we carried out additional infection studies with H5pg4100 Ad wild type virus infected cells (Fig. 8C). Again, we monitored SPOC1 and H3K9me3 protein levels and detected loss of H3K9me3, most presumably due to SPOC1 reduction by proteasomal degradation via E1B-55K/E4orf6 E3 ubiquitin ligases (Fig. 8C, lane 2). As pVII was reported to occupy Ad DNA during the immediate early phase of infection [36], [37], [38], [39], our results indicate that during early Ad infection, pVII likely protects its genome from SPOC1 mediated repressive silencing, prior to the onset of transcription by pVII removal from the genome [40], [41] including loss of pVII-bound repressive factor SPOC1. To determine whether SPOC1 is a key player in antiviral defense in general, and whether strategies to restrain SPOC1 are conserved among different human pathogenic viruses, we tested protein expression in permissive cell lines infected with HSV-1 (herpes simplex virus type 1; Fig. 9A), HSV-2 (herpes simplex virus type 2; Fig. 9A), HIV-1 (human immunodeficiency virus type 1; Fig. 9B) or HCV (human hepatitis C virus; Fig. 9C). To evaluate efficient infection, we monitored viral protein levels of HSV-1 nucleocapsid protein crossreacting with HSV-2 nuclear protein (Fig. 9A), HIV-1 p24 (Fig. 9B) or HCV NS5A (Fig. 9C). Intriguingly, we detected significantly reduced SPOC1 protein levels in all virus-infected cells (Fig. 9). Taken together, functional inhibition of the antiviral host factor SPOC1 by Ad early proteins E1B-55K and E4orf6 can also be achieved by other viral factors expressed in the course of HSV-1, HSV-2, HIV-1 and HCV infection. Viruses exploit cellular pathways for their own benefit, often achieved by providing high affinity binding sites on viral factors that recruit key regulatory proteins from cellular pathways to outcompete their physiological binding partner. This strategy allows viruses to infect cells and establish efficient replication with nothing more than the incoming viral genome and viral capsids. Besides transport of the genome to appropriate replication sites, viruses also need to assure decondensation and transcriptional activation of the viral genome, which in most cases has been packed and stored in the most economical way. Once the viral genome becomes transcriptionally activated new viral proteins are synthesized providing the virus with the capacity to reprogram the cell for viral replication. All of the very early steps during viral infection represent an essential moment in establishing productive infection of all human pathogenic viruses, but are as yet poorly characterized. In this report, we show that the cellular protein SPOC1 is tightly regulated in the course of productive Ad infection, identifying SPOC1 as a antiviral restriction factor in cellular defense. Ad possesses strategies to neutralize SPOC1 via an E1B-55K/E4orf6-dependent proteasomal degradation pathway. Moreover, functional inhibition of SPOC1 seems to be conserved among different human pathogenic viruses. Despite the well-characterized functions of Ad early genes and core/capsid proteins, it is still unclear how Ad transcription is initiated in detail. To summarise our findings we have put together a scheme of how various factors may interact at early stages after Ad infection (Fig. 10). Initially, the genome enters the cell as a highly condensed, transcriptionally inactive nucleoprotein complex, assembled with capsid and core proteins pVI and pVII, identified to recruit cellular factors to the Ad genome Fig. 10; [2], [42]. Recently, we reported transactivating properties of protein VI involving a conserved PPxY motif required for binding to ubiquitin ligases of the Nedd4 family of E3 ubiquitin ligases, prior to Ad-dependent depletion of Daxx/ATRX dependent transcriptional restriction Fig. 10; [2]. The Ad major core protein VII remains bound to the Ad genome during the early phase of infection and is subsequently released due to transcription Fig. 10; [37]; however the duration and amount of pVII complexed with the viral genome is still unclear. Moreover, it also remains elusive whether complete disassociation of pVII from viral DNA is required for active transcription. Nevertheless, pVII is the most abundant structural component of the viral core, is strongly associated with viral DNA in a sequence-independent manner [43], and shares homology with the N-terminal regulatory tail of histone H3 [35]. When this viral factor is imported into the nucleus together with the viral genome, it apparently packages the incoming viral DNA into chromatin-like structures Fig. 10; [37], [44], [45], [46], [47]. SPOC1 is a nuclear PHD-protein, predicted to bind H3K4me2/3 and to regulate chromatin-specific interactions [20], [25]. Therefore, SPOC1 is dynamically associated with chromatin, and plays a major role in chromosome condensation to regulate proper cell division [20]. It is proposed that H3K4me2/3-containing chromatin is converted into more compact chromatin by SPOC1-mediated increase of H3K9 KMTs (lysine methyltransferases) and H3K9me3 [19]. We observed an association between SPOC1 and viral promoter regions; and with Ad core protein pVII in the absence of any other viral protein, and during Ad infection (Fig. 7B; Fig. 8; Fig. 10). We also observe SPOC1-mediated pVII stabilization, 48 hours post transfection (Fig. 8) and hypothesize that pVII cooperation with viral DNA protects the incoming viral genome from immediate early checkpoint signaling. In accordance with recent observations by Karen and co-workers, we believe that pVII prevents the onset of DNA repair signals prior to efficient viral gene expression [36]. These ideas are further supported by recent evidence that SPOC1 expression levels have a strong impact on DDR, DNA repair, and cellular radiosensitivity [19]. It was observed that SPOC1 and associated factors are recruited to DSBs to modulate chromatin structure as well as DDR. Due to SPOC1 binding and stabilization of repressive factors, i. e. H3K9 KMTs and H3K9me3, this antiviral protein helps conversion into a compact chromatin status (Fig. 10). Conversely, loss of SPOC1 releases several factors and repressive histone marks, promoting chromatin decondensation. In the early phase of Ad productive infection SPOC1 is recruited to the viral genome and Ad5 replication centers (Fig. 4; Fig. 7B). The SPOC1-rich chromatin environment promotes the activation and recruitment kinetics of DDR transducers (γH2AX, pATM) [19]. However, the Ad core protein pVII masks the genome and antagonizes recognition by host cell repair factors (Fig. 10). Subsequently, core protein VII and pVII bound SPOC1 is released from viral DNA prior to synthesis of Ad replication proteins E2A-DBP, resulting in recognition of viral origins of DNA replication. The replication proteins presumably compete with components of DDR for binding to Ad sequences to form a functional preinitiation complex prior to DDR being inhibited by E1B-55K, E4orf6 and E4orf3. Interestingly, the correlation between SPOC1 protein levels and H3K9me3 implies that Ad may regulate expression of several H3K9 KMTs. H3K9me3 is not only an epigenetic mark characteristic of heterochromatin, but is also the direct binding platform of cellular and/or viral factors that regulate heterochromatin compaction and spreading [48], [49]. In line with this, Gupta and coworkers recently showed that Ad-dependent E1B-55K/E4orf6 E3 ubiquitin ligase induces proteasomal degradation of the acetyltransferase TIP60 [10]. TIP60 chromodomain binding to H3K9me3 at DSBs is required for its activation [50], and is a prerequisite for efficient ATM activation by acetylation. E1B-55K/E4orf6 might interfere with TIP60 binding to H3K9me3 and its activation, possibly contributing to the observed reduction in activated ATM and delayed DDR. Furthermore, Ad-mediated SPOC1 (Fig. 1 and 2) and TIP60 depletion would promote chromatin relaxation and decreases accessibility of H3K9me3 for TIP60, leading to reduction in ATM activation. So far, several histone H3K9 KMTs were shown to be modulated by SPOC1 (SETDB1, G9A, GLP) [19]. After the onset of viral transcription and efficient E1B-55K/E4orf6 expression, we observe efficient degradation of SPOC1 host factor (Fig. 1 and 2), promoting efficient reduction of H3K9me3 repressive histone marks (Fig. 8). Our findings provide hints that this complex might be altered in stability and activity due to Ad-dependent proteasomal degradation of SPOC1 (Fig. 1,2 and 8), since these KMTs form a multi-subunit complex that is destabilized by depleting the individual components [51]. In this report, we show that SPOC1 is tightly regulated in the course of productive Ad infection to neutralize host cell defense processes. SPOC1 represents a novel antiviral restriction factor, being associated with Ad major core protein VII and therefore being removed from the viral genome during the very early phase of infection prior to the onset of gene expression during productive virus life cycle (Fig. 10). In other words SPOC1 apparently plays a biphasic role during Ad life cycle by regulating determinants of chromatin compaction and damage response immediately after viral genome entry into the host nucleus, prior to activation of Ad transcription leading to its proteasomal degradation by newly expressed viral proteins. Besides Ad, different human pathogenic DNA and RNA viruses can impose reduction on SPOC1 protein levels. Our findings provide a model of how viruses might antagonize intrinsic SPOC1-mediated antiviral responses of their host cells, and create novel awareness for general antiviral restriction factors. These future insights into the immune evasion strategies acquired by viruses and other human pathogens mediated within the host will contribute to identifying new therapeutic strategies and targets to limit or prevent human pathogenic virus-mediated diseases and mortality of patients. HEK293 [52], H1299 [53], SPOC1-inducible U2OS and DLD1 cells [19] were grown in Dulbecco' s modified Eagle' s medium supplemented with 10% fetal calf serum (FCS), 100 U of penicillin and 100 µg of streptomycin per ml in a 5% CO2 atmosphere at 37°C. For HepaRG cells, the media was supplemented with 5 µg/ml of bovine insulin and 0. 5 µM of hydrocortisone. Ad proteins were expressed from their respective complementary DNAs under the control of the CMV immediate-early promoter, derived from the pcDNA3 vector (Invitrogen) to express Ad wild type E1B-55K and E4orf6 [54], [55]. cDNAs encoding E1B-55K-products from each Ad type were cloned as described previously [56], [57], [58]. Transient transfections with luciferase reporter constructs were performed as described previously [2]. pRE-Luc and the RGC-firefly luciferase reporter plasmids pGalTK-Luc have been described previously [59]. SPOC1 wild type protein was expressed from pcDNA4TO-SPOC1 constructs. H5pg4100 served as the wild type Ad5 parental virus in these studies [60]. The mutant viruses H5pm4149 and H5pm4154 were generated as described recently [61]. Both viruses carry stop codons and do not express the respective viral protein [62]. E4orf6 BC-box mutant virus H5pm4139 was generated, resulting in a drastically reduced ability to associate with the Ad5 E3 ubiquitin ligase complex compared to the E4orf6 protein from wild type virus [61]. Viruses were propagated, titrated and infected as described previously [63]. Virus yield was determined by quantitative E2A-72K immunofluorescence staining and viral DNA replication was monitored by quantitative PCR exactly as described previously [13]. ChIP analysis was performed as described previously [2]. The average Ct-value was determined from triplicate reactions and normalized with standard curves for each primer pair. The identities of the products obtained were confirmed by melting curve analysis. For dual luciferase assays, subconfluent cells were transfected prior to preparation of total cell extracts (48 h) [13]. RGC firefly luciferase and pGL-H2A-promoter activity was assayed with lysed extract in an automated luminometer (Berthold Technologies). All samples were normalized for transfection efficiency by measuring Renilla-luciferase activity. All experiments shown were performed in triplicate and data are presented as mean values. Subconfluent cells were infected with wild-type virus and harvested at 24 h p. i. . Total RNA was isolated with Trizol reagent (Invitrogen) as described by the manufacturer. The amount of total RNA was measured and one microgram of RNA was reverse transcribed using the Transcriptor High Fidelity cDNA Synthesis Sample Kit from Roche including anchored-oligo (dT) 18 primer specific to the poly (A) +RNA. Quantitative real-time PCR was performed with a first strand method in a Rotor-Gene 6000 (Corbett Life Sciences, Sydney, Australia) in 0. 5 ml reaction tubes containing a 1/100 dilution of the cDNA template, 10 pmol/µl of each synthetic oligonucleotide primer, 12. 5 µl/sample Power SYBR Green PCR Master Mix (Applied Biosystems). The PCR conditions were as follows: 10 min at 95°C, 55 cycles of 30 s at 95°C, 30 s at 55 to 62°C (depending upon the primer set) and 30 s at 72°C. The average Ct value was determined from triplicate reactions and levels of viral mRNA relative to cellular 18S rRNA were calculated as described recently [13]. The identities of the products obtained were confirmed by melting curve analysis. For protein analysis cells were resuspended in RIPA buffer as described previously [64]. After 1 h on ice, the lysates were sonicated and the insoluble debris was pelleted at 15,000×g/4°C. For immunoprecipitation and immunoblotting protein lysates were treated as described recently [2]. Primary Ab specific for Ad proteins used in this study included E1B-55K mab 2A6 [65], E2A-72K mouse mab B6-8 [66], E4orf6 mab RSA3 [67], rabbit polyclonal serum against protein VI [68] and anti-pVII rabbit polyclonal antibody (generously provided by Dan Engel, University of Virginia). To evaluate efficient infection with different RNA and DNA viruses primary antibodies specific for HSV-1 nucleocapsid protein (monoclonal mouse mab H1. 4; Acris antibodies) crossreacting with HSV-2 nuclear protein, HIV-1 p24 hybridoma 183-H12-5C [69] and HCV NS5A (monoclonal mab 2F6/G11 from immunological and biochemical test systems) were used. Primary antibodies specific for cellular proteins included SPOC1 rabbit polyclonal CR56 and rat mab [20], rabbit polyclonal ab specific for histone variant H3K9me3 (Upstate), Mre11 rabbit polyclonal antibody pNB 100–142 (Novus Biologicals, Inc.), p53 rabbit ab FL393 (Santa Cruz Biotechnology, Inc. [70]), polyclonal rabbit antibody raised against SAF-A protein [71] and ß-actin mouse mab AC-15 (Sigma-Aldrich, Inc.). HA-epitopes were detected with rat monoclonal 3F10 (Roche). Secondary Ab conjugated to horseradish peroxidase (HRP) to detect proteins by immunoblotting were anti-rabbit IgG, anti-rat IgG and anti-mouse IgG (Jackson/Dianova). Cells were prepared and analyzed as described recently [57]. Images were cropped using Adobe Photoshop CS4 and assembled with Adobe Illustrator CS4.

Viruses have acquired functions that target and modulate host cell signaling and diverse regulatory cascades, leading to efficient viral propagation. During the course of productive infection, Ad gene products manipulate destruction pathways to prevent viral clearance or cell death prior to viral genome amplification and release of progeny. Recently, we reported that chromatin formation and cellular SWI/SNF chromatin remodeling processes play a key role in Ad transcriptional regulation. Here, we observe for the first time that SPOC1, identified as a regulator of DNA damage response and chromatin structure, plays an essential role in restricting Ad gene expression and progeny production. This host cell antiviral mechanism is efficiently counteracted by tight association with the major core protein pVII bound to the incoming viral genome. Subsequently, SPOC1 undergoes proteasomal degradation via the Ad E1B-55K/E4orf6-dependent, Cullin-based E3 ubiquitin ligase complex. We also show that other viruses from RNA and DNA families also induce efficient degradation of SPOC1. These analyses of evasion strategies acquired by viruses and other human pathogens should provide important insights into factors manipulating the epigenetic environment to potentially inactivate, or amplify host cell immune responses, since detailed molecular mechanisms and the full repertoire of cellular targets still remain elusive.

Abstract Introduction Results Discussion Materials and Methods

2013

SPOC1-Mediated Antiviral Host Cell Response Is Antagonized Early in Human Adenovirus Type 5 Infection

Cell type specification is a fundamental process that all cells must carry out to ensure appropriate behaviors in response to environmental stimuli. In fungi, cell identity is critical for defining “sexes” known as mating types and is controlled by components of mating type (MAT) loci. MAT–encoded genes function to define sexes via two distinct paradigms: 1) by controlling transcription of components common to both sexes, or 2) by expressing specially encoded factors (pheromones and their receptors) that differ between mating types. The human fungal pathogen Cryptococcus neoformans has two mating types (a and α) that are specified by an extremely unusual MAT locus. The complex architecture of this locus makes it impossible to predict which paradigm governs mating type. To identify the mechanism by which the C. neoformans sexes are determined, we created strains in which the pheromone and pheromone receptor from one mating type (a) replaced the pheromone and pheromone receptor of the other (α). We discovered that these “αa” cells effectively adopt a new mating type (that of a cells); they sense and respond to α factor, they elicit a mating response from α cells, and they fuse with α cells. In addition, αa cells lose the α cell type-specific response to pheromone and do not form germ tubes, instead remaining spherical like a cells. Finally, we discovered that exogenous expression of the diploid/dikaryon-specific transcription factor Sxi2a could then promote complete sexual development in crosses between α and αa strains. These data reveal that cell identity in C. neoformans is controlled fully by three kinds of MAT–encoded proteins: pheromones, pheromone receptors, and homeodomain proteins. Our findings establish the mechanisms for maintenance of distinct cell types and subsequent developmental behaviors in this unusual human fungal pathogen. One of the most important processes that occurs in cells is the specification of cell type, and it is by this process that cells adopt the genetic state that governs their subsequent behaviors. One mechanism of cell type-specification is the expression of genes encoded only in a given cell type. For example, X and Y chromosomes determine male and female sexes because they encode different genes, such as SRY, a Y-specific protein whose expression establishes male identity [1]. In a similar fashion, the sexes of fungi, known as “mating types, ” are also determined through the expression of cell type-specific genes. The most well-characterized cell identity determination system is that of the budding yeast Saccharomyces cerevisiae. In S. cerevisiae the two haploid cell types, a and α, are distinguished from one another by the actions of specific transcription factors encoded at the mating-type (MAT) locus [2]. MATa encodes the homeodomain transcription factor a1, and MATα encodes α1 and α2, an α-domain protein and a homeodomain protein, respectively. The actions of a1, α1, and α2 govern control of haploid cell behavior through the differential expression of a-specific, and α-specific genes, including pheromone and pheromone receptor genes [3]. It is through a pheromone-pheromone receptor system that cell-cell communication occurs, and cells of opposite mating types can sense one another. Specifically, a cells secrete mating factor a pheromone (MFa), which binds to a receptor (Ste3) on the surface of α cells, and α cells secrete MFα pheromone, which is sensed by a receptor on the surface of a cells (Ste2). In response to the presence of the pheromone of a mating partner, cells undergo a cell cycle arrest and subsequent morphological changes to prepare for mating. After cell fusion, a1 and α2 act in concert to regulate haploid-specific genes, specifying the diploid a/α cell, thus completing a MAT-controlled regulatory circuit [2], [3]. A similar cell-cell recognition process occurs in many other fungi, including the corn smut, Ustilago maydis. In U. maydis specific pheromones and receptors are expressed in different cell types; however, in contrast to S. cerevisiae, these genes are not under the transcriptional control of the homeodomain proteins of the traditional MAT locus [4]. Instead, distinct alleles of the pheromones and their receptors are encoded within a separate MAT locus, and these alleles are sufficiently distinct from one another to confer cell type-specificity. In this case, haploid cells expressing distinct pheromones and receptors from the pheromone MAT locus sense and respond to partners of other mating types and fuse. Once compatible mating types have fused, two transcriptional regulators, bE and bW, which are encoded at the second MAT locus, regulate a transcriptional cascade that promotes further sexual development [4]–[6]. In a related, clinically important human pathogen, Cryptococcus neoformans, the determinants of haploid cell identity are unknown. C. neoformans contains a single MAT locus that is over 100 kb in size and contains 23 genes, some of which have been found to be involved in sexual development and others that appear to be essential “housekeeping” genes [7]. This locus represents an evolutionary transition from the two separate MAT loci found in basidiomycete fungi like U. maydis, to the single MAT locus found in ascomycetes [8]; it is unclear how components in this “fused” MAT locus function to specify haploid cell type. In the C. neoformans MAT locus, there are five genes in each mating type that represent the classic MAT components found in basidiomycete MAT loci. They include the homeodomain transcription factors SXI2a and SXI1α, the six pheromone genes MFa1-3 and MFα1-3, and the pheromone receptors STE3a and STE3α. The transcription factors, Sxi1α and Sxi2a, do not appear to play any role in haploid cells, including establishment of haploid cell identity. In contrast, Sxi1α and Sxi2a are both necessary and sufficient to specify the dikaryotic state following cell fusion and ensure that sexual development continues [9], [10]. Conversely, the pheromones and pheromone receptors of C. neoformans have been shown previously to be necessary for haploid cell behaviors, such as sensing and responding to a mating partner [11]–[16]. What is not clear is whether pheromones and pheromone receptors are sufficient to confer haploid cell identity (as in U. maydis), or whether the actions of other regulators are required to establish the a and α mating types (as in S. cerevisiae). To test the hypothesis that pheromones and pheromone receptors alone are sufficient to confer haploid cell identity in C. neoformans, we carried out a “swapping” experiment in which components from the a mating-type (STE3a and MFa1) were relocated into a strain of the opposite mating type, α (in which STE3α and all three copies of MFα had been deleted), and the effects of these modifications on mating and development were examined. The results presented here reveal that pheromones and their receptors are both necessary and sufficient to specify haploid cell identity. That is, strains that harbor a receptor and pheromone from the opposite mating type are capable of sensing and fusing with wild type cells of the same mating type. Furthermore, the fused cells are incapable of progressing through sexual development in the absence of both Sxi transcription factors; however, exogenous addition of SXI2a facilitates complete sexual development. These findings reveal that control of haploid cell identity in C. neoformans is mediated by MAT-encoded cell-cell communication components that are not mating type-specific in their downstream effects, and establishes a model for the maintenance of distinct cell types and developmental behaviors in this human fungal pathogen. To test the possibility that haploid cell identity in C. neoformans is controlled by pheromones and pheromone receptors, a strain where the pheromone and pheromone receptor genes from one mating type were replaced with those from the opposite mating type was constructed. Specifically, we created an altered α strain by functionally replacing the STE3α receptor gene and the three MFα pheromone genes of an α cell with the STE3a receptor gene and three copies of the MFa1 gene from an a cell (Figure 1). To construct the altered α strain, α cells in which all three copies of the MFα gene had been deleted (α mfαΔ), were transformed with a construct to replace the endogenous copy of STE3α with STE3a (Figure 2A) [13]. This replacement resulted in a complete deletion of the STE3α open reading frame and the precise insertion of STE3a under the control of the endogenous STE3α promoter. This strain, designated αmfαΔ, STE3a was confirmed by PCR (data not shown) and Southern analysis (Figure 2A) to contain a single, targeted integration of the STE3a gene at the STE3α locus. This αmfαΔ, STE3a strain was then transformed with a construct expressing MFa1 (under the control of the MFα1 promoter) designed to integrate randomly into the genome (Figure 2B). Multiple transformants of the resulting strain, αa, were assessed by Southern blot analysis, and several strains containing integrated copies of the MFa construct were carried forward in further analyses. All of the transformants confirmed to contain the MFa construct exhibited identical behaviors in subsequent phenotype analyses. A representative αa transformant containing three copies of the MFa1 gene is shown in Figure 2B. To assess the expression patterns of pheromones and their receptors in the constructed strains, northern blot analysis was carried out. Haploid strains were grown on V8 medium for 24 hours, and their relative transcript levels were examined (Figure 2C). The STE3a and STE3α transcripts were detected in wild type a and α cells, respectively, and the STE3α transcript was also present in the αmfαΔ strain. However, as expected, in the αmfαΔ, STE3a strain where the STE3a ORF had replaced the STE3α ORF, only STE3a transcript was observed. In summary, the transcripts detected from the constructed strains were in accordance with the predicted expression patterns for each of the strains tested based on their genotypes. While pheromone transcript was not detected in either haploid cell type (2C, bottom two panels, lanes 1 and 2), this result was not surprising because pheromone genes are transcribed at low levels in haploid cells grown on minimal medium [13]. Furthermore, pheromone transcript was not observed in strains where these genes had been deleted (αmfαΔ and αmfαΔ, STE3a) (lanes 3 and 4). However, when the MFa1 gene was introduced into the αmfαΔ, STE3a strain (resulting in the αa strain), high levels of MFa transcript were detected, and this high level of expression occurred in all of the MFa transformants (lane 5 and data not shown). Because MFa1, in this case, is under the control of the MFα1 promoter (defined here as 1 kb of sequence upstream of the translational start site), a likely explanation for our findings is that normal repression of MFa1 in haploid cells is disrupted. That is, the elements required for wild type levels of expression were likely not included in our construct, resulting in higher levels of MFa1 gene expression. Because pheromones expressed from a non-native promoter do not interfere with development and the mating response, the αa strain was used in our subsequent analyses [13], [14]. To evaluate how the series of modified α strains interacted with mating partners under sexual development conditions, crosses using various test strains were carried out and assessed microscopically for the presence of fusants after 18 hours on V8 medium. The cells used in each cross were stained with either rhodamine (Alexa Fluor 594) or fluorescein (Alexa Fluor 488) prior to mixing with partner cells so that the original mating types could be discerned after cell fusion. In crosses between wild type a and α strains, fusants were identified as “dumbbell” shaped cells, consisting of one a cell (red) connected to one α cell (green) via a conjugation tube (Figure 3A, panel 1). In contrast, no fusants could be identified in crosses between wild type a cells (red) and any modified α cells (green) (panels 2–4). Specifically, in crosses between a cells and αmfαΔ cells, the αmfαΔ cells responded to the presence of a cells by forming conjugation tubes; however, in the absence of α pheromone production, the frequency of fusion with a cells is predicted to be very low, and consistent with this expectation, we did not observe any fusants. Accordingly, in strains no longer harboring the STE3α receptor (αmfαΔ, STE3a and αa), the cells appear to neither form conjugation tubes (in the presence of a cells) nor fuse with a cells (panels 3 & 4). These results indicate that modified α strains, in which the STE3α receptor gene has been deleted, do not respond to or mate with a cells. In parallel experiments, wild type α strains were crossed with the modified α strains. In a control cross, differentially labeled α cells (either red or green) were crossed and evaluated for the presence of fusants. As expected for wild type cells of the same mating type, α cells did not respond to one another or fuse (panel 5). Fusants were also not detected in crosses between wild type α cells (red) and the modified αmfαΔ and αmfαΔ, STE3a (green) strains (panels 6 & 7, respectively). It was not until a wild type α cell (red) was crossed with the altered αa strain (green) that fusants could be recovered at roughly wild type a x α levels (panel 8). Subsequent quantitative fusion assays confirmed that α x αa fusion levels are comparable to those of wild type a x α crosses (data not shown). This finding indicates that, as expected, pheromones and pheromone receptors are required for wild type fusion between a and α cells. It also reveals that the altered αa strain fuses only with wild type α cells, indicating that by simply replacing the α pheromones and receptors with those from a cells, cell identify was switched from α to a. Moreover, that change was sufficient to allow the altered αa cells to fuse with wild type α cells as though they were of opposite mating types. In addition, the introduction of pheromone and pheromone receptor from a cells into α cells changes not only their ability to fuse but also the initial mating response. That is, αa strains do not produce germ tubes in response to wild type α cells, instead appearing to behave like wild type a cells (Figure 3B). These findings indicate that pheromones and pheromone receptors specifically control both the ability to sense a mating partner and the subsequent morphological response. To assess the ability of the series of modified α strains to undergo sexual development, crosses were carried out and assessed for the production of filaments, basidia, and spores. In crosses between wild type a and α strains, florid filamentation, basidia formation, and sporulation were all visible at the periphery of the cross (Figure 4A, panel 1). However, when the modified α strains (αmfαΔ, αmfαΔ, STE3a, and αa) were crossed with either wild type a or α cells, none of the combinations resulted in wild type sexual development (panels 2–8). The a x αmfαΔ cross (panel 2) revealed a significant reduction in sexual development, and this result is consistent with previously published data [13]. In addition, aberrant filaments were observed in the α x αa cross (panel 8). This limited, mutant filamentation was very close in appearance to crosses between strains containing deletions of SXI2a (panel 9). In crosses between wild type α strains and sxi2aΔ mutants, the strains fuse at wild type levels; however, only aberrant filaments are formed, and they do not progress through sexual development [9]. This phenotype occurs because both Sxi1α and Sxi2a must be present to specify the fused state and initiate either diploid-specific or dikaryon-specific developmental programs. Because the products of an α by αa cross contain only the Sxi1α developmental regulator (and not the Sxi2a regulator found only in a cells), such crosses would not be expected to progress through sexual development and would be expected to exhibit a sxi2aΔ cross phenotype. This is, in fact, the case. Conversely, if the Sxi proteins, pheromone, and pheromone receptors were necessary and sufficient to specify cellular identify (dikaryon vs. haploid), then simply supplying the SXI2a developmental regulator to an α by αa cross would result in complete sexual development. To test the hypothesis that α x αa crosses were simply in need of SXI2a to continue through sexual development, we carried out crosses in which exogenous SXI2a was provided. Because strains containing both SXI1α and SXI2a are self-filamentous under sexual development conditions, we crossed the αa strain with an α sxi1αΔ strain carrying an integrated copy of SXI2a under the control of a constitutive promoter (GPD1). This cross resulted in the restoration of complete sexual development (panel 10). In addition, fluorescence and high-resolution microscopy show that dikaryotic filaments basidia, and spore chains were formed in αa x α + SXI2a crosses (Figures 4B, 4C). Because complete sexual development occurred between cells with nearly identical genomes (αa x α + SXI2a), no additional information from a cells is required for this process. This finding was further supported by the results of assays for diploid formation. Previous studies have shown that a/α diploids are self-filamentous on V8 medium, undergoing sexual development [17]. To test the ability of the αa strain to form diploids, crosses of αa by wild type a or α cells were carried out. αa cells fused to form mononucleate cells (presumed diploids) with wild type α cells, but not with wild type a cells, consistent with αa cells exhibiting a cell behavior. In this assay, αa cells (ura5, NEOR) were mixed with either wild type α or a cells (URA5) and incubated on V8. After 24 hours at room temperature, the co-cultures were plated under double selection (Ura− + G418) at 37°C to select for mononucleate diploids [17]. No colonies grew from the a x αa crosses, indicating that no a/αa diploids were formed. In contrast, abundant colonies were recovered from the α x αa crosses, indicating that αa/α diploids were formed. The resulting αa/α strains were determined to be mononucleate (data not shown) and were evaluated for the ability to undergo sexual development by incubation on V8 at room temperature (Figure 4D). As in α x αa crosses, sexual development was not observed (panel 2); however, the addition of the SXI2a gene to the αa/α strain resulted in full sexual development on V8, similar to that of a wild type a/α diploid (panels 3 and 1, respectively). Taken together, these results demonstrate that cell identity and sexual development are controlled entirely by three kinds of genes in C. neoformans: pheromones, pheromone receptors, and homeodomain proteins. Monokaryotic or α fruiting is another form of sexual development that takes place in C. neoformans; however, this process is specific to α cells, which form filaments, basidia, and spores under severe nutrient limitation and desiccation [10], [18], [19]. Although fruiting is known to be an α specific process, the factors responsible for α fruiting are unknown. To assess whether changing haploid cell identity influences α fruiting, fruiting assays were carried out with wild type a, wild type α, and αa strains. Strains were incubated on filament agar for 14 days at room temperature in the dark and evaluated for the production of filaments and spores. We observed that the αa strain undergoes α fruiting that is indistinguishable from wild type α cells (Figure 5), indicating that this α-specific process does not require α-specific pheromones or pheromone receptor, and is not influenced by the presence of a-specific pheromones and pheromone receptor. There is a long history of investigating pheromone/pheromone receptor function in fungi, particularly in S. cerevisiae, and myriad experiments have been carried out to understand the roles of two phylogenetically unrelated receptors STE2 (α factor receptor) and STE3 (a factor receptor). Through altered expression experiments it was found that mating type is controlled by the identity of the expressed surface receptor. Experiments where the receptor expression patterns were reversed in a and α cells led to changes in mating behaviors. For example, α cells that express only STE2 (and not STE3), can sense and respond to α pheromone and therefore mate as a cells and vice versa. Thus, in S. cerevisiae, both a and α cells contain the STE2 and STE3 genes (and both pheromone genes), and it is the actions of transcription factors that impart differential expression patterns that determine cell identity [23], [24]. This is in sharp contrast to other fungi, like U. maydis, in which different cell types encode distinct pheromones and pheromone receptors. Experiments in U. maydis where the pheromones and pheromone receptors have been “swapped” between mating types demonstrate that mis-expression of pheromones and pheromone receptors is sufficient to alter cell identity. In this case, the pheromone receptors have distinct binding specificities but have descended from a common ancestor [4]. This is also the case for other basidiomycetes, including S. commune and C. cinerea. Within these fungi, different mating types encode distinct alleles of pheromones and pheromone receptors; pheromone and pheromone receptors do not govern mating, and cell fusion occurs independently of pheromone signaling. Instead, pheromones and pheromone receptors control post-fusion sexual development [25]–[27]. These examples speak to the wide variety of strategies employed by pheromones and pheromone receptors to control sexual development in fungi and emphasize the need to investigate pheromone/pheromone receptor function on a case-by-case basis. In C. neoformans, both a and α cells contain mating type-specific pheromone genes (MFa1,2, and 3 in a cells, and MFα1,2, and 3 in α cells), but the regulation of these genes differs between mating types. In a cells, MFa pheromone is highly induced on V8 agar (or under conditions of nutrient limitation) in the absence of a mating partner. In addition to induction under nutrient limiting conditions, MFα pheromone expression is also activated by factors secreted by a cells (MFa pheromone) [13], and expression of the MFα genes is dramatically upregulated upon exposure to synthetic MFa pheromone [12], [13]. In the experiments carried out here, expression of the MFa1 gene in αa cells is driven by the MFα1 promoter. The MFa1 construct was originally engineered in this fashion to avoid any cell type-specific regulation that might be imposed upon an a-specific promoter in an α cell. However, in αa strains where MFa1 gene expression was driven by the MFa1 promoter, identical phenotypes and expression patterns were observed. Attempts to test the roles of the pheromone receptors were not as straight forward, and initially, the STE3a cassette was transformed into a wild type α strain to replace the STE3α gene at its native locus. However, after multiple attempts, no transformants were recovered (Ekena, Giles, and Hull, unpublished data). Transformants were recovered only from strains in which the pheromone genes were deleted previously. These findings, while far from conclusive, intimate that cells expressing both STE3a and α pheromone may engage in autocrine signaling that affects cell growth. A similar observation has been made in S. cerevisiae, in which a strain that expresses both a pheromone and its cognate receptor undergoes G1 cell cycle arrest [24]. In S. cerevisiae this arrest is transient; however, if a similar arrest is occurring in C. neoformans, the cells do not appear to recover. This is in contrast to U. maydis where cells that have been engineered to express both receptor alleles do not arrest, are fully viable, and are mating competent. It remains to be determined if and how autocrine signaling may be functioning in C. neoformans. In C. neoformans, a and α cells carry out distinct functions to initiate the mating process. For example, in response to nitrogen limitation, a cells secrete MFa pheromone, and α cells respond to the presence of the MFa pheromone by formation of germ tubes. This morphological response is specific to α cells, and only these cells (and not a cells) form germ tubes as an initial mating response [11], [28], [29]. Interestingly, αa cells not only recognize and fuse with α cells (as would be expected for an a cell), but they also exhibit the morphology of an a cell, and do not produce germ tubes. This finding indicates that the morphological response to pheromone is mediated by the pheromone receptor (not downstream signaling events) and is independent of other mating type-specific factors. Therefore, pheromones and pheromone receptors alone dictate both the detection of a mating partner and the initial morphological response of haploid cells in C. neoformans. Interestingly, however, changes in mating identity did not influence the α-specific process of α fruiting. Monokaryotic (or α) fruiting is a process that is specific to α cells and was recently shown to be a form of sexual development [19]. While the mechanisms underlying α fruiting are relatively undefined, it is well established that the α fruiting pathway is independent of Sxi1α and Sxi2a, and that this process is attenuated in strains in which STE3α or the MFα genes have been deleted [12], [13], [19]. Our data reveal that although α fruiting is primarily an α-specific process, it occurs regardless of whether pheromones and pheromone receptor are derived from a or α cells. In summary, while the α fruiting process occurs despite the identity of pheromones and their receptors, these factors are wholly responsible for determining the initial mating response and haploid cell identity. The findings presented here establish a complete cell identity determination profile for C. neoformans and enhance our understanding of the myriad strategies by which fungi specify mating types and undergo sexual development. To generate the αmfαΔ, STE3a strain (CHY1900), the αmfαΔ (WSC17) strain was biolistically transformed with a fragment containing STE3a and a nourseothricin resistance (NATR) cassette that was flanked by ∼1 kb of sequence from upstream and downstream of the STE3α open reading frame (ORF) [13]. Nourseothricin resistant transformants were screened by PCR for proper integration of the deletion/insertion construct into the STE3α locus, and putative deletion/insertion strains were confirmed by Southern blot analysis [30]. To generate the αa strain (CHY1901), the αmfαΔ, STE3a strain was biolistically transformed with a randomly integrating fragment of DNA containing MFa1 (under the control of the MFα1 promoter) and a neomycin resistance (NEOR) cassette. Neomycin resistant transformants were screened for copy number of the insertion using Southern blot analysis [30]. The αa/α strain (CHY2049) was generated by crossing the 5-FOA resistant αa strain (CHY1517) by wild type α cells (JEC21) and selecting for fusants at 37°C on minimal medium containing neomycin. To generate the αa/αSXI2a strain (CHY2096), 5-FOA resistant αa/α cells were biolistically transformed with a randomly integrating fragment of DNA containing the SXI2a (under the control of the GPD1 promoter) and URA5 genes. RNA was prepared from C. neoformans cells using a hot phenol extraction [30]. Strains were grown on solid V8 medium for 24 hours at room temperature. Northern blots were carried out according to standard protocols with 10 µg of total RNA used for each sample [30]. The glycerol-3-phosphate dehydrogenase gene (GPD1) probe was PCR-generated using CHO651 (CGTCGTTGAATCTACCGGTG) & CHO652 (CACCAGCAATGTAAGAGATG). All other probes were generated by PCR using the following oligonucleotides: CHO2030 (CCCCGACTATCCCTTTTGGAATCTCACTGC) & CHO2031 (GGCGAACAGTTCTTCGGGATATTGTGATACC) for STE3a, CHO2032 (GCACGACCTCAGCCTCGTCATTTTCAGCGG) & CHO2033 (CCGTATCCAGCAGCAATGATCGTCAGC) for STE3α, CHO2052 (GACTTACTCTTGCGTTATTGCTTAAAGTGGG) & CHO1917 (GAAAAGAGGTACGAGTAGAT) for MFa1, and CHO2053 (TTGTGTCATCGCCTAGACCCAACGTCCCC) & CHO2054 (CCATCTAAACAAGTCCCATACGCTTCGTTACC) for MFα1. Radiolabeled probes (Decaprime II kit; Ambion) were used in hybridization reactions as described previously [31]. Hybridizations and washes were carried out at 65°C as described previously [30]. All strains used were of the serotype D background (Table 1) and were handled using standard techniques and media as described previously [32], [33]. Sexual development assays were conducted on 5% V8 medium at room temperature in the dark for 2–4 days and were evaluated by observing the periphery of test spots. The mating tester strains used were JEC20 (a) and JEC21 (α), and crosses were photographed at 100X magnification [18]. Spores were photographed at 400X magnification. Fruiting assays were carried out by growing cells on filament agar (0. 67% yeast nitrogen base (without amino acids or ammonium sulfate) containing 100 nM (NH4) 2SO4 at room temperature in the dark for 14 days. The periphery of test spots were photographed at 100X magnification. Fluorescent microscopy was carried out on a Zeiss Axioskop 2 fluorescent microscope fitted with an Axiocam MRM REV3 digital camera and corresponding AV4 software. Light microscopy was carried out using a Zeiss Axioplan microscope fitted with a 10X long working distance objective. Photographs were taken with a Nikon Coolpix 5400 camera mounted on the microscope. To label fusants, a and α cells were grown to stationary phase in yeast extract peptone dextrose (YPD) liquid medium and were suspended in 1X phosphate buffered saline (PBS). Carboxylic acid, succinimidyl-ester Alexa Fluor 488 or 594 dyes (Invitrogen, Carlsbad CA) were coupled to a or α yeast cells in 100 mM potassium phosphate buffer pH 7. 5 for 30 min at room temperature. After washing, differentially labeled cells were mixed in 1X PBS and spotted onto V8 agar. Spots were incubated at room temperature for 6 hours and were then resuspended into a mounting solution of 1X PBS containing 0. 4 µg/ml calcofluor white MR2 (Sigma-Aldrich). For filament staining, crosses were incubated on a thin layer of V8 media on a microscope slide and incubated at room temperature for 24 hours. Cells were fixed with 2. 5% gluteraldehyde/4% paraformaldehyde for 30 minutes. Cells were washed twice with 1X PBS containing 0. 1% Triton X-100, once with 1X PBS, and incubated in 1X PBS containing 0. 4 µg/mL calcofluor white to stain septa, and 1nM Sytox green (Invitrogen) to stain nuclei for 20 minutes. Cells were washed in 5% glycerol with 20% DABCO antifade (Sigma) prior to visualization.

All organisms that undergo sexual reproduction carry out specific mechanisms to establish different sexes. In fungi, sexual identity is typically determined by components housed within specialized regions of chromosomes known as mating-type (MAT) loci. MAT–encoded genes function to define sexes via two distinct paradigms: 1) by controlling transcription of components common to both sexes, or 2) by expressing specially encoded factors (pheromones and their receptors) that differ between mating types. The two mating types of Cryptococcus neoformans (a and α) are specified by an extremely unusual MAT locus. The unique architecture of this locus makes it impossible to predict which paradigm governs mating type. To identify the mechanism by which the C. neoformans sexes are determined, we created an α strain where the pheromones and pheromone receptor were replaced with the analogous genes from an a strain. We discovered that the resulting strain (αa) now behaves as if it is an a. It senses and responds to α cells, mates with α cells, and no longer exhibits other α-specific behaviors. Our data show that replacement of two and only two genes completely alters the sexual identity of α cells, establishing pheromones and their receptors as the determinants of sexual identity in C. neoformans.

Abstract Introduction Results Discussion Materials and Methods

genetics and genomics/microbial evolution and genomics cell biology microbiology evolutionary biology molecular biology

2010

Allelic Exchange of Pheromones and Their Receptors Reprograms Sexual Identity in Cryptococcus neoformans

Human genome-wide association studies (GWAS) have shown that genetic variation at >130 gene loci is associated with type 2 diabetes (T2D). We asked if the expression of the candidate T2D-associated genes within these loci is regulated by a common locus in pancreatic islets. Using an obese F2 mouse intercross segregating for T2D, we show that the expression of ~40% of the T2D-associated genes is linked to a broad region on mouse chromosome (Chr) 2. As all but 9 of these genes are not physically located on Chr 2, linkage to Chr 2 suggests a genomic factor (s) located on Chr 2 regulates their expression in trans. The transcription factor Nfatc2 is physically located on Chr 2 and its expression demonstrates cis linkage; i. e. , its expression maps to itself. When conditioned on the expression of Nfatc2, linkage for the T2D-associated genes was greatly diminished, supporting Nfatc2 as a driver of their expression. Plasma insulin also showed linkage to the same broad region on Chr 2. Overexpression of a constitutively active (ca) form of Nfatc2 induced β-cell proliferation in mouse and human islets, and transcriptionally regulated more than half of the T2D-associated genes. Overexpression of either ca-Nfatc2 or ca-Nfatc1 in mouse islets enhanced insulin secretion, whereas only ca-Nfatc2 was able to promote β-cell proliferation, suggesting distinct molecular pathways mediating insulin secretion vs. β-cell proliferation are regulated by NFAT. Our results suggest that many of the T2D-associated genes are downstream transcriptional targets of NFAT, and may act coordinately in a pathway through which NFAT regulates β-cell proliferation in both mouse and human islets. Human genome-wide association studies (GWAS) have identified genetic variation at >130 loci that is associated with various traits linked to the development of T2D. The vast majority of the T2D-associated SNPs occur in intergenic or intronic regions, implying that they are involved in gene regulation. Specific candidate genes have been suggested to mediate the association between the T2D-associated SNPs and T2D risk. Recent studies have shown that many of the T2D-associated candidate genes primarily affect the health and function of pancreatic islet cells [1]. Thus, understanding the relationship between their functional role in pancreatic islets, and proximity to T2D-associated loci is critically important. Genetic association is defined as a causal relationship between genetic variation at a locus and a disease phenotype. However, despite the large number of loci that are associated with T2D, collectively they account for only ~10% of the genetic variability of the disease [2]. This “hidden heritability” has been attributed to the inability to detect rare alleles and/or epistasis [3]. It is also possible that dysregulation of pathways leading to T2D occur in many different ways and thus do not reflect genetic variation at a single locus. We and others have explored the high heritability of mRNA abundance [4–7]. In these studies, we have identified expression quantitative trait loci (eQTLs), which control the expression level or abundance of one or more mRNAs as quantitative traits. These traits show high heritability and provide a means to identify sets of co-regulated genes that are often associated with common physiological pathways. When the expression of a group of mRNAs map to a common locus, we can hypothesize that the associated genes are co-regulated at that locus. We previously identified eQTLs in an obese F2 mouse population derived from diabetes-susceptible and diabetes-resistant founder strains [8]. We performed genome-wide expression profiling from pancreatic islets and determined the genetic architecture of all islet mRNA transcripts showing heritability in this F2 population. Approximately 10% of the transcripts that showed genetic regulation in mouse islets, demonstrated linkage to the region in the genome where the gene encoding the transcript physically resides; i. e. cis or proximal eQTLs. The remaining ~90% mapped to a region where the encoding gene does not reside (i. e. trans or distal eQTLs). Often, trans-eQTLs co-mapped to a common locus, giving rise to trans-eQTL hot spots that potentially link to a common regulator. We recently used this approach to identify a bile acid transporter in pancreatic islets as a strong candidate for an islet eQTL hot spot on the distal end of Chr 6 [8]. Here we asked if the expression of T2D-associated candidate genes from human GWAS are subject to coordinate regulation in pancreatic islets. Our model system was an F2 intercross between a diabetes-resistant (B6) and a diabetes-susceptible (BTBR) mouse strain [9]. Approximately 500 F2 mice were made genetically obese in response to the leptin mutation (Lepob/ob), and all mice were sacrificed at 10 weeks of age in order to collect their pancreatic islets for transcriptomic profiling. An interactive database of our gene expression and diabetes-related clinical phenotypes is provided at http: //diabetes. wisc. edu/. We identified Nfatc2 as a strong candidate for the transcriptional regulation of the T2D-associated genes, and responsible for an underlying islet eQTL hot spot on Chr 2. Many of the T2D-associated candidate genes whose expression mapped in trans to a broad region on Chr 2, were regulated by Nfatc2 in both mouse and human islets. In addition to transcriptionally regulating the expression of a large number of the GWAS gene candidates, the overexpression of Nfatc2 induced β-cell proliferation. Our results suggest that Nfatc2 is a key regulator of β-cell proliferation and function that may in part reflect the coordinate regulation within islets of candidate genes associated with T2D in human studies. In order to ask if T2D-associated gene candidates from human studies showed coordinate regulation in mouse islets, we first generated a list of gene candidates, and identified their mouse homologues. A catalog of published human GWAS loci was downloaded from the National Human Genome Research Institute (http: //ebi. ac. uk/gwas). Approximately 300 separate entries (corresponding to 136 distinct genomic loci) were associated with the Disease/Trait, “Type 2 diabetes”, resulting in a list of 150 reported genes (S1 Table). We identified 141 mouse homologues to these human genes, 129 of which were represented on the gene array used for the islet eQTL analysis in our F2 mouse study. We emphasize that these are reported gene candidates that are linked to the genomic loci associated to human T2D risk. Importantly, a disease-associated SNP may not necessarily influence the nearest gene, as is the case for obesity-associated SNPs within the FTO gene, which appear to influence the expression of a more distal gene, IRX3 [10,11]. However, for our eQTL analysis, we included all genes reported to be associated with T2D. Approximately 60% of the transcripts encoded by mouse homologues of the T2D human GWAS candidates showed a significant eQTL (LOD ≥ 5) at one or more loci across the genome (Fig 1A). The vast majority of these eQTLs were distal, or trans-eQTLs, implying a regulatory factor controlling the expression of each transcript (S2 Table). Of the 205 eQTLs identified genome-wide, ~88% were trans-eQTLs with an average LOD score of ~9 (genome-wide P = 0. 05). There were 25 eQTLs that were proximal, or cis to the encoding gene; the average LOD score for these cis-eQTL was ~40. Four loci were identified where >10 GWAS-associated eQTLs co-mapped to the same locus; chromosomes 2,12,13 and 17. A broad locus on Chr 2 showed the greatest number of co-mapping eQTLs, corresponding to >50 GWAS candidate genes. The gene that has the strongest association with T2D in humans is TCF7L2 [12]. In the mouse, the gene for Tcf7l2 is located on Chr 19 (Chr 10 in human), yet its expression maps in trans to Chr 2 at ~148 Mbp with a LOD score of ~9. Co-mapping with Tcf7l2, were ~50 other GWAS candidate genes, the majority of them mapping in trans to a broad region on Chr 2 (S1 Fig). The genotype dependence of the GWAS eQTLs showed that approximately half of them decreased expression with BTBR alleles; the remaining half showed increased expression with BTBR alleles (S2 Fig). In addition to the trans-eQTLs for the GWAS candidate genes, the strongest linkage to fasting plasma insulin in our F2 intercross occurred at a locus on Chr 2, ~148 Mb with a LOD of ~10 (Fig 1B). One possible explanation for a group of genes demonstrating co-mapping trans-eQTLs is that the expression of a transcription factor is perturbed by a cis-acting variant, which in turn leads to a change in the expression of its target genes. The gene for the transcription factor Nfatc2 is located at ~168 Mbp on Chr 2 and demonstrates a strong cis-eQTL with a LOD of ~80 (Fig 1C). To test whether variation in the GWAS candidate eQTLs can be explained by variation in Nfatc2 expression, we considered a conditional QTL mapping analysis. In a conditional QTL analysis, a LOD profile is considered after regressing out the effect of another factor (e. g. another QTL or covariate) to assess the effect of the LOD profile once the effect of the factor has been removed. Specifically, for each GWAS eQTL, we first evaluated the LOD profile after regressing out the effect of Nfatc2 expression. Upon doing so, the LOD score for 49 of the 54 GWAS-associated eQTLs was reduced on average by ~6 units (Fig 1A; see also S2 Table). We also evaluated the LOD profile of insulin after regressing out the effect of Nfatc2 expression and again observed a significant decrease in the LOD profile for insulin (Fig 1B; see also S1B Fig). The dependence of the trans-eQTLs for the GWAS genes and the insulin QTL on Nfatc2 is consistent with a pathway by which Nfatc2 regulates the expression of the GWAS genes, which in turn influences plasma insulin in our obese B6: BTBR F2 mice. To determine the specificity of Nfatc2’s effect on the GWAS eQTLs, we evaluated the potential influence of other transcription factors that demonstrated cis linkage on Chr 2. There are 36 genes with cis-eQTLs on Chr 2 that are known transcription factors, or annotated as playing a role in gene regulation (e. g. , transcription or DNA binding). Conditional scans for the GWAS eQTLs were separately performed on each of these Chr 2 cis-eQTLs, and an overall influence score was computed. Nfatc2 was the top-ranked transcription factor for having the largest impact on the co-mapping GWAS eQTLs (S3 Table). Our conditional analysis therefore supports Nfatc2 as the most likely regulatory gene candidate mediating the trans-QTLs for the GWAS genes, as well as the QTL for plasma insulin. The expression of Nfatc2 is strongly linked to a single nucleotide polymorphism (SNP; rs3024096, ~167 Mb) between B6 and BTBR. This SNP results in reduced expression of Nfatc2, associated with the BTBR allele, yielding a ~2-fold difference between homozygous B6 vs. homozygous BTBR at the Nfatc2 gene locus (S3 Fig). Further, sequencing of the BTBR genome identified two SNPs within the Nfatc2 gene that yield a change in protein sequence relative to B6. A highly conserved proline (Pro251) is converted to a leucine (rs259322485), and a leucine (Leu267) is converted to a proline (rs27289000). These two residues flank the nuclear localization sequence, and thus may affect nuclear import of Nfatc2 following de-phosphorylation by calcineurin. Using LocusZoom [13], a visualization tool of publically-available GWAS data, we asked if SNPs near human NFATC2 and NFATC1 genes are associated with diabetes-related phenotypes (S4 Fig). Within a small interval of both genes, SNPs were identified that showed nominal association with fasting insulin levels in >50,000 non-diabetic individuals [14]. The nominal P-value for the SNPs associated with fasting insulin are ~10−4–10−5; these would not reach statistical significance once corrected for a genome-wide query, which usually require P-values of less than 5 x 10−8. However, by performing queries for a small number of SNPs, the penalty for multiple tests is greatly diminished, suggesting the fasting insulin-associated SNPs at the NFAT loci are significant. Our results showing that trans-eQTL linkages of the GWAS candidate genes to the Nfatc2 locus in mouse is consistent with NFATC2 association with diabetes traits in humans. Previously, we performed RNA-sequencing of mouse islets with sufficient depth to quantify isoform-specific expression of all genes [15]. Islets from B6 mice express 3 isoforms of Nfatc2; variants 1,2 and 4 in relative proportions of ~41%, ~5% and ~54% respectively (S4 Table). Six isoforms of Nfatc1 are expressed (variants 1–6), with variants 1,5 and 6 comprising >99% of the expressed isoforms. The most abundantly expressed Nfatc1 variant (variant 1) is a smaller protein (703 amino acids) than the two abundantly expressed isoforms of Nfatc2 (923 and 927 for variants 4 and 1 respectively). Among the four NFATc genes, Nfatc3 showed the highest expression, whereas Nfatc4 showed the lowest expression. We examined the effects of the ca-NFATs on β-cell proliferation, insulin secretion, and the regulation of gene expression, including the T2D GWAS genes, in isolated mouse and human islets. We used the well-characterized constitutively active (ca) form of Nfatc2, where 12 serine residues are mutated to alanine in the N-terminal regulatory domain of the protein. In addition, we evaluated the effects of ca-Nfatc1, which had 17 serine residues mutated to alanine [16,17]. We employed adenoviruses to overexpress the constitutively active mutants of mouse Nfatc1 and Nfatc2. To mimic the relative abundance of the endogenously expressed NFAT isoforms, the viruses were generated from variant 1 for each gene; 703 amino acids for Nfatc1 and 927 amino acids for Nfatc2. The protein sequences for the corresponding isoforms of mouse and human Nfatc1 and Nfatc2 are ~86% and ~90% identical respectively; all of the Ser residues that were mutated to Ala residues in each of the ca-NFAT mutants are identical between mouse and human (S5 Fig). The adenoviruses increased the expression (S6A Fig) and protein levels (S6B Fig) of ca-Nfatc1 and ca-Nfatc2 in mouse islets. The overexpression of either ca-mutant did not induce the expression of the other; i. e. , ca-Nfatc1 did not induce endogenous Nfatc2 and ca-Nfatc2 did not induce endogenous Nfatc1. The expression of Nfatc4, the NFATc gene with the lowest overall endogenous expression, was induced ~3. 5 (P < 10−4) and 4. 8-fold (P < 0. 02) in response to ca-Nfatc1 and ca-Nfatc2 respectively, although it remained the lowest expressed gene, despite the induction evoked by the ca-NFATs. Overexpression of ca-Nfatc2 induced a ~20-fold and ~3-fold increase in the incorporation of [3H]-thymidine into DNA in mouse and human islets, respectively (Fig 2A). Whereas ca-Nfatc2 was a very potent mitogen in mouse islets, ca-Nfatc1 did not induce cellular proliferation in mouse islets. In contrast, ca-Nfatc1 was effective to stimulate proliferation in human islets, yielding a ~2-fold increase in [3H]-thymidine incorporation into human islet DNA. To further investigate the effect of the ca-NFAT isoforms on islet cellular proliferation, we analyzed the cell cycle profile of dispersed mouse and human islet cells by flow cytometry. Both ca-Nfatc1 and ca-Nfatc2 significantly increased the proportion of human islet cells in the S-phase of the cell cycle (Fig 2B). In mouse islets, ca-Nfatc2 increased the proportion of cells in S-phase from ~2% to ~14%; ca-Nfatc1 did not stimulate S-phase progression in mouse islets, consistent with our measurements of [3H]-thymidine incorporation into DNA in human vs. mouse islets (Fig 2A). To identify the cell type within the islet that is induced to replicate by ca-NFAT, we used the incorporation of BrdU into DNA to mark newly proliferated cells, coupled with hormone-specific immunohistochemistry to label α-cells (glucagon) vs. β-cells (insulin). As illustrated by the islet sections in Fig 2C, ca-Nfatc1 and ca-Nfatc2 enhanced the proliferation of the β-cells in human islets. In mouse islets, only ca-Nfatc2 was sufficient to promote cellular proliferation, and did so predominantly in β-cells. To extend our observations that ca-Nfatc2 induces β-cell proliferation, mouse islets were evaluated for two additional markers of proliferation; Ki67, a nuclear antigen present during all phases of the cell cycle, and pHH3 (S10), a selective marker for M-phase progression. Consistent with cell quiescence, control islets showed minimal BrdU incorporation, and no Ki67 (S7A Fig) or pHH3 (S10) (S7B Fig) immunoreactivity. In contrast, ca-NFATc2 induced BrdU incorporation in β-cells, many of which were positive for Ki67 and pHH3 (S10). The expression of 53BP1, a mediator of DNA damage response [18], was not different between control vs. Nfatc2-treated islets (S7C Fig). No BrdU+/53BP1+ nuclei were observed in response to ca-Nfatc2. These results suggest that ca-NFAT expression selectively promotes cellular proliferation through M-phase progression and not DNA-damage repair pathways. To determine whether ca-NFAT influences β-cell function, we measured insulin secretion following ca-NFAT overexpression. In mouse islets, ca-Nfatc1 and ca-Nfatc2 significantly enhanced insulin secretion in response to high glucose (16. 7 mM), or a depolarizing concentration of KCl (40 mM) in the presence of low glucose (Fig 3A). The ca-NFAT-induced increase in insulin secretion from mouse islets was not due to a change in islet insulin content (Fig 3B), suggesting that the insulin secretory cascade was specifically enhanced by ca-NFAT. In contrast to mouse islets, glucose or KCl-induced insulin secretion from human islets was not improved, but more importantly, the overexpression of either ca-NFAT did not cause any functional impairment in islet function (Fig 3C). Total insulin content in human islets was unaffected by ca-NFAT (Fig 3D). To identify target genes that potentially mediate the effect of NFAT on β-cell proliferation and insulin secretion, we performed RNA sequencing of mouse islets 48 hr after overexpressing ca-Nfatc1, ca-Nfatc2, or GFP. Of the ~20,300 transcripts identified, ~8,800 were differentially expressed (DE) by one or both ca-NFATs compared to GFP. These DE genes tended to follow one of 4 distinct patterns of regulation (Fig 4): A) DE in response to ca-Nfatc1 alone; B) DE in response to ca-Nfatc2 alone; C) DE in response to both ca-Nfatc1 and ca-Nfatc2 equal in magnitude and direction; or D) DE in response to both ca-Nfatc1 and ca-Nfatc2, but unequal in magnitude or direction. Consistent with our observation that only ca-Nfatc2 is a potent β-cell mitogen in mouse islets, one DE gene set (Fig 4B) was highly enriched (P < 10−17) for genes associated with the cell cycle, including Mki67, the cyclins E1, E2 and A2, mini-chromosome maintenance (Mcm) 2,3, 4,5 and 7, Survivin, Aurka, and Foxm1. Most of these cell cycle associated genes (~85) included in this DE pattern were induced ~2-fold or more in response to ca-Nfatc2, while remaining unaffected by ca-Nfatc1. There was one other DE gene set that was significantly enriched for genes associated with the cell cycle (P < 10−4); genes that were induced by ca-Nfatc2, but suppressed by ca-Nfatc1 (Fig 4D). This gene set included Aurkb, E2F1,3 and 7, Tcf19, Ezh2, Pola1, and Cdk4, among others. The expression of these and ~20 additional cell cycle genes identified in this DE pattern were all suppressed in response to ca-Nfatc1 and induced in response to ca-Nfatc2. To further investigate the significance of the small set of cell cycle genes that were induced by ca-Nfatc2 but suppressed by ca-Nfatc1 in mouse islets (Fig 5A), we asked if they were regulated by the ca-NFAT mutants in human islets (Fig 5B). In human islets, the expression of many of the cell cycle genes was induced by ca-Nfatc1, in parallel with the ability of ca-Nfatc1 to induce β-cell proliferation (Fig 2). In mouse and human islets, these genes were induced by ca-Nfatc2. In addition to the gene expression changes illustrated in Fig 5, several other cell cycle associated transcripts were identified that showed differential regulation by the ca-NFATs in mouse vs. human islets, including Ccnd1, Cdkn1c, Plk5 and Nr4a1 (S8 Fig). Some of these genes are known negative regulators of cell cycle progression (e. g. , Plk5, Cdkn1c), and were induced by ca-Nfatc1 in mouse islets, but not in human islets. In summary, one or more of these key cell cycle genes that showed differential regulation by ca-Nfatc1 in mouse and human islets may explain its ability to promote islet cell proliferation. Ca-NFAT significantly regulated the expression of 80 T2D GWAS gene candidates in mouse islets (Fig 6). Some genes showed exclusive regulation by one isoform (Fig 6A), whereas others are regulated by both isoforms (Fig 6B & 6C). For example, the expression of Tcf19, Kif11, Prc1, Rnd3, Grb14, and Cenpw, all genes associated with cell cycle regulation, were induced by ca-Nfatc2, while remaining unchanged in response to ca-Nfatc1 (Fig 6A). Likewise, St6gal1, Tmem163, Tgfbr3, Map3k1 and Tgfbr3 were all exclusively suppressed by ca-Nfatc2. One or more of these genes exclusively regulated by ca-Nfatc2 may be involved in the selective effect we observed for ca-Nfatc2 on β-cell proliferation in mouse islets. In contrast to genes showing selective regulation by one of the ca-NFATs, many genes were responsive to both NFAT isoforms. For example, the expression of Tcf7l2, the gene with the strongest association with human T2D, was induced ~3-fold in response to either of the ca-NFATs (Fig 6C). Other genes induced by both ca-Nfatc1 and ca-Nfatc2, included Wfs1, Camk1d, and Lgr5, whereas Hhex, Slc30a8, Tspan8 and Pparg were suppressed by both ca-NFATs. One or more of these genes may be involved in the effects that we observed in response to either ca-NFAT on insulin secretion from mouse islets. Of the 50 GWAS genes that showed an islet eQTL on Chr 2,34 of them (~70%) were transcriptionally regulated by one or both of the ca-NFATs. We selected a number of GWAS genes that were transcriptionally regulated by one or both of the ca-NFATs in mouse islets, and asked if they were regulated by ca-NFAT in human islets. For our selection, we focused on genes that showed differential regulation by ca-Nfatc1 vs. ca-Nfatc2 in mouse islets; i. e. , regulated by one isoform, but not the other (Fig 6A), as well as several genes showing equal regulation by the two ca-NFATs (Fig 6B & 6D). Similar to our observations for TCF19 (Fig 5B), other cell cycle associated genes were upregulated in response to ca-NFAT in human islets. For example, the expression of the cell cycle regulatory genes KIF11 and PRC1 was induced by both ca-Nfatc1 and ca-Nfatc2 in human islets (Fig 6D), whereas in mouse islets, these genes were only regulated by ca-Nfatc2 (Fig 6A). Other examples included SLC44A3 (a member of a choline transporter family) and PROX1 (a member of the homebox transcription factor family), which were regulated by both ca-NFATs in human islets. Most genes that were suppressed (e. g. , TRP53INP1, ACHE, VEGA) or induced (e. g. , UBE2E2) by ca-NFAT in mouse islets were similarly regulated in human islets (S9 Fig). In summary, our results suggest that a number of the T2D GWAS genes are transcriptional targets (direct or indirect) of the NFAT signaling pathway in both mouse and human islets, and may play a critical role as intermediate traits that mediate the mechanisms by which NFAT affects pancreatic islet function and health. In T2D, there is insufficient insulin to meet the increased demand resulting from insulin resistance that is usually induced by obesity. The gene candidates near the loci that have been identified in human genetic studies for T2D are thought to predominantly exert their effects on T2D susceptibility in pancreatic islets [1,5, 19,20]. Many of these gene candidates control nutrient sensing, insulin secretion, β-cell proliferation, and β-cell survival [21–26]. Our study asked if the expression of the T2D GWAS gene candidates map to specific genomic loci as trans-eQTLs in pancreatic islets isolated from an F2 intercross between mouse strains that differ in their susceptibility to diabetes. We show that ~40% of the mRNAs encoded by human T2D GWAS genes mapped as trans-eQTLs to a broad region on Chr 2 (Fig 1A). Such co-mapping of the GWAS genes was not observed elsewhere in the genome. All eQTL and diabetes-related clinical data is available at our interactive web site; http: //diabetes. wisc. edu/. The NFAT family of transcription factors is composed of 5 members, Nfatc1-4 and Nfat5, and are expressed in pancreatic islets where they are thought to integrate calcium signals to coordinate gene expression and regulate growth, differentiation and cellular response to environmental cues [27–31]. The NFAT signaling pathway has previously been implicated in the regulation of β-cell development, proliferation and function [32–35]. Cytoplasmic subunits of the Nfatc sub-family are substrates for the calcium-activated serine/threonine phosphatase calcineurin. Calcineurin has been previously shown to be essential for normal β-cell proliferation and function [33–35]. Pancreatic islets express all four Nfatc isoforms. However, their relative abundance and contribution to islet function is not fully understood. We show that the overexpression of a constitutively active form of Nfatc2 is sufficient to drive β-cell proliferation in both mouse and human islets (Fig 2), and to stimulate insulin secretion from mouse islets (Fig 3). In addition, Nfatc2 regulates the islet expression of a large proportion of gene candidates identified in human GWAS as having a genetic association with T2D (Fig 6A–6C). The regulation of these genes by Nfatc2 may underlie the trans-mapping behavior to the Nfatc2 locus that we observed in our mouse genetic study. In addition to the T2D GWAS genes, our strongest linkage for plasma insulin occurred at this locus. When multiple traits (expression and clinical) co-map to a locus, i. e. , demonstrate common genetic architecture, it is possible that one driver mediates the trans-linkage of the multiple co-mapping traits. Using Nfatc2 as a conditional co-variate in our QTL analysis, we show that the LOD profiles for the vast majority of the GWAS genes and plasma insulin were dependent upon Nfatc2 (Fig 1A & 1B). Nfatc2 became our top candidate for the driver of the GWAS trans-eQTLs, as its own expression demonstrated a strong cis-eQTL, implying local genetic variation directly influenced its expression, along with the GWAS genes and plasma insulin. We identified a SNP associated with reduced expression of Nfatc2, and two coding SNPs that change different amino acid residues in NFAT1, the product of the Nfatc2 gene, in BTBR mice. Our current study utilized a constitutively active form of NFAT1 that corresponds to the protein sequence for B6 mice. It is possible that these coding differences between the B6 and BTBR NFAT1 proteins influence activation by calcineurin, effective translocation to the nucleus, selective binding to potential nuclear partners for transcriptional gene regulation, or more than one of these steps. Future studies are necessary to delineate the relative contribution of the eQTL SNP vs. the coding SNPs on the results reported here. We show that NFAT regulates the expression of the GWAS genes in human islets (Fig 6D). A number of GWAS genes were similarly regulated by ca-Nfatc1 and ca-Nfatc2 in mouse and human islets (e. g. RBMS1, TP53INP1, PROX1, TCF7L2, HUNK, THADA, VEGFA and SLC44A3). Those that showed differential regulation in mouse vs. human islets, included those associated with the cell cycle; KIF11, PRC1, TCF19. These genes were induced by both ca-NFATs in human, but only by ca-Nfatc2 in mouse, in parallel with enhanced β-cell proliferation. Finally, expression of three GWAS candidates showed opposite regulatory effects in response to ca-Nfatc2 in mouse vs. human islets; ST6GAL1, HHEX and MAP3K1 were upregulated in human, while down-regulated in mouse in response to ca-Nfatc2. Some of these genes may underlie the differential effects we observed on β-cell proliferation or insulin secretion in mouse vs. human islets. Several groups have recently reported the development of small molecule inhibitors of DYRK1A, the priming kinase that initially phosphorylates the NFATc isoforms (Nfatc1-4), leading to their nuclear exclusion and subsequent inactivation [36–38]. When applied to human or rodent islets, the DYRK1A inhibitors promote β-cell proliferation, and restore normal blood glucose levels in several mouse models of diabetes. These results are consistent with our observations and highlight the importance of the NFAT signaling pathway in regulating β-cell proliferation in rodent and human islets. Interestingly, the DYRK1A inhibitors appear to promote the nuclear accumulation of all four Nfatc isoforms [37,39]. However, whether one particular isoform mediates the β-cell proliferative effects of the DYRK1A inhibitors, or whether there is redundancy within the NFAT family remains to be determined. Our study utilized adenovirus to overexpress the ca-NFAT mutants in mouse and human islets. We acknowledge two important points; 1) adenovirus-mediated overexpression typically yields very high levels of expression, raising concerns about non-physiological, off-target effects; and 2) the virus utilized a promoter that does not yield cell-type specific expression, making it difficult to assign NFAT’s action on a particular cell within the islet (e. g. , α-cells or β-cells). We observed isoform-selective effects on β-cell proliferation in mouse islets (Fig 2 and S7 Fig). In mouse islets, both ca-NFATs enhanced glucose and KCl-stimulated insulin secretion, but did not affect basal insulin secretion (Fig 3). Laffitte and colleagues reported that activation of endogenous NFAT signaling in response to small molecule suppression of Dyrk1a in rat islets [37], yielded transcriptional changes similar to what we report here, including cell cycle regulatory genes (e. g. , Ccna1, Aurkb, Cdk1, Dtl, Mki67, and Ccnb2) and T2D GWAS candidates (e. g. , Pparg, Kcnk16, Cenpw, Tcf19, Prc1, and Kif11). Finally, a small collection of cell cycle-related transcripts that were suppressed in mouse islets in response to ca-Nfatc1, were induced in human islets (Fig 5), in parallel with increased cellular proliferation. Our results suggest that despite the high level of expression achieved with adenovirus, the effects we observed on gene regulation and cellular proliferation represent the physiological functions of Nfatc2 in its normal abundance. Our linkage study, which identified the GWAS hotspot on Chr 2, utilized whole-islet RNA; i. e. , islet eQTLs could potentially reflect a specific cell type, or a mix of cells within the islet. It is likely that some of the GWAS eQTLs originate in non-β-cells, for example, Hhex, a gene that is selectively expressed in δ-cells [40]. Hhex demonstrates linkage to the GWAS hotspot on Chr 2 (S1 Fig), and is transcriptionally regulated by Nfatc1 and Nfatc2 in mouse and human islets (Fig 6). Mouse islet expression data recently provided by Huising and colleagues [41] has shown that Nfatc1 and Nfatc2 are expressed in all 3 major cell types within the islet; α-cells, δ-cells and β-cells. Thus, using the non-cell type specific promoter to study the ca-NFATs provided a tool to overexpress the construct in all islet cells. Finally, it is also possible that cell-type specific expression of ca-NFAT would nonetheless affect non-transduced cells through the production and secretion of soluble factors, similar to what has recently been reported for β-cell specific expression of Pdx-1 in rat and human islets [42]. Future studies that would utilize targeted expression of Nfatc1 and Nfatc2 are required to fully assess their cell-type specific roles to gene regulation and islet function. Mouse Nfatc1 and Nfatc2 demonstrated an unexpected species selectivity in their ability to promote β-cell proliferation; whereas Nfatc2 was effective in both mouse and human islets, Nfatc1 was only effective in human islets (Fig 2). This result runs counter to the widely-held observation that it is easier to induce rodent than human β-cells to proliferate [33,43–45]. An interesting clue to the underlying mechanism was provided by a small set of cell cycle regulatory genes that were induced by ca-Nfatc2, but suppressed, or not altered, by ca-Nfatc1 in mouse islets (Fig 4D). Some of these genes have been shown to be sufficient to induce β-cell proliferation, including Ccna2, Ccnb1, Ccnb2, Aurkb and Cdk1 [46–48]. That we found these genes to be induced or suppressed in response to ca-Nfatc1 in human vs. mouse islets (Fig 5) strongly supports their role in β-cell proliferation. However, future studies are required to determine if the effects we report for Nfatc2 in mouse islets are sufficient to expand β-cell mass, as has been demonstrated for Nfatc1 [35]. Using deep RNA-sequencing, we found that multiple isoforms of Nfatc1 and Nfatc2 were expressed in mouse islets. The most abundantly expressed Nfatc1 transcript (variant 1) yields isoform A, a protein consisting of 703 amino acids. In contrast, the most abundantly expressed Nfatc2 transcripts (variants 1 and 4), yield protein isoforms A and D, consisting of 923 and 927 amino acids respectively. The C-terminal domain of the NFATc protein family is thought to mediate protein-protein interactions with nuclear binding partners, facilitating gene regulation [49–51]. The ca-mutants that we studied correspond to isoform A for both Nfatc1 and Nfatc2. Nfatc2 contains an additional 220 C-terminal amino acids, which may influence which nuclear binding partners are recruited, accounting for the difference in proliferation induction we observed in mouse islets. Kim and colleagues demonstrated that a constitutively active human NFATC1 isoform containing 716 amino acids successfully restored β-cell mass and cell cycle regulation in a β-cell specific calcineurin knockout mouse [35]. This suggests that species differences may allow for the recruitment of different NFAT nuclear binding partners, and potentially explain why mouse ca-Nfatc2 leads to a ~3 vs. ~20-fold increase in β-cell proliferation in human and mouse islets, respectively. Our data suggests that genes targeted by Nfatc1 and Nfatc2 have distinct effects on β-cell proliferation and insulin secretion. Whereas only ca-Nfatc2 induced β-cell proliferation in mouse islets, both ca-NFATs were equally effective in promoting insulin secretion from mouse islets (Fig 3). Several genes known to play a role in the insulin secretory cascade were transcriptionally regulated by the two ca-NFATs, including increased expression of Tcf7l2, Munc13-1, as well as L-, T- and N-type Ca2+ channels, and reduced expression of the K+ channel, Kir6. 2, as well as the grehlin receptor, Ghsr, and Ucn3, both of which have been shown to mediate a negative feedback on insulin secretion between δ-cells and β-cells [41,52]. These NFAT-induced transcriptional changes may in part underlie some of the stimulatory effects of the ca-NFATs on insulin secretion that we observe in mouse islets. Neither ca-NFAT isoform exerted a deleterious effect on insulin secretion from human islets (Fig 3C), while significantly enhancing β-cell proliferation (Fig 2). One difference we and others [47,53] have observed between freshly-isolated rodent islets versus human islets is the level of basal insulin secretion when the islets are maintained in low glucose. In mouse islets, basal insulin secretion is ~0. 07 ± 0. 02% of insulin content, whereas in human islets basal secretion can be as high as ~0. 6 ± 0. 1% of insulin content (Fig 3). This nearly 10-fold elevation in basal insulin secretion may in part contribute to a significant difference in islet insulin content observed in human and mouse islets; ~181 and ~275 ng insulin/islet, respectively (P < 10−5). These differences may have affected our ability to observe a stimulatory effect of the ca-NFATs on insulin secretion from human islets. Similar to our studies on Nfatc2, Newgard and colleagues have reported that the in vitro overexpression of the transcription factor Nkx6. 1 is sufficient to induce β-cell proliferation and stimulate insulin secretion in isolated rodent and human islets [47]. However, in vivo overexpression [54] and inactivation [55] studies have raised questions about Nkx6. 1’s role in β-cell proliferation and mass in adult mice. However, recent studies by Sander and colleagues have shown that Nkx6. 1 is required for post-natal β-cell expansion [56] and maintenance of β-cell identity [57]. The effects of Nkx6. 1 are mediated, at least in part, through two members of the Nr4a nuclear receptor family, Nr4a1 and Nr4a3, which are both necessary and sufficient for Nkx6. 1 to regulate β-cell proliferation [47,58]. In our study of mouse islets, the two ca-NFATs differentially regulated the Nr4a genes; whereas Nr4a3 was induced by either ca-Nfatc1 or ca-Nfatc2, Nr4a1 and Nr4a2 were exclusively induced in response to ca-Nfatc2. Finally, neither ca-NFATc1 nor ca-NFATc2 regulated the expression of Nkx6. 1, suggesting multiple pathways converge to regulate the expression of the Nr4a nuclear receptor gene family. The GWAS gene candidates were originally identified because their proximity to SNP variants associated with T2D risk. In our mouse islet study, genetic variation occurred at the Nfatc2 locus, yielding a strong cis-eQTL, and not necessarily at the same loci as the human studies. This critical difference suggests that the GWAS gene candidates were responsive to a common regulator, Nfatc2, in our islet eQTL study. Importantly, genetic variation associated with the GWAS genes was not required for our observations. However, the existence of a driver that can regulate a large number of these genes helps explain how the genes might make a substantial contribution to T2D when dysregulated while only making a modest individual contribution when expressed as an allelic variant. All animal studies were conducted at the University of Wisconsin, were preapproved by the University’s Research Animal Resource Center, and were in compliance with all NIH animal welfare guidelines. The B6: BTBR F2 intercross was generated as previously described [6]. The Leptinob allele was first bred into C57BL/6J (B6) and BTBR T+ tf/J (BTBR) mice [59]. B6ob/+ and BTBRob/+ mice were then bred to generate F1ob/ob mice. Due to infertility that results from leptin deficiency [60,61], at 4 weeks of age the F1ob/ob mice received subcutaneous transplants of white-adipose tissue from leptin wild-type littermates, allowing us to breed the F1ob/ob mice to generate F2ob/ob mice. Approximately 550 F2ob/ob mice were generated; about half male and half female. All mice were maintained on normal rodent chow diet (Purina 5008) and sacrificed at 10 weeks of age. Pancreatic islets were isolated from the F2ob/ob mice as previously described [59] and total RNA was isolated and used for gene expression profiling using a custom Agilent mouse gene expression microarray consisting of ~40,000 60-mer oligonucleotides corresponding to all known genes [6,59]. Oligonucleotide intensities were normalized to the intensity measured for an islet RNA reference pool that was constructed from all F2ob/ob mice, and are reported as the log10 of the ratio of each individual mouse relative to the reference pool. All F2ob/ob mice were genotyped with a 5K mouse SNP array (Affymetrix), which identified ~2,000 SNPs that were polymorphic (i. e. , informative) between B6 and BTBR mice, and spread uniformly throughout the mouse genome. These informative SNPs, along with a set of pseudo-markers inserted within intervals flanked by the informative SNPs, were used for QTL mapping of the expression traits as well as plasma insulin as described previously [8]. Expression traits were first transformed into normal quantiles, and then used for single-QTL genome-wide scans [62], allowing for microarray batch and sex as additive and interactive co-variates, respectively. For each expression trait, we focused on the single highest LOD score per chromosome, with a LOD threshold for genome-wide significance of ≥ 5 (P < 0. 05, trait-wise). Single-QTL analysis was performed on plasma insulin in the same manner as the expression traits, following transformation of the plasma insulin values to normal quantiles. All human islets were received through the Integrated Islet Distribution Program (IIDP); see S5 Table for donor demographics, as well as the studies conducted on each islet preparation. Upon arrival, human islets were cultured overnight in RPMI containing 8 mM glucose, supplemented with penicillin (100 Units/ml) and streptomycin (100 μg/ml) (Pen/Strep), and 10% heat-inactivated FBS. All mouse islets were harvested from our colony of B6 mice housed within the Biochemistry Department’s vivarium at the University of Wisconsin as described previously [59]. Prior to the application of adenovirus, the islets were treated (~3–4 mins) with a calcium and magnesium-free Hanks Balanced Salt Solution containing 2 mM EGTA. Following exposure to the zero divalent cation solution, adenovirus was applied to the islets in RPMI with 8 mM glucose, supplemented with Pen/Strep, without FBS, for 15 mins in a 200 μl volume, followed by transfer to a 60 mm non-TC treated culture dish containing 3. 5 ml of the same culture media (without FBS), and maintained overnight at 37°C. Assuming ~1,000 cells per islet, we added the viruses with an MOI of ~200. Approximately 18 hr after the addition of adenovirus, the islets were washed with fresh media containing RPMI with 8 mM glucose, supplemented with Pen/Strep and 10% heat-inactivated FBS. Islets were maintained in this medium until used as indicated for each study. Four methods were employed to monitor cellular proliferation in mouse and human islets: 1) incorporation of radioactive thymidine into islet DNA; 2) FACS-based separation of disrupted islets into distinct cell cycle phases; 3) incorporation of BrdU into newly replicated islet cells that were co-stained with insulin or glucagon to identify β-cells vs. α-cells, respectively; and 4) immuno-cytochemical detection of the proliferation markers, Ki67 or pHH3 (S10). For the thymidine incorporation measure, 1 μCi/ml of [3H]-thymidine (PerkinElmer) was added to the culture medium 48 hr after the initial infection of the islets with the adenoviruses. Eighteen hours later, the islets were disrupted and DNA was precipitated with 0. 5 ml of ice-cold 10% trichloroacetic acid, followed by re-suspension with 0. 1 ml of 0. 3 N NaOH. The total amount of [3H]-thymidine incorporated into the islet DNA was measured by liquid scintillation counting and normalized to the amount of total cellular protein by Bradford assay. For the FACS-based separation of the islet cells into respective cell cycle phases, islets were disrupted with the StemPro accutase cell dissociation reagent (Life Technologies, A11105-01) at 37°C for 15 minutes. Islet clusters that remained were gently triturated to facilitate disruption to single cells. Islets cells were transduced with Ad-ca-Nfatc1, Ad-ca-Nfatc2 or Ad-LacZ control viruses at MOI of 2. 5 for 4 hr. The disrupted islet cells were re-suspended in culture medium (RPMI medium supplemented with 10% fetal bovine serum and Pen/Strep) and maintained as indicated. Following adenovirus treatment (48 hr), the dispersed human or mouse islet cells were fixed with 100% ethanol and stained with propidium iodide (PI) to provide a quantitative measure of cellular DNA content. Following overnight PI staining (40 μg/ml at 4°C) cells were filtered through a 40-μm sieve, and flow cytometry was performed on a four-laser, LSRII (BD Biosciences). Data were analyzed with ModFit software to estimate the proportion of cells in G0/G1, S and G2/M phases of the cell cycle [63]. To identify the islet cell types that were induced to proliferate in response to ca-NFAT, mouse and human islets were incubated with BrdU (10 μM) added to the culture medium 48 hrs after the initial infection with adenovirus. Islets were removed from culture, washed, and fixed with formalin (10%, 3 hrs, 4°C). The formalin was then removed and the islets were maintained in PBS overnight at 4°C. To aid in their visualization during sectioning, Affi-Gel Blue Gel beads (Biorad) were added to the formalin-fixed islets. The islet and blue bead mixture was fixed in 2% Agar, 1% formalin, followed by paraffin embedding. Sections were dewaxed in xylene (10 mins), rehydrated in decreasing EtOH, boiled in unmasking solution (Vector Labs, H-3300) for 13 mins and then cooled at RT for 10 mins. To reduce non-specific labeling, sections were blocked with Dako Protein Block (X0909) for 30 min, followed by a PBS wash. Guinea pig anti-insulin (Sigma, I-8510), mouse anti-BrdU (Calbiochem, NA-61), goat anti-glucagon (Cell Signaling, 2760S), rabbit anti-Ki67 (Thermo Scientific, 1906s), rabbit anti-pHH3 (S10) (Cell Signaling, 9701S), or rabbit anti-53BP1 (Bethyl Labs, IHC00001) antibodies were added to diluent solution (Dako, S3022), applied to the islet sections and incubated at 4°C overnight. Slides were washed with PBS before the addition of secondary antibodies; Cy3-conjugated anti-guinea pig (Jackson ImmunoResearch); alexafluor 488 anti-rabbit and alexafluor 647 anti-mouse (Life Technologies). The slides were incubated with secondary antibodies (30 min, RT in the dark), washed and then allowed to air dry. Staining was preserved and nuclei identified by adding a drop of Vectashield with DAPI (H-1200, Vector Labs) to each tissue section. Islet sections were imaged with a Nikon AR1 confocal microscope equipped with EZ C1 software (Nikon Corp. , Japan). To evaluate the effect of ca-NFAT on the function of human and mouse islets, we monitored insulin secretion evoked by a variety of insulin secretagogues, as previously described [64]. Media and cellular insulin was measured by ELISA as described [59]. Cellular insulin is extracted via acidified ethanol (0. 18 N HCl, 70% EtOH in water). Insulin secretion is expressed as percent of total insulin (media plus cellular insulin values). Islets were collected from ~14 week-old B6 male mice and used for adenoviral transduction using Ad-LacZ, Ad-ca-Nfatc1 or Ad-ca-Nfatc2 as described above. 48 hr after transduction, islets were lysed using RIPA lysis buffer (Abcam, ab156034). Whole islet lysates (40 μg total protein) from each virus treatment were used to determine NFAT2 (Nfatc1 gene product) and NFAT1 (Nfatc2 gene product) protein levels; NFAT2 antibody (Thermo, MA3-024), NFAT1 antibody (Cell Signaling, 5861S). Total RNA was isolated from human or mouse islets using the RNeasy spin columns according to the manufacturer’s directions (Qiagen). RNA quantity was determined using a Nanodrop (Thermo Scientific) and the quality assessed by a Bioanalyzer (Agilent). Total mouse RNA was used for RNA-sequencing as described below. Total RNA from human islets was converted to cDNA using a high capacity reverse transcriptase kit (ABI). The cDNA was diluted with water to a final concentration of 4 ng/μl, and used for quantitative gene expression measurement with a Qiagen Rotorgene qRT-PCR machine. Gene-selective primer pairs were generated by “Primer Quest” (IDT). All primer sequences are contained within S6 Table. On each of 5 separate days ~1,200 islets were pooled from 6 B6 mice and used for adenoviral treatment as described above to overexpress GFP, ca-Nfatc1 or ca-Nfatc2. 48 hr later, whole islet RNA was isolated using RNeasy purification columns (Qiagen), quantified (Nanodrop) and integrity verified (Agilent) prior to sequencing. The 15 separate RNA samples (N = 5 each for GFP, ca-Nfatc1 and ca-Nfatc2) were bar-coded and randomized for multiplexing across three lanes of an Illumina HiSeq 2000, which yielded ~24M total RNA-sequencing single, paired-end reads/sample (101 bp length). Based on simulation results from Dewey and colleagues, the median percent error is well-controlled (<10%, and in most cases < 5%) with 24M, 101 bp reads in mouse [65]. The RNA-sequencing reads were mapped via bowtie [66] against refseq mm10 reference. Gene and isoform expression values were then estimated via RSEM [65]. Expected counts were normalized using median-by-ratio normalization [67]. Genes and isoform values with 75th quantiles > 10 were used for further analyses. EB-seq [68], a newly-developed multiple condition model, was used to classify genes and isoforms into 5 distinct patterns of differential expression, DE, (C1, ca-Nfatc1; C2, ca-Nfatc2): Pattern 1: C1 = C2 = GFP (no DE); Pattern 2: C1 ≠ C2 = GFP (DE for C1 only); Pattern 3: C2 ≠ C1 = GFP (DE for C2 only); Pattern 4: C1 = C2 ≠ GFP (DE for C1 and C2 equally); and Pattern 5: C1 ≠ C2 ≠ GFP (DE for C1 and C2 unequally). For each gene or isoform, EB-seq computes a posterior probability (PP) associated for each expression pattern. The higher the PP (Pattern k), the more likely that gene/isoform is following Pattern k. To identify DE genes for either C1 or C2, we used a threshold of PP (Pattern 1) < 0. 01 (i. e. , >99% confidence DE for C1, C2 or both). Genes/isoforms illustrated in Fig 4 have PP (Pattern 2–5) > 0. 75. S7 Table contains RSEM-normalized expression values, as well as the PP values for patterns 1–5 for all genes; raw expression values are also available GEO (GSE73697). RNA-Seq data was collected from B6 mouse islets and used to estimate the abundances of endogenously expressed isoforms. Approximately 400 islets were isolated from two B6 mice. Total RNA was isolated from the islets using RNeasy purification columns (Qiagen, Hilden, Germany), verified for integrity on a 2100 Bioanalyzer System (Agilent, Santa Clara, California), and prepared for sequencing using the Illumina TruSeq RNA Sample Prep Rev. A protocol (Illumina, San Diego, CA). Approximately 94 million paired-end reads were sequenced on an Illumina HiSeq 2000 (2 x 101 bp, ~350 bp total size). Isoforms were quantified by running RSEM (version 1. 2. 4) [65] and Bowtie1 (version 0. 12. 7) [66] in paired-end mode and using a synthetically reconstructed transcriptome derived from the mm9 reference genome and RefSeq gene models (downloaded from UCSC browser June 24th, 2013, NR entries removed). The Bowtie-RSEM pipeline directly maps RNA reads to annotated transcripts (isoforms), which has been show to provide better quantification accuracy for known transcripts compared to pipelines that uses splice aligner to map reads to the whole genome [65,69]. All other RSEM parameters used were default. Estimated isoform abundances are reported in Transcripts Per Million reads (TPM) [65], as well as Fragments Per Kilobase of transcript per Million mapped reads (FPKM) [70]. The TPM value represents the expected number of reads—per million reads collected—derived from an isoform, as based on the RSEM model of isoform-quantification [65]. Raw expression values are also available at GEO (GSE76477). To investigate the extent to which variation in Nfatc2 expression explains variation in GWAS eQTLs, for each GWAS eQTL, we recalculated the LOD score profile while adjusting for Nfatc2 expression as a covariate. This type of conditional analysis effectively removes the effect of Nfatc2 on the GWAS eQTLs. To evaluate the specificity of the effect of Nfatc2 on the GWAS eQTLs, we repeated the conditional analysis for each gene having a significant cis-eQTL on Chr 2 (303 cis-eQTLs). We determined the number of significant eQTLs (LOD ≥ 5. 0) on Chr 2 for all GWAS genes before and after conditioning on each of the cis-eQTLs. A summary score was then computed for each cis-eQTL based on the reduction in the number of GWAS eQTL following conditioning. S3 Table provides a list of the summary scores for all cis-eQTL on Chr 2. All expression data has been deposited to the Gene Expression Omnibus (GEO) at NCBI (GSE73697 and GSE76477).

Genome-wide association studies (GWAS) and linkage studies provide a powerful way to establish a causal connection between a gene locus and a physiological or pathophysiological phenotype. We wondered if candidate genes associated with type 2 diabetes in human populations, in addition to being causal for the disease, could also be intermediate traits in a pathway leading to disease. In addition, we wished to know if there were any regulatory loci that could coordinately drive the expression of these genes in pancreatic islets and thus complete a pathway; i. e. Driver → GWAS candidate expression → type 2 diabetes. Using data from a mouse intercross between a diabetes-susceptible and a diabetes-resistant mouse strain, we found that the expression of ~40% of >130 candidate GWAS genes genetically mapped to a hot spot on mouse chromosome 2. Using a variety of statistical methods, we identified the transcription factor Nfatc2 as the candidate driver. Follow-up experiments showed that overexpression of Nfatc2 does indeed affect the expression of the GWAS genes and regulates β-cell proliferation and insulin secretion. The work shows that in addition to being causal, GWAS candidate genes can be intermediate traits in a pathway leading to disease. Model organisms can be used to explore these novel causal pathways.

Abstract Introduction Results Discussion Methods

genome-wide association studies medicine and health sciences diabetic endocrinology gene regulation cell cycle and cell division cell processes hormones endocrine physiology genome analysis mammalian genomics insulin endocrinology gene expression genetic loci animal genomics biochemistry cell biology physiology genetics biology and life sciences genomics computational biology insulin secretion human genetics

2016

The Transcription Factor Nfatc2 Regulates β-Cell Proliferation and Genes Associated with Type 2 Diabetes in Mouse and Human Islets

Laiwu District is recognized as a hyper-endemic region for scrub typhus in Shandong Province, but the seriousness of this problem has been neglected in public health circles. A disability-adjusted life years (DALYs) approach was adopted to measure the burden of scrub typhus in Laiwu, China during the period 2006 to 2012. A multiple seasonal autoregressive integrated moving average model (SARIMA) was used to identify the most suitable forecasting model for scrub typhus in Laiwu. Results showed that the disease burden of scrub typhus is increasing yearly in Laiwu, and which is higher in females than males. For both females and males, DALY rates were highest for the 60–69 age group. Of all the SARIMA models tested, the SARIMA (2,1, 0) (0,1, 0) 12 model was the best fit for scrub typhus cases in Laiwu. Human infections occurred mainly in autumn with peaks in October. Females, especially those of 60 to 69 years of age, were at highest risk of developing scrub typhus in Laiwu, China. The SARIMA (2,1, 0) (0,1, 0) 12 model was the best fit forecasting model for scrub typhus in Laiwu, China. These data are useful for developing public health education and intervention programs to reduce disease. Scrub typhus, also known as tsutsugamushi disease, is a zoonosis transmitted by chigger bites (larval trombiculid mites) and infection with Orientia tsutsugamushi (O. tsutsugamushi), a Gram-negative obligate intracellular bacterium. Globally, scrub typhus is distributed widely in the Pacific region of Asia. It is prevalent in a triangle from Northern Japan and far-eastern Russia in the north, to Northern Australia in the south, to Pakistan and Afghanistan in the west, and also involves islands of the western Pacific and Indian Oceans [1]–[3]. Scrub typhus has existed in China for thousands of years. Before 1986, the disease was found only in Southern China, with epidemics occurring mainly in summer [4]. Beginning in 1986, scrub typhus was also reported in areas of Northern China, such as Shandong Province, Tianjin City, Heilongjiang Province, Shanxi Province and Hebei Province [5]. Clinical infections in Northern China seemed to occur mainly in autumn and winter [4]. Shandong Province is most noted to the one of the most serious foci for scrub typhus in Northern China [5]. In 1986, the first outbreak of scrub typhus occurred in Linyi District, followed other outbreaks in Jinan in 1988, in Jining in 1996, in Yantai in 1997, in Weifang and in Tai′an in 2000 [6]. From 2006 to 2012, a total of 2337 scrub typhus cases were notified in the Diseases Reporting Information System of Shandong Center for Disease Control and Prevention. At present, 13 of 17 districts have reported scrub typhus cases in Shandong Province where Laiwu is the district of the highest scrub typhus incidence. In 1999, the earliest cases of scrub typhus were documented in Laiwu District [7]. Since 2006, scrub typhus cases have been included in the Notifiable Infectious Diseases System managed by Shandong Center for Disease Control and Prevention. From 2006 to 2012,308 cases were detected in Laiwu (134 males and 174 females). The average annual incidence of scrub typhus in Laiwu (3. 59/100,000) was approximately ten times of that in entire Shandong Province (0. 35/100,000). However, researchers and health authorities have more often focused their attentions to Linyi, the initial focus of scrub typhus in Shandong, where the annual incidence was 0. 89/100,000. As a new focus of scrub typhus in Northern China, Laiwu (117°19′-117°58′E, 36°02′-36°33′N) lies in the center of Shandong Province, an important geographical position in Shandong. The population of Laiwu was 1,226,393. Incidence, prevalence, duration and mortality indicators were most frequently used to estimate the burden of disease [8]. Of available indexes, we highlight disability-adjusted life years (DALYs) to measure the disease burden of scrub typhus in Laiwu, China. The DALYs metric was jointly developed by the World Bank, Harvard School of Public Health and the World Health Organization (WHO) for the Global Burden of Disease and Injury Study (GBD) [9]. The European Centre for Disease Prevention and Control (ECDC) had adopted an incidence- and pathogen-based DALYs approach to measure the Burden of Communicable Diseases in Europe Project (BCoDE) across European Member States [10]–[12]. DALY, a summary metric of population health, is often used to identify health gaps by measuring the state of a population' s health compared to a normative goal which is for individuals to live the standard life expectancy in full health. One DALY means one-year loss of ‘healthy’ life. It integrates disease-specific mortality, morbidity and severity together, and quantifies morbidity associated with different clinical outcomes by assigning disability weights on a scale between 0 and 1, where by 0 means no morbidity and 1 means death [13]. DALYs and DALY rates (DALYs per 100,000 population or DALYs per 1000 population) [8], [14], are often used to compare disease burden of the same disease among regions, areas, countries, districts [15], or disease burden of different diseases in the same place [16]. Moreover, DALYs and DALY rate could be used to identify high risk population for targeted interventions or intervention prioritization. In this research, DALYs and DALY rate were adopted to evaluate the disease burden of scrub typhus in Laiwu, China. Epidemic modeling and forecasting is more recognized as an essential tool in preventing and controlling infectious disease. Cases of scrub typhus in Laiwu, China seem to have considerable variation during the period 2006 to 2012. A seasonal time series autoregressive integrated moving average (SARIMA) modeling introduced by Box and Jenkins [17] is most useful in examining data for seasonal or periodic fluctuations that recur with about the same intensity each year. SARIMA models have been successfully applied for forecasting economic, marketing, and social problems. While this model has the advantage of accurate forecasting over short periods, it has a limitation that at least 50 observations are needed [17]. In this study, we sought to identify the most suitable SARIMA model we could use to predict scrub typhus cases in Laiwu over time. By combined the results of disease burden measured by DALYs and DALY rate with a SARIMA forecasting model of scrub typhus, we hoped to identify high-risk populations and interventions or intervention prioritization of scrub typhus. The research could help health authorities to prevent and control of scrub typhus efficiently. The dataset of the human scrub typhus cases in Laiwu, China from 2006 to 2012 was obtained from the Diseases Reporting Information System of Shandong Center for Disease Control and Prevention. The notification system recorded the detailed information for the scrub typhus cases, including gender, age, dates of symptom onset and diagnosis, and recovery or death. The symptom onset date was used in this study, and it was thought to be more useful than the date of diagnosis or the date of notification. The cases recorded were the anonymized in this study. Population data for Laiwu was obtained from the Laiwu Statistical Bureau and stratified by age group and gender [18]. In this research, we adopted the incidence- and pathogen-based DALYs approach [10]–[13] to estimate the disease burden of scrub typhus in Laiwu, China. This approach has been previously used in the BcoDE project by ECDC. Fig. 1 shows a disease outcome tree for Orientia tsutsugamushi infection. All cases of scrub typhus in this study recovered for their illness. DALYs were the sum of years of life lost due to premature death (YLLs) and years lived with disability (YLDs) [19]. According to the outcome tree, YLDs were calculated for each health outcome (l) by multiplying the number of incident cases (n) with the disability weight (w) for a specific health outcome (l), and the duration of the disabling condition (t) [see equation (1) ]. All input parameters in both YLDs and YLLs formula were chosen to be age (a) and sex (s) dependent when such information was available, where a stands for age at infection and ã for age at onset of a condition or death [11], [12]. (1) To estimate the YLLs for those health outcomes (l) that can lead to death, the number of fatal cases (d) for a specific health outcome (i) for an infection acquired at age (a) is multiplied by the remaining life expectancy (e) at age ã [see equation (2) ]. (2) In this research, the Coale and Demeny West Level 26 Life Table adopted in many disease burden studies was used in the calculation of YLLs to assure comparability to other disease burden assessments [20]. Life expectancy for males and females at birth were set at 80 and 82. 5 years, respectively. The average durations of scrub typhus in different age groups and gender were achieved by DisMod II software developed for the calculation of GBD [21]. DisMod II is powered by two basic inputs: the make-up of a region' s population by gender and age, and the overall mortality rate for each demographic group. This model stratifies the relationships between a set of indicators relevant by age and gender: incidence, prevalence, remission, mortality, duration, case fatality, and RR mortality (the relative risk on total mortality) [22]. It requires powering with a minimum of three of these variables (by age groups and gender) and these three variables allow the prediction of the others [22]. Comparing the internal epidemiological consistency of estimates of incidences, prevalence, duration and mortality, we found that when inputting the variables were incidence, remission (100%), and case fatality of scrub typhus clarified by age and gender, the outputs fitted well. Thus, the average durations for scrub typhus in gender and different age groups were obtained (Table 1). No paper has previously reported the disability weight of scrub typhus, including “Global Burden of Disease 2004 Update: Disability Weights for Diseases and Conditions, ” the widely adopted document about disability weights of disease [13], [16], [23]. Scrub typhus and dengue fever have so many similarities in duration, high-risk population, signs and symptoms, pathogenesis and prognosis [1], [24]–[26] that disability weight of dengue fever (0. 197) [13] was referred in the research of scrub typhus. According to BCoDE-project, these raw incidence data should be corrected for underestimation by pathogen-specific multiplication factors (MF), representing either correction for underestimation in one step, or separate correction for under-ascertainment and underreporting in two steps [12], [27]. The overall extent of underestimation can be explained by two major effects represented by under-ascertainment and under-reporting [27]. Since one person infected with Orientia tsutsugamushi will demonstrate acute symptoms, such as fever, rash, eschar and swollen lymph nodes, the person is likely to visit a doctor, and due to physical awareness, scrub typhus is relatively easy to diagnosis in Laiwu, China. Moreover, as a notifiable disease in Laiwu, scrub typhus must be reported to the Diseases Reporting Information System of Shandong Center for Disease Control and Prevention. So, under-ascertainment and under-reporting of scrub typhus in Laiwu was thought to be minimal and not included in the research. We used “1” as the MF value of scrub typhus in Laiwu, China in the research. Many time series data contain seasonal periodic components. To deal with seasonality, a general multiplicative SARIMA model was extended from the ARIMA model (Box et al) [28]. This building process of the SARIMA (p, d, q) (P, D, Q) s model was designed to take advantage of associations in the sequentially lagged relationships which usually exist in periodically collected data. Seven main parameters are selected when fitting a SARIMA model: p, the order of process autoregression; d, the order of difference; q, the order of process moving average; P, D and Q, the corresponding seasonal orders; s, the length of seasonal period. If d is nonzero, a general differencing can be used to remove trend. If D is not zero, seasonal differencing can be used to remove seasonality. In order to construct and validate the model, the database of scrub typhus was verified by dividing the data file into two data sets, i. e. the data between January and December 2006–2011 and data between January and December 2012. The original series of scrub typhus cases was not a stable time series, so it was necessary to make it stable by differential. After general difference of 1 order, followed by seasonal difference of 1 order and length of seasonal period was 12, the series was satisfied with stability. Thus d = 1, D = 1, S = 12. Then the order of autoregression and moving average were identified using autocorrelation function (ACF) and partial autocorrelation function (PACF) of the differenced series. The most suitable model was selected on the basis of Normalized Bayesian Information Criteria (BIC) [29] and Ljung-Box test. Lower values of Normalized BIC and Ljung-Box test (higher significant) were preferable. Furthermore, Ljung-Box test was performed to test if ACF of the residuals at different lag times were significantly different from zero, where no different from zero was expected [30]. Compare the predicted values in 2012 using the most suitable SARIMA model with the number of scrub typhus cases notified in 2012 to validate the forecasting ability of the model. The analyses were carried out with STATA version 12. 0 (Stata Corporation, College Station, USA). No cases of death were reported in Laiwu from 2006 to 2012, so DALYs of scrub typhus in Laiwu was equal to YLDs. From 2006 to 2012, DALYs of scrub typhus were 5,10,10,5, 8,10 and 13, respectively. The average annual DALYs of scrub typhus was 9, and DALYs were higher in females (5) than males (4) (Table 2). DALY rates (DALYs/100,000) of scrub typhus were 0. 3663,0. 7933,0. 8106,0. 4437,0. 6179,0. 7883 and 1. 0618 from 2006 to 2012, respectively. The average annual DALY rate was 0. 6974 DALYs/100,000. The DALY rate was higher among females (0. 8019 DALYs/1 00,000) than among males (0. 5961DALYs/100,000). Moreover, DALY rates in females and males were both highest for the 60–69 years age group (Table 3). Fig. 2 shows the breakdown of average DALY rates of scrub typhus in different age groups and gender in Laiwu. Since all cases of scrub typhus were cured in Laiwu, no sequelae were recorded. The annual incidence (/100,000) of scrub typhus was 1. 88,4. 08,4. 16,2. 28,3. 18,4. 08, and 5. 46 from 2006 to 2012, respectively. The average annual incidence was 3. 59/100,000. There was no statistical difference between incidence of scrub typhus in males and females each year (2006–2012) (P>0. 05). Related x2 and P values were showed in Table 4. But, the average annual incidence in females (4. 11/100,000) was higher than males (3. 07/100,000) (x2 = 5. 850, P = 0. 016). The average annual incidence was the highestin 60–69 age group (Table 4). The trend of DALY rates of scrub typhus was consistent with incidences from 2006 to 2012 in Laiwu, China. Since 2009, DALY rates and the incidences were both increasing year by year (Table 3,4). Fig. 2 showed that in 60–69 age group, the DALY rate in females was sharply higher than males. The series of notified cases was a non-stationary series. Therefore, by taking 1 order general difference, followed by 1 order seasonal difference and length of seasonal period was 12, the time series of scrub typhus cases was corrected into stationary series. Fig. 3 shows the autocorrelation function (ACF) and partial autocorrelation function (PACF) of scrub typhus cases in Laiwu, China after differencing. Based on the distribution characteristic, we conducted several models, SARIMA (2,1, 0) (0,1, 1) 12, SARIMA (1,1, 0) (1,1, 1) 12, SARIMA (0,1, 0) (2,1, 1) 12, SARIMA (2,1, 0) (0,1, 0) 12, SARIMA (1,1, 0) (1,1, 0) 12, SARIMA (0,1, 0) (2,1, 0) 12, SARIMA (2,1, 1) (0,1, 0) 12 and SARIMA (2,1, 1) (0,1, 0) 12. Of all the models tested, the SARIMA (2,1, 0) (0,1, 0) 12 model was the best fit for the data (Table 5). Moreover, the Ljung-Box test suggested that the ACF of residuals for the model at different lag times was not significantly different from zero, i. e. the residuals of the SARIMA (2,1, 0) (0,1, 0) 12 model was satisfied with white noise. The stationary residuals provided the evidence that the SARIMA (2,1, 0) (0,1, 0) 12 model was adequate. All the coefficients of the SARIMA (2,1, 0) (0,1, 0) 12model were significant (Table 5,6). The equation of the SARIMA was. All the number of notified cases from Jan 2012 to Dec 2012 were in the 95% confidence interval of the forecasting values by the SARIMA (2,1, 0) (0,1, 0) 12 model. The model was used to predict values from January to December 2012 for validation. The notified cases and fitted cases by the best fitted SARIMA model from 2006 to 2011, and the actual cases and predicted cases from January to December 2012, were illustrated in Fig. 4. It showed that the predicted values could follow the upturn and downturn of the observed series reasonably well. In addition, the fitted values appeared some negative values, which was a common case with the series with too many zeros as observed values in the series of scrub typhus cases. The SARIMA (2,1, 0) (0,1, 0) 12 model showed that the prevailing disease occurred mainly in autumn and peaked in October. Incidence is often used as one measure index of disease frequency, and since its calculation only concerns the cases number and basic population data, it must be standardized by some standard population when compared between different areas or different times. However, DALYs consider both survival time and life quality caused by a certain disease and its calculation requires age and sex dependent incident cases, fatal cases, severity of disease and life expectancy. Therefore, DALYs can demonstrate the threat of some disease to the whole population, and it is more objective and comprehensive than incidence [23]. In addition, standardization issues have been covered inherently in the calculation of DALYs [8], [14]–[16], [31], [32], such as different age groups and gender are assigned with different life expectancies and different diseases are assigned with different disability weights. So, DALYs and DALY rate could be compared among areas or diseases directly. From 2006 to 2012, the average annual DALYs of scrub typhus was 9, and DALY rate was 0. 6974 DALYs/100,000 in Laiwu District. The average annual DALYs and DALY rate of scrub typhus were 15 and 0. 1504 DALYs/100,000 in Linyi District, the initial foci of scrub typhus in Shandong Province. Populations of Linyi and Laiwu were 10,039,435 and 1,226,393, respectively. So, even though DALYs of scrub typhus was higher in Linyi than Laiwu, DALY rate was apparently higher in Laiwu than Linyin. Therefore, considering the limited health resources, more attention should be paid to Laiwu, a new focus of scrub typhus. No cases of death were reported in Laiwu, China from 2006 to 2012, so the YLLs were 0 and the DALYs were equal to YLDs. There were several reasons why DALYs were adopted rather than solely YLDs. Worldwide acceptance of DALYs and DALY rate enabled international comparisons or ranking of areas by disease burden [8], [15], [16], [33]–[43]. DALYs represent unified summary estimate of disease burden that takes mortality as well as morbidity into consideration, i. e. , DALYs and DALY rate would be more useful in comparing the disease burden among different diseases and different areas than other indexes. In addition, the levels of diagnosis and treatment of scrub typhus in Laiwu, China were high enough to cure all the patients, but there were still death cases in some other areas in China [44], [45]. For example, there were death cases of scrub typhus reported in Guangzhou, China in 2012. When comparing the disease burden of scrub typhus between these two districts, as a single metric DALYs is better than YLDs [44]. The DALYs of scrub typhus were equal to YLDs in Laiwu demonstrated that more attention should be paid to the unhealthy conditions of scrub typhus. Since 2009, DALY rates and incidences of scrub typhus have both been increasing year by year (Table 3 and Table 4). With global warming, the predicted scenarios of increased temperature and rainfall were also causing concern for increases in vector-borne diseases, particularly, endemic arboviruses [46]. In 2010, Kim SH and Jang JY [47] reported that the incidence of scrub typhus in hyper-endemic region during the outbreak period was positively correlated with temperature and humidity during the summer. In addition, scrub typhus was transmitted to humans by the bites of chiggers which are mainly active in forest clearings, riverbanks, and grassy regions [47], [48]. With the increased urbanization and higher green coverage rate, people have more opportunity to contact chiggers. In 2012, there were several death cases of scrub typhus reported in Guangzhou, China and these patients had lied, stood, or sat on lawn in gardens before onset of the disease [44]. Green coverage rate achieved 42. 19% and park green space per capita was 18. 49 square meters in Laiwu, China in 2012 [49]. More suitable environment for chiggers' multiplication, more cases might appear in Laiwu. From 15 to 80 years old, the average annual DALY rate and incidence of scrub typhus in females were both higher than males (Table 3, Table 4). Moreover, in the 60–69 years age group, females had a sharply higher DALY rate (2. 6095 DALYs/100,000) than males (1. 5689 DALYs/100,000) (Fig. 2). Thus, females, especially those from 60 to 69 years old, were the highest risk population for scrub typhus in Laiwu, China. As a zoonosis, scrub typhus was infected with O. tsutsugamushi, and rodents were the main host of O. tsutsugamushi. With increasing touch chance with O. tsutsugamushi, people in fields were easier to suffer from the disease [50]. Being a modern industrial area, Laiwu had become one home of numerous iron and steel businesses in China. Therefore, with more and more young and males moving from rural to urban areas, more and more elderly and females now engage in agriculture activity [18]. Additionally, with sparse awareness of the disease, the elderly have a greater chance of contact with chiggers [18]. Hence, health education efforts regards scrub typhus should be focused upon high risk groups like females who are 60–69 years old, especially those who live in countryside. SARIMA modeling was useful for interpreting and applying surveillance data in disease control and prevention [51], [52]. As with many infectious diseases, the time series data of scrub typhus in Laiwu, China showed the components of trend and seasonal pattern. One of the most recognized disadvantages of this approach is the necessity of a large amount of data (i. e. , a minimum of 50 observations) to build a reasonable SARIMA model [53]. In this research, monthly number of cases from 2006 to 2011 (72 observations) was used to build the SARIMA model, and monthly cases during the period of January to December in 2012 were used to validate the corresponding SARIMA model. The chosen SARIMA (2,1, 0) (0,1, 0) 12 model fit the observations well and the residual series were satisfied with white noises. Therefore, the SARIMA (2,1, 0) (0,1, 0) 12 model could be used to forecast the monthly cases of scrub typhus in Laiwu, China. The prevailing disease occurred in autumn and peaked in October, which suggests that education and other protective measures should occur just before October. In the research, we adopted the incidence- and pathogen-based DALYs approach used by BcoDE Project to estimate the disease burden of scrub typhus rather than prevalence-based DALYs method presented in the GBD 2010. In case of an infectious pathogen, and in particular for priority settings of intervention to prevent primary infection, incidence is a more appropriate input for the DALYs metric than prevalence with the reason that only with the initial start of the infection is it possible to include all the disease squeals that result from the infection [12]. In addition, the incident cases of scrub typhus in Laiwu could be obtained from surveillance systems accurately. Considering above mentioned, the incidence- and pathogen-based DALYs were adopted in this research. Limitations of this research should be acknowledged. The occurrence of scrub typhus was influenced by many factors such as pathogen prevalence, mites, rodents, human being' s activities, and social or environmental factors. While building the SARIMA model, factors mentioned above were not considered. In addition, our research regarding disease burden and time series analysis of scrub typhus only focused on Laiwu, China. The results may not generalize to the most of China' s population. The disease burden of scrub typhus has increased year-by-year in Laiwu, China. Public health authorities should make concerted efforts to control and prevent the disease. The results of disease burden can also assist authorities in identifying the high-risk population. The SARIMA (2,1, 0) (0,1, 0) 12 model developed in the research could offer prediction of scrub typhus monthly cases in Laiwu. Combined with disease burden measurement and SARIMA forecasting, these data should help official as a decision support tool in a scrub typhus risk management program and in planning various prevention efforts.

Scrub typhus, also known as tsutsugamushi disease, is a zoonosis transmitted by chigger bites (larval trombiculid mites) and the pathogen Orientia tsutsugamushi (O. tsutsugamushi), a Gram-negative obligate intracellular bacterium. It is distributed widely in the Pacific regions of Asia, and the islands of the western Pacific and Indian Oceans. People with outdoor activities that involve contact with grasses or shrubs are at highest risk. Scrub typhus has existed in Southern China for thousands of years, but it has been noted to spread from the South to the North of China in recent decades. Though this research we studied the disease burden of scrub typhus with disability-adjusted life years (DALYs), and developed a forecasting time series model for human clinical disease in Laiwu, China. Results demonstrated that the disease burden of scrub typhus was increasing year by year in Laiwu, and it was higher in females than males. Moreover, DALY rates in females and males were highest for persons in the 60–69 years age group. Of all the seasonal autoregressive integrated moving average (SARIMA) models tested, the SARIMA (2,1, 0) (0,1, 0) 12 model was the best fit for scrub typhus cases in Laiwu. The disease occurred mainly in autumn, with a peak in October.

Abstract Introduction Materials and Methods Results Discussion

medicine and health sciences plant science infectious disease epidemiology epidemiology plant pathology biology and life sciences

2015

Burden of Disease Measured by Disability-Adjusted Life Years and a Disease Forecasting Time Series Model of Scrub Typhus in Laiwu, China

Cutaneous infection by Mycobacterium ulcerans, also known as Buruli ulcer (BU), represents the third most common mycobacterial disease in the world after tuberculosis and leprosy. Data on the burden of BU disease in the Democratic Republic of Congo are scanty. This study aimed to estimate the prevalence rate and the distribution of BU in the Songololo Territory, and to assess the coverage of the existing hospital-based reporting system. We conducted a cross-sectional survey (July–August 2008) using the door-to-door method simultaneously in the two rural health zones (RHZ) of the Songololo Territory (RHZ of Kimpese and Nsona-Mpangu), each containing twenty health areas. Cases were defined clinically as active BU and inactive BU in accordance with WHO-case definitions. We detected 775 BU patients (259 active and 516 inactive) in a total population of 237,418 inhabitants. The overall prevalence of BU in Songololo Territory was 3. 3/1000 inhabitants, varying from 0 to 27. 5/1000 between health areas. Of the 259 patients with active BU, 18 (7%) had been reported in the hospital-based reporting system at Kimpese in the 6–8 months prior to the survey. The survey demonstrated a huge variation of prevalence between health areas in Songololo Territory and gross underreporting of BU cases in the hospital-based reporting system. Data obtained may contribute to better targeted and improved BU control interventions, and serve as a baseline for future assessments of the control program. Cutaneous infection by Mycobacterium ulcerans, also known as Buruli ulcer (BU), represents the third most common mycobacterial disease in the world after tuberculosis and leprosy [1]. In Africa, children under 15 years old have the highest incidence, but healthy persons of all ages, races, and socioeconomic classes are susceptible [2], [3]. Rates of infection among males and females are equal [3]. BU most affects the extremities [2], [4], and is diagnosed in the majority of patients at the ulcerative stage [5]. The disease has a scattered focal distribution within endemic regions, which impedes accurate estimation of disease burden [5], [6]. BU is considered as one of the Neglected Tropical Diseases (NTDs) with a poorly known global prevalence [7], and mainly affects remote rural African communities [8]. A recent review on prevalence [9] reported that, of the estimated 7,000 cases of BU reported annually worldwide, more than 4,000 cases occur in Sub-Saharan Africa. The largest numbers of reported BU cases are from the West African countries of Côte d' Ivoire (about 2,000 cases annually), Benin and Ghana, each reporting about 1,000 cases a year [3]. Various prevalence rates (Table 1) have been reported from different endemic regions in Sub-Saharan Africa [6], [10]–[13]. In the Democratic Republic of Congo (DRC), more than 500 BU cases had been reported before 1980 [14]. The first BU case reports in the Province of Bas-Congo were published in the 1960s and 1970s [15]–[17]. However, in-depth interviews of former patients conducted in the Bas-Congo by Meyers et al. strongly supported the concept that BU was an ancient disease in that region [14]. After 1980, there was a silent period of 20 years without any cases reported in the scientific literature [14]. A national hospital-based survey conducted in 2004 identified 487 clinically suspected cases of BU from six provinces [18]. Between 2002–2004, an apparent resurgence of BU was reported in Songololo Territory [4], known to be the main focus of BU in the country [17]. Since the end of 2004, the General Reference Hospital (GRH) of the Institut Médical Evangélique (IME) of Kimpese launched a specialized BU program sponsored by American Leprosy Missions, offering in-patient treatment free-of-charge and supplementary aid. A recent study has shown a strong increase in the number of admitted BU cases at the IME Hospital after the start of the BU Control Project [19]. Although the number of BU cases admitted in the hospital was rising, data on the exact prevalence and the extent of the disease in the region was lacking. We set up a study to obtain relevant information for planning subsequent control activities, and to provide baseline data for future control program assessments. This study aimed (i) to assess the prevalence and the geographic distribution of BU, (ii) to determine the epidemiologic characteristics of BU, and (iii) to determine the project coverage in Songololo Territory, the target endemic region of the project. The Congolese Ministry of Health granted approval to conduct the survey. We obtained ethical clearance for this study from the Institutional Review Board of IME (N° IME/CS/01/2008). All patients, or their guardian in the case of minors, provided written informed consent for all diagnostic and treatment procedures and publication of any or all images derived from the management of the patient, including clinical photographs that might reveal patient identity. After informed consent had been given, data were recorded on a Community BU Form recommended by WHO. Patient care was free of charge. The case search covered two rural health zones (RHZ), Kimpese and Nsona-Mpangu, both located in Songololo Territory (Figure 1), one of ten territories of Bas-Congo Province. It is situated in the District of Cataractes and covers an area of 8,190 Km2, approximately 15. 2% of the total surface of the province, with a population of 237,418 inhabitants in 2008 (enumeration conducted on December 2007 by the Central Offices of the 2 RHZ). An average of 6 persons per household was used as a regional estimate, giving a total of 39,569 households to be visited by 80 community health workers (CHW). Songololo Territory is limited in the north by the Congo River, in the west by Sekebanza Territory, in the east by Mbanza-Ngungu Territory and in the south by the northern border of Angola. Each RHZ is subdivided into 20 health areas (Table S1 & Table S2). The primary level of health care facilities includes the Rural Health Posts (HP), Health Centres (HC) and Reference Health Centres (RHC), and the secondary level is represented by the GRH. We conducted a cross-sectional survey (July–August 2008) using the door-to-door method simultaneously in the two RHZ of the Songololo Territory (i. e. , Kimpese and Nsona-Mpangu), each containing twenty health areas. Cases were defined clinically as active BU and inactive (healed) BU in accordance with WHO-case definitions [20]. We defined functional limitation as any reduction in the range of motion of one or more joints, and assessed it by clinical observation. Lesions were considered as mixed forms when the simultaneous presence of different forms of disease, including bone and joint involvement, in the same patient was noted. In addition, we defined as simple ulcerative forms (SUF) the ulcerative lesions not associated with other clinical lesions such as papule, nodule, plaque, edema or osteomyelitis at the same site. Lesions were categorized as follows: A single lesion <5 cm (Category I); a single lesion 5–15 cm (Category II); a single lesion >15 cm, multiple lesions, and lesions at critical sites (face, breast and genitalia) or osteomyelitis (Category III). The status of relapse was assessed by questioning the patients, or their guardian in the case of minors, on the history of the lesion, and defined as the reappearance of an ulcer or another form of the disease at the original site of the lesion or elsewhere during the 12 months that followed the end of the previous treatment (antibiotics and/or surgery). This study was conducted in two phases: a preparatory phase and an investigation phase. During the preparatory four-week phase (June 2008), the purpose of the study was explained to the local political and health authorities, and their approval was obtained. Then, 80 CHW, i. e. , 40 per RHZ, were trained in the use of the survey tools (BU community form, pictorial document to recognize BU) and in the identification of suspected BU cases in their communities. We also trained six physicians (working in the RHC of both RHZ), two nurse-supervisors of the leprosy and tuberculosis program (LT), and 40 head nurses (in charge of peripheral health areas), in active case-finding of BU cases and in the use of the survey tools. For the survey, each RHZ was provided with 1 motor bike, 1 Global Position System device, 4 digital photo cameras, 30 bicycles (at least 1 for each health area), 25 megaphones (at least 1 for each health area), drugs and required medical and laboratory consumables. The investigation phase was divided in two periods. The first period (two to three weeks depending on health area) consisted of making an inventory of all BU-like cases by the CHW, using the door-to-door approach in all villages and in each section of two cities in Songololo Territory (Songololo city and Kimpese city). The recommendation to CHW was to visit 40 households per day. A pictorial document, showing different clinical manifestations of BU, was presented to the head of the household or his/her representative asking if any household members presented similar lesions. If the head of the household was not present, the household was revisited once. The second period (6 weeks) included the clinical validation of suspected BU cases by trained health professionals. The eight validation teams were each composed of two people: firstly, a team member of the BU Project (physician or nurse), or another physician, or a LT supervisor, and secondly one of the head nurses. The diagnostic confirmation process of suspected cases involved the collection of swabs from ulcerative lesions and fine needle aspirates from non-ulcerative lesions, followed by laboratory analyses (bacteriology and/or molecular biology) according to WHO recommendations [20]. The initial direct smear examinations for acid-fast bacilli were made at the IME/Kimpese laboratory, followed by in vitro culture for M. ulcerans. Samples were sent in tubes to the “Institut National de Recherche Biomédicale” in Kinshasa, DRC, where PCR for the detection of M. ulcerans DNA was performed, according to WHO recommendations [20]. The external quality control was conducted by the Unit of Mycobacteriology of the Institute of Tropical Medicine in Antwerp, Belgium. The study was carried out simultaneously in the different health areas of both RHZ. Data were recorded on a standardized Case Registry Form elaborated by WHO (BU02), entered into an Excel database (Microsoft Corporation, Redmond, WA) and analyzed with Epi-Info version 3. 3. 2 (Centers for Diseases Control and Prevention, Atlanta, GA). The Pearson chi-square test was used to compare proportions with a significance level set at 5%, and the Fisher' s exact test when an expected cell value was less than 5. Coverage was calculated as the number of active cases detected who had visited the BU reference center in IME Hospital. We produced the distribution maps of BU in Songololo Territory using ArcGIS 9. 2 (ESRI, Redlands, CA, USA). The CHW visited a total of 39,044 households distributed across 9 sections of two cities (Kimpese and Songololo), 46 hamlets and camps, and 547 villages of the Songololo Territory. The estimated coverage of the study was 98. 6%. During the household visits, the CHW inventoried 2,516 persons with BU-like lesions, among which 775 (30. 8%) were validated in a second step as probable cases of BU, all forms included (i. e. , 259 with active and 516 with inactive lesions). A total of 72 out of 241 (30%) patients with active lesions in whom a sample could be taken were confirmed by at least one positive laboratory test for M. ulcerans. The overall prevalence of BU (active and inactive) in Songololo Territory was 3. 3/1000 inhabitants, varying from 0 to 27. 5/1000 between health areas, while the prevalence of active BU was 1. 1/1000 inhabitants with the minimum of 0. 3/1000 when only active, laboratory confirmed BU, was considered. Table 2 shows the prevalence of different BU forms in both RHZ of Songololo Territory, and the distribution per health area is presented in Figures 2,3, S1 and S2. The overall prevalence for the RHZ of Kimpese was 2. 6 per 1000 inhabitants and could vary between health areas from 0. 1 (Kimbanguiste) to 24. 4 (Mukimbungu). The prevalence of BU active forms was 1 per 1000 inhabitants, varying between health areas from 0. 1 (Kimbanguiste) to 5. 7 (Mukimbungu). The health areas of Mukimbungu and Kasi, located in the North of the RHZ of Kimpese, are the most endemic, representing together 60% of the identified patients during the survey (Table S1). Sixty percent of the identified patients in the RHZ of Nsona-Mpangu were from 3 health areas, Kisonga, Nkamuna, and Songololo (Table S2). The overall prevalence in this RHZ was 4. 4 per 1000 inhabitants, varying from 0 (health areas Nduizi, Nkenge and Pala Bala) to 27. 5 (Kisonga). The prevalence of active forms of BU was 1. 3 per 1000 inhabitants, varying between health areas from 0 (Nduizi, Nkenge and Pala Bala) to 3. 8 (Kisonga). The age distribution of all cases ranged from 2 to 94 years (Median 27, Interquartile range (IQR) 14–44) with no significant differences between active and inactive cases. The supplementary tables provide the detailed information. We observed a predominance of female gender (60%) among the recorded cases. Among the 259 patients with active lesions, no sex difference was observed, as 130 (50. 2%) were female. The proportion of new cases was far higher (94%) than the relapses. The ages ranged from 2 to 94 years (Median 27 years; IQR 11–47 years), and the distributions in the two RHZ were similar. Among these 259 patients, 192 (74%) had ulcerative lesions and 62 (23. 9%) were diagnosed with functional joint limitations. Lesions on the limbs were predominant, representing 90% of the sites of lesions. Regarding the patients' categorization, 48. 8% were in category I, 31. 5% category II, and 19. 7% category III. The proportion of patients with ulcerative lesions was higher (p<0. 001) in the RHZ of Kimpese (83%) compared to the RHZ Nsona-Mpangu (63. 6%). Less than half of the patients of the RHZ of Kimpese (41. 2%) and more than half (57. 6%) in the RHZ of Nsona-Mpangu were in category I (p = 0. 031) (Table S3). Female patients predominated amongst active confirmed cases compared to unconfirmed cases; on the other hand, male patients were more frequent in active unconfirmed patients (p = 0. 029). No differences in the age distribution were observed between active confirmed and unconfirmed patients. The lower limb locations were significantly more frequent amongst active unconfirmed patients (p<0. 001). Upper limb sites predominated (p<0. 001) amongst active confirmed patients (Table 3). Features of active cases in the two RHZ were quite similar, with a few exceptions. The ulcerated forms (p<0. 001) and functional limitations on diagnosis (p<0. 001) predominated in the RHZ of Kimpese. Features of inactive cases in the two RHZ were similar but functional limitations were more often observed in the RHZ of Kimpese (p = 0. 005) (Table S4). Only 25 BU patients were admitted and notified at the General Hospital IME/Kimpese between January and August 2008, amongst which 18 were still under treatment for active BU during the survey. Thus, 93% of all active BU patients at the time of the community survey were not captured by the hospital-based reporting system, corresponding to a ratio of 1 reported case for approximately 13 unreported cases. The present study is the first exhaustive population-based survey in DRC aiming to assess the prevalence and distribution of BU in a well-circumscribed endemic region. The survey demonstrated a huge variation in prevalence between health areas and gross underreporting of BU cases in Songololo Territory, compared with the ongoing hospital-based reporting system. Case-definition during the survey was essentially clinical. Case validation was performed by physicians from the BU project and physicians working in the area, well-trained in BU diagnosis, assisted by either a nurse from the BU project or a LT-supervisor, with the nurse responsible for the health area. We are aware of the limitations of clinical diagnosis, which is dependent on the range of experience of health professionals. This may account for certain non-BU cases included in this study. In endemic regions, depending on the clinical stage of the disease, BU may be confused with many other conditions such as nodular onchocerciasis, cyst, lipoma, lymphadenitis, phagedenic tropical ulcer, pyomyositis, necrotizing fasciitis [20], [21], to name a few. Our study showed that 72 out of the 241 (30%) patients who were tested, were confirmed in the laboratory. The low confirmation rate is mostly due to the relatively high number (almost half) of the ulcers being in an advanced stage of healing. Likewise, the technical problems encountered by peripheral health professionals when sampling non-ulcerated lesions and wounds, where mixtures of traditional herbs had been applied, may have played a role. Nevertheless, lesions due to another etiology misclassified as BU cannot be excluded, as lower limb locations were significantly more frequent among active unconfirmed patients. Indeed, among 92 clinically suspected patients recruited from the RHZ of Nsona Mpangu, Kibadi et al. found 31 (33. 7%) PCR negative patients and among them, 25 with histopathological features not compatible with BU (chronic inflammation and bacterial infections due to gram positive cocci) [22]. Despite these limitations, we suggest that our results reflect the endemicity of BU in Songololo Territory reasonably well. In fact, the areas previously established as most endemic were corroborated through this survey, as were the non- or hypoendemic areas [15]–[17], [4]. When considering only active lesions, no sex difference was observed, similar to findings in other studies [2], [11], [12], [23], [24], although our study showed a predominance of females among all cases detected (active and inactive), because among inactive cases, 64. 9% (335/516) were females and only 35. 1% (181/516) were males. Females predominated also among active confirmed BU cases. This preponderance may be due to time itself, or the fact that the population was predominantly female. When referring to the national census figures (July 2008 estimates), for a total population of 66,514,504 inhabitants, 50. 3% were female and 49. 7% male. Among the 259 patients with active lesions, the majority (66%) were over age 15, similar to previous findings in the same area [19]. Ages observed in this survey were higher than found in other disease-endemic countries [2], [10], [12], [25]. The median ages for both RHZ were similar with the median age of 25 years found in Ghana [11], and relatively high when compared to the 15. 5 years observed in Cameroon [13]. The predominant clinical presentation was an ulcerative lesion in 192 cases (74%). This is consistent with studies in Côte d' Ivoire [10] and Cameroon [12], [13], while the percentage of ulcerative lesions was lower in some other studies, for example, 48. 5% in Ghana [11], approximately 50% in Benin from 1997 to 2001 [26] and 57. 5% in 2004 in the same country [6]. Of the 259 active cases, 62 (23. 9%) were diagnosed with joint functional limitations, similar to previous findings in the same area [19], and in other African endemic regions [6], [12]. The general finding of limbs being most affected was confirmed in this study [2], [11]–[13], [23], [24]. The results presented in Table 3 shows that nearly 50% of the BU patients had category I lesions. A similar observation was made in the District of Akonolinga, Cameroon [13]. Ambulatory treatment, based on antibiotic therapy in the primary health care facility, is indicated for this category of patients. Indeed, most category I and some category II lesions may heal completely with antibiotic treatment alone [3], [27]. The introduction of antibiotic therapy [28] has shifted the balance between surgical treatment, mainly limited to reference centers, and antibiotics administered at the most peripheral level of the health system [3]. The clinical presentation of BU was different in the two health zones (Table S3). The degree of functional limitation was significantly higher in patients in Kimpese and they had more often ulcerated lesions. We speculate that this difference is most likely due to differences in health seeking behavior, with higher patient delays in Kimpese, notwithstanding the fact that they were living at shorter distance from the IME hospital. In recent years, an influential religious sect has been a factor in the reluctance to seek medical care in the Kimpese area. Although the number of BU patients admitted at the hospital has increased in recent years, the survey results have demonstrated that the coverage of the population at risk was still insufficient. Of the 259 patients with active BU, 18 (7%) had been reported in the hospital-based reporting system. Porten et al. reported a coverage of 16%, limited to the area close to the Akonolinga hospital in Cameroon, where Médecins Sans Frontières (MSF) opened a BU programme in 2002. The need for improved access to care in the high prevalence areas was emphasized [13]. In the same area, Grietens et al. found that despite the significant reduction in costs for medical care, hospital treatment for BU often remained financially and socially unaffordable for patients and their households, leading to the abandonment of biomedical treatment, or avoiding it altogether. They concluded in their study that from a socio-economic perspective, a decentralized treatment system may limit the impoverishment of households caused by a long hospitalization period [29]. We agree with this opinion because bringing treatment as close as possible to the communities will have a significant mitigating impact on the socio-economic repercussions of BU. The survey demonstrated large variations in prevalence between health areas within an endemic health zone consistent with previous studies in other African BU-endemic regions [6], [12], [13]. Tables S1 and S2 show that in both RHZ, 60% of patients were respectively identified from 2 out of 20 health areas (Mukimbungu, Kasi) in the RHZ of Kimpese and 3 out of 20 health areas (Kisonga, Nkamuna, Songololo) in the RHZ of Nsona-Mpangu. Therefore, priority in case detection should be given to the most endemic health areas. A close collaboration with the provincial Leprosy & Tuberculosis control officers may facilitate the integration of BU activities at the primary health care centers. In fact, the use of the same case-confirmation network or the organization of integrated supervisions would help to reduce the BU intervention costs. Data obtained in this survey may contribute to better targeted and improved BU control interventions, and serve as a baseline for future assessments of the control program.

Buruli ulcer (BU) is a necrotizing bacterial infection of skin, subcutaneous tissue and bone, caused by an environmental pathogen, Mycobacterium ulcerans. BU is considered as one of the Neglected Tropical Diseases with a poorly known global prevalence, and mainly affects remote rural African communities. Data on the burden of BU disease in the Democratic Republic of Congo (DRC) are scanty. The present study is the first exhaustive survey in DRC on the frequency of BU in the community. The survey demonstrated large variations in prevalence between health areas in Songololo Territory. Moreover, our data showed that the BU cases in the hospital-based reporting system reflect only the tip of the iceberg of the true active BU prevalence. Indeed, only one in thirteen active BU cases was notified at the hospital at Kimpese in the 6–8 months prior to the survey. The present data will serve as a baseline assessment for the evaluation of control interventions in the study area, and, more generally, this study aims to raise awareness about the issue of underdetection of BU and the importance of increasing access to diagnosis and care. As such, we hope the study will contribute to improved control of BU.

Abstract Introduction Methods Results Discussion

2013

Burden of Mycobacterium ulcerans Disease (Buruli Ulcer) and the Underreporting Ratio in the Territory of Songololo, Democratic Republic of Congo

Hepatitis delta virus (HDV) increases morbidity in Hepatitis B virus (HBV) -infected patients. In the mid-eighties, an outbreak of HDV fulminant hepatitis (FH) in the Central African Republic (CAR) killed 88% of patients hospitalized in Bangui. We evaluated infections with HBV and HDV among students and pregnant women, 25 years after the fulminant hepatitis (FH) outbreak to determine (i) the prevalence of HBV and HDV infection in this population, (ii) the clinical risk factors for HBV and/or HDV infections, and (iii) to characterize and compare the strains from the FH outbreak in the 1980s to the 2010 HBV–HDV strains. We performed a cross sectional study with historical comparison on FH-stored samples (n = 179) from 159 patients and dried blood-spots from volunteer students and pregnant women groups (n = 2172). We analyzed risk factors potentially associated with HBV and HDV. Previous HBV infection (presence of anti-HBc) occurred in 345/1290 students (26. 7%) and 186/870 pregnant women (21. 4%) (p = 0. 005), including 110 students (8. 8%) and 71 pregnant women (8. 2%), who were also HBsAg-positive (p = 0. 824). HDV infection occurred more frequently in pregnant women (n = 13; 18. 8%) than students (n = 6; 5. 4%) (p = 0. 010). Infection in childhood was probably the main HBV risk factor. The risk factors for HDV infection were age (p = 0. 040), transfusion (p = 0. 039), and a tendency for tattooing (p = 0. 055) and absence of condom use (p = 0. 049). HBV-E and HDV-1 were highly prevalent during both the FH outbreak and the 2010 screening project. For historical samples, due to storage conditions and despite several attempts, we could only obtain partial HDV amplification representing 25% of the full-length genome. The HDV-1 mid-eighties FH-strains did not form a specific clade and were affiliated to two different HDV-1 African subgenotypes, one of which also includes the 2010 HDV-1 strains. In the Central African Republic, these findings indicate a high prevalence of previous and current HBV-E and HDV-1 infections both in the mid-eighties fulminant hepatitis outbreak and among asymptomatic young adults in 2010, and reinforce the need for universal HBV vaccination and the prevention of HDV transmission among HBsAg-positive patients through blood or sexual routes. Hepatitis B virus (HBV) infection can be prevented by vaccination, which can obviate progression of chronic liver disease to cirrhosis, liver failure and hepatocellular carcinoma. However, owing to limited access to vaccination during the past 30 years in resource-poor countries, more than 2 billion people have been infected with HBV and approximately 257 million remain chronic carriers, especially those infected at birth or in early childhood [1]. Viral hepatitis can be associated with acute severe illness, sometimes leading to fulminant destruction of the liver parenchyma or to chronic inflammatory-based injury, resulting in liver remodeling and regeneration that contributes to fibrosis and early carcinogenic events. Both acute co-infection with HBV and hepatitis D virus (HDV) and superinfection of chronic HBV carriers by HDV can aggravate these clinical spectra [2,3], sometimes leading to fulminant hepatitis (FH) at higher rates than in HBV monoinfection. In the early years after HDV discovery, HDV co- and super-infection profiles accounted respectively for 58% and 42% of delta-associated FH in Western Europe [4]. Furthermore, liver fibrosis accelerates in chronic HBV and HDV infections and HDV may have a role in carcinogenesis [5,6]. A recent meta-analysis studying the prevalence of HDV infection in Sub-Saharan countries indicated that in countries surrounding CAR, such as Cameroon, anti-HD Ab prevalence was up to 14% of HBsAg-positive individuals from a national survey [7]. Furthermore, in the Central African Republic (CAR) among patients followed up in hepatology clinics, the anti-HD prevalence reached 50% and this prevalence was 18% in patients with HCC [7]. Although HDV was identified only 40 years ago [3], it has probably been present in humans for a long time. The virus is ubiquitous, and the prevalence of HDV infection among HBsAg-carriers may vary from 0. 5% to 40%, depending of the geographical area studied [2]. A global estimation of HDV-infected population suggests that 10–20 million HBV-infected people have encountered HDV. The virus has high genetic variability, especially in Africa, Amazonia and Asia [8,9]. Currently, the Deltavirus genus comprises eight major clades, or genotypes, labelled HDV-1 to HDV-8 [9,10]. HDV-1 is ubiquitous while HDV-2 and HDV-4 in Asia and HDV3 in Amazonia are geographically restricted. HDV-5–HDV-8 have a clear African origin but may be found elsewhere, due to slave trading and human migration [9,11]. While HBV is a DNA virus with a reverse transcriptase targeted by antiviral drugs, HDV RNA replicates by hijacking the cellular RNA polymerase II. Therefore, chronic hepatitis due to HBV–HDV is the most difficult-to-treat of all severe virally-induced liver diseases [12–14]. Even though therapeutics are available for chronic HBV or HCV infections, about 15 million HBV–HDV-infected patients worldwide await efficient therapy. Alpha-Interferon or Pegylated-IFN that are currently the sole anti-HDV authorized drugs, may be of limited efficacy in less than one third of infected patients. Furthermore, addition of Adefovir or Tenofovir or Entecavir to Pegylated-IFN does not improve the sustained response rate among HDV-1-infected patients, underlining the importance for developing new approaches [12,14]. Indeed, new strategies, such as viral entry competitors, farnesyl transferase inhibitors and nucleic acid polymers are nowadays promising, even if they are still upstream from a wide clinical use [14]. The severity of HBV–HDV infections could be due to particular evolutionary and/or geographical combinations of HBV–HDV genotypes, and some HBV–HDV pairs may be more pathogenic than others [15,16]. In the mid-1980s, cases of severe acute liver failure in Bangui, Central African Republic (CAR), were associated with an HDV outbreak. At least 124 cases of FH occurred among 154 cases of jaundice at Bangui University Hospital, and 88% of the patients died [17]. The evolution of HDV has not been studied since the outbreak, although the high prevalence of HBV infection [18] and the global instability in CAR constitute a potential background for the spread of HDV that may lead to severe chronic liver disease and increased risk of dying from liver failure. Furthermore, although a national HBV preventive program is still conducted, the civil war in the country might have maintained high levels of HBV and HDV transmission through injuries, blood and sexual contacts, mother-to-child transmission and even unsafe blood transfusion. Indeed, viral hepatitis infections, especially hepatitis B, are likely to continue to be a major problem in the near future in CAR. Therefore, it seems crucial to reappraise the level of HBV and HDV infection in young CAR patients to specify the level of unprotected persons that may benefit from catch-up vaccination, if they are free of infection, and those that might be included in future therapy protocols at a time where the World Health Organisation reflexion encourages specific hepatitis funding among infectious diseases global health strategies. The aims of our study were (i) to determine the prevalence of HBV and HDV infection in young asymptomatic students and pregnant women born at the end or after the end of the FH outbreak, (ii) to determine possible clinical risk factors for HBV and/or HDV infections in this population and (iii) to characterize and compare the strains from the FH outbreak in the 1980s to 2010 HBV–HDV strains. Two cohorts, volunteer students (n = 1298) and pregnant women (n = 874), were prospectively sampled between January and December 2010 (S1 Method and S1 Table). Based on volunteering, students came from almost all public and private high schools in Bangui, the capital of CAR, and all different departments of the University of Bangui (S2 Table). Objectives of the study were carefully explained to high-school directors and to the university general secretary in order to obtain authorisation. The investigators then organized information meetings in each secondary school and faculty before distributing the information notice. Written consent forms were obtained from those who freely accepted to participate to the study. Therefore, this recruitment should be considered as passive. Interested students filled the questionnaire anonymously. Recruitment of pregnant women (PW) was performed in all public health structures (n = 6) with maternity wards in Bangui. In these places, medical staffs, including qualified midwives involved in the Mother-to-Child HIV transmission prevention (MCTP), were responsible for sampling. PW coming for prenatal consultation were sampled (dried blood spot) after an initial pre-test counselling. All consecutive pregnant women that accepted to participate to the study signed the consent form and answered the epidemiological questionnaire. Those who had an HBsAg positive result were then followed at the 2 reference public hospitals: “Communautaire” and “l’Amitié”, under the supervision of one of us (AS). No recruitment occurred in private clinics. In all, 874 consecutive pregnant women agreeing to participate were included. The information asked for all participants included gender, age, place of living in Bangui, occupation, marital status, personal viral hepatitis history and previous HBV vaccination. Risk factors investigated were surgery, dental extraction, blood transfusion, IVDU, tattooing, use of sharp tool, alcohol intake and sexual risk factors including ancient and actual sexual partners and use of condom (S1 Method: French-to-English translated questionnaire). The study protocol was approved by the Scientific Committee of the Faculty of Medicine (CSVPRS, 12/2009; University of Bangui, CAR) and by a French ethical committee (CCPPRB, CPP 09024/2009; Saint-Germain-en-Laye, France) under ANRS guidelines. Written informed consent was obtained from all participants. For subjects under 18, parental consent was obtained. Each participant or parent was individually informed of the results of the serological tests. 1984–1987 outbreak: Archived samples had been conserved in a cold room (–20°C) in Lyon, France, since the FH outbreak [17]. For HBsAg serological studies, we used 179 serum samples from 159 individuals, including patients and close relatives or health-care workers (HCWs), sampled at the same time. Because of insufficient volumes, we conducted molecular analyses for only 133 samples from FH patients and close relatives or HCWs. However, we could not distinguish between these categories because the identifiers were no longer available. Specimen collection in 2010: Dried blood spots (n = 6/individual) were prepared as described previously [18]. Filter papers were sealed and stored at the Institut Pasteur de Bangui at –20°C with desiccant until serological assay. To ensure enough sample for molecular analyses of HBV and HDV, 10 mL of total blood were taken from the HBsAg-positive volunteer students (see below) and conserved at –80°C. Dried blood spots were examined as described previously [18]. All three markers HBsAg, anti-HBc and anti-HBs tests (Abbott-Murex, United Kingdom) were performed according to the manufacturers’ instructions in 2160 individuals’samples. HBsAg-positive results were confirmed in the Murex HBsAg Confirmatory test, Version 3. HBsAg-positive samples from the 1985 (n = 84) and 2010 (n = 181) serum collections were screened for HDV with the anti-HDV total antibody test and HDAg detection tests (diaSorin, Italy). From the 114 HBsAg-positive students contacted, 82 individuals agreed to give serum samples for molecular analyses, including all HDV-infected students. Studies were performed at the University Paris 13, Bobigny, France, as previously described [19,20]. To avoid contamination with recent samples, archived samples were extracted and amplified in a separate laboratory in Lyon, France. Initially, Macherey-Nagel extraction procedures were used for 40 samples using RNA and DNA extraction kits to explore both HBV and HDV genotypes, respectively. For subsequent samples, due to the conservation conditions, all efforts were focused on HDV RNA detection and characterization. Total nucleic acids were extracted by an automated procedure (VERSANT® kPCR, Siemens Healthcare). Control water extraction was included for every 10 samples. A 400 base pair HDV fragment was amplified according to reference [9]. (dx. doi. org/10. 17504/protocols. io. mtdc6i6) with HiFi Kapa polymerase (KapaBiosystems). This RT-PCR procedure has been shown to be of higher sensitivity than the RT-qPCR approach [21] and was chosen to minimize the risk of underestimation of the results because of the sub-optimal conservation conditions of the historic samples (-20°C during 25 years). To amplify the full-length L-HD gene, we designed a 5’ primer corresponding to the seven first codons of HDV-1 sequences. For HBV genotyping, PreS1 and PreC-C genes were amplified according to [19] (dx. doi. org/10. 17504/protocols. io. m8jc9un and dx. doi. org/10. 17504/protocols. io. mixc4fn, respectively). Unfortunately, despite many attempts, full-length sequence analysis of mid-eighties FH-associated strains was unsuccessful. Sequence analysis and phylogenetic studies were conducted with different models and Bayesian reconstruction, as described previously [9,10,20]. Sample size was calculated with the following formula taking into account potential cluster effects: N0 = Dz2pq/i2 where N0 is the sample size, p the expected HBV prevalence, q = 1-p, D the cluster effect (set at 1), and i the acceptable precision (set at 2%). Considering a theoretical HBsAg prevalence among students of 15% [18], the calculated sample size is 1257 (z = 1. 96). For pregnant women, the required sample size based on a theoretical prevalence of 13% is 1087. Results are presented as numbers and percentages for qualitative variables, and as means +/- standard deviations (SD) for quantitative variables. Pearson Chi-squared or, when necessary, Fisher exact tests were used to compare categorical data between the different groups. For continuous data, Student’s t test was used. Multivariate analysis using logistic regression was performed to identify factors potentially associated with HDV. A p value <0. 05 was considered statistically significant. Statistical analyses were performed using SPSS for Windows Version 19. 0 (IBM Corp. , Armonk, NY, USA). Among the 159 individuals, identifiers of their status such as health care worker were no longer available, although four family clusters (n = 6,6, 2,2 members, respectively) could be distinguished. Therefore, to study hepatitis delta markers, we retrospectively focused on samples having an HBsAg positive result (n = 94 patients). For 84 of them (89. 4%), both HDAg and total anti-HD Ab were tested using commercial ELISA tests. Results are described in Table 1. HDAg and anti-HD total Ab tests were both positive in 4/84 samples, HDAg was the only positive delta marker in 24/84 samples and anti-HD total Ab was the only marker in 23/84 samples, making 51/84 (60. 7%) samples that had at least one positive HDV serological marker. These results are to be regarded in the context of the fulminant acute HDV/HBV outbreak, at a time when anti-HD Abs could not be detectable in a large number of samples corresponding to Day1, Day2 or Day3 of the severe acute liver disease. HBV infection was highly prevalent in this highly-educated young asymptomatic population. When anti-HBc was used as a marker of previous HBV infection among 2160 individuals, 345 students (26. 7%) and 186 pregnant women (21. 4%) had encountered HBV at some time (Table 1). HBsAg was present in samples from 110 students (8. 5%) and 71 pregnant women (8. 2%) (p = 0. 824) (Table 1). Of the 531 subjects positive for anti-HBc, 181 (110 + 71) had HBsAg detectable in their serum, indicating a high chronicity rate of 34. 1%. As the risk of becoming a chronic carrier is higher in young children than in adolescents or adults and because no significant known adult risk factors of transmission emerged when comparing HBV-positive versus HBV-negative individuals (Table 2 and S3 Table and S4 Table), this result may indicate that possible major routes for HBV infection were (i) being born from an HBsAg-positive mother or (ii) being infected in the very first years of life. To avoid confounding factors, each cohort was also studied specifically comparing the HBV-non infected versus the HBV-infected individuals (S5 Table and S6 Table). Age was slightly significantly associated to HBV previous or current infection in students, whereas previous hepatitis and alcohol intake were identified as risk factors in the pregnant women group. HDAg and total anti-HD (IgM/IgG) were studied in sera from 110/110 (100%) HBsAg-positive students and 69/71 (97. 2%) HBsAg-positive pregnant women. Five students and 13 pregnant women had anti-HD antibodies in their serum, and one student, without detectable anti-HD, had a positive HDAg result thus considered to have acute hepatitis D. Therefore, the prevalence of HDV infection was 6/113 students (5. 4%) and 13/69 pregnant women (18. 8%). Indeed, whereas HBsAg prevalence was similar between the two groups, HDV prevalence was significantly higher among pregnant women than students (Table 1, p = 0. 010). The possible contributing factors for HDV acquisition were a higher mean age (p<0. 001), higher rates of previous blood transfusion (p< 0. 039) and tattooing (p<0. 055) and a lower rate of condom use (p = 0. 049) (Table 3). Concerning this last point, we could not rule out a confounding factor as pregnant women obviously did not always have condom-protected sex and were significantly more frequently infected with HDV than students. On the other hand, we cannot exclude age as an HDV increased risk factor, as this was found, although non significantly, in each cohort considered independently (S7 Table and S8 Table). Because the end of the fulminant hepatitis outbreak occurred late 1987, older students and pregnant women would have been born when HDV transmission risk was still present. Nucleic acid testing for HBV sequences was achievable for 40 historical samples with sufficient residual volume, and the Pre-C/C region of HBV could be amplified and sequenced from eight samples. Seven sequences were of HBV genotype E and one of genotype A2 (S1A Fig). For the contemporary survey, of the 110 HBs-Ag positive students, 82 (71. 9%) agreed to return for medical examination and virus analyses. The HBV DNA viral loads were quantified, except for two samples, in which amplification failed. In the 80 remaining samples, HBV-DNA was undetectable in seven (8. 75%), at a low level (< 3. 3 log10 IU/mL) in 33 (41. 25%) at a medium level (> 3. 3 < 8. 3 log10 IU/mL) in 18 (22. 5%) and at a high level (> 8. 23 log10 IU/mL) in 22 (27. 5%), indicating that half of the asymptomatic students had moderate-to-high viral replication. Amplification of the Pre-S1 coding region of the HBV genome resulted in 50 interpretable HBV sequences that were submitted to Bayesian analyses with reference sequences (S2A Fig). Tree reconstruction indicated that most of CAR2010 were of HBV genotype E (n = 48) whereas 2 CAR2010 strains corresponded to HBV_DE recombinants (S2A Fig and S1B Fig and S2B Fig). In addition, we then aimed to amplify the PreC/C from 2010 CAR HBV strains and obtained 35 sequences all affiliated to HBV genotype E (S1B Fig). We also tried to amplify the Pre-S1 coding region from the 40 HBsAg positive historical samples for which DNA had been extracted, focusing on HBsAg-positive ones. We could obtain 5 Pre-S1 sequences corresponding to 3 patients with fulminant hepatitis (HF2, HF19 and HF27). Tree reconstruction, focused on HBV D and E genotypes, indicates that all three FH-associated HBV strains corresponded to HBV_E genotype and one of them could cluster with one CAR2010 HBV strain (S2B Fig). Among 88 of the 94 HBsAg-positive historical samples (Table 1) that were available for HDV molecular studies, a 400-bp target encompassing the 3’ coding region of the LHD gene was successfully amplified, cloned and sequenced from 12 HDV-infected patients. Full-length LHD gene amplification from historical samples was successful in only three cases, although several clones were sequenced from each. For the 2010 samples from the 6 HDV-positive students, the 400-bp target could be amplified in five cases from both serum and dried blood spot and from serum only in one case. All sequences from historical and contemporary samples were unambiguously belonging to the HDV-1 genotype (Fig 1). A Bayesian reconstruction consensus tree, obtained using CAR historical and contemporary sample sequences from this study, African (Cameroon, CAR, Chad, Côte d' Ivoire) sequences (S9 Table), and sequences retrieved from databases, is displayed in Fig 1. The database sequences included the eight HDV genotype-prototype sequences (shown in red in Fig 1) and other HDV-1 CAR sequences characterized in 2009 [22]. The topology of the tree clearly indicates that the FH viral sequences do not form a specific historical clade. Thus, all HDV sequences from the HF-mid-eighties strains did not derive from a common ancestor in a single phylum, unrelated to the recent HDV sequences retrieved from asymptomatic students. Instead, one large clade intercalates historical and recent sequences from our study in CAR, plus other Central and West African HDV-1 strains, and sequences from Ethiopia, Somalia and Lebanon characterised previously. Inside this clade, there is one strong sustained branch (posterior probability (pp) = 0. 99) that includes four FH strains (FH12, FH82, FH88 and FH123) and sequences from strains that infect patients in CAR and Chad. All the other FH strain sequences cluster (pp = 1) with isolates from patients originating from CAR, Cameroon, Côte d’Ivoire and Lebanon. Among HDV-1 strains, recent full-length HDV genome analyses have raised the possibility of the existence of at least 4 HDV subgenotypes [23]. Even if we could only obtain partial sequences due to conservation conditions, we reanalysed the HDV sequences and conclude that FH12, FH82 FH88 and FH123 were clearly affiliated to the suggested HDV-1a (Ancestral African) subgenotype. In contrast, the other ancient and recent trains clustered together with the Lebanon strain suggesting that they corresponded to HDV-1b subgenotype stains. Although a formal analysis should have included full- length sequences, these results reinforce the fact that the fulminant hepatitis outbreak in the mid-eighties was linked to very different HDV-1 strains. Translation of the coding region for the C-terminus of HDAg showed a serine residue at position 202 in all sequences (Fig 2), corroborating an African HDV-1 affiliation [20]. FH27 isolate clone 4 sequence had a nucleotide deletion leading to a frameshift mutation in the L-HDAg sequence that abrogates the farnesylation signal. Otherwise, the comparison of L-HDAg amino acid sequences from historical FH samples with recent samples revealed no specific FH-associated signature. To study the persistence of hepatitis B and hepatitis Delta viruses twenty-five years after a major HDV and HBV FH outbreak in an equatorial city in Africa, we conducted a prospective search among young healthy adults in 2010. The study was not designed to specify HBV and HDV prevalence in extensive Bangui populations, but was focused on individuals born at or after the end of the epidemic outbreak, to assess viral long-term continuance. After patient inclusion, questionnaire filling and screening, further actions included volunteer-based HBV vaccination and medical examination, for seronegative individuals and HBsAg-positive carriers, respectively. An important observation is that both HBV and HDV are still actively circulating in CAR. Indeed, both students and pregnant women had a high prevalence of previous HBV infection and HBsAg-chronic carriage. In this survey, including a highly educated population, more than one third of HBV-infected individuals were chronic carriers. It has been estimated that 90% of newborns infected at or shortly after birth will become HBV carriers, although this rate decreases with age, until less than 5–10% for adolescent or adult HBV-infections [24]. In this study, most of the students and pregnant women who were anti-HBc-positive had probably been infected at an early age, indicating the importance of sustaining an anti-HBV vaccine strategy for newborns in CAR. Furthermore, HDV could be amplified from all six students with anti-HD or HDAg, indicating continuing HDV replication and an active reservoir that also could have been prevented by early anti-HBV vaccination. Indeed, an important issue corresponds to the fact that a very low prevalence of earlier anti-HBV vaccination (<2%) was observed and that a failure for students to participate for a de novo free anti-HBV vaccination was evidenced, both indicative of HBV (and HDV) potential acquisition in adulthood. HDV infection was still active in 2010 cohorts, however, due to our study design focused on asymptomatic young adults, no fulminant hepatitis was observed in 2010 infections. By comparison, a study conducted in the Democratic Republic of Congo revealed than HBV-HDV could be found in 26. 1% of acute febrile jaundice samples from a national yellow fever survey from 2003 to 2011 [25]. In CAR, HBV-E and HDV-1 were present in the 1980s and 2010 and no specific fulminant hepatitis-associated HDV strain pattern was clearly evidenced comparing 1980s to 2010 strains. During the FH outbreak in Bangui in the mid-1980s, the disease was associated with specific pathological lesions (spongiocyte-associated hepatitis) involving lympho-plasmocytoid infiltrates, lesions of eosinophilic necrosis and massive macro- or micro-vacuolar steatosis, giving cells a particular" morula-like" type and resulting in a high mortality rate [17,26]. Such strain-associated lesions could be propagated in the woodchuck model [26]. Similar lesions have been described in South American FH outbreaks [27,28] that were linked to HDV-3, usually associated with HBV-F, although HBV-A and HBV-D have also been implicated in Amazonia [29,30]. In Bangui, the outbreak involved HDV-1, here associated with HBV-E. This particular histopathological type of FH, associated with high mortality, was therefore not due to a single HBV–HDV genotypic association. During the outbreak, it may be hypothesised that a recent combination of different HDV-1 subgenotypes with the Twentieth Century-introduced African HBV-E helper, rather than HBV-A or HBV-D, could have resulted in a severe disease. Such an explanation was suggested in a study in Taiwan, where the spread of HDV-1 to the Asian HBV-B and HBV-C genotypes resulted in more FH than the indigenous HDV-2 [31]. Alternatively, there may be in equatorial rainforests some as yet unrecognised cofactor that might increase the number of cases of fatal acute liver failure observed in both CAR and Amazonia [25,32]. In this study, the potential role of specific HDV-1 strains or its HDAg protein in the pathological lesions was looked for. We found no evidence of a specific HDV-1 fulminant associated clade. Indeed, within the limits of the sequences obtained, the 1985 HDV-1 FH strains were indistinguishable from asymptomatic-associated 2010 HDV-1 strains. Due to conservation conditions since the outbreak, we were unfortunately unable to amplify full-length FH-associated HDV genomes, even using many different approaches. This obviously cannot rule out some specific virulence factors that could have been deduced from comparison of full-length genomes between historical fulminant strains and recent asymptomatic-associated strains. The FH strains and their HDAg sequences intermingle with strains isolated from 2010 asymptomatic patients in CAR and with strains isolated elsewhere in Africa. The only exception corresponded to one HDV clone that displayed a frameshift mutation at the 3’ terminus of the L-HDAg gene that would eliminate the farnesylation signal of L-HDAg (Fig 2, FH27 clone 4). This may affect virion assembly and perhaps induces intracellular retention of viral component that may accumulate in infected cells. This variant was observed only once among 45 clones analysed from 12 FH-HDV strains. However, it can be hypothesised that mutants leading to intracellular retention may not easily egress from the infected-cell and might therefore be under-represented in the serum. In addition, they may be potentially cytopathic for the infected hepatocytes. For example, in HBV monoinfection of the human hepatocyte, retention of HBV-PreS1 mutants results in specific pathological lesions known as" ground-glass" hepatocytes [33]. It is disappointing that vaccination against HBV is so uncommon among pregnant women and students with a high education level, even among medical students. Neonatal vaccination (DTP-HepB-Hib) has been practised in CAR since 2008, but coverage is still not universal (http: //www. gavi. org), in part due to civil unrest. At present, there is also no HBV immunotherapy nor immediate vaccination of neonates born outside hospitals, in the absence of qualified midwifes [1]. Indeed, modelisation has suggested that the proportion of new chronic cases arising from mother-to-child transmission may increase by up to 50% in 2030 [34]. Neonatal anti-HBV vaccination is urgently needed, preferably including a dose as near as possible after birth. Ideally, vaccination of as yet unprotected young populations should also be sustained [1,34]. Meanwhile, at least 8% of the young population of Bangui is at risk for further HDV superinfection and related liver damage. The situation outside of the nation’s capital is probably worse as a recent evaluation has estimated an HBsAg rural prevalence of 11. 65% in CAR [35]. The results of this study should clearly contribute to sensitise the health authorities to consider HBV and HDV infections as a major health challenge in the Central African Republic, and could lead to specific surveys, screening and medial actions among populations at risk of HBV acquisition and/or transmission.

In the Central African Republic (CAR), due to 20 years of conflict, the health system has been disorganized. This could contribute to maintenance of high transmission levels of Hepatitis B Virus (HBV) and its satellite Hepatitis Delta Virus (HDV). This work studies the evolution of both infections 25 years after a fulminant hepatitis (FH) outbreak occurring in the mid-1980s associated with HDV superinfection. In young asymptomatic adults, the results show that both HBV and HDV were still actively circulating in CAR in 2010. Indeed, more than one third of HBV-infected individuals were chronic HBV carriers. Furthermore, HDV infection could be spreading among 10% of them through blood and sexual transmission. The past FH outbreak and contemporary infections were both associated with heterogeneous HDV-1 strains, combined with HBV-E. Vaccination against HBV was uncommon among pregnant women and students, even among medical students. The study constitutes warning signals to help CAR health-care reconstruction and underlines the importance of HBV vaccination. The high level of HBV infection creates a background for HDV superinfection. Neonatal HBV vaccination is needed, together with vaccination of unprotected populations. Awareness of health authorities as well as the general public would help reduce HBV and HDV infections.

Abstract Introduction Methods Results Discussion

taxonomy medicine and health sciences body fluids pathology and laboratory medicine pathogens immunology microbiology hepatitis b virus liver diseases viruses preventive medicine phylogenetics data management phylogenetic analysis gastroenterology and hepatology vaccination and immunization research and analysis methods public and occupational health sequence analysis computer and information sciences bioinformatics medical microbiology epidemiology microbial pathogens hepatitis delta virus hepatitis viruses evolutionary systematics liver disease and pregnancy blood anatomy viral pathogens physiology database and informatics methods biology and life sciences evolutionary biology organisms

2018

Hepatitis B and hepatitis D virus infections in the Central African Republic, twenty-five years after a fulminant hepatitis outbreak, indicate continuing spread in asymptomatic young adults

The global incidence of snakebite is estimated at more than 2. 5 million cases annually, with greater than 100,000 deaths. Historically, Myanmar has one of the highest incidences of venomous snakebites. In order to improve the health outcomes of snakebite patients in Myanmar, access to accurate snakebite incidence data is crucial. The last population-based study in Myanmar was conducted more than a decade ago. In 2014, the Ministry of Health and Sports data from health facilities indicated an incidence of about 29. 5 bites/ 100,000 population/year (a total of 15,079 bites). Since data from health facilities lack information about those who do not seek health care from government health services, a new population-based survey was conducted in 2 rural areas of Mandalay region. The survey data were compared to those obtained from healthcare services. 4,276 rural respondents in Kyaukse and Madaya townships in Mandalay Division were recruited using cluster sampling that involved random selection of 150 villages and random sampling of 30 households from each village. One adult member of each household was interviewed using a structured questionnaire. One respondent from each of 4,276 households represented 19,877 residents from 144 villages. 24 people in these households had suffered snakebite during the last one year giving an annual incidence of 116/100,000. During the last ten years, 252 people suffered snakebites. 44. 1% of the victims were women. 14% of the villages reported 4 or more bites during the last ten years, whereas 27% villages reported no snakebites. 92. 4% of the victims recovered fully, 5. 4% died, and 2% suffered long term health issues. One victim was reported to have died from causes unrelated to the snakebite. While there was no statistically significant difference between outcomes for children and adults, 4 of 38 of those under 18 years of age died compared to 7 of 133 adults between 19 to 40 years of age. This incidence reported by the community members points to substantially more snakebites than the number of snakebite patients attending health facilities. This higher incidence points to the need for a nation-wide population-based survey, community education about gaining access to care where antivenom is available, and to the potential need for a larger supply of antivenom and expansion of medical care in rural areas. Snakebite is a major global health issue. The estimated global annual incidence is about 5 million snakebites resulting in between 81,000 and 138,000 deaths [1]. In 1998, Chippaux [2] estimated the incidence as high as 5. 4 million bites and in 2000, White [3] suggested more than 150,000 deaths. In India alone, a well-designed community-based study documented more than 45,000 deaths in 2005 [4]. Many venomous snakebites lead to severe and persisting morbidity resulting from amputation, infection, scarring, stigmatisation and psychological trauma, all resulting in substantial physical, psychological and economic disabilities and hardship. The global burden of snakebite is predominantly in tropical developing countries where manual farming practices predominate. With inadequate access to appropriate health care and the high indirect cost associated with receiving care in hospitals in major urban centres, snakebites victims and their families often face significant economic loss [5,6]. Many snakebite victims do not seek or receive care in the formal health care sector. Epidemiological studies dependent on data from the health care system will therefore fail to capture information about these victims. This results in serious underestimation of the true snakebite incidence. This problem is exacerbated by the fact that snakebites mainly affect people in countries where health systems are evolving and the health management information systems are insufficiently robust to capture information even about those presenting to health facilities. Global burden estimates indicate that South and Southeast Asia and sub-Saharan Africa have the highest snakebite burden in the world [1]. Historically, within Southeast Asia, Myanmar has one of the highest incidences of venomous snakebite. In 2014, Ministry of Health and Sports (MOHS) data from health facilities indicated about 29. 5 bites / 100,000 population / year (a total of 15,079 bites). The most recent community level data available on snakebite in Myanmar were obtained more than a decade ago [7]. For planning an effective response to this important public health issue and to facilitate effective clinical care for all snakebite victims, it is important to re-assess the population-based snakebite incidence in addition to having information about the patients who receive care within the health system. With demographic changes, changes to agricultural practices, education, economic development and associated changes in prevention and access to health care, it is important periodically to re-assess the incidence, morbidity, mortality, community knowledge and practices in relation to prevention and service utilisation. We conducted a survey in two townships near Mandalay to describe snakebite incidence and mortality in rural areas, as an overwhelming majority of snakebites occur in these areas. A secondary aim is to compare these survey data with data from health centre and hospital presentations. The research was conducted with ethics approval from the Ethics Committee, Department of Medical Research, Ministry of Health, Myanmar, and the Human Research Ethics Committee at the University of Adelaide. A survey to assess the population-based snakebite incidence in the Mandalay region was conducted as part of a large health system and community development collaborative project involving the Myanmar Ministries of Health and Industry in conjunction with the University of Adelaide and other Australian and international institutions. This collaborative project aims to improve outcomes of snakebite patients, with a focus on improving the quantity and quality of nationally produced antivenom, its distribution in quantities sufficient for the needs of the various regions of the country, health services development through in-service training and resource improvement, and community education targeting training for prevention, first aid and appropriate use of health services. As part of the larger project, this survey was conducted to measure incidence of snakebite in the previous year, snakebite mortality in the previous 10 years, and snakebite-related knowledge and practices among community members. After consultations with Myanmar colleagues and visits to numerous candidate townships and hospitals, we determined that the appropriate combination of a high incidence of snakebite, accessibility and local support from senior health care workers could be found in the Kyaukse and Madaya Townships of Mandalay Division. The project was in collaboration with the Myanmar MOHS and had approval for these two townships only. MOHS data based on reports from all facilities across the country indicated that Mandalay region and these two townships were representative of the high incidence regions. Incidences based on MOHS data in 2014 were 29. 4/100,000 across the whole country, 38. 3/100,000 across the 7 highest incidence regions (where about 90% of all snakebites in the country occur), and 44. 7/100,000 in Mandalay region. In Myanmar, a ‘township’ comprises not just an urban centre, but also the surrounding rural hinterland that may cover a significant area with many rural villages. All residents of the rural areas of Kyaukse and Madaya townships in Mandalay Division were eligible for inclusion in the survey. We excluded residents who lived in urban areas of Kyaukse and Madaya Townships (16% and 9% respectively). There is a clear distinction in the Myanmar health system between urban and rural areas. The regional health management authority had separate lists for rural and urban areas and that was the basis of distinction. We excluded urban areas as we considered that snakebite would be relatively infrequent in these areas and also that the issues related to accessing treatment would be different. Data from Mandalay General Hospital (MGH) and Mandalay Teaching Hospital (MTH) provide evidence of the relative paucity of snakebite cases in urban areas compared to rural areas. MGH takes patients from the rural regions, whereas the MTH takes patients from the urban areas. In the 12 months from April 2017 till March 2018, MGH admitted a total of 856 patients, and MTH admitted only 153 patients. Also note that many rural bite victims were cared for in Township Hospitals and did not require transfer of care to MGH at any stage. Our estimate, based on the observations across the hospitals is that only about 1 out of 10 rural patients treated at the Township Hospitals required transfer to MGH. The rest were managed at their respective Township Hospitals without onward referral. For each selected household, an adult member was interviewed. History of snakebites and snakebite-related deaths among all household residents covering a period of 10 years prior to the survey was collected. The household was the sampling unit. The sample size was defined as 4,500 households with about 20,000 household members. This is similar to most other recent studies in South and Southeast Asia. The sample size of those studies ranged from about 2,000 to 11,000 households [8,9, 10,11] with the exception of one large study in Sri Lanka where the sample size was 44,136 households [12]. Cluster sampling was used, in line with the method recommended in the World Health Organisation Vaccination Coverage Cluster Surveys [13]. Three stages of sampling involved: (i) stratification by township, (ii) random selection of 75 villages from each of the two townships, (iii) random sampling of 30 households from each village selected from lists of households provided by the local government health departments. Census data on the population of each village was used in stage (ii) so that random sampling of villages was done with probability proportional to size. We used a 150x30 sampling strategy, determined after considering both statistical (precision) and practical (feasibility) issues. Based on a previous survey in Myanmar (7), the anticipated snakebite incidence was 100/100,000. The design effect for snakebite incidence in Myanmar is unknown. Design effects for most variables in UNICEF’s 2009–2010 Myanmar Multiple Indicator Cluster Survey [14] (thirty households per village, as in our survey) were between 1. 5 and 2 and so we used a design effect of 2 in the design of our survey. We felt that it was feasible to conduct interviews in 30 households in a single day and that 150 villages would give good geographical coverage of the study area. The resulting sample size (4,276 households, 19,877 individuals), with correction for cluster sampling methodology, was considered adequate to enable some estimation of snakebite incidence. This sample size is similar to several previous studies of snakebite incidence in Asia (8,9, 10). Data were collected using an interviewer-administered questionnaire that took about 20 minutes to complete. The structured questionnaire (Table 1) included questions about the types of snakes in the area as reported by the respondents. Photos were not shown for the respondents to identify the type of snakes sighted in the area as this method has proved misleading and unreliable. However, patients’ descriptions and use of Burmese names for the different species contributed to our tentative identifications. Snakebites during the last 10 years, first aid and health care use, snakebite outcomes, as well as knowledge and practices for prevention, first aid and treatment. The questionnaires were administered by primary health care workers, midwives, health assistants and public health staff employed at government health centres providing outreach care. The data collectors were trained by the research team and local Project team members, all of whom were native speakers of the Myanmar language. The completed questionnaires were reviewed daily for completeness by the trained data collector supervisors and were then reviewed by the Project team field supervisors for the survey. The questionnaires were double-entered for quality assurance. The statistical analysis was conducted using SPSS (IBM SPSS Statistics Version 24) after the survey data were weighted for respondent sampling probability. Population rates and incidence were adjusted for the design effect of the survey using the SPSS Complex Samples Function. All numerical findings, including 95% confidence intervals, are adjusted for the three stage sampling methodology. In the 4,276 study households, representing 19,877 residents, 24 people had suffered a snakebite during the previous one year, an unadjusted incidence of 123/100,000 and adjusted incidence of 116/100,000 people (95% CI 74/100000–182 /100,000). 252 respondents informed us that at least one person in their households had been bitten by a snake during the previous 10 years (5. 9% of households, 95% CI 4. 8%–7. 3%). Among the victims, 44. 1% were women. Relatively fewer households in Madaya Township had experienced a snakebite (4. 7% of households, 95% CI 3. 3%–6. 7%) during the last 10 years than in Kyaukse Township (7. 3%, 95% CI 6. 7–9. 2%). 111 of 252 (44%) snakebites were reported from 21 villages (14% of the villages) were there had been 4 or more snakebites during the last 10 years. From 40 villages, there were no reports of snakebites, and only 1 snakebite was reported from 38 villages during the previous 10 years, 2 snakebites in 25 villages, and 3 snakebites in 19 villages. Outcomes were reported for 230 victims in the previous 10 years. The respondents informed us that 209 (92. 4%, 95% CI 87. 6%–95. 5%) of these had fully recovered, 14 (5. 4%, 95% CI 2. 9%–10. 0%) died due to that snakebite, and 6 (2%, 95% CI 0. 9–4. 7%) did not recover fully and had long term health issues as a result of the snakebite. It was reported that one victim died from causes unrelated to the snakebite. While there was no statistically significant difference between poor outcomes for children and adults, 4 of 38 victims under 18 years of age died compared to 7 of 133 adults 19 to 40 years of age. A higher proportion of women died (7 out of 87) than of men (7 out of 141). This difference was not statistically significant. The species of snakes reported are listed in Table 2. The respondents informed us that most snakebites occur in and around fields and in the forests (Table 3). Snakebites were reported to occur during all four seasons but peak incidence was believed to during hot periods before the monsoon (Apr-June–Tu Gu—Na Yone season) and during the monsoon period (Table 4). (Note that Tu Gu—Na Yone is shorter than the other three seasons.) The activities most associated with people being bitten were farming and collecting crops (Table 5). Considering the likely better recall about symptoms and treatment for those 24 snakebite victims who were bitten during the one year preceding the survey, the household members were asked questions about symptoms and treatment received. 6 (25%) of those 24 that reported snakebite to one of their family members reported that it was a venomous snake, whereas 3 (12. 5%) reported it was not a venomous snake. Of the 24 respondents 19 could not recall the type of treatment received. Three reported that the victim received anti-venom. There were 2 reports of intravenous fluid being used with one reported to have received oxygen as well. The respondents informed that 8 (33. 3%) of these 24 patients fully recovered. Others did not report if their family member fully recovered or had long term consequences. We acknowledge that the sample size was too small to provide a precise estimate of snakebite incidence. With only 24 snakebites in the past year, the point estimate of 116 snakebites per 100,000 population had wide confidence intervals (74/100,000 to 182/100000). To achieve a precise estimate of snakebite incidence in Myanmar would require a sample size of at least 100,000 people. The lack of detailed information about symptoms, treatment and outcomes including long term consequences is another limitation. Considering difficulty in recalling clinical information by the community members it is challenging to capture such clinical information through a community survey. Research that involves reviewing the data of patients admitted to the township, district and referral hospitals will help capture such clinical information which when used in conjunction with the information about incidence would help refine and strengthen in-service trainings and better define which anti venoms are necessary.

Snakebite is a major health issue affecting large numbers of people, particularly in tropical developing countries. Myanmar has one of the highest incidences of venomous snakebite in the world. Considering changes in demography, development, agricultural practices, knowledge about prevention and preventive practices, regular and accurate assessment of incidence is needed in order to improve public health programs and health services provision to improve health outcomes for snakebite patients. For that purpose, we conducted a large population based survey in Mandalay, which is one of the seven high incidence regions in Myanmar. The survey indicated a substantially higher incidence of snakebites than that suggested by the number of snakebite victims attending health care centres and hospitals. This higher incidence of snakebite has implications for community health service planning, scale of production of antivenom, and the need to improve access to health care centres or hospitals where antivenom is available, and suggests a need for community health education regarding appropriate 1st aid.

Abstract Introduction Methods Results Discussion

medicine and health sciences health education and awareness health services research tropical diseases geographical locations vertebrates animals health care reptiles myanmar rural areas neglected tropical diseases snakebite geography snakes people and places eukaryota asia urban areas health care facilities squamates earth sciences geographic areas biology and life sciences amniotes organisms

2018

Snakebite incidence in two townships in Mandalay Division, Myanmar

Cancer drivers are genomic alterations that provide cells containing them with a selective advantage over their local competitors, whereas neutral passengers do not change the somatic fitness of cells. Cancer-driving mutations are usually discriminated from passenger mutations by their higher degree of recurrence in tumor samples. However, there is increasing evidence that many additional driver mutations may exist that occur at very low frequencies among tumors. This observation has prompted alternative methods for driver detection, including finding groups of mutually exclusive mutations and incorporating prior biological knowledge about gene function or network structure. Dependencies among drivers due to epistatic interactions can also result in low mutation frequencies, but this effect has been ignored in driver detection so far. Here, we present a new computational approach for identifying genomic alterations that occur at low frequencies because they depend on other events. Unlike passengers, these constrained mutations display punctuated patterns of occurrence in time. We test this driver–passenger discrimination approach based on mutation timing in extensive simulation studies, and we apply it to cross-sectional copy number alteration (CNA) data from ovarian cancer, CNA and single-nucleotide variant (SNV) data from breast tumors and SNV data from colorectal cancer. Among the top ranked predicted drivers, we find low-frequency genes that have already been shown to be involved in carcinogenesis, as well as many new candidate drivers. The mutation timing approach is orthogonal and complementary to existing driver prediction methods. It will help identifying from cancer genome data the alterations that drive tumor progression. Carcinogenesis is an evolutionary process driven by the accumulation of advantageous mutations in single cells and the subsequent outgrowth of those cells due to clonal expansion. Mutations in certain genes are present in a large fraction of cancers, such as TP53 mutations; others exhibit high mutation rates in cancers of the same type such as BRCA1 in breast cancer. The functional alterations of these recurrently mutated genes are referred to as hallmarks of cancer [1]. However, cancer genomes contain many more mutations which do not show high degrees of recurrence. There are several reasons for these low rates of recurrence. Firstly, mutations at some loci depend on the presence of mutations at other loci [2]. This dependence may result from an epistatic interaction where a mutation is selectively advantageous only in the context of other mutations. However, cancer diagnoses do not occur at the same point during carcinogenesis. Instead, some tumors are detected very early, some very late, and most tumors are diagnosed at intermediate stages. Therefore, mutations that are highly dependent on other mutations tend to occur late, viz. after the right genetic background has evolved. Hence they are present only in a small fraction of patients. Secondly, mutations in single members of functional groups, such as signaling pathways, are often sufficient to disturb the pathway function. Mutations within those pathways display patterns of mutual exclusivity and low mutation rates across cancer samples, because additional mutations are unlikely as they do not provide an additional selective advantage [3], [4]. Finally, cancer cells accumulate a large number of passenger mutations in the process of carcinogenesis [5]. These mutations are selectively neutral and occur at random, but they are also manifested within a cancer cell population due to their co-occurrence with advantageous driver mutations. The goal of driver–passenger discrimination is to separate these harmless passenger mutations from driver mutations which actually provide a selective advantage and drive tumor growth. Driver–passenger classification approaches fall into three categories. They are either based on (i) mutation frequencies, (ii) mutual exclusivity, or (iii) biological pathway or network information. Mutation frequency-based methods aim at finding either genome-wide or locus-specific mutation frequency cutoffs which are optimal with respect to a given false positive rate or other criterion [6]–[8]. Mutual exclusivity-based approaches try to find sets of genes in which mutations are mutually exclusive, while most of the cancer samples display a mutation in one of the genes in these sets [9]–[11]. The third class of methods relies on enrichment of the driver candidates in annotation databases or in specific subgroups of biological networks [12]. Furthermore, combination of different methods has been used to get high confidence predictions of drivers [13]. Besides computational techniques, experimental approaches are also employed in order to identify driver genes [14]. However, experimental approaches are limited to mutations which show an effect on their own and do not depend on a more complex mutational context. With the exception of mutual exclusivity-based methods, all the above methods assume that the selective advantage of drivers is independent of the mutational and signaling context in which they appear. Mutual exclusivity-based methods assume that the effect of a driver is only present in the absence of other specific mutations. Since some driver genes are dependent on mutations in other genes, their mutation frequency across samples can be low. Here, we describe a computational approach to identify driver genes that have low mutation frequencies due to their conditional occurrence. The method takes cross-sectional binary mutation data as input and aims to separate dependent events from independent ones. Our approach is based on the observation that the probabilities of neutral mutations increase linearly with the total number of mutations, which is used here as a proxy for the timespan between the start of oncogenesis and the detection of the tumor. By contrast, non-neutral mutations that depend on other mutations occur with probabilities displaying a non-linear pattern of punctuated increase (Fig. 1). Extensive simulation studies are used in order to evaluate the best-case performance of the mutation timing approach and to assess its dependency on several factors, including the number of samples, the genotyping error rate, and the variation of the background mutation rates. We analyze real CNA datasets from ovarian and breast cancer as well as SNV datasets from breast and colorectal cancer obtained from the TCGA database. We find a number of known oncogenes and tumor suppressors to be highly ranked as well as genes which have not yet been implicated in tumor development. Furthermore, we find a very low overlap of genes highly ranked by our mutation timing approach and genes which have high degrees of mutation recurrence in the datasets we analyzed, indicating complementarity between our novel and existing approaches. In order to identify candidate cancer driver genes, we propose to detect mutations whose occurrence depends on the presence of other mutations (Fig. 1). The rationale for this approach is that passenger mutations occur independently because they are selectively neutral, whereas the selective advantage of many driver mutations depends on the genetic background they occur in. For example, mutation of KRAS tends to occur after mutation or loss of FAP in colorectal tumorigenesis [2]. To distinguish independent from dependent mutations, we study their rate of occurrence among tumors. Independent mutations occur at a constant rate (Fig. 1E), and our driver–passenger discrimination approach is based on detecting deviation from this behavior. For example, if mutations occur in a linear order, , then mutation can occur only after all its predecessors have occurred, and once this has happened, the probability of mutation increases much faster than in the neutral case (Fig. 1D; S1 Text, section 3. 2). This difference in the rate of change of observing mutations over time is the basis for our gene ranking procedure. For each mutation m, we consider the conditional probability of its occurrence given that in total at most mutations have accumulated, (1) Here is the total number of mutations, which we use as a proxy for the time of observation relative to the onset of tumorigenesis. Thus, we measure time in number of mutations, , and is the probability of observing mutation before time. For independent mutations occurring at identical rates, is a linear function with constant slope, where is the total number of possible mutations (Fig. 1E; S1 Text, Eq. 21). By contrast, for dependent mutations, is non-linear with a sharp increase in a more confined time interval. Furthermore, we demonstrated that if is involved in dependency relations, then has a higher maximal slope than for independent mutations (S1 Text, Theorem 1). This finding shows that mutational dependencies can be detected by considering the steepest slope of. We model the probability using the sigmoid function, where is the location of the inflection point, the slope at, and the mutation frequency over all samples (Fig. 1F; S1 Text, section 4. 1). Mutation timing ranking then ranks genes by decreasing slope values, . Genes that tend to have the same probability of occurrence over time will be ranked low, whereas genes with a narrow window of high occurrence probability are ranked high (Fig. 2). In order to evaluate the performance of the mutation timing ranking, we conducted simulation studies and compared our approach to the baseline frequency-based approach. We ranked the simulated genes by the mutation timing method and by their marginal frequency and computed and compared AUC values for both rankings. We simulated samples according to a continuous-time Conjunctive Bayesian Network model [15], in which dependencies among binary mutational events are represented by a directed acyclic graph (Fig. 1A). After generating the mutation profiles we added noise by flipping every mutation indicator with probability (Methods). The probabilities of all mutations that are early to intermediate in the dependency structure or are independent of the other nodes can be described very well by the sigmoidal approximation (S1 Fig.). Late mutations with low marginal frequencies suffer from higher fluctuations due to small sample sizes. They are affected more by measurement noise (S1 Fig.). In order to evaluate the influence of the sample size on the classifier performance, we drew samples of different sizes from three different dependency networks (Fig. 3). For the error rate, , we used a value of 0. 01. We applied our classifier to sets of size 100,500,1000, and 5000. Increasing the sample size improved the performance of our classifier (Fig. 4A–C), whereas no performance increase was observed for the frequency-based classifier. Without observation error () and a sample size of 500 the classifier works almost perfectly, for all three networks. With increasing error rate, , from 0 to 0. 1 and sample size of 500, the performance of the classifier drops to the level of the frequency-based classifier and below in case of the linear network C (Fig. 4D–F). The performance of the frequency-based classifier is not affected by the error rate. The marginal probabilities of mutations in network B and the marginal mutation probabilities of late events of network C are lower than for network A. Therefore, ranking by mutation timing is much more sensitive to the error rate. In networks B and C, the marginal frequencies of nodes low in the dependency structure are in fact smaller than the error probability in some cases. In these cases, the mutation timing method performs worse than random. This effec is due to the sigmoidal curves getting flattened out by the noise, because the noise is uniform with rising k. The effect of the noise, i. e. , the flattening of the mutation timing curve, is stronger on these low nodes than on the independent nodes because the marginal frequency of the low nodes in the dependency structure is on average smaller than the marginal frequency of the independent nodes. Next, we investigate the influence of the variation of the waiting times of the independent nodes, i. e. , the per-gene evolutionary rates. We varied the ratio between the minimum and maximum of the support of the log-uniform distribution which is used to sample the waiting time parameter for the independent nodes. The geometric mean of the minimum and maximum was kept at 0. 6,0. 25, and 0. 15, respectively. The ratio between the maximum and the minimum of the log-uniform distribution was set to 1,4, 25, and 100. Then, 500 samples were drawn from the networks with an error rate of 0. 01. The performance of the mutation timing classifiers on the datasets of all three networks did not change when varying the variation of the rate parameter of the independent nodes (Fig. 4G–I). The performance of the frequency-based classifier drops due to the increased number of passengers with higher marginal frequency after increasing the variability of the passenger mutation rate. Increasing the number of independent nodes (to 990,989, and 990, respectively) hardly changed the performance of both the mutation timing and the frequency-based classifier (S2 Fig.). The mutation timing classifier shows similar performance improvements over the frequency-based classifier on random networks with 10 connected nodes and 90, respectively 990 unconnected nodes (Fig. 5). The samples size for this analysis was 500 and the observation error was 0. 01. We applied mutation timing ranking to CNAs identified in 569 ovarian cancer samples available from the TCGA database (download on May 24,2013). All analyzed TCGA data in this paper was downloaded via the cBio portal [16]. The CNAs were called jointly for all samples with GISTIC2 and subsequently mapped to genes for the individual tumors [17]. Only amplifications and homozygous deletions were considered. Gains and heterozygous deletions were not considered as CNAs since they are very difficult to detect and prone to high false positives. Out of 9312 CNAs, 80% were present in between 9 and 60 tumors (10% in less and 10% in more). Most tumors had between 55 and 1075 CNAs (10% less and 10% more). For a certain CNA to be considered for ranking in this analysis, it had to be present in at least 35 of the 569 samples (i. e. , a marginal frequency of at least 6%). Some CNAs affect multiple genes. Those genes were grouped and ranked together. The number of CNAs (genes or groups of genes), which were present in at least 35 samples, was 2515 and their sample-wise mutation profiles were unique among all 569 samples. For approximating time (), we used the overall number of CNAs present in each sample. Lowly ranked CNA-altered genes exhibit a shape which resembles a linear function as expected for passengers (Fig. 6, bottom row), while highly ranked genes exhibit a punctuated rise of their mutation frequency indicating deviation from neutrality (Fig. 6, top two rows). Since we do not consider amplifications and losses separately in this analysis, we subsequently analyzed the type of all highly ranked CNAs. All but ZNF426 (rank 10) show clear consistency in their type of alteration, i. e. , they were either mostly deleted or mostly amplified, but not deleted in some samples and amplified in others (S3 Fig.). Among the top-10 ranked genes (and gene groups) out of 2515 (Fig. 6, S1 Table) we find GRAMD4 (rank 2) which has p53-like function and mediates p73-triggered apoptosis [18]. CNAs affecting GRAMD4 are mostly homozygous or heterozygous deletions (S3 Fig.). Furthermore, amplification of UQCRFS1 (rank 5) has been associated with high grade breast cancer [19] and enhanced colony formation in cell lines [20]. C19ORF12 (rank 8) and PLEKHF1 (rank 9) are almost always amplified together. Both genes were found to be essential for cell proliferation [20]. There was no overlap between the 10 most frequent CNAs and our top 10 ranked CNAs, and we found only two overlapping genes between the two top-100 lists, namely ADAMTS10 and the ADIPOR1/KLHL12 group. Next, we applied our ranking scheme to CNA data from 913 breast cancer samples from the TCGA database (download on May 24,2013). CNAs were called and processed in the same way as describe above for ovarian cancer. The percentage of all 8309 CNAs present in between 7 and 85 tumors (10% in less and 10% in more) was 80. Most tumors had between 4 and 622 CNAs (20% less and 10% more). Due to the higher sample number and hence increased power, we considered CNAs which are present at lower marginal frequencies than in the ovarian study. The threshold for consideration was set to 4%, or 37 of the 913 samples. This threshold was passed by 1713 CNAs, which were subsequently ranked. Highly ranked CNAs as well as lowly ranked CNAs follow the sigmoidal shape (with different slopes) very well (S4 Fig. , S2 Table). Interestingly, all top-10 ranked genes showed amplifications (S5 Fig.). The highest ranked gene is GAL. The Galanin signaling cascade has been proposed as a candidate pathway regulating oncogenesis in human squamous cell carcinoma [21]. MRPL21 (rank 6), which is located near GAL in locus 11q13. 2, has been suggested to play a role in carcinogenesis [21]. Similarly, SLC29A2 (rank 4) has been implicated in the carcinogenesis of hepatocellular carcinoma [22]. CPT1A (rank 3) can also be found in the list of highly ranked genes. It promotes cell motility and is therefore thought to increase the risk of metastases [23]. POLD4 (rank 8) has been associated with genomic instability in lung cancer [24]. However, the cancer driving effect of POLD4 was associated with downregulation of this gene [24] and here we find it consistently amplified. Similarly, low levels of KDM2A (rank 9), a JmjC-domain containing histone demethylase, have been associated with carcinogenesis, and we find it consistently amplified [25]. Besides by CNAs, breast cancer progression is also driven by SNVs [26]. Therefore, we applied mutation timing ranking to SNV data from 772 breast cancer samples from the TCGA database (download on May 24,2013). Since SNVs do not exhibit as high marginal frequencies as CNAs, we set the cutoff for SNVs to be ranked to 1% or 8 out of 772. The final number of ranked genes was 256. SNVs do not show the punctuated rise in mutation frequency as pronounced as CNAs (S6 Fig.). Some of the genes display much higher conditional frequencies than marginal frequencies in early time intervals. AOAH and BRCA2 are examples of this behavior (S6 Fig.). This effect could either be due to higher relative sampling fluctuations for SNVs, or it may have a biological reason. For example, tumors containing these mutations could be more aggressive and are therefore diagnosed earlier and do not have the time to accumulate more mutations. The set of top-30 ranked genes contains a number of known drivers (S6 Fig. , S3 Table). AKT1 has been associated to breast, colorectal, and ovarian cancer formation [27]. Germline mutations of BRCA2 are associated with increased risk for developing breast and ovarian cancer [28]. However, also somatic mutations of BRCA2, which are considered here, have been shown to drive cancer progression [4]. Furthermore, known drivers such as CDH1, CTCF, and GATA3 are found within the top-30 ranked genes [26]. Of the top-30 ranked genes 16 have q-values below 0. 2 (Bejamini-Hochberg). Of the rest only two more display q-values below 0. 2 (S3 Table). Dependencies among mutations were initially studied in colorectal cancers and adenomas [2]. However, the number of sequenced colorectal cancer samples in TCGA is still relatively small. Furthermore, it was shown that in colorectal cancer (among others) the number of mutations is correlated with the age of the patient and is therefore a poor measure of tumor age [29]. Nevertheless, we applied mutation timing ranking to SNV data from 223 colorectal cancer samples from the TCGA database (downloaded on July 31,2014). We set the cutoff for SNVs to be ranked to 5% or 12 out of 223. Furthermore, we excluded all samples which supposedly exhibited a mutator phenotype. Therefore, we used a cut-off of 1000 SNVs in the analyzed samples. This resulted in 212 samples used for mutation timing ranking. The final number of ranked genes was 69 (S7 Fig. , S4 Table). The classical colorectal cancer genes APC, KRAS and TP53 (all rank 1) are among the top-predicted constrained genes [30]. Furthermore, known drivers NRAS and PIK3CA (rank 16 and 18, respectively) are highly ranked [31], [32]. Since the number of samples and ranked genes is very small the mutation timing-ranked list is probably unstable. Of the highly-ranked genes in colorectal cancer, APC, KRAS, and TP53 are significant at a false discovery rate of 0. 2 (Benjamini-Hochberg). In order to check for dependencies among highly-ranked cancer drivers we used the oncogenetic tree model, because it is an alternative method to our conjunctive Bayesian network model used here to motivate the mutation timing method [33]. The driver list we used is a union of different driver lists from http: //www. bushmanlab. org/links/genelists accessed on August 5,2014. In the ranked 256 breast cancer SNV-affected gene list, 62 drivers are present (S5 Table). We split the 62 identified drivers into top and bottom half according to their rank in the mutation timing ranking, i. e. , 31 top drivers versus 31 bottom drivers. Subsequently, we learned oncogenetic trees separately for both (top and bottom half) subsets. Since for oncogenetic trees, there is no posterior probability for single edges available and they are very unstable, we bootstrapped the data and relearned the trees 1000 times. We considered not only direct but also indirect dependencies, i. e. , the transitive closure of the trees in this analysis. If we compare dependencies with bootstrap confidence scores above 60% then the number of drivers with at least one dependency in highly-ranked genes versus lowly-ranked genes is 26 versus 2 (, Fisher' s exact test). Thus, there is an enrichment of dependent drivers in the list of highly-ranked drivers. Since CNA driver lists are very sparse we could not perform the same analysis for the mutation timing-ranked CNA genes. Furthermore, the list of ranked genes in colorectal cancer was too short for this type of analysis. We have presented a driver–passenger discrimination method which specifically aims at identifying cancer-driving mutations that are constrained in their occurrence during carcinogenesis. It is known that the selective advantage of several drivers depends on the genetic background they occur in. Hence, genetic constraints that result from epistatic gene interactions have been our main motivation. However, the mutation timing approach does not make any assumption about the biological origin of the constraints. They may be unobserved and can include, for example, epigenetic, transcriptional, post-transcriptional, or environmental constraints. The novelty of the mutation timing approach presented here is that it identifies constrained gene mutations. Several computational methods have been developed for estimating the actual dependency structure among a small set of genes [34]–[37]. Mutation timing ranking can be used to identify these genes that are amenable to modeling cancer progression in more detail. We emphasize that our mutation timing approach does not require estimating the dependency structure, which is a statistically and computationally challenging problem. Instead, we are only detecting deviation from independence of mutations. The computational time complexity of mutation timing depends linearly on the number of genes and samples. Thus, the method can be applied efficiently to datasets of virtually any size, including genome-wide measurements of large patient cohorts. Here, we have applied mutation timing ranking to two different CNA datasets and two SNV datasets on a gene level. Once there are larger datasets available, the method can readily be applied to mutations on an amino acid or nucleotide level. In some cancers about half of the somatic mutations found in tumor samples are assumed to be present already before tumor initiation [29]. In these cases the implicit assumption made here, that mutation accumulation starts about the same time as tumor initiation and the bulk of the mutations occurres during carcinogenesis is violated. Therefore, the results of the mutation timing method might be less stable in those cases. We have developed a method for assessing the significance of the mutation timing-ranked genes based on a weighted permutation test. However, we only applied this test to SNV datasets here. CNAs on the gene level are highly dependent on each other because in some cancers a single CNA might be larger than in others and alter more neighboring genes then in others. It is assumed that only a small number of genes in CNA-altered regions has cancer-driving effects. However, it is very difficult to discriminate which ones are drivers and which ones are passengers. Mutation timing highly-ranked genes have often very similar CNA profiles across tumors. And significance analysis is with our test not possible because it assumes that CNAs are only dependent on timing. Therefore, we get very high numbers of significant genes. In order to make the method more appropriate for CNAs, a reliable method for counting copy number events per tumor is needed. Mutation timing-based driver–passenger discrimination is complementary to existing approaches, including those based on mutation frequencies, mutual exclusivity, and prior network information. Thus, it can also be expected to improve the performance of ensemble methods, which integrate different classifiers [38]. Among the existing approaches, only the frequency cutoff-based approach scales to larger datasets the way mutation timing does, and our simulation studies have highlighted the improvements of mutation timing in the presence of dependencies. Like any other approach, mutation timing has some limitations. One limitation is its difficulty to detect drivers that are already present at very early (measured by the total number of observed mutations) carcinogenesis stages, because generally there will be few samples observed at this time and most of them will already harbor such a mutation. Therefore, we can not observe a steep rise in conditional mutation frequency when increasing the number of mutations we condition on. An example of this behavior is PTEN in the breast cancer CNA analysis. Such a pattern may result, for example, from genetic subgroup structures among tumor samples, where a mutation is an early unconstrained event in one subgroup, but not in the others. However, these very early mutations can easily be identified by their high overall frequency in the dataset, i. e. , by the complementary frequency-based approach. Similarly, unconstrained drivers, i. e. , mutations that increase the fitness of cells independently of the mutational context, can not be detected by mutation timing. They are expected to have a higher rate of occurrence than neutral mutations, but not a different temporal pattern. In general, unconstrained drivers with high mutation frequencies can again be detected by their increased frequencies. An implicit assumption made in the mutation timing as well as the frequency-based method is that changes in evolutionary rates over time do not affect the underlying driver–passenger dependency structures; they only cause the observation time point (relative to the number of observed mutations) to be shifted. The number of available tumor samples and the accuracy of mutation calling are two additional limiting factors. Our simulation studies show that the mutation timing method performs well for datasets with more than 500 samples and genotyping error rates below 1%. Furthermore, as a rule of thumb, marginal frequencies of mutations should at least be twice as high as the average genotyping error rate. While these estimates are based on simulations under simplified conditions, they indicate that with current sample sizes and genotyping error rates, even rare driver mutations can be detected if their occurrences are constraint. Both the CNA and SNV data analyzed here contained a small number of samples with very high numbers of CNAs and SNVs, respectively. This phenomenon is usually referred to as a mutator phenotype. The models used for simulating data in the simulation study do not capture this phenomenon and therefore only represent samples without a mutator phenotype. However, the influence of the samples with the mutator phenotype on the mutation timing method is minor, because the sigmoidal curve fits were weighted by the number of samples exhibiting a specific number of mutations. Most driver–passenger discrimination approaches [6]–[8] assume either constant per-gene or per-base background evolutionary rates and predict as drivers those genes that exhibit significantly higher mutation frequencies in mid to large scale studies. However, these approaches ignore that different evolutionary rates could also be correlated with, for example, epigenomic context [39] and might not be due to a cancer-driving character of the affected gene. Our mutation timing method is invariant to different evolutionary rates, because the slopes we use for ranking reflect only how fast the mutation probability rises to its marginal mutation probability and not how large this marginal mutation probability is. Furthermore, epistatic interactions between cancer driving genes can cause low mutation frequencies and complicate the identification of those drivers. Our mutation timing-based cancer driver ranking approach identifies constrained candidate drivers in a computationally efficient manner. This approach will help in analyzing the upcoming data from large tumor sequencing studies which will be available in the near future. We have analyzed the waiting times of somatic cancer-related events in the presence and absence of order constraints using conjunctive Bayesian network models. Based on the differences we found, we have developed a method for ranking genes according to their likelihood of being contrained and hence being potential drivers. The waiting time analysis and the ranking method are described in detail in the S1 Text. For validation, we used three different mutation networks, denoted as network A, network B, and network C to simulate datasets (Fig. 3). We sampled from a continuous-time Conjunctive Bayesian network [15], in which mutations are constrained by the occurrence of their predecessors in the network. Network A was designed such that the marginal probabilities of observing the mutations are between 0. 1 and 0. 8. For network B, we aimed at lower mutation probabilities and set the mutation rate parameters such that all marginal mutation probabilities were between 0. 05 and 0. 5. Network B had already been used for a simulation study in [35] in order to evaluate an inference method. Network C represents a linear dependency structure with marginal mutation probabilities between 0. 05 and 0. 8. The sampling time parameter (), which is used to stop the time-dependent mutation accumulation process, was set to 1 for all simulation setups. We added data from 90,89, and 90 independent nodes, respectively, such that each network consisted of a total of 100 mutations. The mutation rate parameters for the independent nodes, denoted as, were sampled from a log-uniform distribution as follows: For fixed and, we sampled, and set. We subsequently compromised the datasets by flipping every mutation indicator with a probability of, 0. 01,0. 05,0. 1}. We fitted sigmoid functions weighted by a kernel density estimator of the empirical sample distribution and ranked the features according to the slope of the fitted sigmoidal curves (Figs. 6, S4 and S6, grey lines). Most of the samples had an intermediate number of mutations and therefore contributed considerably to estimating the slope parameter, S. The inflection point was forced to be between 1 and the maximum number of mutations observed in the samples. The sigmoid functions were fitted by weighted least-square using the nls function from the stats package of the statistics software R [40]. In order to compute ROC curves and area under the ROC curve (AUC) values, the cutoff value for the classification was increased continuously. This procedure was repeated 100 times in order to obtain an estimate of the variability. In each iteration, the mutation rate parameters of the independent nodes, , were redrawn, the underlying network and its parameters remained constant. The and values were set to 0. 06 and 6 for network A, 0. 05 and 1. 25 for network B and 0. 015 and 1. 5 for network C. In order to generate random networks we used the following procedure: (1) Start with a set of (ten) unconnected nodes denoted as D. (2) Select a random unconnected node a from D. (3) Select a second random (connected or unconnected) node b from D (4) Decide if a becomes the parent or child node of b with probability 0. 5. (5) Repeat from (2) until all nodes from D have a least one connection. This procedure ensures that all networks are valid posets, i. e, directed acyclic graphs. Six of the ten nodes had the mutation rate parameter set to 2, the rest to 1. The values of and were set to 0. 06 and 6 for the random networks. For the significance analysis of the mutation timing curves, we developed a weighted permutation-based test. The null hypothesis assumes that the probability of having a mutation in a certain gene in a certain sample depends only on the marginal frequency of the mutation of this gene and on the number of mutated genes in this sample. For every gene, a specific null hypothesis is constructed. This null hypothesis keeps the marginal mutation frequency of the gene constant, i. e. , the number of samples, in which it is mutated, is kept constant. Under the null hypothesis, the probability of observing a set of tumors harboring a specific mutation follows a conditional multinomial distribution with parameters proportional to the total number of mutated genes in the tumor samples and the observed number of times the gene was mutated. In each tumor, each gene can be mutated at most once, i. e. , the null distribution is the multinomial conditioned on pairwise mutually exclusive events (tumors). After constructing the null distribution of the mutated genes, we fit a sigmoidal curve analogous to the other genes and record the slope. The whole procedure is repeated 1000 times. The fraction of times the slope of the randomized column was smaller than the slope of gene i is the p-value of gene i. The R code for mutation timing-based gene ranking as well as the CNV and SNV cancer data analyzed here are available at http: //www. cbg. ethz. ch/software/mutationtiming.

Cancer genome sequencing projects result in vast amounts of cancer mutation data. However, our understanding of which mutations are driving tumor growth and which are selectively neutral is lagging behind. Functional interactions among mutations can result in mutational dependencies, and these mutations then display low marginal mutation frequencies across tumor samples complicating the identification of these drivers. Here, we present a simple method for calling candidate driver mutations by discriminating dependent mutations from independent ones based on their dynamical patterns of occurrence. The gene ranking procedure measures deviation from neutral mutation timing patterns. We demonstrate, for different types of cancers and genetic alterations, improvement over classical frequency-based approaches if drivers do not occur independently, and we show complementarity to other approaches.

Abstract Introduction Results Discussion Methods

oncology statistical models medicine and health sciences mathematics statistics (mathematics) basic cancer research biology and life sciences physical sciences computational biology evolutionary biology genomics statistics evolutionary processes

2015

Identification of Constrained Cancer Driver Genes Based on Mutation Timing

In the Indian subcontinent, Visceral leishmaniasis is endemic in a geographical area coinciding with the Lower Gangetic Plain, at low altitude. VL occurring in residents of hill districts is therefore often considered the result of Leishmania donovani infection during travel. Early 2014 we conducted an outbreak investigation in Okhaldhunga and Bhojpur districts in the Nepal hills where increasing number of VL cases have been reported. A house-to-house survey in six villages documented retrospectively 35 cases of Visceral Leishmaniasis (VL). Anti-Leishmania antibodies were found in 22/23 past-VL cases, in 40/416 (9. 6%) persons without VL and in 12/155 (7. 7%) domestic animals. An age- and sex- matched case-control study showed that exposure to known VL-endemic regions was no risk factor for VL, but having a VL case in the neighbourhood was. SSU-rDNA PCR for Leishmania sp. was positive in 24 (5%) of the human, in 18 (12%) of the animal samples and in 16 (14%) bloodfed female Phlebotomus argentipes sand flies. L. donovani was confirmed in two asymptomatic individuals and in one sand fly through hsp70-based sequencing. This is epidemiological and entomological evidence for ongoing local transmission of L. donovani in villages at an altitude above 600 meters in Nepal, in districts considered hitherto non-endemic for VL. The VL Elimination Initiative in Nepal should therefore consider extending its surveillance and control activities in order to assure VL elimination, and the risk map for VL should be redesigned. Visceral Leishmaniasis (VL) is a vector-borne parasitic disease that is fatal if left untreated. With an estimated 162,000 to 313,000 new cases per year in northeast India, Nepal and Bangladesh, VL poses a public health problem in the Indian subcontinent (ISC) [1]. In this region, VL is caused by Leishmania donovani and transmitted by Phlebotomus argentipes with humans as the only reservoir [2]. In endemic foci, infected domestic animals have been encountered, clustering with asymptomatic human infections, but their role in transmission is not established [3]. The habitat of the sand fly vector depends on biotic (vegetation and availability of human and/or animal blood meals) and abiotic (temperature and precipitation) factors, specific for each species. In the case of P. argentipes, these conditions are met in the plains of the river Ganges. The range of this species is considered to be limited to altitudes below 700 meters above the sea level [4–6]. In Nepal, the disease is endemic in 12 districts in the south-eastern plains, known as the Terai, where eight million people are at risk. Sporadic VL cases were reported in the hills at altitudes of 1,000 meters above sea level [7–8], but none of these have been firmly linked to local transmission. As the travel history of these cases was not documented, they may possibly have contracted VL in the Terai plains. Transmission of VL in the foothills of the Himalaya at altitudes ranging from 1,300 to 3,000 m has been suggested elsewhere, in India [9–12] and in Bhutan [13]—though not unequivocally confirmed as local. The transmission cycle requires a large vector population, but only few of these studies include entomological information [13–14]. Ozbel et al. suggested that not the altitude itself, but the gradient of habitats, relief and climate it offers, exerts the structuring effect on sand fly range [15]. Since 2000, VL cases have been reported in increasing numbers from the hilly regions in eastern Nepal. This challenges the ongoing VL elimination program—as they could constitute a reservoir of future re-emergence if local transmission is confirmed. We report an outbreak investigation conducted early 2014 in this region. We reviewed the epidemiological surveillance data of the Ministry of Health & Population and the patient database of the BPKIHS hospital for the period 2000–2013. All VL cases from the last five years who named a hill district as their residence at the time of admission were listed and the two most affected districts, namely Bhojpur and Okhaldhunga, were selected for field investigation. In each district we selected three villages on the basis of the total number of VL cases since 2005, combined with the number of cases reported during each of the previous two years, accessibility on foot and local support by health authorities and community. The selected locations were Thakle Jakme, Thakle Richuwa and Mathilo-Richuwa in Okhaldhunga district, and Bhojpur village ward no. 3, Dalgaun ward no. 3, and Manebhanjyang ward no. 9 from Bhojpur district. The following case definitions were used: current VL case: fever for at least two weeks and positivity in a rapid diagnostic test for VL (Kala-azar Detect TM Rapid Test; InBios International, Seattle, WA); past case of VL: history of treatment for VL, corroborated by prescriptions and/or case records from the health facility; past VL death: any death related to a febrile illness of longer than two weeks duration and at least one more VL-specific sign as verified by a verbal autopsy procedure. The investigation took place in March 2014 in Okhaldhunga district, and in May 2014 in Bhojpur. We conducted an exhaustive household census combined with a serosurvey, a case-control study, and an entomological survey. Household census. All households were geo-referenced by a Global Positioning System (GPS) device, specifying longitude, latitude and altitude of the houses. Trained field workers interviewed each head of household on age and gender of present and absent family members, occupation, household characteristics, and livestock ownership, as well as current or past history of VL in the family including deaths. For each VL case, disease history, health care-seeking behavior, place of treatment, drugs and outcome were recorded. Case-control study. We interviewed each current or past VL case identified in the census with a semi-structured questionnaire to assess risk factors. For each case we selected four age- and sex-matched controls (age matching based on age at time of VL diagnosis) living in the same village without history of VL in their lifetime, and applied the same questionnaire. The power of this nested case-control study was calculated prior to the study with the factor “contact with known VL endemic area” as primary exposure of interest, and VL as the outcome. Based on the expectation to find at least 35 cases, a ratio of four controls per case allowed to detect an odds ratio of three with a power of 80% at a significance level of 5%. The risk factor questionnaire focused on both individual and household characteristics such as educational status, occupation, travel history, sleeping habits, bed net use, cases and deaths due to VL in the family and in the immediate neighborhood, socio-economic factors, animal ownership and housing characteristics. Serological study. All subjects aged ≥ 2 years who lived permanently in the study clusters—or their guardians—were invited to give consent for the draw of a 2 ml venous blood sample. A trained laboratory technician collected a 2 ml venous blood sample that was divided over an EDTA tube, a serum tube and onto a pre-printed Whatman # 3 filter paper. The dried filter papers were placed in plastic bags containing silica gel. All samples were transported in a cold box at 4–8°C until storage at -20°C at BPKIHS. At the same time, veterinary technicians took blood samples from domestic animals such as bovines, goats, sheep, pigs, and dogs. Entomology. Timing of the insect collections (March-May) purposely coincided with the second annual peak of P. argentipes density in the plains [16]. Sand flies were captured in-door in eight households including houses of past VL patients, in each of the six study clusters. Trained insect collectors supervised by an experienced medical entomologist installed CDC light traps before dusk inside the house and/or cattle shed for two consecutive nights and collected them between 5: 00 and 6: 00 AM the next morning. During that visit they also caught resting sand flies by mouth aspiration for 15 minutes in the households and the cattle sheds. Collected specimens were preserved the same day in 80% ethanol and transported to the entomology laboratory at BPKIHS Dharan for examination under a binocular dissecting microscope. Insect species were morphologically identified according to the Lewis key [17] and female P. argentipes were separated from other insects and pooled by household in a cryotube with 80% alcohol. During further processing for molecular analyses, the source material was completely blinded with regard to species, sex and feeding/gravid status. Direct agglutination test (DAT). DAT was performed using a freeze-dried antigen of fixed, trypsin-treated and stained promastigotes of Leishmania donovani obtained from ITM-Antwerp as described by Jacquet et al. [18]. A DAT titre ≥ 1: 1600 in capillary blood was taken as marker of Leishmania donovani infection in humans [19]. PCR-based detection of Leishmania sp and species identification. DNA was extracted from all blood samples and sand flies, using the QiaAmp DNA mini kit (Qiagen, Hilden, Germany). DNA from 200 μl blood or single sand flies was eluted in 50 μl AE buffer. A diagnostic Leishmania PCR based on the small-subunit ribosomal DNA of Leishmania, (SSU-rDNA) was performed on 2 μl DNA of each sample, with the inclusion of an internal control amplicon, essentially as described in Odiwuor et al. [20]. A sample was considered to contain Leishmania sp. if the PCR scored positive, i. e. if an amplicon of around 115 bp was seen on an ethidium-bromide stained agarose gel. Other samples were scored negative when only the internal control amplicon was successfully amplified, or invalid in case the internal control amplicon was not detected even after repeating the PCR. To verify the Leishmania sp. identity of the SSU-rDNA amplicons, they were sequenced and compared with publically available sequences. In order to further type the parasites to the L. donovani species level, the heat-shock protein 70 gene (hsp70) was partially amplified, using HSP70-F as outer and HSP70-N as hemi-nested PCR [21], followed by sequence analysis [22–23]. All sand flies that were positive for the SSU-rDNA PCR were molecularly identified by sequencing the barcoding fragment of the mitochondrial cytochrome oxidase I (COI) as described by Versteirt et al. [24]. Sequences were submitted to the online identification tool of the Barcoding of Life Database (BOLD) to identify the sand fly [25]. We used R statistical software version 3. 0. 0. [26] for data analysis. For each potential risk factor we computed Odds Ratios by means of a logistic regression adjustment analysis with matching variables (age, gender, village) forced into the regression model. To retain only the independent predictors of being a VL case, statistically significant risk factors in the univariate analyses (P < 0. 05) were entered in a multivariate analysis. The final model and adjusted Odds Ratio’s resulted from a stepwise selection in a logistic regression analysis with matching (age, gender, village) adjustment [27]. Ethical clearance for the study was obtained from the Institutional Ethical Review Boards of BPKIHS and the Nepal Health Research Council, Kathmandu in Nepal, the Ethical Review Board of the Institute of Tropical Diseases in Antwerp, and the University Hospital of Antwerp in Belgium. Community consent was sought and obtained during meetings with local health facilities’ staff, village authorities, and village assemblies. Individual informed consent was obtained in writing, prior to blood sampling and interviewing. For children, a parent or guardian provided written informed consent. We retrieved 101 past VL cases in the document review, 38 from Okhaldhunga and 63 from Bhojpur district. The distribution over time of these cases is shown in Fig 1. In Okhaldhunga cases were only reported after 2010, when the mission hospital there started outreach activities, including screening for VL. The characteristics of the six study clusters are shown in Table 1. The characteristics of the households are included as supporting information (S1 Fig and S1, S2 and S3 Tables). No active VL cases were found in the household survey, but 21 percent (26/122) of the households had at least one confirmed case of VL in the past. We ascertained a total of 35 confirmed VL cases, of which four had occurred before the year 2000, and one of them more than 50 years ago. Age at time of VL ranged from 1 to 52 years, most frequent in the youngest age group (Q1/Median/Q3 was 6,17 and 28 years resp.) (see S2 Fig in supplementary material). Of the 31 VL cases reported since 2000,22 (71%) were reported in the last five years. Fourteen of these 31 had their diagnostic workup at the BPKIHS-hospital in Dharan where they were parasitologically confirmed as VL. In one of them, who was enrolled in a research project, L. donovani was confirmed at species level through hsp70 sequencing of an isolate cultured from a bone marrow aspirate. Of the 35 confirmed VL cases, four had been fatal, three of which were children below five years old. Treatment comprised Sodium-Antimony-Gluconate (13), miltefosine (5) or amphotericin B (17). An additional four deaths were recorded that could possibly be attributed to VL according to the verbal autopsy, and two of them had a confirmed VL case among other family members. These four persons were not included as VL cases in further analysis. Table 2 shows the breakdown by year of onset of all the VL cases retrieved during the house-to-house survey. Sex-, age- and village-adjusted Odds Ratios of potential risk factors for being a VL case are given in Table 3. History of travelling to any of the known VL endemic areas was not associated with being a VL case (Odds Ratio 0. 73 [95% CI 0. 27–1. 87]. The strongest risk factor for being a VL case was the occurrence of a VL case in the neighborhood (i. e. within a perimeter of 100 meter from the house). All the 35 (100%) of the VL cases reported such a case in the neighborhood against 52/139 (37 percent) of the controls, leading to an Odds Ratio of positive infinity. Other significant risk factors for VL in univariate analysis were cow dung in front of the house, sleeping on the ground, increasing distance from the river, monthly income, house type and having a case of VL in the family. A multivariate stepwise logistic regression modeling procedure—ignoring the factor “VL case in neighborhood”—retained only sleeping on the ground, (OR 5. 65 [95% CI 1. 12–30. 43], cow dung in front of the house (OR 4. 70 [95% CI 1. 21–20. 06]) and distance from the river (OR 0. 06 per additional km from river [95% CI 0. 001–0. 92]) as independent predictors. As these ORs could be confounded by the effect of having a VL case in the neighborhood, a factor that could not be included in the multivariate modeling approach because of its infinite OR, we did a stratified analysis on a restricted data set that only included those cases and controls with a VL case in the neighborhood. Beyond that factor only two lifestyle-related risk factors remain statistically significant: cow dung in front of the house (OR 4. 72 [95% CI 1. 64–15. 26] and sleeping on the ground (OR 4. 74 [95% CI 1. 18–21. 5]). Overall, 14. 1% (95% CI: 10. 8–17. 3) of the study population (62/441) was positive in DAT varying from 2. 5% to 50% according to study cluster. Excluding the past VL cases, seroprevalence in the asymptomatic persons was 9. 6% (95% CI: 6. 8–12. 4) varying between villages from 0 to 42%. There was no significant difference in seroprevalence according to sex and age (Fig 2). Blood samples were collected from 155 domestic animals. Ten goats, one cow and one buffalo had DAT titers just above the cut-off (one (goat) = 1/6400, all others = 1/1600). Of all goats sampled (n = 83), this represented 12%. DNA was extracted from human (n = 441) and animal blood samples (n = 155). SSU-rDNA was positive in 24 (5%) of the human samples, and in 18 (12%) of the animal samples. The amplicon sequence was obtained from five humans, four goats, and one sheep, and in all cases, the sequence was compatible with Leishmania sp. An hsp70 sequence could be obtained from two of the human samples, which confirmed the parasite identity as Leishmania donovani. The hsp70 gene from the animal samples could not be sequenced, likely because of the low parasitemia in these samples, together with the lower sensitivity of the hsp70 PCR (vs the SSU-rDNA PCR). The total number of sand flies captured was 281, morphologically identified and segregated into P. argentipes, Sergentomyia spp. and other Phlebotomus spp. Amongst the P. argentipes, 82 were male, 29 unfed female, 108 fed female, and 6 gravid female. DNA extraction was successful in 271 sand flies and 16 specimens (14,5% of the female, fed sand flies) were found positive by SSU-rDNA PCR. In 11 of these the sequence could be obtained and was found compatible with Leishmania sp. The hsp70 sequence was obtained from one of these, confirming infection with L. donovani. 14 out of 16 SSU-rDNA PCR positive sand flies, including the one infected with L. donovani, were molecularly identified as P. argentipes, and all of them were female and fed. The other two were confirmed to be Phlebotomus spp. , but could not be molecularly identified to species level as their COI sequence was not available in BOLD. Retracing the origin of the 16 sand flies, they came from 10 houses, i. e. two households from Thakle Jakma, five from Thakle Richuwa, two from Mathilo (Okhaldhunga district), and one from a household in Manebhanjyang (Bhojpur). Five of these households effectively had had a VL case in the past, one had asymptomatic DAT positives only, and four had neither (though not all family members gave a blood sample). Fig 3 shows the dendrogram with the location of the three hsp70-positive samples. Our data show there is strong evidence for local transmission of VL in the hill districts of eastern Nepal. Exposure to known endemic areas in the plains of Nepal or India was not a risk factor for VL. On the contrary, the strongest risk factor for being a VL case in the hills was having a VL case in the close neighborhood of the house (< 100 m). Seventy-five percent of the VL cases (26/35) had no reported exposure to known endemic areas in Nepal or India whatsoever. VL cases actually had travelled less than controls to those areas, though this difference was not statistically significant (adjusted OR. 0. 73; 95% C. I. 0. 27–1. 87; P = 0. 524). Entomological and microbiological findings meanwhile strongly point towards local transmission. We caught P. argentipes sand flies in five of the six villages, at altitudes up to 1500 m above sea level, way above the 700 meter limit suggested in literature [4–6]. In one of the P. argentipes samples, we were able to demonstrate the presence of L. donovani DNA. While these entomological data suggest the high likelihood of P. argentipes as the vector of VL in the study areas, it does not unequivocally prove that the species is the vector in this habitat, which would require the demonstration of transmissible parasite stages (e. g. by microscopical observation of Leishmania in the sand flies, or detection of stage-specific antigens or RNA) [28]. In the asymptomatic human samples, 24 (5. 4%) were positive with SSU-rDNA PCR. This assay was designed for detection of Leishmania sp, but can also give positive results with some monoxenous trypanosomatids like Crithidia and Leptomonas [29]. The latter protozoa are occasionally reported in humans, essentially in cases of immuno-suppression due to HIV or co-infection with L. donovani [30] Two of the samples were confirmed as L. donovani, in three additional ones a Leishmania sp. SSU-rDNA signature was demonstrated. Moreover, a Leishmania SSU-rDNA sequence was confirmed in 11 out of 16 SSU-rDNA PCR positive sand flies, one of which was shown to be L. donovani. Among the samples that could not be sequenced, we cannot strictly exclude the occurrence of Leptomonas sp. It should be noted that some trypanosomatids share the Leishmania SSU-rDNA sequence, but as far as we know these are not endemic in the Indian subcontinent. In all, these arguments point to active Leishmania transmission in the sampled areas. Finally, the large proportion of the inhabitants showing antibodies to the DAT is strongly supportive for local transmission. DAT has been shown to be negative in healthy populations from Indian states without VL transmission [31]. The DAT prevalence in individuals without a history of VL was 9. 6% in our household study, comparable with prevalences in villages with recent outbreaks in the endemic areas of the Terai [32]. DAT positivity was present here in all age groups, and not related with travel: among the 139 controls in the case-control study, there were nine asymptomatic DAT positive cases, but only two of them had a history of traveling to VL-endemic territory. Furthermore, there was a clear link between DAT positivity and the presence of VL case (s) in the household, an epidemiological feature of endemic transmission [33]: in our study, having a VL case in the family represented an OR of 3. 2 (95% C. I. : 1. 643–6. 361) (p = 0. 0007) to be DAT-positive. We observed a substantial proportion of domestic animals with positive serological titers and PCR-positivity, as we did previously in the Terai plains of Nepal [34,35]. For the DAT we chose to use one single, relatively high cut-off of 1: 1600, regardless of the animal species (e. g. in dogs, 1: 800 is the cut-off commonly used), to increase specificity. The DAT seroprevalence (12%) and SSU-rDNA PCR positivity (12%, Leishmania sp confirmed by sequencing in 4 samples) amongst goats clearly indicates local exposure to Leishmania parasites, as all goats are bred locally. It is tempting to conclude to L. donovani transmission as the causal agent, as domestic animals are also exposed to P. argentipes bites, who are opportunistic feeders. Nonetheless, circulation of animal Leishmania spp. , can be another explanation as titers were low and none of the PCR positives could be confirmed as L. donovani. We conclude there is local transmission of Leishmania donovani ongoing in the hilly districts of Nepal, based on three arguments. Firstly, there were several VL cases in permanent residents of settlements in the hills, most of them relatively recent, who had not travelled to known endemic areas. Secondly, a considerable number of asymptomatic residents have been exposed to Leishmania sp. as shown by DAT serology and PCR and in two of them L. donovani was confirmed. Thirdly, P. argentipes sand flies are present in these hilly districts. A substantial number were infected with Leishmania sp. , and we confirmed L. donovani at species level in one specimen. Further work is required for a finer genotyping of L. donovani and to examine links with parasite populations encountered in the lowlands. The hill districts Okhaldhunga and Bhojpur in Nepal are hitherto not considered as endemic for VL transmission in Nepal, which implies that the health staff is not specifically trained, diagnostic and treatment facilities are not available, reporting is not standardized, and there are no prevention campaigns. Those engaged in VL elimination should pay attention to these new geographical presentations and reconsider the existing risk maps.

Visceral leishmaniasis is a neglected but deadly disease occurring in north-eastern India, the south-eastern lowland of Nepal, and the Ganges delta in Bangladesh; all part of the Lower Gangetic plains. Districts at higher altitude, such as those situated in the foothills of the Himalaya in Nepal, are considered non-endemic. Consequently, diagnostic, therapeutic and surveillance facilities are not available, and sporadic cases of VL occurring in residents of these districts are considered the result of Leishmania donovani infection during travel. This parasite is transmitted from man to man through the bite of a sand fly, Phlebotomus argentipes. To investigate the increasing number of cases reported from some villages in the Nepal hills, we interviewed former VL cases, as well as their healthy co-villagers, on their history of travel, we checked their blood for signs of infection, and we set traps to capture sand flies. We found that many residents had been infected, symptomatically as well as asymptomatically, regardless of travelling to endemic areas. Moreover, we managed to capture sand flies and found DNA of leishmania parasites inside. This proves that there is indeed ongoing local transmission in hill districts and that surveillance and control activities should be extended.

Abstract Introduction Methods Results Discussion

2015

Transmission of Leishmania donovani in the Hills of Eastern Nepal, an Outbreak Investigation in Okhaldhunga and Bhojpur Districts

Fossil records indicate that life appeared in marine environments ∼3. 5 billion years ago (Gyr) and transitioned to terrestrial ecosystems nearly 2. 5 Gyr. Sequence analysis suggests that “hydrobacteria” and “terrabacteria” might have diverged as early as 3 Gyr. Bacteria of the genus Azospirillum are associated with roots of terrestrial plants; however, virtually all their close relatives are aquatic. We obtained genome sequences of two Azospirillum species and analyzed their gene origins. While most Azospirillum house-keeping genes have orthologs in its close aquatic relatives, this lineage has obtained nearly half of its genome from terrestrial organisms. The majority of genes encoding functions critical for association with plants are among horizontally transferred genes. Our results show that transition of some aquatic bacteria to terrestrial habitats occurred much later than the suggested initial divergence of hydro- and terrabacterial clades. The birth of the genus Azospirillum approximately coincided with the emergence of vascular plants on land. Fossil records indicate that life appeared in marine environments ∼3. 5–3. 8 billion years ago (Gyr) [1] and transitioned to terrestrial ecosystems ∼2. 6 Gyr [2]. The lack of fossil records for bacteria makes it difficult to assess the timing of their transition to terrestrial environments; however sequence analysis suggests that a large clade of prokaryotic phyla (termed “terrabacteria”) might have evolved on land as early as 3 Gyr, with some lineages later reinvading marine habitats [3]. For example, cyanobacteria belong to the terrabacterial clade, but one of its well-studied representatives, Prochlorococcus, is the dominant primary producer in the oceans [4]. Bacteria of the genus Azospirillum are found primarily in terrestrial habitats, where they colonize roots of important cereals and other grasses and promote plant growth by several mechanisms including nitrogen fixation and phytohormone secretion [5], [6]. Azospirillum belong to proteobacteria, one of the largest groups of “hydrobacteria”, a clade of prokaryotes that originated in marine environments [3]. Nearly all known representatives of its family Rhodospirillaceae are found in aquatic habitats (Figure 1 and Table S1) suggesting that Azospirillum represents a lineage which might have transitioned to terrestrial environments much later than the Precambrian split of “hydrobacteria” and “terrabacteria”. To obtain insight into how bacteria transitioned from marine to terrestrial environments, we sequenced two well studied species, A. brasilense and A. lipoferum, and a third genome of an undefined Azospirillum species became available while we were carrying out this work [7]. Horizontal gene transfer has been long recognized as a major evolutionary force in prokaryotes [12]. Its role in the emergence of new pathogens and adaptation to environmental changes is well documented [32]. While other recent studies indicate that HGT levels in natural environments may reach as much as 20% of a bacterial genome [33], our data suggest that HGT has affected nearly 50% of the Azospirillum genomes, in close association with dramatic changes in lifestyle necessary for transition from aquatic to terrestrial environments and association with plants. Emergence of these globally distributed plant-associated bacteria, which appear to coincide with radiation of land plants and root development, likely has dramatically changed the soil ecosystem. The genome of Azospirillum lipoferum 4B was sequenced by the whole random shotgun method with a mixture of ∼12X coverage of Sanger reads, obtained from three different libraries, and ∼18X coverage of 454 reads. Two plasmid libraries of 3 kb (A) and 10 kb (B), obtained by mechanical shearing with a Hydroshear device (GeneMachines, San Carlos, California, USA), were constructed at Genoscope (Evry, France) into pcDNA2. 1 (Invitrogen) and into the pCNS home vector (pSU18 modified, Bartolome et al. [34]), respectively. Large inserts (40 kb) (C) were introduced into the PmlI site of pCC1FOS. Sequencing with vector-based primers was carried out using the ABI 3730 Applera Sequencer. A total of 95904 (A), 35520 (B) and 15360 (C) reads were analysed and assembled with 504591 reads obtained with Genome Sequencer FLX (Roche Applied Science). The Arachne “HybridAssemble” version (Broad institute, MA) combining 454 contigs with Sanger reads was used for assembly. To validate the assembly, the Mekano interface (Genoscope), based on visualization of clone links inside and between contigs, was used to check the clones coverage and misassemblies. In addition, the consensus was confirmed using Consed functionalities (www. phrap. org), notably the consensus quality and the high quality discrepancies. The finishing step was achieved by PCR, primer walks and transposon bomb libraries and a total of 5460 sequences (58,602 and 4800 respectively) were needed for gap closure and quality assessment. The genome of strain Azospirillum brasilense Sp245 was sequenced by the whole random shotgun method with a mixture of ∼10X coverage of Sanger reads obtained from three different libraries and ∼25X coverage of 454 reads. A plasmid library of 3 kb, obtained by mechanical shearing with a Hydroshear device (GeneMachines, San Carlos, California, USA), were constructed at Plant Genome Mapping Laboratory (University of Georgia, USA) into pcDNA2. 1 vector (Invitrogen). Large inserts (40 kb) were introduced into the PmlI site of pCC1FOS. Sequencing with vector-based primers was carried out using the ABI 3730 Applera Sequencer. The Arachne “HybridAssemble” version combining 454 contigs with Sanger reads was used for assembly. Contig scaffolds were created using Sequencher (Gene Codes) and validated using clone link inside and between contigs. AMIGene software [35] was used to predict coding sequences (CDSs) that were submitted to automatic functional annotation [36]. The resulting 6233 A. lipoferum 4B CDSs and 7848 A. brasilense Sp245 CDSs were assigned a unique identifier prefixed with “AZOLI” or “AZOBR” according to their respective genomes. Putative orthologs and synteny groups were computed between the sequenced genomes and 650 other complete genomes downloaded from the RefSeq database (NCBI) using the procedure described in Vallenet et al. [36]. Manual validation of the automatic annotation was performed using the MaGe (Magnifying Genomes) interface. IS finder (www-is. biotoul. fr) was used to annotate insertion sequences [37]. The A. lipoferum 4B nucleotide sequence and annotation data have been deposited to EMBL databank under accession numbers: FQ311868 (chromosome), FQ311869 (p1), FQ311870 (p2), FQ311871 (p3), FQ311872 (p4), FQ311873 (p5), FQ311874 (p6). The A. brasilense Sp245 nucleotide sequence and annotation data have been deposited at EMBL databank under accession numbers: HE577327 (chromosome), HE577328 (p1), HE577329 (p2), HE577330 (p3), HE577331 (p4), HE577332 (p5), HE577333 (p6). In addition, all the data (i. e. , syntactic and functional annotations, and results of comparative analysis) were stored in a relational database, called AzospirilluScope [36], which is publicly available at http: //www. genoscope. cns. fr/agc/mage/microscope/about/collabprojects. php? P_id=39. BLAST searches were performed using NCBI toolkit version 2. 2. 24+ [38]. Multiple sequence alignments were built using the L-INS-i algorithm of MAFFT [39] with default parameters. Phylogenetic tree construction was performed using PhyML [40] with default parameters unless otherwise specified. 16S rRNA sequences were retrieved from the Ribosomal Database Project [41]. A concatenated ribosomal protein tree was constructed from sequenced members of alpha-proteobacteria with a 98% 16S rRNA sequence identity cutoff to limit overrepresentation. The following ribosomal proteins were used: L3, L5, L11, L13, L14, S3, S7, S9, S11, and S17. The proteins were identified using corresponding Pfam models and HMMER [42] searches against the genomes of sequenced alpha-proteobacteria selected above. The sequences were aligned and concatenated. GBlocks [43] with default parameters was used to reduce the number of low information columns. The tree was constructed using PhyML with the following options: empirical amino acid frequencies, 4 substitution categories, estimated gamma distribution parameter, and NNI tree topology search. Protein sequences queries from all 3 Azospirillum genomes were used in BLAST searches against the non-redundant microbial genome set constructed by Wuichet and Zhulin [26] supplemented with sequenced members of Rhodospirillales absent in the original set (Acetobacter pasteurianus IFO 3283-01, alpha proteobacterium BAL199, Magnetospirillum gryphiswaldense MSR-1, and Magnetospirillum magnetotacticum MS-1). E-value cutoff of 10∧−4 was used. Only the first occurrence of each species was used in ancestry assignment. Proteins were assigned as being ancestral or horizontally transferred, with varying degrees of confidence, based on the presence of members of Rhodospirillales and Rhodospirillaceae in the top eight BLAST hits. Ancestral assignment was based on the top 8 hits, based on the number of Rhodospirillaceae genomes in the database: 2 Azospirillum, 3 Magnetospirillum, 2 Rhodospirillum, and Nisaea sp. BAL199, excluding the organism on which ancestry assignment is being performed. High confidence ancestral proteins have at least 6 of the top 8 species belonging to Rhodospirillales or all but 1, if the BLAST result had less than 8 species. This rule allows for 1–2 independent events of HGT from Rhodospirillales to other distantly related species. Medium confidence ancestral proteins have at least 4 Rhodospirillaceae in the top 8. Low confidence ancestral proteins have at least 1 Rhodospirillaceae in the top 8, excluding hits to other Azospirillum genomes. High confidence horizontally transferred proteins have 0 hits to Rhodospirillales in the top 10, excluding hits to other Azospirillum genomes. Medium confidence horizontally transferred proteins have 0 hits to Rhodospirillales in the top 5, excluding hits to other Azospirillum genomes. Low confidence horizontally transferred proteins have 0 hits to Rhodospirillaceae in the top 8, excluding hits to other Azospirillum genomes. Unassigned proteins either have no BLAST hits outside Azospirillum, or simultaneously classify as medium confidence horizontally transferred and medium or low confidence ancestral. Bidirectional BLAST was used to identify orthologs of the putative glycoside hydrolase (GH) genes. Phyml package was used to confirm evolutionary relationships and visualize the results. Domain architectures were obtained through Pfam [53] search for each protein. Then information from CAZy [54] and recent analysis [55] was used to assign putative activities of the predicted GHs. Chemotaxis proteins were identified in genomic datasets as previously described [56]. Using CheA sequences from a recent chemotaxis system classification analysis [26], alignments of the P3–P5 regions of CheA were built for each class and for the entire set of CheA sequences. Each alignment was made non-redundant so that no pair of sequences shared more than 80% sequence identity. Hidden Markov Models (HMMs) were built from each non-redundant alignment and used to create library via the HMMER3 software package (version HMMER 3. 0b3) [42] and default parameters. The rhizosphere CheA sequences from a recent study [25] were run against the CheA HMM library. Unclassified sequences (Unc) are those with top hits to the full CheA set HMM rather than a class-specific HMM. The remaining sequences were assigned to the class of the top scoring HMM. Azospirillum strains and control strains (Dickeya dadantii 3937 as a positive control, A. tumefaciens NT1 as a negative control) were cultured for 16 h in liquid AB minimal medium [57] containing 0. 2% malate and 1 mg/L biotin. An aliquot of 107 cells (for Dickeya dadantii 3937) or 2. 107 cells (for all other strains) was deposited on top of AB plates containing 0. 1% carboxymethylcellulose instead of malate. Plates were incubated for 5 days before being stained as previously described [58]. A 211-bp cpaB (AZOBR_p460079) internal fragment was amplified by PCR with primers F6678 (GCGTGGACCTGATCCTGAC) and F6679 (GTGACCGTCTCGCTCTGAC) and subcloned into pGEM-T easy (Promega). White colonies were screened by PCR with primers F6678 and F6679 for correct insertion in pGEM-T easy, resulting in pR3. 37. The insert of plasmid pR3. 37 was digested with NotI and cloned into the NotI site of pKNOCK-Km [59], resulting in pR3. 39 after transfer into chemically-competent cells of E. coli S17. 1 λpir. pR3. 39 was introduced into A. brasilense Sp245 by biparental mating. Transconjugants resulting from a single recombination event of pR3. 39 were selected on AB medium containing 0. 2% malate, ampicillin (100 mg/mL) and kanamycin (40 mg/mL). The correct insertion of pKNOCK into cpaB was confirmed by PCR with primers (F6678 and F5595 TGTCCAGATAGCCCAGTAGC, located on pKNOCK) and sequencing of the PCR amplicon. Sp245 and Sp245cpaB were labelled with pMP2444 [60] allowing the constitutive expression of EGFP. The strains were grown in NFB* (Nitrogen free broth containing 0. 025% of LB) with appropriate antibiotics in glass tubes containing a cover-slide, under a mild lateral agitation for 6 days. After the incubation, the liquid and the cover-slide were removed from the tubes and the biofilm formed at the air/liquid interface was colored by 0. 1% crystal violet. After two washings with distilled water, crystal violet was solubilized by ethanol and quantified by spectrophotometry at 570 nm. The experiment was performed twice in triplicate. In parallel, the colonization of the glass cover-slide was monitored by confocal laser scanning microscopy (510 Meta microscope; Carl Zeiss S. A. S.) equipped with an argon-krypton laser, detectors, and filter sets for green fluorescence (i. e. , 488 nm for excitation and 510 to 531 nm for detection). Series of horizontal (x-y) optical sections with a thickness of 1 µm were taken throughout the full length of the Sp245 and Sp245cpaB biofilms. Three dimensional reconstructions of biofilms were performed using LSM software release 3. 5 (Carl Zeiss S. A. S.).

Genome sequencing and analysis of plant-associated beneficial soil bacteria Azospirillum spp. reveals that these organisms transitioned from aquatic to terrestrial environments significantly later than the suggested major Precambrian divergence of aquatic and terrestrial bacteria. Separation of Azospirillum from their close aquatic relatives coincided with the emergence of vascular plants on land. Nearly half of the Azospirillum genome has been acquired horizontally, from distantly related terrestrial bacteria. The majority of horizontally acquired genes encode functions that are critical for adaptation to the rhizosphere and interaction with host plants.

Abstract Introduction Results/Discussion Materials and Methods

genome sequencing genome complexity genome evolution microbial evolution biology genomics comparative genomics microbiology genetics and genomics

2011

Azospirillum Genomes Reveal Transition of Bacteria from Aquatic to Terrestrial Environments

Many questions about the genetic basis of complex traits remain unanswered. This is in part due to the low statistical power of traditional genetic mapping studies. We used a statistically powerful approach, extreme QTL mapping (X-QTL), to identify the genetic basis of resistance to 13 chemicals in all 6 pairwise crosses of four ecologically and genetically diverse yeast strains, and we detected a total of more than 800 loci. We found that the number of loci detected in each experiment was primarily a function of the trait (explaining 46% of the variance) rather than the cross (11%), suggesting that the level of genetic complexity is a consistent property of a trait across different genetic backgrounds. Further, we observed that most loci had trait-specific effects, although a small number of loci with effects in many conditions were identified. We used the patterns of resistance and susceptibility alleles in the four parent strains to make inferences about the allele frequency spectrum of functional variants. We also observed evidence of more complex allelic series at a number of loci, as well as strain-specific signatures of selection. These results improve our understanding of complex traits in yeast and have implications for study design in other organisms. Most traits of agricultural, evolutionary, and medical significance are genetically complex, involving multiple genes that interact with one another and the environment [1]. Despite decades of effort, our understanding of how such traits are specified at the genetic level remains incomplete [2]. Studies in model organisms can provide fundamental insights into the genetic basis of complex traits that are applicable to other species, including humans [3]. However, such studies typically detect only a small fraction of the loci that contribute to a trait due to low statistical power [4]. To improve genetic mapping of complex traits in Saccharomyces cerevisiae, we recently developed extreme QTL mapping (X-QTL), which is a bulk segregant mapping technique that employs millions of cross progeny [5]. X-QTL involves three key steps: generation of very large segregating populations, isolation of cross progeny with extreme trait values, and quantitative measurement of pooled allele frequencies across the genome in these phenotypically extreme individuals [5]. To make the pools of segregants that are the starting point for X-QTL, we use selectable markers to obtain an effectively unlimited number of progeny from a cross of two strains. We then employ selection-based phenotyping to isolate large numbers of segregants with extreme trait values from populations that contain millions of cross progeny. DNA is extracted from pools of phenotypically extreme segregants, and the allele frequencies of markers throughout these individuals' genomes are determined using custom microarrays or next generation sequencing. In an X-QTL experiment, a locus that influences a trait is expected to show an allele frequency skew in the direction of the parental allele that contributes to a more extreme trait value. By applying X-QTL to a number of chemical resistance phenotypes in a single cross of the lab strain BY4716 and the vineyard strain RM11-1a (hereafter, BY and RM, respectively), we were able to show that large numbers of loci can underlie quantitative trait variation between S. cerevisiae isolates [5]. Following our publication, another group observed similar results in a different cross [6], suggesting that high genetic complexity may be a common feature of heritable trait variation among yeast strains. Here, we examined how genetic complexity varies among strains and crosses. We used X-QTL to identify the genetic basis of resistance to 13 diverse chemicals in all 6 pairwise crosses of strains BY, RM, YJM789, and YPS163. YJM789 (hereafter, YJM) is derived from a clinical isolate, and YPS163 (hereafter, YPS) is an oak strain. These 4 strains are highly diverged at the sequence level [7], [8], [9], [10], [11] and exhibit a wide range of heritable phenotypic differences [12], [13], [14], [15], [16], [17], [18], [19]. Because of the statistical power gained by using very large mapping populations, we detected approximately an order of magnitude more loci than did previous studies involving multiple crosses of yeast strains [15], [17], [20], allowing us to gain deeper insights into the genetic architecture and evolution of complex traits in S. cerevisiae. We previously noted that levels of genetic complexity underlying heritable variation in growth differed among chemical conditions in a single cross [5]. Here, we sought to determine the generality of our previous finding by examining additional crosses. We first generated the strains and microarrays to conduct X-QTL in all 6 pairwise crosses of the BY, RM, YJM, and YPS strains (Materials and Methods). Because the statistical power of X-QTL is dependent on effective enrichment of highly resistant cross progeny in a segregating pool, and the crosses vary in their genetic compositions, leading to different distributions of resistance among the progeny of each cross, we used dose-response experiments to determine cross-specific, highly selective drug concentrations for each of 13 diverse chemicals that resulted in similar selection intensities for all crosses (Materials and Methods; File S1). Once the selective doses were determined, we conducted one X-QTL experiment for each chemical and cross combination. We observed substantial variation in the number of loci detected in different conditions and crosses (Figure 1). Across all 78 X-QTL experiments, we identified 837 total peaks at a False Discovery Rate (FDR) of 1%, or an average of 10. 7 peaks per trait per cross (Figure 1; Figure S1A–S1M). Both the chemical and the cross had significant effects on the number of peaks detected in an X-QTL experiment (ANOVA, chemical effect F = 5. 27, d. f. = 12, p = 5. 67×10−6; cross effect F = 3. 14, d. f. = 5, p = 0. 014), with the effect of the chemical (partial R2 = 0. 46) being much larger than the effect of the cross (partial R2 = 0. 11). An ANOVA testing the effects of chemical and strain resulted in a similar effect of chemical on the number of detected peaks (partial R2 = 0. 46; F = 4. 52, d. f. = 12, p = 3. 51×10−5), but no strain had a significant effect on its own (partial R2<0. 02; F<2. 5, d. f. = 1, p>0. 12; Materials and Methods). Consistent with a comparatively small effect of strain background on genetic complexity, only one trait showed a significant excess of peaks in crosses involving any one strain: crosses in which RM was one of the parents had an excess of peaks in diamide (χ2 = 22. 44, d. f. = 1, Bonferroni-corrected p = 1. 97×10−4; Figure 1). These results suggest that genetic complexity in yeast is mainly a property of the trait being examined rather than of the strain background. For each trait, we expected to detect loci at the same genomic positions in different crosses sharing a parent. To identify only the distinct loci affecting each trait, we performed a grouping procedure on the peaks identified in all crosses for a given chemical condition. We found 411 distinct loci (an average of 32 loci per condition), with a minimum of 8 loci for growth in cycloheximide and a maximum of 57 loci for growth in zeocin (Figure 1 and Figure 2A). We then examined the extent to which these loci showed effects on growth in multiple conditions. For a range of genomic window sizes, we considered peaks detected for multiple chemicals within a window to correspond to the same underlying locus, and counted the number of conditions in which the locus showed an effect. With 50-kilobase (kb) windows, we found that 40% of the distinct loci had effects in only one condition, 29% had effects in two conditions, 11% had effects in three conditions, and only 20% had effects in four or more conditions (Figure 2B; Materials and Methods). Although the numbers differed across window sizes, the general observation that most of the detected loci had effects in a relatively small number of the tested conditions, and only a small number of loci showed effects across a large number of conditions, held over the entire range of plausible sizes. With 50 kb windows, three loci exhibited effects in more conditions than expected by chance (Materials and Methods). These loci were located on Chromosome V near the X-QTL control marker CAN1, Chromosome X near ENT3, RSF2, and VPS70, and Chromosome XIV near the pleiotropic gene MKT1. We next examined the patterns of detection of loci for each trait across the six crosses. With four strains, two simple patterns are possible at bi-allelic loci: one strain can carry an allele that confers susceptibility or resistance relative to the allele carried by the other three strains, or two strains can carry the more susceptible allele and two strains the more resistant allele. We refer to these cases as “allelic singletons” and “allelic doubletons, ” respectively. These two cases should give rise to different patterns of peaks: peaks with a consistent direction of effect in all three crosses involving one strain for allelic singletons, and peaks with specific effect directions in four specific crosses for allelic doubletons (Table S1; Table 1). Allowing for false-negative peaks, 135 of the 411 distinct loci showed patterns consistent with allelic singletons, and 28 showed patterns of peaks consistent with allelic doubletons (Table S1; Table 1). We attempted to narrow the number of candidate genes for each of the bi-allelic loci by scanning the parental genome sequences for SNP alleles that are found in the four strains in a pattern consistent with the peaks. Using this approach, we found an average of 10 candidate genes per locus, with a range of 1 to 18 genes. Further restricting the list of candidate genes to those that carry nonsynonymous polymorphisms with appropriate allelic patterns reduced the average number to 6 per locus. We attempted to validate the genes underlying some of these loci by constructing allele replacement strains, and found reproducible evidence that HXT6 and RED1 harbor functional polymorphisms that confer growth differences in rich medium and tunicamycin, respectively (Figure S2; Materials and Methods). HXT6 is a high affinity glucose transporter [21], suggesting that variability in glucose uptake may contribute to growth differences among the strains. The effect of RED1 on tunicamycin resistance is less clear, as this gene is thought to be involved in chromosome segregation [21], and tunicamycin affects the unfolded protein response. We also constructed allele replacement strains for two other genes: NUP157, which lies within a copper sulfate resistance locus with the resistance allele coming from BY, and PTK1, which lies within a paraquat resistance locus with the resistance allele coming from YPS. However, we obtained inconsistent results for NUP157 and PTK1: the allele replacements produced effects on resistance that were in the opposite direction from those seen in the X-QTL selections, and also caused growth defects on standard rich medium, suggesting that we did not identify the right candidate genes for these loci. In addition to the simple bi-allelic patterns, we observed other more complex patterns of peaks (Figure 2A). Some of these are consistent with the presence of allelic series, in which either three or four alleles with different phenotypic effects are present among the four strains; we observed 29 examples involving at least 3 alleles and 9 examples that can only be explained by the presence of 4 different alleles (Table S2). The other 210 loci (51% of all loci) showed patterns of peaks that were not easily interpretable in terms of specific allelic classes. This probably reflects a mixture of false negatives in which a peak was present but not detected in a given cross, and cross-specific effects due to non-additive interactions and linkage between loci. The allele frequency spectrum of causal loci is critical for the design of genetic mapping studies and for understanding sources of missing heritability in natural populations, including humans. As discussed above, we were able to distinguish and enumerate two simple allelic classes—singletons and doubletons. We used a maximum likelihood approach that accounted for false negatives to estimate the ratio of allelic singletons to doubletons. We estimated the peak detection rate to be 51%, with a 95% confidence interval of 39%–62%, and the ratio of allelic singletons to doubletons to be 3. 03, with a 95% confidence interval of 1. 7–5. 3 (Figure 3A; Figure S3). This result suggests that despite the high statistical power of X-QTL, a substantial fraction of loci with weaker effects likely still go undetected in any one cross. Interestingly, the estimate of the ratio of allelic singletons to doubletons is similar to that observed for nonsynonymous polymorphisms in the genomes of the parent strains (2. 97), and is shifted toward singletons relative to both the neutral expectation of 2. 67 and the observed ratio of 2. 57 for 109,585 SNPs genome-wide (Figure 3A). Thus, the frequency spectrum of variants that contribute to complex trait variation in yeast appears to be mildly shifted toward lower frequencies by purifying selection, but, given the wide confidence interval for the estimated ratio of allelic singletons to doubletons, we cannot rule out that the variant frequencies follow the neutral spectrum. Several lines of evidence suggest that lineage-specific selection or demography has shaped variation among the four strains. We observed an excess of allelic singletons at detected loci for BY and RM, and a deficit for YJM and YPS, relative to the numbers of singleton SNPs in the parent genomes (χ2 = 35. 98; d. f. = 3, p<0. 0001; Figure 3B). The laboratory strain BY also exhibits other signatures of selection for both general and chemical-specific resistance. For instance, BY carries a marginally significant excess of allelic singletons that confer resistance relative to the other three strains (Fisher' s exact test, Bonferroni-corrected p = 0. 06; Figure 3C; Table 1). In addition, trait-specific sign tests [22] identified one significant result: an excess of copper sulfate resistance alleles contributed by BY in the BYxRM cross (18 loci with BY carrying the resistance allele and 2 loci with RM carrying the resistance allele; binomial test, Bonferroni-corrected p = 0. 031; Figure 3D). Interestingly, BY is among the most copper-resistant S. cerevisiae strains [23], [24], and our data suggest that this resistance in BY may be the result of selection, possibly due to the use of high levels of copper or another chemical with similar effects in standard growth media. However, the BYxYJM and BYxYPS crosses do not show significant excess of BY alleles, and RM is also among the more highly copper-resistant strains [23], making the excess of BY resistance alleles in the BYxRM cross difficult to explain. Overall, our results are consistent with previous analyses that have shown lab strains isogenic to BY exhibit high evolutionary rates relative to other yeast isolates [25], probably due to both relaxed purifying selection [26] and adaptation [26], [27]. We have shown that variation in chemical resistance among yeast strains is typically due to a large number of underlying loci. The level of genetic complexity, as measured by the number of loci detected, is largely a property of each resistance trait, although it is also affected to a lesser extent by the choice of parent strains. The total number of distinct loci detected for a trait in these crosses among four strains ranged from 8 to 57, and these numbers substantially exceeded those seen in any one cross. These observations suggest that the total number of loci affecting certain resistance traits in S. cerevisiae can be very large, since many of them will have escaped detection because they don' t vary among the four parent strains examined here, have effect sizes that are too small, or are too closely linked to be resolved as separate loci by our mapping technique. Our results suggest that the functional variants underlying complex traits are broadly distributed across the frequency spectrum from rare to common alleles, and that many loci harbor more than two allelic variants. These findings provide multiple non-exclusive explanations for the sources of the “missing heritability” of complex traits, and illustrate the power of a simple model system for probing genetic complexity. The Synthetic Genetic Array marker system [28] was used to generate MATa haploid pools as previously reported [5], with the exception that thialysine and the dominant sensitive LYP1/lyp1Δ marker system were not employed. All six pairwise crosses of BY, RM, YJM, and YPS were made, with one strain in a cross having the genotype MATα can1Δ: : STE2pr-SpHIS5 his3Δ and the other having the genotype MATa his3Δ. In notation describing crosses (e. g. , BYxRM), we first list the MATα and then the MATa parent. The selection experiments used for X-QTL were conducted as previously described [5]. The drug doses used in the selections, which were determined by plating millions of cells across a range of drug doses and finding a concentration at which 300 to 1,000 colonies could be resolved, are given in File S1. Each experiment was conducted once, as we previously found that biological replicates conducted on the same day produced highly similar results [5]. Microarrays were designed from the BY genome sequence obtained from the Saccharomyces Genome Database (http: //www. yeastgenome. org/) and from assemblies of the RM, YJM, and YPS genomes obtained from the Saccharomyces Genome Resequencing Project [10]. Note that the YPS606 genome was used to design the YPS array, as YPS606 is isogenic to YPS163. We aligned the genomes chromosome-by-chromosome using Fast Statistical Alignment (FSA) [29]. These multiple sequence alignments were filtered for SNPs using the following criteria: i) all 4 strains had to have been sequenced at a position and ii) all 4 strains had to have a specific base called (i. e. A, C, G, or T) at the position. These SNPs were then used for microarray design, as well as for downstream population-genetic analyses. Cross-specific microarrays were designed using only bi-allelic SNPs. Probes were chosen to have a length between 21 and 27 nucleotides and a melting temperature between 54 and 56°C as described previously [5], [30]. One probe was designed for each allele of a SNP, and the two probes for a SNP were randomly positioned on the microarray. Probes were targeted to regions where only one SNP would be covered by the probes. Markers were chosen to provide near-uniform coverage of the genome. The arrays were tested using control DNA from both parents and the heterozygous diploid to ensure that they could discriminate the two alleles of a SNP. All hybridizations and processing was done as previously described [5]. All microarray data is available in the Princeton University MicroArray database (http: //puma. princeton. edu/). The processed log10 hybridization intensities are included in Files S2, S3, S4, S5, S6, S7. For a given SNP, the difference in the log10 ratios of the intensities of the MATα and MATa parent-specific probes on a single array was computed (subsequently referred to as a ‘log10 intensity difference’), and this metric was used in downstream analyses. Background allele frequency changes that occur during pool construction were removed from the data for each X-QTL selection. This was done separately for each SNP by subtracting the average log10 intensity difference obtained in seven cross-specific control experiments from the log10 intensity difference observed in an X-QTL selection. A peak detection algorithm was then employed that used a Savitzky-Golay filter to smooth the data within sliding windows of 100 probes. This smoothing approach was used to preserve local maxima in the data. Loci were called at a 1% FDR threshold, where the number of false discoveries was determined by running the peak caller on the control data using a range of thresholds, and the total number of discoveries was determined by running the peak caller on the selection data at the same thresholds used to analyze the controls. Thresholds were set by examining the quantiles of log10 intensity differences observed for every 100 SNP genomic window on an array, and taking the median interquantile range between the x and 1-x quantiles, where x ranged from 0. 005 to 0. 45. We found that setting x as 0. 045 resulted in a 1% FDR. Peak calling and all other statistical analyses were conducted in R (http: //www. r-project. org/). The peak caller and an associated function library are included in Files S8 and S9. The detected peaks are listed in File S10. The test for cross effect was conducted using the model y = chemical+cross, while the test for strain effect was conducted using the model y = chemical+strain1+strain2+strain3. Implementing the second test required specifying the design matrix for the strain effect. Each row in the design matrix represented a single X-QTL experiment from a particular combination of chemical and cross. Entries in the design matrix were parameterized as follows: a strain had a value of −1,1, or 0 if it was the MATa parent, the MATα parent, or not a parent in a particular experiment, respectively. Only three strains were included in the test, because the information for the fourth could be obtained from the other three. To ensure that results were not dependent on the three included strains, we conducted the test with all four possible combinations of the three strains and reported the maximum partial R2 and F values, and the minimum p value in the text. We first conducted χ2 tests in which single strains were examined. This test has two categories – one that is the sum of the peaks detected in the three crosses involving the query strain and another that is the sum of the peaks detected in the other three crosses. The expectation is that each of these classes will contain half of the peaks detected for a trait. We then conducted χ2 tests in which two strains were examined. The first category here is the sum of the peaks detected in the four crosses involving the two strains, while the second is sum of the peaks detected in the other two crosses. Here, the expectation is that the first category will contain two-thirds of the peaks, while the second will contain one-third of the peaks. Peaks identified across the six crosses for a single trait were grouped into distinct loci. We started with the most strongly selected peak on each chromosome and grouped with it all peaks that occurred within a 200-kilobase window surrounding it. This window size accommodated the grouping of peaks that exhibited weak but significant allele frequency changes, and may result in the underestimation of the total number of loci due to the overgrouping of peaks. Remaining peaks were grouped into distinct loci using additional iterations of the procedure until all peaks identified for a trait were members of a group. We divided the genome into equally sized bins ranging from 20 to 100 kb and counted the number of distinct loci that fell into each bin. A bin was considered to have an excess of distinct loci if the number present in it exceeded the number expected by chance from a Poisson distribution, given the number of distinct loci divided by the total number of bins and a multiple testing correction for the number of bins. With the 50 kb bin size reported in the text, 8 or more distinct loci were required to be present in a bin for the bin to be considered significant. The distinct loci identified for each trait were used to classify singletons and doubletons. The specific patterns used to identify the allelic classes are described in Table S1. We focused on exact pattern matches and on patterns that were missing an expected peak at a given locus in one cross. A number of distinct loci had peaks detected in four or more crosses, but did not conform to the patterns expected for allelic doubletons. We considered these loci as allelic series, and for each of these putative series we determined the possible logical relationships of the parent alleles to each other. These relationships are reported in Table S2. For each bi-allelic locus, we evaluated a 30 kb interval centered on its estimated position for polymorphisms that segregated among the parent strains in the same pattern as the X-QTL peaks. Any gene that harbored a polymorphism in the coding region or in the immediate upstream and downstream regions was considered a candidate. The candidate genes are listed in File S11. To generate the replacement strains, we used the allele replacement technique described by Storici et al. [31]. This method is a two-step process that involves knocking out a gene with a selectable marker cassette, and then replacing the selectable marker cassette with a different allele of the gene. We made each allele replacement strain once in one parental background, and then compared the phenotypes of the strains to their progenitors. For the two genes that exhibited the expected phenotypic effect, we made a second version of the allele replacement strain to validate the presence of functional variation in the gene. The observed counts of exact-match allelic singletons and doubletons and near-exact-match allelic singletons and doubletons were modelled using two parameters: the detection rate of peaks (α) and the ratio of singletons to doubletons (β). The formulae underlying this computation are provided in Text S1. The likelihood of each combination of parameter values was examined across a two-dimensional grid of parameter values using χ2 tests with 3 degrees of freedom. The likelihood reached a maximum at α = 0. 51 and β = 3. 03. We obtained 95% confidence intervals for α and β by using the χ2 distribution with 3 degrees of freedom and identifying the χ2 value for the 95% quantile. We then identified parameter combinations that produced an χ2 value below this threshold (7. 81), and determined the minimum and maximum values of α and β that satisfied this condition.

Most heritable traits of agricultural, evolutionary, and medical significance are specified by multiple genetic loci. Despite decades of research, we have only a limited understanding of the genetic basis of such complex traits. Studies in model organisms have the potential to provide fundamental insights into this research area, but most genetic mapping studies in these species have had low statistical power to detect multiple loci with small effects. Using a technique in which we employed millions of cross progeny in genetic mapping, we previously showed that resistance to chemicals has a highly complex genetic basis in a cross of a lab strain and a wine strain of the budding yeast Saccharomyces cerevisiae. Because we only examined a single cross, it was unclear how general our findings were. Here, we expand our work to all six possible crosses of four strains—the two isolates we used in our last study, as well as an isolate from an immunocompromised human being and an isolate from the sap of an oak tree. Our results based on these four ecologically and genetically distinct S. cerevisiae strains suggest that resistance to chemicals commonly exhibits a highly complex genetic basis among yeast isolates.

Abstract Introduction Results/Discussion Materials and Methods

genetics biology genomics evolutionary biology genetics and genomics

2012

Genetic Architecture of Highly Complex Chemical Resistance Traits across Four Yeast Strains

The improved capacity to acquire quantitative data in a clinical setting has generally failed to improve outcomes in acutely ill patients, suggesting a need for advances in computer-supported data interpretation and decision making. In particular, the application of mathematical models of experimentally elucidated physiological mechanisms could augment the interpretation of quantitative, patient-specific information and help to better target therapy. Yet, such models are typically complex and nonlinear, a reality that often precludes the identification of unique parameters and states of the model that best represent available data. Hypothesizing that this non-uniqueness can convey useful information, we implemented a simplified simulation of a common differential diagnostic process (hypotension in an acute care setting), using a combination of a mathematical model of the cardiovascular system, a stochastic measurement model, and Bayesian inference techniques to quantify parameter and state uncertainty. The output of this procedure is a probability density function on the space of model parameters and initial conditions for a particular patient, based on prior population information together with patient-specific clinical observations. We show that multimodal posterior probability density functions arise naturally, even when unimodal and uninformative priors are used. The peaks of these densities correspond to clinically relevant differential diagnoses and can, in the simplified simulation setting, be constrained to a single diagnosis by assimilating additional observations from dynamical interventions (e. g. , fluid challenge). We conclude that the ill-posedness of the inverse problem in quantitative physiology is not merely a technical obstacle, but rather reflects clinical reality and, when addressed adequately in the solution process, provides a novel link between mathematically described physiological knowledge and the clinical concept of differential diagnoses. We outline possible steps toward translating this computational approach to the bedside, to supplement today' s evidence-based medicine with a quantitatively founded model-based medicine that integrates mechanistic knowledge with patient-specific information. The potential of computer-based, algorithmic support for medical decision making in data-rich environments, and in particular in the context of evidence-based practice, was recognized early on and has been pursued extensively [3,7–11]. Few of these efforts, which mostly have consisted of rule-based expert systems, statistical models, or approaches driven by machine learning ideas such as dynamic Bayesian or artificial neural networks, have reached a sufficient level of practicality and usefulness to be accepted into the day-to-day practice of acute care medicine [8,12,13]. These tools either attempt to formalize empirical knowledge already available to a physician (expert systems) or to capitalize on statistical associations of phenomena and inherent structures of the available dataset. All largely fail to make direct and quantitative use of known causalities and dynamics in the physiologic systems underlying the observed pathophysiology, which are typically characterized by basic science investigations. A promising approach to incorporating this knowledge into the medical decision making process would be to use mathematical models of physiologic mechanisms to map clinical observations to quantitative hypotheses about physiologic conditions, leading to improved insight into current patient status and, eventually, predictions about responses to therapeutic interventions. While complex mathematical models of physiology in general, and the cardiovascular system and its control in particular, have a long history and are still actively being developed [14–23], their translation to clinically useful tools has proved challenging. Early examples of using mathematical models to quantify “hidden” parameters based on clinical measurements include the pioneering work of Bergman et al. in the late 1970s on glucose control and insulin sensitivity [24]. More recent work in the same field has focused on accurately quantifying the uncertainty arising in the resulting parameter estimation problems using current methodology, such as Markov chain Monte Carlo approaches [25]. In the critical care environment, extremely simple models of the cardiovascular system have been in use for decades and are implemented in commercially available products, examples being the electric circuit analog of the systemic circulation used to calculate total peripheral resistance, which can then become a therapeutic target, or pulse contour analysis, which attempts a model-based assessment of systemic flow from arterial pressure waveforms [26–28]. The clinical application of mathematical models of physiology to date has failed to extend to models of sufficient complexity to significantly help alleviate the previously discussed problem of information overload in the diagnostic process. We contend that a key obstacle preventing the successful clinical use of available mathematical models has been the lack of a robust solution to the inverse problem. That is, any physiologically reasonable mathematical model of components of the human body will typically be nonlinear and have a large number of parameters. Despite the complexity of such models, if the user fixes the parameter values and initial values of the physiological states in a model, then the model can be simulated to obtain time courses of the physiological states (solving the forward problem; Figure 1). However, the corresponding inverse problem, i. e. , inferring parameters and starting conditions of state variables from measured physiological data (http: //www. ipgp. jussieu. fr/~tarantola/Files/Professional/Books/InverseProblemTheory. pdf) [29] will usually be ill-posed in the sense of Hadamard, meaning that it does not admit a unique solution that depends continuously on the data [30–32] (Figure 1). This ill-posedness is directly related to the concept of system identifiability in both the statistical and engineering senses of the term. The most popular approaches to the inverse problem in physiology, such as nonlinear least squares, which seeks a maximum likelihood estimate by minimizing the sum of squared residuals, inherently assume the existence of a unique “best” solution. The ill-posedness of the inverse problem corresponds to a violation of this assumption, which often causes solution approaches such as least squares to fail completely, in spite of regularization of the underlying nonlinear programming problem, or to give meaningless or even misleading results. More recent work attempts to quantify the uncertainty of resulting parameter estimates [25]. However, given the uncertainty stemming from the fundamental ill-posedness of the inverse problem, together with additional uncertainty from measurement error and model stochasticity, which affect both forward and inverse problems (Figure 1), a fully probabilistic approach to the inverse problem in quantitative physiology seems appropriate. We hypothesize that the ill-posedness of the inverse problem is not merely a technical obstacle but reflects clinical reality in the sense that an experienced physician is rarely certain about a patient' s status, despite a large number of available observations. More typically, the physician entertains an evolving differential diagnosis, consisting of a list of hypotheses of varying likelihoods about the physiological mechanisms underlying available observations, updated and ranked according to current observations. We therefore propose to approach the inverse problem in such a way that uncertainty from all sources is quantitatively reflected by the solution, which will consequently take the form of a (typically multimodal) probability distribution on parameter and state space. This distribution will represent the relative likelihoods of the possible values of the physiological elements that these parameters and states represent, in the patient for whom the clinical observations are made. To explore the feasibility of such an approach, we combine a mechanistic model of cardiovascular physiology with a stochastic model of the observation process and Bayesian inference techniques to infer a posterior probability distribution on parameter and state space from prior (population-level and individual) knowledge and quantitative observations. We illustrate these ideas in a simplified simulation of a clinically relevant differential diagnostic procedure and examine the relationship between the obtained posterior probability density functions and pertinent qualitative differential diagnostic concepts. When only one blood pressure measurement was made, the probable parameter/state range represented a continuum of various combinations of contractility and hydration status. As expected, high-precision measurements (σ = 10 mm Hg; Figures 5A and 6A) led to more concentrated probability density functions than low-precision measurements (σ = 30 mm Hg; Figures 5B and 6B), independent of the type of prior used. Two peaks, corresponding to the differential diagnoses of “heart failure” (low contractility, normal-to-high total intravascular volume) and “hypovolemia” (normal contractility, low intravascular volume), can be discerned in the case when blood pressure was measured with high precision for both Gaussian and uniform priors (Figures 5A2 and 6A2). When the measurement was less precise, the peak corresponding to “heart failure” was nearly absent with a Gaussian prior, but not with a uniform prior (Figures 5B2 and 6B2). To illustrate the additional diagnostic knowledge gained from perturbing the system, we simulated a fluid challenge [33]. Depending on the system' s response to the intravenous administration of 1,500 ml of fluid, the updated posterior densities on parameter space were altered significantly (Figures 5A3,5A4,5B3,5B4,6A3,6A4,6B3, and 6B4). Specifically, for a high-precision measurement, a fluid challenge differentiated between cardiac causes of hypotension (“heart failure”; low contractility, low responsiveness to volume resuscitation; Figures 5A3 and 6A3) and lack of intravascular volume as cause (“hypovolemia”; normal or high contractility, high responsiveness to volume resuscitation; Figures 5A4 and 6A4). With low-precision measurements, the failure to restore blood pressure following the fluid challenge did not eliminate hypovolemia as the cause of hypotension (Figures 5B3 and 6B3). While the clinician often wonders whether there is a preferred sequence of diagnostic challenges for ascertaining an accurate diagnosis, the order of fluid challenges of different sizes, with consecutive assimilation of intermediate observations, had little effect on the final posteriors in this highly simplified setting. When we allowed three parameters to vary, the posterior distributions became truly multimodal. We depict two different visualizations of the grid points accounting for 95% of the total probability mass of posterior densities for the scenarios described earlier (Figure 7A and 7C), as well as for a more ambiguous post-resuscitation observation of 50 mm Hg (Figure 7B). As can be seen, the assimilated observations are still sufficient to meaningfully constrain the probable region in parameter/state space. In the poor (30 mm Hg post-resuscitation) response to volume scenario, an additional probability concentration appears. This additional probability mass corresponds to the possibility of shock induced by severely decreased peripheral resistance, which corresponds to the differential diagnostic possibility of a failure of vasomotor tone, as observed in septic, anaphylactic, or neurogenic shock states. For intermediate values of the post-resuscitation observation, the structure becomes even richer (Figure 7B), while the post-resuscitation observation of 70 mm Hg (good response) concentrates probability mass in a region of low intravascular volume (Figure 7C). While a greatly simplified physiological representation, our mathematical model of the cardiovascular system fulfills its design objectives: to be qualitatively correct in its response to variations in hydration status and myocardial contractility while incorporating enough homeostatic mechanisms to create realistic ambiguity in the identification of parameter values underlying observed states. The conversion of the discrete dynamical system representing the sequential filling and emptying of the heart (and the resulting “history awareness” of the system) into a compact system of ODEs that preserves the physiologic meaning of parameters of the discrete system is, to our knowledge, novel. Physiologically constrained cardiovascular simulations done by previous authors have typically involved either simulating intra-beat dynamics, which rapidly becomes computationally prohibitive, using a more ad hoc approximation of the Starling mechanism at the expense of physiological interpretability of parameters, or resorting to a beat-to-beat discrete time representation (e. g. , [20]). Our model is therefore particularly suited for simulation scenarios where an accurate description of intra-beat details is not required, yet a continuous form of inter-beat dynamics that preserves parameter meanings is desired. Our model derivation aims to achieve a reasonable compromise between representing all known mechanisms in full physical detail, which leads to challenges of simulation expense and intractability of the inverse problem, and model reduction, which may result in loss of physiological accuracy and interpretability. The need for such a trade-off is typical when modeling complex biological systems. From our perspective, making use of domain-specific knowledge to arrive at meaningfully interpretable model reductions whenever possible, and resorting to multiscale models with a hierarchical arrangement of submodels of different granularity and timescales when the assimilation of data on very different spatial and temporal scales is desired, may be the most promising way to address this issue. The ideal level of model complexity will generally depend both on the amount of data available for assimilation and the specific application intended. To what extent the growing theoretical understanding of model selection based on information theoretical measures [34] can be leveraged to facilitate or partially automate this process for physiological applications is an interesting topic for further investigation. As illustrated by this proof-of-concept implementation, the proposed methodology holds promise as a tool for integrating existing mechanistic knowledge and data generated by measurements in a clinical setting into a quantitative assessment of patient status. Our approach offers a means to achieve this integration in a way that not only incorporates all available data, but also quantifies the remaining uncertainty, thus avoiding unjustified claims of high certainty that could prove disastrous in a clinical setting. In particular, the clinical construct of differential diagnoses of different likelihoods is reflected in the observed multimodality of posterior probability distributions (Figures 5–7). More generally, representations of probability densities of states and parameters provide a natural setting for linking mathematical models of different scales and levels of detail since the distribution of states of some detailed small-scale models (“microstates”) may naturally determine a value or distribution of values for parameters of larger scale/lumped parameter models. The proposed approach aims to map clinical syndromes described by a set of observations to configurations of physiologically meaningful pre-observation states and parameters appearing within a mathematical model. Based on the physiological knowledge embodied in the model, certain regions in parameter and state space may in turn be associated with differential diagnoses, similar to the conditions of “hypovolemia, ” “heart failure, ” and “sepsis” in our simplified example. When this linkage is possible, the quantitative nature of the method presented here allows for the estimation and refinement of probability values associated with certain diagnoses. This, to our knowledge, is the first time that such a high-level concept central to clinical decision making is shown to emerge naturally from the combination of sequential observations, diagnostic challenges, and physiological principles. Moreover, we believe that the methods presented herein open novel avenues for exploring theoretical aspects of clinical epistemology, independent of practical applications. Since measurement characteristics are described stochastically, the method we demonstrate is not fundamentally limited to assimilating data from device-based quantitative measurements, but can also make use of rather qualitative clinical observations such as quality of peripheral perfusion, presence of lung rales, or altered mental status, provided reasonably informative densities on system states or parameters conditional on such observations can be defined. Similarly, genomic information can be naturally incorporated, since it can provide probability distributions of physiological parameters conditional on individuals' genomes. To what extent a combination of several subjective (or inaccurate) observations may exploit physiological coupling of observables and yield informative posterior distributions corresponding, for example, to a carefully performed clinical examination is a matter of current investigation. While modifying the order of physiological challenges did not have a tangible impact on diagnostic discrimination in our limited exploration, we anticipate that order generally matters, as a system' s response to a perturbation can be highly dependent on the system state at the time that the perturbation is delivered. That is, an initial diagnostic challenge will alter the state of the underlying system, which may impact its response to a subsequent challenge. Our approach could allow for a theoretical exploration of how to optimize the selection, and order, of diagnostic challenges for maximal information gain in the context of specific clinical scenarios. The simulations presented here illustrate the importance of congruence between the accuracy of observations and the level of information included in prior distributions. In particular, the inappropriate use of informative priors can be misleading in this context. In our results, for example, the combination of informative Gaussian priors with inaccurate observations effectively eliminates the physiologically reasonable “heart failure” peak in both the posteriors after a single observation of low blood pressure and the post-resuscitation posteriors for the low-response case, while the peak is still clearly evident in the case of uniform priors (Figures 5B2,5B3,6B2, and 6B3). This example demonstrates that a conscious choice needs to be made as to whether an interpretation based on population-level probabilities (corresponding to the use of informative priors) or an unbiased assessment of physiological possibilities (corresponding to the use of uniform priors) is more appropriate, when only few, low-quality measurements are available. Whether an optimal degree of incorporation of population-based information exists and how such an optimum could be defined are highly relevant issues that remain to be explored. Our observations suggest that when addressing inverse problems in quantitative physiology, the traditional approach of requiring a unique optimal solution may be misleading and introduce unnecessary information loss, at least in situations where the additional computational burden of characterizing the posterior distributions more fully is not prohibitive. Model reduction to eliminate perceived “overparametrization” may weaken the correspondence of components of a physiologically faithful mathematical model with components of the actual physiological system it describes. Furthermore, we hypothesize that the ambiguity of the mapping from observation to parameter space is at least partially due to a characteristic particular to physiological systems, namely their ability to tightly control certain system states via highly tuned, and often nested, internal regulatory feedback control mechanisms, such as the baroreflex in our simple example. In situations where such a controlled variable is observed, the ambiguity in the mapping from observation to parameter space is naturally exacerbated since perturbations of the observable will be compensated by alterations of other, possibly unobserved system states, as has already been proposed in a neurophysiological context [35]. Since a living organism is a system that maintains a state of dynamical equilibrium at energy expenditure, this phenomenon is likely to be the rule rather than the exception. A more precise formulation of this qualitative observation is the subject of current investigation. We believe that our approach provides a conceptually new quantitative framework for a theoretical description of the development of differential diagnoses, which may potentially be harnessed to improve this process. Eventually, this methodology could be extended to an outcome prediction tool and could help to optimize diagnostic and therapeutic interventions in individual patients. Its practical implementation will require broad interdisciplinary collaborations, because of the significant challenges involved. We nevertheless believe that the potential gains in diagnostic effectiveness and efficiency that can be made by taking a quantitative approach to uncertainty, based on our ever-growing mechanistic understanding of physiology, will make the effort worthwhile. The model we developed was designed to be computationally and conceptually simple while achieving a good qualitative reproduction of system responses to alterations in contractility and hydration status. It consists of a continuous representation of the monoventricular heart as a pump, connected to a representation of the systemic circulation with the large blood vessels treated as linear capacitors (Windkessel model) and with arterial pressure controlled by a physiological feedback loop (baroreflex [39], Figure 2). The pulmonary circulation is excluded for simplicity, since the perturbations to be studied in our example are not directly related to it. The physiological variables and parameters used in the following exposition are summarized in Table 1. The heart as a pump. At a basic level, the heart acts analogously to a piston pump. During each heartbeat, blood from the venous side of the circulation fills the ventricle (“piston”) during the filling phase (diastole). When the cardiac rhythm generator (anatomically, the sinoatrial node) triggers myocardial contraction, the heart starts to contract, increasing the pressure in the ventricle. This process leads to the ejection of blood toward the arterial side of the circulation, and thus emptying of the ventricle, as soon as intraventricular pressure exceeds the pressure on the arterial side of the circulation. Simplifying the underlying physiology somewhat, one key factor for the amount of volume entering the ventricle in diastole is the so-called preload, which is related to the pressure in the large veins immediately upstream of the heart. How much blood is ejected during systole depends on the so-called afterload, corresponding to the pressure in the large arteries downstream of the heart, the strength of the contraction of the heart muscle (myocardial contractility), and the extent to which the ventricle was filled during diastole. The amount of force the myocardium can develop depends on its current level of stretch, which gives rise to a relationship between the amounts of filling during diastole and ejection during systole, termed the Starling mechanism. We developed an ODE model of the monoventricular heart, omitting pulmonary circulation, by considering a single-cycle representation of the emptying (ejection) and filling of the ventricle. Systole. The model of ejection was based on the experimentally observed linearity of the relationship between stroke work WS, which refers to the work performed by the heart during ejection, and end-diastolic volume VED over a wide range of volumes [40]. This linear relation takes the form where the slope factor cPRSW is termed the preload recruitable stroke work. The volume axis intercept of this relationship has been found to be equivalent to the volume at which the passive intraventricular pressure is 0 mm Hg, [40]. Approximating stroke work as pure volume work performed from VED to the end-systolic ventricular volume VES against the arterial pressure Pa yields where PED is the intraventricular pressure at the end of diastole, and is the stroke volume. Based on the finding that the ventricular volume will not usually decrease below, we define the end-systolic volume as a function of the end-diastolic volume as follows: The expression for is obtained by equating the expressions for WS in Equations 1 and 2, solving for VS, and substituting the result into Equation 3 to obtain Note that is a continuous function of VED since, if, the limit of as PED approaches Pa from below is smaller than. Diastole. To complete one stroke cycle, we derive an expression for the end-diastolic volume as a function of the end-systolic volume. Ventricular filling is modeled as a simple passive filling through the linear inflow resistance Rvalve, driven by the difference between pressure in the central veins PCVP and ventricular pressure PLV (VLV), through the ODE In Equation 6, the dependence of ventricular pressure on ventricular volume is governed by the experimentally characterized [40] exponential relationship: Under the assumption of constant PCVP, Equation 6 is of the general form with constants which resolves to by quadrature. By letting t = 0 at the beginning of diastole and eliminating the unknown constant C using end-systolic volume VES as the initial condition, we obtain where the final expression is numerically advantageous since it avoids floating point overflow in the exponential terms. At a given heart rate fHR, and assuming an approximately constant duration of systole TSys (physiologically, the duration of diastole is much more strongly affected by alterations in heart rate than the duration of systole [41]), the end-diastolic volume will therefore be with V (t) given by Equation 11. If PCVP exceeds the intraventricular pressure at the beginning of diastole, then passive filling can occur and Equation 12 provides the desired expression for VED as a function of VES, through the VES-dependency of Equation 11. Otherwise, no filling will occur. The overall expression for VED as a function of VES is thus note that is a continuous function of VES since the limit of as PLV (VES) approaches PCVP from below is VES. Joining systole and diastole. We can now define a discrete dynamical system describing the beat-to-beat evolution of VED (or, similarly, VES). Specifically, given the current end-diastolic volume, we can use Equation 4 to compute and use Equation 13 to obtain. Together, these steps yield To obtain a continuous dynamical system amenable to coupling with continuous representations of the physiologic control loops and simulation with available ODE software over long time intervals, we converted VES and VED to state variables of a continuous time system. This was done by setting their rates of change to the average rates of change over an entire cardiac cycle that would occur during one iteration of the discrete time system for the current VES and VED values, to obtain The discrete system (Equation 14) and the continuous system (Equation 15) share identical sets of fixed points. Indeed, fixed points of the discrete system (Equation 14) are given by and by applying to Equation 16, we have thus at fixed points of the discrete system. Conversely, inspection shows that fixed points of Equation 15 satisfy Equation 16 and hence are fixed points of Equation 14 as well. The relationship between the stability of fixed points of Equation 14 and the stability of fixed points of Equation 15 is not obvious, since the discrete system (Equation 14) treats systole and diastole sequentially, while VES and VED co-evolve under Equation 15. However, linearization shows that any fixed point is stable with respect to both systems if, whereas the fixed point is unstable with respect to both systems if. (The stability condition can be evaluated at all points in (VES, VED) space, except for on the finite collection of lines where either derivative fails to exist (see Equations 4 and 13). In the case that, the fixed point is stable with respect to the continuous time system and unstable with respect to the discrete, suggesting that the continuous approximation in theory has the potential to eliminate instabilities inherent in the beat-to-beat dynamics. However, preliminary numerical explorations suggest that is significantly smaller than one for the relevant parameter range. When the continuous system (Equation 15) is embedded in the complete circulation model (described below), the full system very quickly settles to a stable fixed point (Figure 8). The systemic circulation. The circulation is represented by a simple Windkessel model. It consists of linear compliances representing the large arterial vessels of volume Va and venous vessels of volume Vv with respective pressures where α is “a” or “v” and where is the respective unstressed volumes, i. e. , the in general non-zero volume at which the pressure in the respective compartment will be 0 mm Hg. These pressures appear in the equation that links the arterial and venous compartments through a linear resistor, representing the total peripheral resistance RTPR regulating the arterio–venous capillary blood flow IC, namely The veno–arterial flow (cardiac output) ICO generated by the heart is given by the product of the heart rate fHR and the volume VS ejected per beat, which by Equation 3 takes the form Assuming conservation of volume at the nodes, the evolution of arterial and venous volumes is described by the following differential equations: where Iexternal represents a possible external blood withdrawal or fluid infusion to or from the venous compartment. Baroreflex control of blood pressure. Baroreflex control of blood pressure, which is one of the key regulatory mechanisms in cardiovascular homeostatis, is implemented based on the established representation of the central processing component of the baroreceptor sensor input as a combination of a sigmoidal nonlinearity (logistic function, in our case) with a linear system [15,18]. For simplicity, we reduced baroreflex activity to a single activating (sympathetic) output instead of the more physiologically accurate balance of stimulating (sympathetic) and inhibiting (parasympathetic) outputs. Since our model of the heart is designed to represent timescales significantly larger than a single beat, the linear part of the baroreflex feedback loop is simplified to display first-order low-pass characteristics with a time constant on the order of the slowest actuator response (unstressed venous volume control). Pure delays associated with the neural transmission of baroreflex signals are neglected. Under these assumptions, the temporal evolution of the stimulating output from baroreflex central processing is governed by the differential equation The stimulating output S (t) of the feedback loop acts on heart rate fHR, total peripheral resistance RTPR, myocardial contractility cPRSW, and unstressed venous volume effectors/actuators to adjust blood pressure according to its current deviation from the set point, based on the linear transformations where α = fHR, RTPR, or cPRSW, and The form of Equation 23, in particular, arises since the venous capacitance vessels contract, reducing their unstressed volume, in response to drops in blood pressure. Combining Equations 15 and 17– 23, and writing out dependencies relevant to the coupling of the system explicitly, we obtain a system of five ODEs: It should be noted that for Iexternal = 0, conservation of total intravascular volume would allow for elimination of one state variable (either Va or Vv) to obtain a four-dimensional system. We chose to leave the system in the above form, however, to preserve direct correspondence between anatomical entities and mathematical representation, at the cost of some loss in computational efficiency. With regard to the coupling between equations, it should be noted that the sympathetic nervous system activity S, which serves as a central control mechanism critical for functional cardiovascular system homeostasis, links together all components, while the coupling between the equations describing heart and circulation reflects the cyclical structure of the cardiac action and the circulation. Numerical solution of system Equation 24, as well as all other algorithms used for this work, was implemented in the MATLAB 7 (The MathWorks) programming environment, using the ode15s solver for numerical integration. The source code used in generating results is available in Text S1. Parameter selection. The parameters = 2. 03 mm Hg, = 7. 14 ml, and = 0. 066 ml−1 describing the ventricular pressure–volume relationship were estimated from experimental data for the left ventricle from [40] using the Levenberg-Marquard nonlinear least squares algorithm (Figure 9). The remaining parameter values and ranges for variables, as well as the respective sources, are given in Table 2 (for meaning of parameters, see Table 1). When no explicit source is given, assignment was based on the authors' perception of physiologically reasonable values for the simplified system, without claiming quantitative accuracy. For the purpose of this exposition, we are lumping the subset of n initial conditions (states) and parameters with which the inference is concerned into one product vector space with elements x ∈ X ℜn. Additionally, for simplicity, we assume a fully deterministic model M: X →, x ↦; M (x) of the physiological process that gives a mapping from X to a finite dimensional observation space with elements y ∈ Y ℜm. Extending the methodology to stochastic process models is straightforward but computationally more burdensome. Generation of prior densities. In a practical application, prior probability densities on parameter/initial condition space would have to be inferred from a finite set of observations of a population, which in itself constitutes an ill-posed inverse problem. For simplicity, we assume a known prior distribution fX: X → [0,1] on parameter/initial condition space and compute an approximation to the prior density on observation space consistent with the mechanistic model by evaluating using a Monte Carlo approach. Specifically, we draw N = 102dim (X) samples xi from fX and compute the corresponding observations yi = M (xi). A histogram with 50 bins/dimension is computed for approximate evaluation of fY. In the two-dimensional simulations, the parameter/initial condition space X consisted of the total intravascular volume Vtotal = Va + Vv, corresponding to hydration status, and a positive scaling constant c applied to the cardiac contractility response range. Initial conditions for compartmental volumes were generated from Vtotal by setting them proportional to the unstressed volume of the respective compartment and ensuring that observations were taken only after the system had equilibrated. In the three-dimensional setting, an additional scaling factor for the arteriolar resistance range was introduced. The prior densities fX for the informative/Gaussian case were independent Gaussian distributions with mean one and standard deviation 0. 5 for the scaling constants, while the mean of the distribution of total intravascular volume was chosen to correspond to a level of sympathetic nervous activity S (t) ≈ 0. 5 in steady state with both scaling factors at one, with a standard deviation of 1,000 ml. All distributions were truncated at zero by repeating the sampling if a negative value was drawn. fX for the uniform/non-informative case were independent uniform distributions on the intervals two standard deviations above and below the means of the corresponding Gaussian distributions, again truncated at zero. Sequential assimilation of observations. Sequential updating of the posterior densities was accomplished based on the standard Bayesian calculation of the conditional density where the stochastic measurement model g: Y × Y → [0,1], (y, ytrue) ↦ g (y, ytrue) describes the distribution of observations as a function of the true value of the observable ytrue. In Equation 26, the prior distributions fX and fY are either the original prior distributions described in the previous paragraph, if the first observation is being assimilated, or the posterior distribution on X obtained in the preceding assimilation step, approximated using weighted Gaussian kernel density estimation as implemented in the KDE toolbox for MATLAB [42], and its corresponding consistent distribution on Y, computed as described in the previous paragraph, in all other cases. For all of the above steps, visual comparisons between simulations with different and incommensurate grid resolutions were performed to ascertain that the observed multimodality of the distributions did not in fact constitute an artifact resulting from aliasing or other numerical effects.

Although quantitative physiology has developed numerous mathematical descriptions of components of the human body, their application in clinical medicine has been limited to a few mostly primitive and physiologically inaccurate models. One reason for this is that the inverse problem of identifying unknown model parameters and states from prior knowledge and clinical observations does not usually have a unique solution. Hypothesizing that this non-uniqueness might actually convey clinically useful information, we used a simplified mathematical model of the cardiovascular system and its control, in combination with Bayesian inference techniques, to simulate the diagnosis of low blood pressure in an acute care setting. The inference procedure yielded a distribution of physiologically interpretable model parameters and states that exhibited multiple peaks. The key observation was that these peaks corresponded directly to clinically relevant differential diagnoses, enabling a quantitative, probabilistic assessment of the simulated patient' s condition. We conclude that the proposed probabilistic approach to the inverse problem in quantitative physiology may not only be useful for quantitative interpretation of clinical data, and eventually allow model-based prediction and therapy optimization, but also provides a novel link between mathematically described physiological mechanisms and the clinical concept of differential diagnoses based on patient-specific information.

Abstract Introduction Results Discussion Methods

mathematics physiology dog computational biology homo (human) cardiovascular disorders

2007

From Inverse Problems in Mathematical Physiology to Quantitative Differential Diagnoses

The cellular immune system screens peptides presented by host cells on MHC molecules to assess if the cells are infected. In this study we examined whether the presented peptides contain enough information for a proper self/nonself assessment by comparing the presented human (self) and bacterial or viral (nonself) peptides on a large number of MHC molecules. For all MHC molecules tested, only a small fraction of the presented nonself peptides from 174 species of bacteria and 1000 viral proteomes (0. 2%) is shown to be identical to a presented self peptide. Next, we use available data on T-cell receptor-peptide-MHC interactions to estimate how well T-cells distinguish between similar peptides. The recognition of a peptide-MHC by the T-cell receptor is flexible, and as a result, about one-third of the presented nonself peptides is expected to be indistinguishable (by T-cells) from presented self peptides. This suggests that T-cells are expected to remain tolerant for a large fraction of the presented nonself peptides, which provides an explanation for the “holes in the T-cell repertoire” that are found for a large fraction of foreign epitopes. Additionally, this overlap with self increases the need for efficient self tolerance, as many self-similar nonself peptides could initiate an autoimmune response. Degenerate recognition of peptide-MHC-I complexes by T-cells thus creates large and potentially dangerous overlaps between self and nonself. The recognition of peptide-MHC-I complexes (pMHC) by the T-cell receptor (TCR) is required for effector T-cells to kill an infected cell. Although some MHC-I molecules have a preference to present pathogen-derived peptides [1], pMHC are formed with both self and nonself peptides. Therefore, to allow CD8 T-cells of the cellular immune system to discriminate self from nonself, presented nonself peptides should be different from presented self peptides. What would happen if a nonself peptide is so similar to a self peptide that it is recognized by the same T-cell (we will call such peptides “overlapping peptides”)? Firstly, an effector T-cell response to an overlapping peptide, could cause T-cell mediated autoimmune disease, such as type 1 diabetes [2]–[4] or multiple sclerosis [5], [6]. Secondly, to avoid autoimmunity, T-cells recognizing self-pMHCs are tolerized during negative selection [7]. Due to this self tolerance, overlapping nonself peptides should fail to elicit a T-cell response, and this may limit the number of pathogen-derived peptides that are available for an immune response and hence the chance to control a pathogen [8], [9]. Assarsson et al. showed that of the MHC-I presented vaccinia derived peptides are not recognized by T-cells [10]. Similarly, for HIV-1-derived peptides predicted to be presented on the well-studied HLA-A*0201 molecule, only as been reported to elicit a T-cell response [9]. Taken together, these studies suggest large “holes” in the T-cell repertoire [8], [11], which could be caused by overlaps with self pMHCs. We have previously shown that on HLA-A2 molecules only a minute fraction () of the presented nonself peptides are identical to presented self peptides [12]. Such a small overlap cannot cause the large holes in the T-cell repertoire. However, at that time there was too little data available on T-cell recognition of pMHCs, to study its impact on the self/nonself overlap. It is well established that T-cells are cross-reactive and can recognize similar, and sometimes even unrelated, peptides presented on the same MHC molecule [13]. The principles of TCR-pMHC interactions that allow for this flexibility are not fully understood. CTL recognition-studies using peptide libraries with altered peptide ligands [9], [14]–[18] and pMHC-TCR structures [19], [20] allow some inferences to be made. The middle (P4–P6) part of the peptide forms the core of the interaction [9], [14]–[20], where the majority of amino acid substitutions (with exception of those with very similar amino acids) tend to perturb pMHC recognition. Other positions in the peptide, although not in direct contact with the TCR, can still be important for the TCR-pMHC interaction if they affect the configuration of the P4–P6 residues [14], or MHC-binding [21]. In most cases, the N-terminal position (P1) of the peptide is unimportant for the TCR-pMHC interaction [9], [14], [16]–[20]. Given these new insights, we here extend our previous investigations on self/nonself overlaps by including the T-cell recognition of pMHCs. In addition, we analyze the self/nonself overlap of peptides presented on several HLA-A and HLA-B molecules, to estimate the degree of variance among different MHC-I molecules. Using high-quality predictors of the MHC-I presentation pathway [22]–[25], we show that presented peptides derived from nonself are in almost all cases () distinct from presented self peptides, for all common MHC molecules. This result is in agreement with our original observation that most peptides with a length of nine amino acids (9 mers) of unrelated species are unique [12]. However, the cross-reactivity of T-cell recognition is shown to increase the self/nonself overlap between sufficiently similar peptides to about one-third. Our results suggest an explanation for the observed holes in the T-cell repertoire during an infection, and we show that our self/nonself overlap estimates can be used to distinguish immunogenic from non-immunogenic pMHCs. Moreover, the estimates of self/nonself overlap demonstrate that the risk of autoimmunity due to molecular mimicry with pathogens is nonnegligible. MHC class I molecules shape CD8 T-cell responses via the presentation of peptides derived from intracellular proteins. These peptides are short: most MHC-I molecules prefer to bind peptides of 9 amino acids (9 mers). To investigate how similar self and nonself peptides are, the human and a large number of nonself proteomes (data selection is detailed in Methods) were cut into fragments of various lengths (1–20 amino acids long) and peptides that occur both in self and nonself proteomes were identified (i. e. without considering MHC-I presentation). The fraction of foreign peptides that are also present in the human proteome defines the “overlap”, i. e. the chance that a randomly chosen nonself peptide is identical to a self peptide. For small peptides shorter than five amino acids, the overlap is 100%, since almost every 5mer is present in the human proteome (see Figure 1). For longer peptides the overlap decreases rapidly, and at a length of 9 amino acids the average overlap is only 0. 20% for viruses (between 0–0. 5% for 95% of all viruses) and 0. 19% for bacteria (0. 1–0. 4% for 95% of all bacteria). These results are in excellent agreement with our previous estimates based on a much smaller set of nonself proteomes [12]. To conclude, 9 mers contain enough information to discriminate self from nonself, i. e. the chance that a nonself 9mer overlaps with a self 9mer is only 0. 2%. Surprisingly, the overlaps do not decrease much further for peptides longer than 9 mers (see Figure 1). To characterize these overlapping sequences further, for each human protein we counted the number of viruses or bacteria that has at least one overlapping 9mer peptide. The proteins where this number was larger than expected (p0. 01, see Methods) were analyzed by a functional annotation cluster analysis [26], [27]. This analysis showed that bacterial 9 mers tend to overlap with human proteins of mitochondrial origin, which is in line with the bacterial origin of mitochondria [28]. In addition, proteins involved in metabolic processes that might be common to bacteria and humans had more overlapping 9 mers (see Table S1). For viruses, the overlap is largest with nuclear proteins and transcription factors that are possibly acquired via horizontal gene transfer to modulate host cellular processes (see Table S1). In order to test the effects of homologous sequences or convergent evolution on self/nonself overlaps, sequences were shuffled before examining the overlap to break up any overlap that might be the result of these effects. Indeed, this shows that a far majority of the overlaps were due to these homologous sequences as the overlaps in shuffled sequences are much lower than the actual overlaps (Figure 1, in stars). Only peptides that are presented on an MHC-I molecule, i. e. about 1–3% of all 9 mers [10], can be recognized by T-cells. Due to the binding preferences of different MHC-I molecules, the self/nonself overlap of MHC-I presented peptides can be different per MHC-I molecule and does not need to be the same as the overlap based on all 9 mers. For instance, we recently showed that certain MHC-I molecules have a preference for pathogen-specific peptides [1]; such a preference should decrease the self/nonself overlap for that MHC-I molecule. To estimate the self/nonself overlap of MHC-I presented peptides, an in silico approach was undertaken using state-of-the-art MHC-I pathway predictors [22]–[25] (see Methods). For a large set of common human MHC-I molecules (13 HLA-A molecules and 15 HLA-B molecules, see Methods for selection criteria), the presented peptides in the human proteome and a large set of nonself proteomes were predicted. To define presented peptides we made use of the well-studied HLA-A*0201 molecule. For this molecule an IC50 value of 500 nM is often taken as threshold to separate the binders from non-binders. Applying this threshold to all self peptides we find that HLA-A*0201 has a specificity of 2. 3%, i. e. 2. 3% of the tested peptides would be binders. For other HLA molecules we determined “scaled” binding thresholds, so that they have the same specificity as HLA-A*0201, i. e. they present 2. 3% of all self peptides. Next, the overlap between presented self and nonself peptides was enumerated per MHC-I molecule, by comparing for each HLA molecule, self and nonself peptides presented on that HLA molecule. On average, only 0. 15% of the MHC-I presented nonself peptides is identical to a presented self peptide (see Figure 2A, left). The average overlap of MHC-I presented peptides is somewhat smaller than the overlap of all 9 mers in the proteome (0. 2%, see Figure 1), which is in agreement with the fact that many MHC-I molecules have a slight preference for pathogen-derived peptides [1]. The maximal overlap of 0. 33%, which is still very low, was found for peptides presented by HLA-B*5401. These results demonstrate that for all common human MHC-I molecules, only a minute fraction of the presented nonself peptides is identical to a presented self peptide. By using scaled binding thresholds, we take the conservative assumption that different HLA molecules have similar specificities, this does not have to be so. The self/nonself overlaps were also calculated by using a fixed binding threshold of 500 nM, which leads to different specificities for different HLA molecules. In this case, the self/nonself overlap determined for peptides presented on different HLA molecules remained as low as when scaled thresholds were used (see Figure 2A, right). So far, we only considered identical self and nonself peptides as overlaps. However, also non-identical MHC-I presented peptides can be recognized by the same T-cell [13]. This cross-reactivity is partly due to the fact that not all the residues on a presented peptide are accessible for the TCR. For example, most MHC-I molecules have two binding pockets that bind positions 2 and 9 (i. e. anchor-residues) of the presented peptide. These anchor-residues are hidden in the binding pocket of an MHC-I molecule, and are not exposed to the TCR [29]. Recently, we analyzed the T-cell recognition of the HIV-1 derived SLFNTVATL peptide presented on HLA-A*02 and suggested that not only the anchor-residues (P2 and P9), but also the first position (P1) of the presented peptide, hardly affects T-cell recognition [9]. Furthermore, at the remaining six middle positions (P3–8), some amino acid substitutions did not perturb T-cell recognition, especially those between amino acids with similar physical-chemical properties. TCR recognition was most stringent at the fifth position (P5), where only a Threonine-to-Serine substitution did not affect recognition [9]. To see if other TCR-pMHC contacts follow the same interaction-“rules”, all non-redundant TCR-pMHC-I structures found in the PDB-database (www. pdb. org [30]) encompassing a 9mer (n = 9, see Methods for selection criteria) were studied. In agreement with Frankild et al. [9], the majority of interactions in these structures involved the middle positions of the presented peptide (Figure 3). Several other reports on TCR-pMHC structures, and on different T-cell clones, confirm the degeneracy at the first position, and confirm that substitutions among similar amino acids are allowed in other positions [14]–[20]. Our structural analysis suggests that the third position has less contacts with the TCR than the other middle positions (Figure 3). However, Tynan et al. [14] show examples in which position 3 is important for T-cell recognition. Therefore, we conservatively assume that the third position is as important for T-cell recognition as the other middle positions (P4–8). Given these data, we studied how much of presented nonself can be discriminated from presented self by T-cells. First, the self/nonself overlaps were determined on those positions recognized by T-cells, i. e. the middle positions (P3–8) of MHC-I presented peptides. The self/nonself overlap of these 6mer fragments is on average 18 times higher than the overlap based on all positions (i. e. , 2. 7% for scaled thresholds and 1. 7% for fixed thresholds see Figure 2B). This increase in the overlaps is mainly due to excluding the first position: if only both anchor positions are discarded, the overlap determined on the non-anchor positions (P1 and P3–8) remains low (i. e. 0. 4% on average, see Table 1 and Figure S1). Similarly, if only one of the anchor positions and position P1 are discarded, the overlap is much higher (Table S2). We showed previously that highly specific anchor-positions of MHC molecules do not have to be exposed to the TCR to contribute to self/nonself discrimination because T-cells are MHC restricted [12]. For instance, HLA-A*0101 has a very specific preference for Tyrosine at the second anchor position (P9), and even if an HLA-A*0101 restricted T-cell is not interacting with this amino acid, all presented peptides it can possibly respond to must have a Tyrosine at position 9. Next, overall self/nonself overlaps were estimated with a novel model of degenerate T-cell binding. As above, T-cells were assumed to bind to the middle positions (P3–8) of the MHC-I presented peptides only. In addition, the degeneracy was modeled by considering two peptides as overlapping if they have mismatches in maximally two regions. We allow one mismatch at the N-terminal side of the fifth position (P1–4) and one at the C-terminal side of that position (P6–9) (see Methods). Moreover, only mismatches between amino acids having similar peptide-protein interaction properties were allowed, as such conservative substitutions have been shown to have a limited influence on T-cell recognition [9], [13]–[15]. The similarity between amino acids was derived from the PMBEC amino acid substitution matrix, that is based on peptide-MHC interactions and therefore specifically tailored to estimate the influence of amino acid substitutions on peptide-protein interactions [31]. We refer to this new overlap as the “degenerate” overlap. The degenerate self/nonself overlap is much higher than the identical overlaps of P3–8, on average 29% (see Figure 2C, left). These results can be ascribed to the degenerate nature of T-cell recognition: when using an alternative model of TCR recognition described by Frankild et al. , the “peptide similarity score”-method (see Methods) [9], similarly high self/nonself overlaps were observed (results not shown). The self/nonself overlaps based on middle positions of the presented peptide (P3–8), determined using fixed binding thresholds were very similar to the overlap based on scaled thresholds (see Figures 2C, right), though more varied and somewhat lower. This is a result of the differences in the specificities of HLA molecules. The specificity determines the fraction of presented self and nonself peptides, which in turn influences the chance of finding a self/nonself overlap. One can explain this intuitively as the following: if an MHC molecule is very specific, it presents a small set of self peptides. For every presented nonself peptide, the chance of having an overlap with self would then become smaller. Therefore, there is a strong correlation between binding specificity and self/nonself overlaps (see Figure S2). Furthermore, we tested the robustness of our results for various methods of peptide binding predictions, measures of amino acid similarity, and assumptions on T-cell recognition (summarized in Table S2). In all cases did degenerate T-cell recognition lead to a high self/nonself overlap of. Despite the high overlaps, our assumptions on the degenerate T-cell recognition can be considered conservative. For example, position 3 of the presented peptide tends to have few interactions with the TCR (see Figure 3) and our model should probably allow more mismatches at this position. Furthermore, many peptides with more than two substitutions at the middle positions (P3–8) have been shown to be cross-reactive [9]. If we assume that only a fraction of the self proteins provides a source of presented peptides, our estimates on self/nonself overlap decrease proportionally (see Table 1 and Table S2). Cole et al. [21] recently showed that in some cases, the anchor residues are involved in T-cell recognition. This observation might be more of an exception rather than the general mode of T-cell recognition, as in most cases T-cell recognition has been described to be less specific and not influenced by the anchor residues [9], [13], [14], [19], [29]. Recent estimates on T-cell crossreactivity confirm that our model remains conservative. Ishizuka et al. tested the T-cell recognition of 30. 000 unrelated MHC-I presented peptides using human and Murine T-cell clones, and found a single cross-reactive response, which suggested a cross-reactivity level of (1/30000) [32]. Typical T-cell precursor frequencies in a mouse are 1/100000 [33]–[35], i. e. on average 1 in a 100. 000 T-cells are expected to recognize a particular pMHC, and 1 in a 100. 000 pMHCs are expected to be recognized by a single T-cell clone. In other words, precursor frequency and cross-reactivity are similar concepts reflecting the specificity of a T-cell [36]. In our degenerate T-cell recognition model, single T-cells recognize only one in 2. 7 million () pMHCs (see Methods). Since this is much more specific than the experimental estimates, we think that our degenerate self/nonself overlap of about one-third is conservative and underestimates the actual overlap. Although these estimates on cross-reactive overlaps remain relatively crude, our results show that the degenerate recognition of MHC-I presented peptides by T-cells has a profound effect on self/nonself discrimination. This reconfirms that deletion of self reactive T-cells is important, as many of them would be activated during an infection and induce an autoimmune response. As a consequence, we estimate that about a third () of the foreign pMHCs is expected not to trigger an immune response. To test this prediction, the self/nonself overlap of HIV-1 derived peptides presented on HLA-A*0201 was studied to see if our model can account for the observed poor immunogenicity of these peptides. The presentation of, and T-cell responses to, HIV-1 derived peptides presented on HLA-A*0201 has been the subject of extensive investigations. Because it is such an intensively studied system, the lack of a reported T-cell response for one of the predicted pMHCs can be used as a reasonable indication for the lack of immunogenicity of that pMHC [9]. One explanation for the lack of immunogenicity is an overlap of the epitope with a self pMHC, and hence the self tolerance of the corresponding T-cell clone. We tested this by comparing overlaps of immunogenic and non-immunogenic HIV-1 pMHCs with self (see Methods). Only 4 of the 33 immunogenic pMHC (12%) were found to overlap with self according to our degenerate T-cell recognition model using the PMBEC similarity matrix. A significantly higher fraction of non-immunogenic pMHC, i. e. 18 of 54 (33%), overlapped with self (Chi-square test: p = 0. 027) (see Table 2), which is comparable to the overlaps reported by Frankild, using a different model for self-similarity but the same pMHCs [9]. We extended the analysis of self/nonself overlaps to vaccinia-derived peptides presented in HLA-A*02-transgenic mice for which Assarsson et al. [10] have determined the immunogenicity (see Methods). The overlap between (murine) self and immunogenic peptides is again lower than the self overlap of non-immunogenic peptides, although not significant due to the small number of data points (see Table 2). These results are also valid for other HLA molecules: using data provided by Perez et al. [37] on non-HLA-A*0201 presented HIV-1 peptides we found the same trend, that immunogenic peptides have less self/nonself overlaps than their non-immunogenic counterparts (see Table 2, and Methods). Finally, we analyzed immunogenic/non-immunogenic pMHCs derived from the IEDB [38] that were presented on the same HLA molecule (see Methods for selection criteria). The number of immunogenic and non-immunogenic pMHCs was large enough only for HLA-A*0201, and therefore the self/nonself overlaps of these sets were compared. Again, we found significantly less self overlaps among immunogenic peptides than non-immunogenic ones (Chi-square test: ; see Table 2). These results on the HLA-A*0201 presented HIV-1 and IEDB peptides are robust to the model assumptions: In all alternative overlap models described in Table S2, the number of overlaps with self was smaller for immunogenic pMHCs than for non-immunogenic pMHCs. This difference was always significant for the large set of IEDB peptides, for the smaller set of HIV-1 peptides a significant difference was not always observed (data not shown). Thus, in various data sets and model assumptions we find a correlation between pMHCs being immunogenic and their overlap with self, but these correlations only become significant for HLA-A*0201 where there is enough data. Summarizing, high self/nonself overlaps can explain the observed large “holes” in the T-cell repertoire [8], [11], and play an important role in determining the immunogenicity of foreign pMHCs. Previously, we have shown that the few epitopes sampled from a pathogens proteome are likely to be unique and are not expected to be present in the host (human) proteome [12]. Here, we extend this study by investigating a much larger set of nonself proteomes and a larger set of common HLA molecules. From this analysis we conclude that the pMHC of all common HLA-A and HLA-B molecules carry enough information for self/nonself discrimination, as a small minority (0. 1% to 0. 3%) of nonself derived peptides is expected to be identical to presented self-peptides. However, if the degenerate T-cell recognition of pMHCs is taken into account, the results change drastically. The cross-reactive recognition by T-cells results in a much higher self/nonself overlap of that is robust to various assumptions on degenerate T-cell recognition (see Table S2), i. e. in the “eyes” of a T-cell, about a third of the epitopes is expected to be similar to a self peptide presented on the same MHC-I molecule. Such a large overlap is expected to have a strong effect on the immunogenicity of pathogen-derived epitopes. One might intuitively think that the high self/nonself overlap estimates are in disagreement with the exquisite specificity of T-cell recognition. However, in our “degenerate” model of the middle positions (P3–8) with maximally 2 conservative mismatches, an individual T-cell recognizes only one in 2. 7 million pMHCs. This level of specificity is much higher than experimental measurements of about one in 100. 000 [32]–[35]. Therefore, we think that our current self/nonself overlap estimates are conservative. Could longer peptides be a solution for the high self/nonself overlaps caused by degenerate T-cell recognition? Given that T-cells cannot use all the information that is present in an MHC-I presented 9mer, we do not expect that the presentation of longer peptides would make much difference. Even though a longer peptide would contain more information, if that is not detected by the T-cells it would not improve self/nonself discrimination. Alternatively, MHC binding could be more specific at for instance position 1, thus preserving self/nonself information as now happens at the anchor positions. The disadvantage of more specific binding motifs would be the reduced presentation of foreign peptides and more opportunities for a virus to escape MHC presentation. Another consequence of a high self/nonself overlap could be high risk of autoimmunity. The identification of self antigens targeted in autoimmune diseases remains an enormous challenge, and our method of identifying overlapping peptides could possible help to narrow the search for these auto antigens. This requires a thorough understanding of the pathogens that might trigger a particular autoimmune disease and the corresponding HLA risk factors. Unfortunately, only for few autoimmune diseases sufficient data is available to extract such associations. For instance, Epstein Barr virus and HLA-B*4402 are associated with multiple sclerosis [39], [40], and HTLV-1 and HLA-B*5401 are associated with HAM/TSP [41]. We are currently searching the overlaps between the presented peptides of these viruses and the human self peptides presented on these HLA molecules for potential CTL targets in these autoimmune diseases (work in progress). The predicted self/nonself overlap varies between HLA molecules (see Figure 2), and two factors explain most of this variation. First, some HLA molecules have a preference for peptides derived from organisms with a low G+C content [1], which seems to be a universal signature for pathogenicity [42]. HLA molecules with such a preference for presenting nonself (e. g. HLA-A*2301) have a lower self/nonself overlap than other HLA molecules, because they present peptides that are less likely to occur in the human proteome. Second, the usage of additional (auxiliary or atypical) anchors at positions that also interact with the TCR increases the chance that presented peptides overlap according to our model. For example, HLA-B*0801 with atypical anchors at the third and fifth position will present more peptides that overlap at position three and five, and has the highest estimated self/nonself overlap (see Figure 2C). Indeed, a strong correlation between the use of additional anchors (see Methods) and self/nonself overlaps is found (Spearman Rank test: correlation = 0. 88, , not shown). Possibly, peptides presented on HLA-B*0801 have more specific TCR-interactions at the conventional anchor positions (P2 and P9) than in our T-cell recognition model, leading to an overestimate of the self/nonself overlap for this HLA molecule and others with atypical anchors. If the degenerate self/nonself overlap is not based on the middle positions of the presented peptide (P3–8), but on an HLA molecule specific choice of the six least specific positions (see Methods), the overlaps are however very comparable to an overlap based on the middle positions (see Table S2). Our estimates on self/nonself overlaps can explain why MHC-I restricted cellular immune responses to a pathogen are more narrow than the (predicted) number of pMHCs for that organism [9], [10]. We show that about one-third of the nonself pMHC should not elicit T-cell responses because they overlap with a self pMHC, i. e. this explains the large “holes” found in the T-cell repertoire [8], [9], [11]. We validated this prediction by comparing the overlaps of immunogenic and non-immunogenic pMHC from HIV-1, vaccinia or the IEDB, and showed that the number of self overlaps is significantly higher for non-immunogenic pMHC than for immunogenic pMHC. Still, a fraction of the immunogenic pMHCs were predicted to be overlapping with self, possibly because not all self-proteins induce tolerance or because regulatory processes are overridden during some viral infections causing autoimmunity [43]. In addition, an improved understanding of the rules of T-cell recognition could result in an even better distinction between overlapping/non-overlapping, and non-immunogenic/immunogenic pMHCs. This would be important in vaccine design and the understanding of immunogenicity in cellular immune responses. Human, Murine, viral and bacterial proteomes were downloaded via http: //www. ebi. ac. uk, the human proteome in May 2008, bacterial and viral proteomes in October 2008 and the Mouse proteome in January 2011. Only human and mouse proteins that have been shown at the protein or transcript level were included in the “self” data set. Redundant bacterial proteomes were removed by selecting only one strain per species, which resulted in 174 species of bacteria. 1000 non-redundant viral proteomes were selected with a maximum similarity of 80%. The similarity between viruses was determined as the number of exact matches in an all-to-all alignment of proteome sequences using BLASTP 2. 2. 18 relative to the smallest virus. Human viruses were selected based on the reported host information in the downloaded proteome, or on the term ‘human’ in their species name (e. g. Human Immunodeficiency Virus). A list of all bacteria and viruses used in this study is available upon request. The peptides presented on a certain MHC-I molecule can be predicted by simulating three key-processes of MHC-I presentation, i. e. proteasomal cleavage, TAP transport and peptide-MHC-I binding. The combination of proteasomal cleavage and TAP-transport determines which peptides reach the ER to potentially bind MHC-I. This process was predicted using NetChop Cterm3. 0 [22], [23]. Peptide-MHC-I binding was predicted using NetMHC-3. 2, an improved version of NetMHC-3. 0, that was shown to perform best in a large benchmark study of Peters et al. [24], [25]. The fraction of nonself peptides that overlap with a self peptide presented on an MHC-I molecule depends on the number of self peptides that is predicted to bind to this MHC-I molecule. Because we want to compare the self/nonself overlap of different MHC-I molecules, we have chosen to exclude the variance in the number of presented self peptides by using scaled thresholds, i. e. , the number of self peptides predicted to bind to each MHC molecules is scaled to be similar. Unfortunately, this procedure will eliminate the variation as a result of possible differences in specificity among MHC molecules. For each MHC molecule the threshold was set such that the presented fraction of self was similar to that on HLA-A*0201 with a 500 nM threshold (2. 3%) [44], [45]. This results in on average 250. 492 self pMHCs, 3. 750. 428 bacterial and 196. 265 viral pMHCs, per HLA molecule. Alternatively, we repeated the analysis with a fixed threshold of 500 nM (see Figure 2 and Table S2). In order to exclude HLA molecules with too similar binding motifs from our analysis, we selected the most frequent HLA molecule available in NetMHC-3. 2 at two digit resolution. This resulted in a set of 13 HLA-A and 15 HLA-B molecules. All results were checked for consistency with two other MHC-I binding prediction methods, NetMHCpan-2 [46] and a Stabilized Matrix Method (SMM) -based MHC-binding prediction tool [47], for HLA-A*0101, HLA-A*0201, HLA-A*0301, HLA-B*0702, HLA-B*0801 and HLA-B*3501. Note that for the HLA molecules that we have included in our analysis the average AUC for NetMHC and NetMHCpan predictions is 0. 809 and 0. 812, respectively [48]. As expected, similar results were obtained with NetMHCpan, but also when using SMMs (Table S2). Per MHC-I molecule, the set of presented 9 mers derived from viral or bacterial (nonself) proteomes and that from the human (self) proteome were compared to see how much these sets overlap. In the self/nonself overlap determination for vaccinia-derived pMHC from Assarsson et al. [10], the Mouse proteome was used as self. Overlaps were determined in different ways. First a “complete overlap” was determined as the exact match of all positions of the 9mer (positions 1–9, as in Figure 2A). Second, a “middle positions 6mer overlap” was defined as an exact match of the amino acids at positions 3–8 (as in Figure 2B). Third, the “non-anchor 7mer overlap” was determined as the exact match of the amino acids at position 1 and 3–8 (as in Figure S1). Finally, a “degenerate overlap” was determined by allowing two amino acid mismatches. Amino acid mismatches were not allowed at the most specifically recognized position 5. Moreover, we reasoned that two amino acid substitutions close-by would be more likely to abolish T-cell recognition. Therefore, only a single mismatch was allowed at the positions N-terminal from position 5 (P1–P4) and at the positions C-terminal (P6–P9) from position 5. Finally, only mismatches between amino acids with similar peptide-protein interaction properties were allowed. Following Kim et al. , amino acids were considered similar if their absolute covariance was greater than 0. 05 in the PMBEC matrix [31]. The PMBEC matrix is based on measured binding affinities between peptides libraries and MHC-I molecules, and was shown to capture similarity features common to substitution matrices such as BLOSUM50, and outperform other matrices when used as a Bayesian prior in MHC-I binding predictor training [31]. Furthermore, repeating our analysis using a positive score in the BLOSUM62 or BLOSUM50 matrix to identify allowed mismatches, similar results were found (Table S2). The self/nonself overlap is the chance a nonself pMHC overlaps with self, and was calculated by dividing the total number of overlaps in all nonself proteomes by the total number of pMHCs in all nonself proteomes. The self/nonself overlap was determined for bacteria and viruses separately, and the average of these two self/nonself overlaps is presented throughout the paper. Additionally, self/nonself overlaps were estimated using the “peptide similarity score”-method described in detail by Frankild et al. [9]. In this method the similarity between two peptides is determined using the BLOSUM35 amino acid substitution matrix and all positions of the compared peptides. The similarity score is subsequently scaled to the minimal and maximal similarity scores for the reference peptide, in order to normalize for the intrinsic similarity that a certain peptide has to all other peptides. If for instance the BLOSUM35 similarity score between peptide A and peptide B is 3, and the minimum and maximum possible similarities for any peptide with peptide A are 1 and 11, respectively, the peptide similarity score is (see [9] for a full description of the method). Frankild et al. showed that a self similarity score of 0. 85 tends to separate too self-similar, and hence non-immunogenic, from immunogenic HIV-epitopes [9]. This analysis and an analysis of cross-reactive peptides from literature was used for verification of this method [9]. We used the same threshold when determining overlaps with this “peptide similarity score”-method, i. e. nonself peptides with a similarity score exceeding 0. 85 with a self peptide are considered as overlapping. The cross-reactivity in our degenerate overlap model of T-cell recognition (described above) was determined in order to compare it with experimentally determined levels. For every possible 9mer peptide, the number of variants at the T-cell recognized middle positions (P3–8) was determined that would be recognized by the same T-cell in our degenerate overlap model. In other words, for every combination of amino acids at P3–8 we performed an exhaustive search to determine how many other combinations would also be recognized. On average, 24 of such combinations were found. Thus, given the number of possible variants at positions P3–8 (), the cross-reactivity in our model is), which is 1 in 2. 7 million or. Four sets of pMHCs were obtained for which the immunogenicity had been determined previously. The first set of HIV-1 derived peptides presented on HLA-A02 was determined by Frankild et al. [9], who predicted which HIV-1 peptides were presented on HLA-A02 and then defined the ones as immunogenic if there was at least one report of a T-cell response in a patient in the Los Alamos Database. Because HIV-1 responses for the most frequent HLA-A*02 molecule are studied extensively, we defined all other peptides as non-immunogenic. Thus, 33 immunogenic and 54 non-immunogenic HIV-1 derived peptides were defined using this strategy. The second set is derived from Assarsson et al. [10], who tested the immunogenicity of vaccinia derived peptides in a humanized mouse-system expressing HLA-A*02. We classified the 9 mers shown to be naturally processed and immunogenic (termed “Dominant” and “Subdominant”) as immunogenic peptides, and non-immunogenic peptides (termed “Negative”) were classified as such. This resulted in the selection of 18 immunogenic and 26 non-immunogenic vaccinia derived peptides. The third data set is derived from Perez et al [37], who measured the T-cell response in HIV-1 patients to a set of HIV-1 peptides. The patients were HLA class I genotyped [37]. We only considered responses to 9mer peptides with a predicted binding affinity of less than 500 nM, to only one of the patients HLA-A and HLA-B molecules. Binding predictions were done with NetMHCpan-2 [46]. The virus in every patient was sequenced by Perez et al. [37], and we excluded all T-cell responses in which the peptide that was used for testing the T-cell response was not encoded by the viral genome. Only peptides presented on HLA molecules other than HLA-A*0201 were selected since HLA-A*0201 presented HIV-1 peptides were already compared in the data set derived from Frankild et al [9]. Peptide-HLA combinations with only negative T-cell responses measured by Perez et al. were classified as non-immunogenic (n = 13), all other peptide-HLA combinations were classified as immunogenic (n = 9). The fourth data set was derived from the IEDB [38], by downloading all entries that describe a T-cell response assay to a 9mer peptide presented on one of the HLA molecules in our test set, performed in a human subject upon infection. Only peptide-HLA combinations in which the predicted binding affinity was less than 500 nM were considered. Furthermore, we required that the assayed T-cells were not re-stimulated in vitro, and that the peptide was used in the T-cell response assay. Peptide-HLA combinations were classified as immunogenic if a “Positive (-High) ” or “Positive-Low” T-cell response was measured, and classified as non-immunogenic if the T-cell response was always reported to be “negative”. We were able to classify more than 20 immunogenic and 20 non-immunogenic peptides only for HLA-A*0201 (i. e. 197 immunogenic and 592 non-immunogenic peptides). For all HLA molecules, we predicted the binding of 1. 000. 000 random peptides with equal amino acid frequencies using NetMHC-3. 2 and the thresholds described above. The Shannon entropy was determined per position on the predicted binders, per HLA molecule, and used as a measure of selectivity. Based on this selectivity, the six least specific positions were determined for each HLA molecule to use in the “allele specific” analysis of degenerate self/nonself overlaps (Table S2). Additional anchor selectivity was calculated as the sum of the entropy at the non-anchor positions (P1 and P3–8), per HLA molecule. An HLA molecule was defined to have additional anchors if the additional anchor selectivity was larger than 25% of the sum of entropy at all positions (P1–9) for an HLA molecule. Structures of HLA-I-9mer-TCR-complexes were downloaded in August 2011 from the PDB-database (www. pdb. org [30]). After redundancy reduction we selected nine structures for further analysis: 1AO7,1BD2,1LP9,1MI5,2ESV, 3GSN, 3KPR, 3O4L and 2F53 [49]–[57]. The selected structures consist of HLA-A*02 (n = 6), HLA-B*08, HLA-B*44 and HLA-E molecules. Per peptide position the number of TCR contacts was determined as the number of TCR amino acids within a 5. 0 Å distance. For each structure, we determined per peptide position the fraction of TCR contacts relative to all peptide-TCR contacts in that structure. Boxplots of these fractions are shown in Figure 3. Statistical tests were performed using the stats-package from the scipy-module in Python. A Permutation test was also done in Python, using the shuffle function in the random-package from the numpy-module, to identify human proteins that have more than expected peptides that overlap with viruses or bacteria. The permutation test was performed as follows: per human protein, we counted the number of viruses or bacteria that overlap with a 9mer peptide in this protein. These counts were normalized by the length of the protein, i. e. the number of overlapping viruses or bacteria was divided by the protein length. In 1000 permutations, per human protein a number of overlapping viruses or bacteria was drawn based on the expected fraction of overlaps and given the protein length. If the actual number of overlaps was higher than the number in all 1000 permutations, the human protein was selected as a protein with a significantly high number of viral or bacterial overlaps. A similar analysis was performed to identify proteins with more than expected HLA-B*5401 ligands. First, per protein the number of HLA-B*5401 binding peptides was predicted as described above. Next, this prediction was compared in 1000 permutations where a number of binding peptides was drawn based on the specificity of HLA-B*5401 (i. e. 2. 3% as described above). If the actual number of binding peptides was higher than the number in all 1000 permutations, the protein was selected as a protein with a significantly high number HLA-B*5401 ligands.

Human cells sample short peptides from endogenous proteins, and present them to the immune system via HLA class I molecules on the cell surface. T-cells scan the presented peptides and need to discriminate foreign (nonself) peptides from human (self) peptides. We show that this is a difficult task, despite the exquisite specificity of T-cells. We estimate, using HLA-peptide binding predictions and T-cell recognition models, that almost a third of the nonself peptide-HLA complexes is so similar to a self peptide-HLA that a T-cell cannot tell them apart. Since T-cells have to ignore self peptides to prevent autoimmunity, we estimate that at least a third of the foreign peptides has to be ignored as well, and therefore fails to evoke an immune response. Foreign peptides that are never used in immune responses, have been referred to as the “holes in the repertoire”. Since the sizes of the holes we predict agree with those that were previously found, our conjecture is that the holes are entirely due to similarity with self peptides. We test this conjecture with public data on HIV-1 and vaccinia responses, and confirm that self similarity is a major determinant of the immune response to nonself peptides.

Abstract Introduction Results Discussion Methods

medicine vaccination infectious diseases immune cells clinical immunology immunity immunology biology computational biology microbiology

2012

Degenerate T-cell Recognition of Peptides on MHC Molecules Creates Large Holes in the T-cell Repertoire

Cereal endosperm represents 60% of the calories consumed by human beings worldwide. In addition, cereals also serve as the primary feedstock for livestock. However, the regulatory mechanism of cereal endosperm and seed development is largely unknown. Polycomb complex has been shown to play a key role in the regulation of endosperm development in Arabidopsis, but its role in cereal endosperm development remains obscure. Additionally, the enzyme activities of the polycomb complexes have not been demonstrated in plants. Here we purified the rice OsFIE2-polycomb complex using tandem affinity purification and demonstrated its specific H3 methyltransferase activity. We found that the OsFIE2 gene product was responsible for H3K27me3 production specifically in vivo. Genetic studies showed that a reduction of OsFIE2 expression led to smaller seeds, partially filled seeds, and partial loss of seed dormancy. Gene expression and proteomics analyses found that the starch synthesis rate limiting step enzyme and multiple storage proteins are down-regulated in OsFIE2 reduction lines. Genome wide ChIP–Seq data analysis shows that H3K27me3 is associated with many genes in the young seeds. The H3K27me3 modification and gene expression in a key helix-loop-helix transcription factor is shown to be regulated by OsFIE2. Our results suggest that OsFIE2-polycomb complex positively regulates rice endosperm development and grain filling via a mechanism highly different from that in Arabidopsis. Rice (Oryza sativa) serves as the staple food for over half of the world' s population. The main body of rice grain is endosperm, which is consumed as food. Endosperm stores energy primarily in the form of starch, storage proteins, and lipids. As in other angiosperms, rice seed development requires initiation signals from double fertilization. During double fertilization, one sperm cell fuses with the egg cell to develop into the diploid embryo while the other sperm cell fuses with the central cell of the female gametophyte to develop into the triploid endosperm. Before fertilization, premature divisions of the egg and the central cells are suppressed. Substantial progress has been made in understanding the role of the polycomb group (PcG) genes in the repression of central cell division before fertilization in Arabidopsis. Polycomb group (PcG) genes, first discovered in Drosophila melanogaster, play an important role in maintaining the repressed state of homeotic (HOX) genes by posttranslational modifications of histones [1]. PcG genes also control a large number of genes regulating many cellular functions and developmental pathways, such as cell proliferation, stem cell maintenance, imprinting and cancer [2], [3]. PcG proteins form three principle types of multiprotein complexes, Polycomb Repressive Complex 1 (PRC1), Polycomb Repressive Complex 2 (PRC2) and Pho RC. PcG complexes have been identified in different organisms, including Drosophila melanogaster and Arabidopsis thaliana [3]–[9]. In Drosophila melanogaster, PRC2 complex consists of E (Z), ESC, Su (Z) 12 and p55 [10]. The identification and characterization of PRC2 complex genes in plants indicated remarkable structural and functional conservation of the PRC2 complexes between plants and animals. Genetic and molecular studies suggested the presence of three PRC2-like complexes in Arabidopsis thaliana: the Fertilization Independent Seed (FIS), Embryonic Flower (EMF) and Vernalization (VRN) complexes [3], [8], [9], [11]–[17]. The EMF complex, which probably contains CLF/SWN, EMF2, FIE and MSI1, promotes vegetative development by repressing the transcription of flowering activators such as Flowering Locus T (FT) and Agamous-like 19 (AGL19) [18]–[20]. The VRN complex is involved in cold-induced epigenetic silencing of Arabidopsis Flowering locus C (FLC). The VRN complex is purified biochemically by tandem affinity purification method [3]. It includes VRN2, SWINGER (E (Z) homolog), FIE (ESC homolog), MSI1 (p55 homolog), VRN5, VIN3, and VEL1 [3], [4], [13], [14], [21], [22]. The VRN complex associates with PHD-finger proteins for better H3K27me3 deposition. Mathematical modeling showed that a polycomb-based switch underlying quantitative epigenetic memory [23]. Mutations of FIS complex genes (MEA, FIS2, FIE and MSI1) result in the formation of multinucleate central cell that develops to the point of cellularization in the absence of double fertilization. The partially developed seed like structures eventually atrophy. Multinucleate central cell that develops to an endosperm was capable of nourishing an embryo [24]. Mutations in Arabidopsis FIS genes also show an interesting phenotype after fertilization, including endosperm over-proliferation, embryo arrest and abortion. In FIS mutants, endosperm nuclei continue to divide even after the wild type endosperm stopped to replicate and develops large balloon like seeds with delayed cellularization in the endosperm. Gradually, the endosperm collapses and the seed aborts. It is not clear whether the endosperm and embryo phenotypes are both directly due to the FIS gene mutation or whether one is the primary defect and the other is a downstream event [25]. Mutations in Arabidopsis FIS genes also show parent-of-origin effect on seed development [5]–[8], [26]. Every seed that inherits a maternal mutant FIS allele aborts regardless of the presence of a wild type paternal allele [5], [27]–[30]. MEADEA (FIS1) and FIS2 are imprinted genes and expressed maternally throughout endosperm development in Arabidopsis [5], [31]–[33]. FIE (FIS3) maternal allele is expressed in the early endosperm [12], [16]. A recent study using the FIE mutant showed that the PRC2 complex is essential for the transition from embryonic phase to the seedling stage [34]. The FIE mutant show delayed germination. 40% of the homozygous FIE mutants failed to germinate after 20 days while the wildtype germinated within 2 days. In addition, polycomb group proteins are required to couple seed coat initiation to fertilization [35]. Cereal crops, such as Maize, barley and Rice, have multiple homologs of PRC2 core complex genes. Two homologs of ESC are identified in maize, ZmFie1 and ZmFie2. The ZmFie1 is expressed only in endosperm and imprinted during endosperm development, whereas ZmFie2 is expressed in the egg cell and more intensively in the central cell [36]–[38]. Maize has three E (Z) homologs: MEZ1, MEZ2, and MEZ3 [39], [40]. Four polycomb gene homologs are identified in barley, HvFIE, HvE (Z), HvSu (Z) 12a and HvSu (Z) 12b [41]. Genome wide analysis studies indicated the presence of multiple homologs of ESC, E (Z), and Su (Z) 12 in Rice genome [42]. Rice has two ESC homologs (OsFIE1 (Os08g04290) and OsFIE2 (Os08g04270) ), two E (Z) homologs (OsSET1 (Os03g19480) and OsCLF (Os06g16390) ) and two Su (Z) 12 homologs (OsEMF2a (Os04g08034) and OsEMF2b (Os09g13630) ) [42]–[45]. Interestingly, not all the FIS complex functions identified in Arabidopsis are conserved in other plants [42], [46]. Both imprinted genes MEA and FIS2 are involved in the central cell repression and endosperm development in Arabidopsis. But rice, maize and barley genomes do not have MEA and FIS2 gene orthologs [42], [46], [47]. In addition, all the reported rice polycomb genes are widely expressed in different tissues including endosperm except that OsFIE1 is imprinted and the maternal allele is expressed specifically in endosperm [42], [45]. Mutations in Arabidopsis FIS genes show autonomous endosperm development whereas T-DNA insertion line in OsFIE1 gene did not result in autonomous central cell proliferation in rice [42], suggesting that even though OsFIE1 is imprinted in rice endosperm, it is not involved in the repression of central cell proliferation. Knockdown of ZmFIE1 and ZmFIE2 genes in maize also produced no autonomous central cell proliferation [46]. In sexual Hieracium pilosella, RNAi lines of HFIE failed to show the autonomous central cell proliferation [48]. But down-regulation of HFIE in sexual Hieracium resulted in seed abortion after fertilization. Similarly, specific HFIE down regulation in the apomictic Hieracium resulted in embryo abortion and defective endosperm, indicating that HFIE gene is important for the development of viable seed in both sexual and apomictic Hieracium [48]. The presence of homologs of Arabidopsis polycomb genes in rice, maize and barley genomes indicates the conservation of polycomb genes between monocot and dicot plants. Whereas, the lack of some important seed specific FIS genes (MEA and FIS2) in rice, maize and barley genomes suggests that there might be different functional regulatory mechanisms involved in the embryo and endosperm development in monocots. Despite the extensive genetic studies on polycomb complex genes during the seed development in plants, biochemical characterization of the polycomb complex is still very limited except that the VRN complex has been purified via tandem affinity purification in Arabidopsis. The enzyme activities of the PcG complex have not been demonstrated in vitro for plants. To elucidate the biological function of the polycomb complex and its molecular mechanism in the regulation of gene expression, we purified the polycomb complex, identified the components of the complex via mass spectrometry analysis, and detected the enzyme activity of the complex by in vitro assay. In addition, we generated RNAi and over expression lines of OsFIE2. The phenotype of the mutant lines clearly demonstrated an essential role of the complex in the regulation of endosperm development, grain filling and seed dormancy. The regulatory mechanisms are different from those in Arabidopsis. Transcription and proteomic analysis of selected nutrient metabolic pathway genes showed that many of them are subject to OsFIE2 regulation, directly or indirectly. Finally, the H3K27me3 binding sites in young endosperm were identified using ChIP-Seq approach. Our results suggest that the polycomb group genes control seed development and grain filling by regulating a large number of genes. The FIE proteins belong to a family of WD-repeat proteins that promote protein-protein interaction in various multiprotein complexes. They are necessary for the viable seed formation in both Arabidopsis and Hieracium. To identify the proteins associated with rice OsFIE2, the tandem affinity purification (TAP) approach was used [3], [49]–[57]. The TAP tag used in this study contains protein A, the TEV recognition site and the calmodulin-binding protein (CBP) domain [49]–[51]. The final N-Terminal TAP-Tagged transgenic rice plants and cell suspension cultures expressing the TAP-Tagged OsFIE2 were generated. The expression of TAP-tagged OsFIE2 protein was verified by Western blots in multiple transgenic lines (Figure 1A-i). A cell culture system established from the most highly expressed transgenic line 04A was selected for future studies. Different steps invloved in the TAP purification method are depicted in Figure 1B. In the first step of the tandem purification, protein A domain of OsFIE2 TAP tag was bound to the IgG sepharose beads. The TAP tagged proteins enrichment to the IgG beads were revealed by Western blot analysis (Figure 1A-ii). The IgG bound proteins were eluted by digestion with AcTEV protease. The AcTEV cut eluate was incubated with CAM agarose beads. The protein enrichment by the CAM beads was detected by Western blot analysis (Figure 1A-iii). Finally, OsFIE2-TAP complex proteins were eluted with buffer containing EGTA. The eluate was TCA precipitated and processed for mass spectrometry analyses after trypsin digestion. The complex proteins were identified using LC-MS/MS mass spectrometry analyses with the protein identification criteria of minimal two peptides as we reported previously [58]–[60]. The OsFIE2 complex includes OsFIE2 (homolog of ESC), OsCLF (homolog of E (Z) ), OsSET1 (homolog of E (Z) ), and OsEMF2b (homolog of Su (Z) 12) (Table 1). The histone H4 protein was identified in this study (Table 1). It is reported that Drosophila histone H4 strongly binds to polycomb protein p55 [61]. The RbAp48/46, mammalian polycomb protein, also binds to the histone H4 [62], [63]. Whether H4 is a stable component of the rice OsFIE2 complex remains to be further verified although it is possible. Two other proteins identified in our purification were elongation factor 1-alpha and heat shock cognate 70 kDa protein 2. These two proteins had been identified in an affinity purification using the same TAP system in rice and were considered as nonspecific interaction [53]. We also identified the heat shock proteins during the purification of another protein using the same TAP system (Mujahid and Peng, unpublished results). Therefore, we concluded that these two proteins were not the interactive proteins of the OsFIE2-TAP complex. To understand the structure of the OsFIE2-PRC2 complex, we carried E. coli two-hybrid assays for pairwise interactions of the four polycomb subunits. E. coli two-hybrid assay has some advantages over yeast two-hybrid such as E. coli grows much faster than yeast and it can be transformed with higher efficiency so larger number of interactions can be rapidly screened. In addition, E. coli two-hybrid assay reduces the chance that the host harbors a eukaryotic homologue of one of the interacting protein partners [64]. Our results showed that OsFIE2 interacted with OsSET1 strongly (Figure 1A-iv). However, no other interactions were detected using the E. coli two-hybrid system. The results support that OsSET1 is a component of the OsFIE2 complex and suggest that post translational modification (s), other cellular component (s) such as long non-coding RNAs, or the nuclear environment is required to reveal other protein-protein interactions within the OsFIE2 complex. Mammalian and Drosophila PRC2 complexes are shown to methylate H3K27 (histone H3 at lysine 27) and, to a lesser extent, H3K9 both in vivo and in vitro [65]–[68]. The substrates include H3K27, H3K27me, H3K27me2, and H3K9me. Although VRN complexes have been purified in Arabidopsis, no enzyme activities are reported in vitro. We carried out in vitro histone methyltransferase assay using the TAP purified OsFIE2 protein complex as enzyme, chicken core histones as substrate, and S-adenosyl-[methyl-3H]-l-methionine as methyl donor. The TAP purified OsFIE2-PRC2 complex demonstrated strong methyltransferase activity against histone H3 only (Figure 2A), indicating that our purified OsFIE2-PRC2 complex is a functional complex with H3 specific enzyme activity. To further illustrate the enzyme specificity of the OsFIE2-PRC2 complex, we examined the H3K27 modification state in the OsFIE2 overexpression line and the OsFIE2 RNAi lines. Our results showed that H3K27me3 was reduced to about 54% in the RNAi line NO3 (Figure 2b). Meanwhile, H3K27me3 level had increased to about 152% in the overexpression line 04A compared with the wild-type. In contrast, the content of H3K27me and H3K27me2 had no detectable change, suggesting that OsFIE2 PRC2 complex specifically catalyzed the conversion of H3K27me2 to H3K27me3. We also examined H3K9me2 and H3K9me3 levels in OsFIE2 overexpression and RNAi lines (Figure 2B). The results indicated that H3K9 methylation state was not affected by OsFIE2 overexpression or reduction, suggesting that H3K9me and H3K9me2 are not the primary substrates of the OsFIE2-PRC2 complex at genome level compared with H3K27me3. However, we can not rule out the possibility of the gene specific effect of OsFIE2 PRC on H3K9me and H3K9me2 in the rice genome. To further validate the results, we tested other RNAi and overexpression lines. Similar results and conclusions were obtained (Figure S1). To study the biological role of OsFIE2 in rice, we made OsFIE2-RNAi constructs and obtained twenty independent RNAi lines using Agrobacterium mediated transformation. Six independent RNAi transgenic lines (No3, No8, No9, No10, No13 and No15) with different OsFIE2 expression levels as revealed by quantitative real time PCR (Figure 3A) were selected for further analysis. No 8 and No 9 had less than 20% of the expression level of the wild-type. No 3 and No 10 had about 40% to 50% expression levels and No15 and No 13 had about 70% to 90% expression levels. We categorized Lines No8 and No9 as severe RNAi lines and the others as weak RNAi lines. The vegetative development of the weak RNAi plants was the same as the wildtype. Phenotype evaluation of seed development in OsFIE2-RNAi T1, the endosperm was at T2 generation, and generations following revealed substantial differences compared with the wild-type. The ratio of mature and fully filled seeds was substantially reduced in RNAi lines (Figure 3B and Table 2). Fully filled mature seeds were produced in a ratio ranging from 66% to 81% in the weak RNAi lines, while the wild-type was about 88%. Interestingly, per thousand seed weight of the fully filled seeds in the weak RNAi lines was smaller in all the tested lines (Figure 4E, 4F and Table 2). The rest of the seeds were partially filled or not filled (Figure 4B). For the partially filled seeds, we found that some of them lost dormancy with seed germination before the seed mature (Figure 4D). We carried out a series of sectioning to examine the seed development in both wildtype and the RNAi lines. Defective shrunken endosperm tissue was observed in the partially filled seeds of RNAi lines (Figure 4H, 4J and K). In addition, we observed big embryos in the partially filled seeds without clear germination, which might be due to partial germination or uncontrolled embryo development (Figure 4J and 4K). Overall, the embryo sizes in the partially filled seeds increased to different degrees. Interestingly, the Arabidopsis fie mutants showed substantial delay in germination in contrast to loss of dormancy [34]. The T0 plants of the two severe OsFIE2-RNAi lines (No8 and No9) did not produce viable seeds. We failed to obtain the reproductive progeny of the two transgenic lines although we successfully maintained the plants by asexually amplifying new tillers. To verify the phenotype observed in the severe lines, we carried out a second batch of transformation and generated another 18 independent lines. One line displayed severe phenotype but also failed to produce viable seeds. Since all the three severe RNAi lines failed to produce viable progeny, we could not verify if the observed phenotype is inheritable meiotically. Therefore, the phenotypes of the severe lines are not discussed further in this manuscript. The small seeds and partially filled seed phenotype were repeatable in the weak RNAi lines generated in the second batch (data not shown). To examine pollen quality, we carried out pollen viability test using iodine-potassium iodide staining method (Figure 3C). We found that the viable pollen ratio was about 87% in wild-type and about 86% in weak RNAi lines, suggesting that the pollen grains developed normally. Since autonomous endosperm development was reported in Arabidopsis FIS/FIE mutants, we emasculated anthers before anthesis in over 100 spikelets in the RNAi severe and weak lines, respectively. After examination with microscopy and the naked eye, no sign of autonomous endosperm development was observed in the absence of fertilization event. Meanwhile, we examined the expression of the OsFIE2 gene and other polycomb group genes in anthers with reverse transcription PCR to study if these genes subject to imprinting regulation as in Arabidopsis. The OsFIE2 gene was highly expressed while OsFIE1 gene expression was not detected in anthers (Figure 5A). We also examined the expression of other OsFIE2-PRC2 complex genes and OsFIE1 in other plant tissues, including leaf, sheath, stem nodes, young and mature flowers, panicles, and endosperm tissues in three different stages. All the OsFIE2 complex genes were highly expressed in these tested tissues except that OsCLF might have minor quantity changes in five days old endosperm tissues (Figure 5B). In contrast, OsFIE1 was only detected in five days and ten days old endosperm tissues. Luo et al. (2011) [45] carried out a genome wide analysis of imprinted genes in rice endosperm. They found that only OsFIE1 but not other polycomb genes were imprinted in rice endosperm. Given that OsFIE2 was highly expressed in anther but not OsFIE1, our results were consistent with their report (Luo et al. 2011) [45] that OsFIE1 was probably paternal imprinted but not OsFIE2. The dominant negative nature of RNAi line prevented us to further test if OsFIE2 is imprinted by reciprocal crosses. Surprisingly, the endosperm specific OsFIE1 mutant had no endosperm phenotype [42], but the endosperm unspecific gene OsFIE2 RNAi lines all had reduced seed size and partially filled phenotype. We compared OsFIE1 and OsFIE2 expression level in endosperm with real-time PCR (Figure 5C). The results showed that the OsFIE1 expression level was only about 4% of the expression level of OsFIE2 in endosperm. The OsFIE1 and OsFIE2 proteins share 72% of overall amino acid sequence identities. If the N-terminal 82 amino acid stretch specific for OsFIE1 is excluded, the two proteins share about 85% amino acid sequence identities. If OsFIE1 and OsFIE2 are functionally redundant, our results would imply that OsFIE1 level is too low to replace the function of OsFIE2 because a reduction of OsFIE2 expression to 20% could not produce viable progeny but the OsFIE1 expression is only 4% of the OsFIE2gene. In contrast, OsFIE2 level is high enough to replace OsFIE1 function. Therefore, mutation on OsFIE1 gene may not show phenotype, but mutation on the OsFIE2 gene will show phenotype. Due to the high homology between OsFIE2 and OsFIE1, we tested if the OsFIE1 expression was also reduced in the OsFIE2 RNAi lines. Our results showed that OsFIE1 expression was not affected in the OsFIE2 RNAi lines (Figure S3). We also generated multiple OsFIE2 overexpression transgenic lines. We found no obvious phenotype in all the transgenic plants and the endosperm and embryo development were normal as wild type (data not shown). Given that the mature seeds of the RNAi lines are much smaller than the wild-type and many seeds are only partially filled, we examined the expression levels of selected starch biosynthesis pathway genes in the endosperm using quantitative real-time PCR. The healthy looking developing seeds of weak RNAi lines in the grain filling stage were selected for the experiment. The genes tested included ADP-glucose pyrophosphorylase small subunit 1 (AGPS1) -AK073146, AGPS2a-AK071826, AGPS2b-AK103906, AGPL1-AK100910, AGPL2-AK071497 and AGPL3-AK069296. Among them, the expression of AGPS2b reduced to 9% compared with the wild type (Figure 6). Other genes had no substantial change. Our results indicated that the AGPS2b gene, encoding the rate limiting step enzyme in the starch biosynthesis pathway [69] is subjected to the regulation of OsFIE2 either directly or indirectly. During the seed filling stage, massive amounts of storage proteins are accumulated in cereal endosperm [70], [71]. To investigate the possible role of OsFIE2 in the regulation of storage protein accumulation, we examined the storage protein content in the endosperm of fully filled mature seeds in the weak RNAi lines NO 3, NO 10, and NO15 using two-dimensional (2-DE) gel analysis. We isolated total proteins from the mature seeds of both OsFIE2-RNAi and wild-type seeds using phenol extraction method as we reported [72]. After 2-DE gel separation and PDQUEST analysis, differentially regulated protein spots were robotically excised and analyzed using MALDI TOF-TOF mass spectrometry. Proteins with reduced accumulation in the RNAi seeds included prolamin, globulin, alpha-amylase and seed allergenic proteins (Figure 7 and Table 3). The results were similar in all three tested RNAi lines. Real-time PCR analysis demonstrated the expression of some genes also reduced at the mRNA level (Figure 6), suggesting a regulation of these genes at the transcriptional level. We also found that the expression of Glutelin A1 (Os01g55690) reduced in RNAi seeds compared with the wild type (Figure 6). Our studies above indicated that OsFIE2 played a critical role in regulating endosperm development and grain filling and the OsFIE2-PRC2 complex specifically catalyzes the methylation of H3K27me2 to H3K27me3. We hypothesized that some key regulatory genes controlling storage starch and protein synthesis in endosperm must be subjected to the regulation of H3K27me3 either directly or indirectly. We have carried out a genome wide ChIP-Seq experiment using 6 to 7 days post pollination endosperm tissues as materials to compare four algorithms used for ChIP-Seq data analysis [73]. The endosperm ChIP-Seq results were verified by ChIP-PCR using 28 selected genes [73] and the antibody used for the experiment (antibodies against H3K27me3, produced by Millipore) is the antibody with 100% specificity for H3K27me3 [74]. Given that the role of H3K27me3 is primary for gene repression and 6 to 7 days post pollination is in the stage before starch and storage protein synthesis, we expected that H3K27me3 should be still associated with genes regulating storage nutrient synthesis in this stage. Therefore, we decided to use the published data set to search for genes regulated by H3K27me3 modification in endosperm. Given that the antibody might bind to non specific sites in the mutant background due to reduced H3K27me3 [34], the results obtained in wild-type background should be the best one could get. Malone et al. (2011) [73] compared four Algorithms for ChIP-Seq data analysis and concluded that the overall biological conclusion of the results obtained by the four algorithms is the same. Since the FindPeak algorithm combines excellent sensitivity with accuracy compared with the other three Algorithms [73], the findPeak results are discussed here. The FindPeak program analyses found that 76 endosperm specific genes, including DNA binding proteins, and 97 nutrition metabolic pathway genes were enriched by antibodies against H3K27me3 in the endosperm ChIP-Seq experiment (Table S1). The total list of enriched peaks is shown in Table S2. The selected examples of the peak profiles are shown in Figure S2. Examples of enriched endosperm specific genes include: MYB protein, putative (LOC_Os02g09670); helix-loop-helix DNA-binding domain containing protein (LOC_Os04g35010); basic region leucine zipper domain containing protein (LOC_Os09g34880); RNA recognition motif containing protein (LOC_Os09g14550); trehalose-6-phosphate synthase (LOC_Os09g25890); OsIAA29-Auxin-responsive Aux/IAA gene family member (LOC_Os11g11430); etc. The ChIP enriched nutrition metabolic pathway genes include: seed specific protein Bn15D1B (LOC_Os04g50970); starch synthase (LOC_Os06g04200); globulin 2 (LOC_Os11g34780); glutelin (LOC_Os02g15090); seed maturation protein PM23 (LOC_Os03g41080); starch binding domain containing protein (LOC_Os05g37450); etc. Our ChIP-Seq results showed that the helix-loop-helix DNA-binding domain containing protein (LOC_Os04g35010) is enriched in the ChIP-Seq experiment for H3K27me3 modification (Figure S2 and Table S1). The gene' s Arabidopsis ortholog RGE1 (AT1G49770) is expressed in the endosperm surrounding region which directly surrounds the developing embryo. It exerts its effect non autonomously in the developing embryo [75], [76]. Loss of RGE1 function results in small embryos. Mutant seedlings are extremely sensitive to desiccation due to the abnormal cuticle. Since we observed large embryo phenotype in our RNAi lines, we checked the expression of the rice LOC_Os4g35010 gene in both wild-type and the RNAi lines in developing seeds. We found that this gene was over expressed in the RNAi lines compared with wild-type as shown in Figure 8. Further ChIP-PCR experiment using antibodies for H3K27me3 showed that the H3K27me3 modification was substantially enriched in this locus in the wild-type but the enrichment substantially reduced in the RNAi lines. Our results confirmed the ChIP-Seq result and suggest that this gene' s H3K27me3 modification level and expression are regulated by OsFIE2 expression. The role of polycomb group genes in plant development has been extensively studied in Arabidopsis. The polycomb complex that regulates vernalization process has been purified in Arabidopsis using Tandem affinity tag approach and the seven putative subunits of the complex have been identified following LC-MS/MS analysis [3], [14]. However, no enzyme activities had been tested in vitro for the PRC2 complexes in plants. Thus, the mechanism of PRC2 complex action has not been demonstrated in plants. Given the unique and essential function of PRC2 complexes in cereal crops, purifying and characterizing the complex in cereals is particularly important. In this report, we purified the OsFIE2-PRC2 protein complex in rice and demonstrated the Histone methyltransferase activity in vitro. We find that OsFIE2 (the ESC homologue) forms a stable protein complex with polycomb proteins OsCLF (the E (Z) homologue), OsSET1 (the E (Z) homologue), and OsEMF2b (the Su (Z) 12 homologue). Our E. coli two hybrid studies show that OsFIE2 may directly interact with OsSET1. The OsFIE2-PRC2 complex genes are well expressed in young and mature endosperm tissues in rice. Along with the PRC2 core components, the histone H4 protein was co-purified with the OsFIE2 protein. Interaction of histone H4 with polycomb proteins was reported in mammals and Drosophila. Mammalian ortholog of p55, RbAp 48/46, directly binds to histone H4 [62], [63]. GST pull-down assay revealed that Drosophila histone H4 strongly and specifically binds to p55 [61]. Therefore, H4 may directly interact with OsFIE2-PRC2 complex although further experiments are required to confirm the interaction. There are four p55 homologous genes in the rice genome [46]. But we did not detect the p55 homologous proteins in purified complex although we successfully detected methyltransferase activity with the purified complex. One possibility is that the quantity of p55 homologous protein is very low in the complex. The mass spectrometer failed to detect it. Alternatively, the polycomb complex in rice has components different from those in other organisms. Nevertheless, it is clear that the rice OsFIE2-PRC2 complex is conserved in overall structure compared with the Drosophila PRC2 complex. Our E. coli two hybrid studies show that OsFIE2 may directly interact with OsSET1. Other interactions were not detected in E. coli two hybrid assays among the components, suggesting that posttranslational modifications or additional components may be involved in complex formation. It is interesting to note that non coding RNAs play a key role in the polycomb complex [77], [78]. But it is still unknown if the RNAs play any critical structure role in the complex. Our expression studies show that all the OsFIE2-PRC2 complex genes are well expressed in both young and mature endosperm tissues of rice. In Mammals and Drosophila, PRC2 complexes were shown to methylate histone H3 at lysine 27 and, to a lesser extent, H3K9 both in vivo and in vitro [65]–[68]. In Arabidopsis vernalization mutant vrn2, encoding a subunit of PRC2 complex, both methylation marks at H3K9 and H3K27 were lost [22], [79]. In two other mutants vrn1 and vin3, the H3K9me2 mark is missing. Vin3 has also been shown to be a component of the PRC2 complex with VRN2 [3], suggesting a role of the PRC2 complex in regulating H3K9 methylation. Therefore, detecting the methyltransferase activity and revealing the substrate specificity of the OsFIE2-PRC2 complex is important for understanding the molecular mechanism of PRC2 complex function in plants. Our studies with the overexpression and RNAi lines demonstrated that manipulation of the OsFIE2 gene expression level had no effect on the cellular level of H3K9me2, H3K9me3, H3K27me and H3K27me2 but resulted in a substantial change in the cellular level of H3K27me3, suggesting that OsFIE2-PRC2 complex is primary for regulating the formation of H3K27me3 at genome level. However, we can not rule out the possibility of the gene specific effect of OsFIE2 PRC on other modifications in the rice genome. The results suggest that the function of the rice OsFIE2-PRC2 complex is different from the Arabidopsis VRN2-PRC2 complex, the later contained more subunits and may have H3K9 methyltransferase activity [3], [22], [79]. The OsFIE2-PRC2 complex is also different from the Arabidopsis endosperm development related MEA-FIE complex because the orthologs of the Arabidopsis MEDEA and FIS2 were not present in the rice genome. The polycomb gene FIE is well studied in Arabidopsis and is required to repress the endosperm development in the absence of fertilization [6], [7], [13], [15], [16]. Loss-of-function mutations of genes in the MEA-FIS- FIE-MSI1 complex can form autonomous diploid endosperm in the absence of fertilization [6], [7]. Mutation in MSI1 gene exhibits very strong phenotype of autonomous endosperm development [8], [26]. It is shown that MEDEA represses expression of Arabidopsis MADS-box gene PHERES1 during the seed development [8], [9]. Interestingly, MEA gene itself is one of the target genes of the FIS complex, indicating a self-imprinting mechanism in Arabidopsis [32], [33], [80]. We found that emasculated OsFIE2-RNAi florets did not show autonomous endosperm development. Similarly, no autonomous endosperm development was observed in rice T-DNA mutants of OsFIE1 and OsEMF2b [42]. The maize RNAi plants of ZmFIE1 and ZmFIE2 also produced no autonomous endosperm in the absence of fertilization [46]. Further, RNAi lines of HFIE in sexual Hieracium pilosella, failed to show the autonomous central cell proliferation [48]. These results, together, suggested that the function of FIE gene is not to suppress endosperm development. The function of the FIE containing PRC2 complex in repressing autonomous endosperm development in Arabidopsis is not conserved among plants. We found that down regulation of OsFIE2 gene in RNAi plants resulted in the production of small seeds and partially filled seeds (Figure 3 and Figure 4). Our results suggested that the OsFIE2 positively regulated endosperm development in rice instead of repressing endosperm development as predicted using Arabidopsis as the model. We found that some of the partially filled seeds germinated before maturation. Cross sectioning of the un-germinated partially filled seeds found that the embryos were bigger than the wild-type. It is still unknown if the large embryo was due to partial germination or other developmental defect. In addition, it is also unknown if the early germination is directly regulated by polycomb complex or it is a secondary effect due to partial filling or others. Interestingly, the Arabidopsis FIE mutant showed delayed seed germination [34]. While the wild-type seeds germinated within 2 days, approximately 40% of the homozygous FIE mutants stayed dormant for the course of the entire experiment, which lasted 20 days. Therefore, the FIE containing polycomb complex in rice and Arabidopsis have distinct roles. While the Arabidopsis complex suppress endosperm development in the early stage and promote seed germination after maturation, the rice OsFIE2 complex, in contrast, is required for seed and endosperm development and suppress early germination of the seeds either directly or indirectly. Interestingly, no phenotype was reported in mutants of the endosperm specific gene OsFIE1, a gene sharing the highest homology with OsFIE2 in the rice genome. Given the high homology of these two genes, these two genes might be functionally redundant. Analysis of the gene expression profile indicates that the expression level of OsFIE2 is about 25 times higher than that of the OsFIE1 gene in the endosperm. If these two genes are truly redundant in function, OsFIE1 mutant will display no phenotype because OsFIE2 can replace OsFIE1 function. In contrast, OsFIE2 mutant will display phenotype because the expression level of OsFIE1 is too low to replace the function of OsFIE2. Indeed, no phenotype was reported for OsFIE1 mutant but striking phenotypes were observed in our OsFIE2 RNAi lines. Our real-time PCR results showed that the starch synthesis rate limiting step enzyme (glucose pyrophosphorylase) subunit OsAGPS2b was down regulated in the OsFIE2 RNAi lines. Meanwhile, proteomics analysis revealed that the expression of prolamin, globulin, alpha-amylase, and seed allergenic proteins are reduced in OsFIE2 RNAi lines (Figure 6, Figure 7, and Table 3). These results suggest that the partially filled and smaller seeds were probably due to a reduction of starch and storage protein synthesis. To understand how the polycomb group genes regulate gene expression in rice endosperm, we analyzed our prior published rice endosperm ChIP-Seq data set. The ChIP-Seq data has been well verified by ChIP-PCR experiments with twenty eight genes [73]. We found that a large number of endosperm specific genes and storage nutrient metabolic pathway genes were subjected to the regulation of H3K27me3 in endosperm (Table S1 and Table S2). However, most genes shown to be down regulated by real-time PCR in the OsFIE2 RNAi lines (Figure 6) are not the direct targets of the H3K27me3, suggesting a complicated gene expression regulation network in endosperm. Seed development is a complex quantitative trait. If the starch and protein synthesis genes were directly regulated by polycomb complex, it would act as a simple Mendelian trait. Therefore, it is understandable that many of the starch and storage protein genes are not the direct targets of OsFIE2 complex. In addition, there are multiple polycomb complexes in plants. It is also possible that many ChIP-Seq identified loci are not the direct target of OsFIE2 complex but the targets of other PRC2 complexes. Arabidopsis helix-loop-helix DNA binding domain containing gene RGE1 (AT1G49770) is specifically expressed in the endosperm surrounding region which directly surrounds the developing embryo. It exerts its effect non autonomously in the developing embryo. Loss of RGE1 function results in small embryos. We found that the expression of rice RGE1 gene was increased and the modification of H3K27me3 at the locus was reduced in the RNAi lines, suggesting that it is a potential target subject to OsFIE2 regulation. Interestingly, while a mutation of the RGE1 gene leads to small embryo in Arabidopsis, we observed large embryo in the RNAi lines showing high expression of the rice RGE1 gene. The MADS box gene OsMADS6 (LOC_Os02g45770) is highly expressed both in flower and endosperm. Prior studies have shown that OsMADS6 is subjected to the regulation of H3K27me3 [71] and our ChIP-seq data analysis results verified that OsMADS6 was enriched by antibodies for H3K27me3 (Table S1 and S2). OsMADS6 controls grain filling by regulating the gene encoding OsAGPS1 (ADP-glucose pyrophosphorylase small subunit1, AK073146), a subunit of the ADP-glucose pyrophosphorylase [71]. Meanwhile, we found that the expression of another subunit gene OsAGPS2b of the glucose pyrophosphorylase enzyme was down regulated in the OsFIE2 RNAi lines. Further, a starch synthase and several storage protein genes were shown to be enriched in the ChIP-Seq experiment. Our results suggest that the OsFIE2 polycomb complex controls endosperm development and grain filling by regulating multiple levels of targets including transcription regulators as well as metabolic pathway genes. The large number of genes subject to the regulation of H3K27me3 in endosperm is consistent with the critical role of OsFIE2 gene in endosperm development and grain filling and suggests a highly sophisticated regulatory network. Rice (Oryza sativa, cultivar Nipponbare) growth conditions were similar to our previous report [81]. All rice plants were grown in the greenhouse of the Department of Biochemistry and Molecular Biology, Mississippi State University, MS, USA. Wild-type Nipponbare (rice subspecies japonica) was used as control. The OsFIE2 full length cDNA clone (AK111761, Jo23058f21) was obtained from National Institute of Agrobiological Sciences, Ibaraki, Japan. The cDNA was amplified using primers 61TAPF and 61TAPR and cloned into pENTR™/D-TOPO (Invitrogen) entry vector. The cDNA clone in the entry vector was transferred to final UbI-NTAP-1300 destination vector [51] by using a single LR clonase recombination reaction (Invitrogen). The final N-Terminal TAP-tagged OsFIE2 construct was transformed to Agrobacterium strain EH105A by electroporation. Rice transformation was performed as described Wu et al. (2003) [82]. Transformed resistant callus was screened using hygromycin. Transgenic plants were regenerated from the resistant callus. Mature seeds from the T1 transgenic plants with high expression of TAP-tagged OsFIE2 were used to induce transgenic callus. Cell suspension cultures were generated from transgenic callus and maintained in suspension medium (3. 2 g/liter Gamborg B5 basal medium (Phytotechnology Laboratories™), 0. 5 g/liter MES, 20 g/liter sucrose, 2 mg/liter 2,4-Dichloro acetic acid, 2 g/liter N-Z-AmineA, PH 5. 7, adjusted with 1 M KOH) at 25°C in darkness by constant shaking (150 rpm) on a gyratory shaker. A portion of coding sequence fragment was amplified using primer set 61RNAiF and 61RNAiR from OsFIE2 cDNA clone (AK111761, Jo23058f21) and cloned into pENTR™/D-TOPO (Invitrogen) entry vector. The amplified fragment in the entry vector was transferred to final pANDA vector [83] by using a single LR clonase recombination reaction (Invitrogen). The final OsFIE2-RNAi construct was transformed to Agrobacterium strain EH105A by electroporation. Rice transformation was performed as described Wu et al. (2003) [82]. Transformed resistant callus was screened by using hygromycin. Transgenic plants were regenerated from the resistant callus. Exponentially growing cell suspension cultures (30 g) were harvested 3 days after subculturing and ground in liquid nitrogen. Protein extracts were prepared in two volumes of extraction buffer (20 mM Tris-Hcl, pH 8. 0,150 mM NaCl, 0. 1% IGEPAL (Sigma), 2. 5 mM EDTA, 2 mM benzamidine, 10 mM β-mercaptoethanol, 20 mM NaF, 2 mM phenyl methanesulfonylfluoride (PMSF), 1% Protease cocktail (Sigma), 10 µM leupeptin (Sigma), and 10 µM 3,4-dichloroisocoumarin (Sigma) ). The suspension was homogenized with the help of Polytron PTA 20 TS homogenizer for 2 min on ice and filtered through two layers of miracloth. The soluble protein fraction was collected by centrifugation twice at 30,000 g for 20 min at 4°C. Affinity purifications were performed as described by Rohila et al. (2006) and Van Leene et al. (2007) [53], [54] with some modifications. The protein extract was incubated with 400 µl IgG sepharose beads (GE Healthcare) for 1 h at 4°C. The IgG beads were loaded on to a polyprep chromatography column (Bio-Rad Laboratories, CA, USA) and washed with 10 ml of extraction buffer lacking protease inhibitors and 10 ml of TEV (Tobacco etch virus) cleavage reaction buffer (10 mM Tris–HCl, pH 8. 0,150 mM NaCl, 0. 1% IGEPAL, 0. 5 mM EDTA, 1 mM DTT). Bound proteins were eluted by digestion with 150 U of Ac TEV enzyme in TEV cleavage reaction buffer containing 1 µM E-64 protease inhibitor for 1 h at 16°C. IgG-eluted fraction was adjusted to 2 mM Cacl2 and diluted with 3 volumes of calmodulin binding buffer (10 mM Tris–HCl pH 8. 0,150 mM NaCl, 1 mM Mg-acetate, 1 mM imidazole, 2 mM CaCl2,0. 1% IGEPAL, 10 mM β-mercaptoethanol) and the fraction was incubated with 400 µl of calmodulin-agarose beads (Stratagene, CA) for 1 h at 4°C. The calmodulin-agarose beads were loaded on to a polyprep chromatography column and washed with 10 ml of calmodulin binding buffer. The protein complexes were eluted with 2 ml of elution buffer (10 mM Tris–HCl pH 8. 0,150 mM NaCl, 1 mM Mg-acetate, 1 mM imidazole, 25 mM EGTA, 0. 1% IGEPAL, 10 mM β-mercaptoethanol) and proteins were precipitated with TCA. The protein pellet was washed with cold acetone. Tandem affinity purified complex proteins were trypsin digested and identified using LC-MS/MS mass spectrometry analysis as described by Chitteti et al. (2008); Tan et al. (2007); Tan et al. (2010) [58], [59], [60]. Protein extracts were prepared according to the TAP protocol. Protein samples were collected in different stages of the TAP purification and separated on a 12% SDS-polyacrylamide gel and transferred on to PVDF (Millipore) membrane. In the first step of purification, presence of TAP-tagged protein was identified by using peroxidase anti peroxidase conjugate (PAP, Sigma) antibody, which is specific to protein A domain of the TAP tag, as previously described Rivas et al. (2002) [56]. In the second step of purification, CBP domain was detected by using biotinylated CAM as previously described Rohila et al. (2004) [51]. To check the methylation status, protein extracts from OsFIE2-RNAi, OsFIE2-overexpression and wild type leaf tissues were separated by 15% SDS-PAGE gel and transferred to immobilon membrane. Immunoblots were performed using antibodies for H3K27me1, H3K27me2, H3K27me3, H3K9me2, H3K9me3 and unmodified H3 (Table S4) by following the standard Western blot procedure [59]. Equal amounts of protein samples were loaded for Western blot analysis. Western blot signals were normalized using unmodified H3 band intensity as a control and quantified with the help of PDQUEST software. Procedures for histone methyltranferase assay were adapted from the reported method Li et al. (2002) [84]. Briefly the assay was carried out by incubating TAP purified OsFIE2 protein complex, 10 µg of chicken core histones (Upstate) and 2 µCi of S-adenosyl-[methyl-3H]-l-methionine (GE Healthcare) in 50 µl of reaction buffer (50 mM Tris–HCl, pH 8. 5,20 mM KCl, 10 mM MgCl2,1 mM CaCl2,10 mM 2-mercaptoethanol, 1 mM dithiothreitol, and 250 mM sucrose) for 2 h at 30°C. The reaction was stopped by addition of SDS loading buffer and boiling at 100°C for 10 minutes. The proteins were separated by 15% SDS-PAGE gel and visualized by coomassie staining and fluorography. Total seed proteins were isolated from the mature seeds of both OsFIE2-RNAi and wild type plants. Mature seeds were ground into fine powder and protein extractions were performed using phenol extraction method as described Chitteti and Peng (2007) and Li et al. (2008) [72], [85]. Eight hundred micrograms of seed proteins were separated by using a standard 2-DE gel electrophoresis as described by Chitteti and Peng (2007) [85]. After PDQUEST analysis, the spots of interest were robotically excised from 2-DE gels by proteome works spot cutter (BioRad). In-gel digestion and MALD TOF-TOF mass spectrometry was performed as described by Chitteti and Peng (2007) [85]. Total RNA extraction from different tissues of rice plants was performed by using Trizol (Invitrogen) according to the manufacturer' s instructions. Three micrograms of total RNA was used for the cDNA synthesis by using Moloney murine leukemia virus (M-MLV) Reverse Transcriptase (Invitrogen). Rice Glyceraldehyde 3-phosphate dehydrogenase (GAPDH) (Loc_Os04g40950) was used as an internal control. PCR products were analyzed by 1% agarose gel electrophoresis. Primers used in the experiment are listed in Table S3. Rice developing grains were fixed in FAA (3. 7% formaldehyde, 5% acetic acid, 50% ethanol) at 4°C for overnight. Samples were dehydrated through graded ethanol series. After infiltration with xylene, samples were embedded in paraffin (Sigma-Aldrich), sectioned at 10 µm, and stained with 0. 1% Eosin Y. Sections were observed and photographed using a bright-field microscope (Nikon). Real-time quantitative PCR analysis was performed with LightCycler 480 Real-Time PCR System (Roche Applied Science), using LightCycler 480 SYBR Green I Master. The relative quantification of transcript levels was performed by using 2−ΔΔC T Method [86]. The rice GAPDH (Loc_Os04g40950) gene was used as an internal control. Primers used in the experiment are listed in the Table S3. E. coli two hybrid assay was carried out according to the instructions given by manufacturers (Agilent Technologies). Recombinant bait (pBT) and target (pTRG) vectors were co-transformed into BacterioMatch II competent cells and positive interactions were screened by using selective and dual selective medium [64]. To examine the pollen viability of OsFIE2-RNAi and wild-type plants, anthers were collected from well developed spikelets just before anthesis. Mature pollen grains were stained with 1% iodine in 3% potassium iodide solution (KI-I2) [87]. Viable pollen grains were stained black and nonviable pollen grains were stained light yellow. The viable pollen grains were examined and counted under microscope. The chromatin was extracted from endosperm following the protocol of Gendrel et al. (2005) [88] with minor modifications as reported recently Malone et al. (2011) [73]. Chromatin immunoprecipitation (ChIP) experiments were performed as reported Malone et al. (2011); Gendrel et al. (2005) [73], [88] using H3K27me3 antibody (Millipore) and protein A agarose/Salmon Sperm DNA. The library preparation and Solexa Sequencing were the same as reported Malone et al. (2011) [73]. Input and ChIP samples were processed following Illumina' s protocol from the ChIP DNA Sample Prep Kit. Briefly, 10 ng input and ChIP enriched DNA was subjected to end repair, addition of “A” bases to 3′ ends, ligation of adapters, agarose gel size selection for fragments with average size about 186 bp, and PCR amplification to produce a DNA library of adapter-modified fragments. DNA sequencing was carried out using the Illumina/Solexa Genome Analyzer sequencing system at a concentration of 2 to 4 pM. Cluster amplification, linearization, blocking and sequencing primer reagents were provided in the Solexa Cluster Amplification kits and were used according to the manufacturer' s specifications. Mapping the short reads is the same as reported Malone et al. (2011) [73]. The generated short reads were mapped onto the genome using SeqMap [89] allowing up to two mismatches between the short read and genome. The Illumina reads were aligned to TIGR version 6 of the rice genome [90]. The alignments were output in ELAND format. Only reads which mapped uniquely to the genome were retained. FindPeaks [91] was used to identify peaks with the mapped reads and the parameters used were the same as reported [73]. All the ChIP-Seq data were deposit to NCBI' s Gene Expression Omnibus (http: //www. ncbi. nlm. nih. gov/geo/) with the deposition number GSE27048 for genome-wide maps of chromatin state in rice endosperm.

Rice is the staple food for over half of the world' s population and an important feedstock for livestock. The rice grain is mainly endosperm tissue. The regulatory mechanism of rice endosperm development is still largely unknown thus far. Understanding the underlying mechanism will lead to crop yield and quality improvement in the long term, besides gaining new knowledge. Polycomb complex is a protein complex with a potential role in endosperm development according to prior publications. In this manuscript, we purified the rice OsFIE2-polycomb protein complex and demonstrated the enzyme activity of the complex. Genetic studies showed that a reduction of polycomb group gene OsFIE2 expression led to smaller seeds, partially filled seeds, and seed germination before seed maturation. Gene expression and proteomics analyses found that the starch synthesis rate limiting step enzyme and multiple storage proteins are down-regulated while a key transcription factor is up-regulated in OsFIE2 reduction lines. In addition, we identified many loci in the rice genome whose histone proteins are modified by the polycomb complex enzyme via a method called ChIP–Seq. Our results demonstrate that OsFIE2-polycomb complex positively regulates rice grain development via a mechanism distinct from that in Arabidopsis and provide new insight into the regulation of rice grain development.

Abstract Introduction Results Discussion Materials and Methods

agriculture biology

2013

Polycomb Group Gene OsFIE2 Regulates Rice (Oryza sativa) Seed Development and Grain Filling via a Mechanism Distinct from Arabidopsis

The genetic impact associated to the Neolithic spread in Europe has been widely debated over the last 20 years. Within this context, ancient DNA studies have provided a more reliable picture by directly analyzing the protagonist populations at different regions in Europe. However, the lack of available data from the original Near Eastern farmers has limited the achieved conclusions, preventing the formulation of continental models of Neolithic expansion. Here we address this issue by presenting mitochondrial DNA data of the original Near-Eastern Neolithic communities with the aim of providing the adequate background for the interpretation of Neolithic genetic data from European samples. Sixty-three skeletons from the Pre Pottery Neolithic B (PPNB) sites of Tell Halula, Tell Ramad and Dja' de El Mughara dating between 8,700–6,600 cal. B. C. were analyzed, and 15 validated mitochondrial DNA profiles were recovered. In order to estimate the demographic contribution of the first farmers to both Central European and Western Mediterranean Neolithic cultures, haplotype and haplogroup diversities in the PPNB sample were compared using phylogeographic and population genetic analyses to available ancient DNA data from human remains belonging to the Linearbandkeramik-Alföldi Vonaldiszes Kerámia and Cardial/Epicardial cultures. We also searched for possible signatures of the original Neolithic expansion over the modern Near Eastern and South European genetic pools, and tried to infer possible routes of expansion by comparing the obtained results to a database of 60 modern populations from both regions. Comparisons performed among the 3 ancient datasets allowed us to identify K and N-derived mitochondrial DNA haplogroups as potential markers of the Neolithic expansion, whose genetic signature would have reached both the Iberian coasts and the Central European plain. Moreover, the observed genetic affinities between the PPNB samples and the modern populations of Cyprus and Crete seem to suggest that the Neolithic was first introduced into Europe through pioneer seafaring colonization. The term “Neolithic” refers to the profound cultural and technical changes that accompanied the transition from a hunter-gatherer subsistence economy to an agro-pastoral producing system [1]. The first Neolithic societies originated 12 to 10 thousand years ago in a region of the Near East traditionally known as the “Fertile Crescent” [2]. From this region the Neolithic technology rapidly expanded to Anatolia reaching the rest of Europe in less than 3,000 years by following two main routes linked to different archaeological cultural complexes. The Danubian route, associated to the Linearbandkeramic (LBK) cultural complex, brought the Neolithic to the central European plains and from there to the British Islands and Scandinavia (Funnel Beaker Cultural Complex) while the Mediterranean one, associated to the Cardial-Impressa cultural complex, spread it along the Mediterranean coast up to the Atlantic façade of Iberia [3]. The nature of the diffusion of the Neolithic and the possible demographic input associated to it have been widely debated. In this regard, two extreme hypotheses representing opposite views have been formulated: the demic diffusion model (DDM) and the cultural diffusion model (CDM) [1], [2], [4], [5]. The former stands up for a “wave of advance” of Neolithic immigrants with subsequent genetic replacement of the hunter-gatherer (Mesolithic) populations while the latter proposes a cultural adoption of Neolithic practices from local populations, implying a genetic continuity since the Palaeolithic. Moreover, integrationist models that involve a different extent of interaction between immigrants and local hunter-gatherers while considering the effect of geographic barriers and agricultural boundary zones, have been also used to explain the transition to the Neolithic process at a more local scale [6]. Genetic analyses from modern and ancient populations have contributed extensively to this debate providing discordant results. Principal component analysis and spatial autocorrelation of allele frequencies of “classic” genetic markers in modern European populations showed a South East to North West cline compatible with a Neolithic DDM. The Neolithic contribution to the modern genetic pool was estimated in this case to be around 27% [7]. The frequency distribution of Y chromosome polymorphisms displayed a similar pattern and haplogroups F*, E3b, G and J2, representing a 22% of extant lineages, were initially identified as the main contributors of the Neolithic spread [8], [9]. However, the analysis of the geographic distribution of the microsatellite diversity of the allegedly Paleolithic haplogroup R1b1b2, has been recently reinterpreted as a signal of substantial demic diffusion [10]. Phylogeographic analyses of another haploid marker, the mitochondrial DNA (mtDNA), in Europe and the Near East initially supported a limited Neolithic genetic contribution of around 9–12% in the Mediterranean and 15–22% in Central Europe [11]. Molecular dating and founder analyses identified then mtDNA haplogroups J, T1 and U3 as the main genetic markers of this expansion, with probable contributions of some other lineages from clusters H and W [12]. However, recent analysis of complete mtDNA sequences from the same region has pictured contradicting results depending on the analysis performed, from all mtDNA haplogroup expansions predating the Neolithic [13] to Neolithic expansions of mtDNA haplogroup H [14]. In the light of these results, the usefulness of modern genetic variability to reconstruct the Neolithic dynamics in Europe has been questioned [15], [16]. First of all, a certain level of genetic differentiation between hunter-gatherers and Near Eastern farmers has to be assumed in order to detect differences between both groups. Secondly, the existence of SE-NW clinal patterns in Europe may reflect the accumulation of small migrations entering the continent rather than a single migratory event [17]. Finally, original population substructure and subsequent processes of genetic drift and founder effects can introduce errors into the estimation of coalescence dates of mitochondrial and Y chromosome haplogroups [18]. In this regard, recent diachronic aDNA analyses of Central European populations have documented a fluctuation in haplogroup frequencies as a result of population bottlenecks and post-Neolithic migratory events [19], [20]. Besides, these estimated haplogroup dates do not necessarily correspond to the time of arrival of the lineages to the region [21]. As a result, the misidentification of genetic variants associated to the Neolithic spread and the effect of post-Neolithic expansions in the genetic make-up of Europe could have introduced important biases in the estimations of the Neolithic component of the European gene pool producing misleading conclusions [22]. During the last decade, ancient DNA analyses of Neolithic populations have provided a more reliable picture of the Neolithic transition process at a local scale. Studies have concentrated at the two edges of the two routes of the Neolithic wave of advance: Central/Northern Europe and the Iberian Peninsula/Southern France. In Central Europe and Scandinavia a DDM has been proposed to explain the observed genetic discontinuity between hunter-gatherers and the first farmer populations [19], [23]–[26]. However, recent analyses have suggested the coexistence of genetically distinct hunter-gatherer and farmer groups during several millennia at the same archaeological site, suggesting that the genetic replacement of hunter-gatherers populations was not complete [20]. In North Eastern Iberia and Southern France contradictory interpretations have been proposed to explain the nature of the Mesolithic-Neolithic transition process. On one hand, mtDNA studies of Cardial Neolithic remains seem to favor a pioneer Near Eastern colonization of North Eastern Spain [27], [28]. On the other hand, mtDNA results of Epicardial, Middle and Late Neolithic populations have been interpreted as a signal of pre-Neolithic legacy [29]–[31]. Dating and cultural differences between the studied groups, the effect of genetic drift at the beginning of the Neolithic and differences in the methods of analysis used (model-based statistical inference vs assignment of mtDNA haplogroup dating categories respectively) could be responsible of the observed differences [12], [27]. Moreover, the Y chromosome diversity of the Epicardial and Late Neolithic datasets has also shown a predominantly Near Eastern influence, suggesting that males and females might have played a differential role in the Neolithic dissemination process [16], [30], [31]. In this framework, the knowledge of the original Neolithic genetic pool in the Near East seems essential to correctly identify the variants associated to the Neolithic spread and also to provide a more global picture of the Neolithic dynamics in Europe. In order to examine the genetic background existing in the first Neolithic communities and its impact over the European genetic pool, we have studied 3 archaeological sites in Syria located in two geographic areas in which agricultural practices were first documented: the middle Euphrates valley and the oasis of Damascus (Figure 1). These sites are dated back to the Pre-pottery Neolithic B period (PPNB). It is during this initial Neolithic phase that animal husbandry first appears, while full-scale agricultural practices are documented in the whole Levant. At the PPNB there is also an increase in the size of the settlements, probably as a response to the population growth derived from the consolidation of the new food-producing economic system [32]. The obtained results allowed us to put into context ancient DNA results from available European Early Neolithic populations, to draft a general model of the Neolithization in Europe and to propose probable routes of expansion of the first Neolithic communities. DNA preservation at the studied samples was assessed at three levels: (1) Estimating the number of copies of the target mtDNA in some of the extracts using a specific Real Time PCR design, (2) estimating the percentage of reproducible Hypervariable Segment I mitochondrial DNA (mtDNA-HVS1) sequences out of all the analyzed samples and (3) computing the miscoding lesions in clone sequences. The average number of mtDNA HVS1 copies per amplified volume of extract was in all cases higher than 1000, with a mean value of 10. 4×104 in Tell Halula and 1. 1×106 in Tell Ramad, corresponding respectively to 7. 44×10−5 and 7. 60×10−4 ng/µl (Table S1). Reproducible mtDNA sequences could be recovered from 24 out of 112 DNA extracts, corresponding to 15 different skeletons from Tell Halula and Tell Ramad (see Table S2). Differences in sample recovery success ratios could be a result of the strict screening approach used –in which samples displaying more than 2 negative amplification results were discarded (30% of the aDNA extracts) - and of the differences in efficiency between the amplification strategy used in both laboratories. The overall ratio of endogenous DNA recovery for the studied remains was 23. 8%. The average number of miscoding lesions per clone and nucleotide in the studied samples was 0. 0078 in Tell Halula and 0. 0047 in Tell Ramad. Individual sample variability ranged from 0. 0000 (sample H3) to 0. 0303 (sample H68) in Tell Halula and from 0. 0006 to 0. 0101 in Tell Ramad, indicating a differential preservation across the samples (Table S3). Damage values per sample are within the range reported by other authors in samples with similar chronology from temperate environments (La Braña: 0. 0116–0. 0163; Can Sadurní: 0. 0054–0. 0632; Chaves: 0. 0092–0. 0872; Sant Pau del Camp: 0. 0000–0. 0133). Reproducible mtDNA HVS1 sequences were obtained from 15 out of 63 skeletons from the archaeological sites of Tell Halula and Tell Ramad (Table 1). The alignments of both the direct sequences and the clones are presented in Table S3. Sequences have been deposited in Genbank (http: //www. ncbi. nlm. nih. gov/genbank) with accession numbers KF601411- KF601425. In 10 cases it was possible to reconstruct the complete haplotype (nucleotide positions, np 16,126–16,369), while the extent of degradation of the remaining 5 samples only allowed the recovery of partial haplotypes. Nine different haplotypes were identified. Two of them were shared between 2 individuals of Tell Halula (16311C) and among 3 individuals of Tell Ramad (16224C 16311C 16366T). Motif 16293C, though, was present at both sites, pointing at a pre-existing common genetic pool in the region. The complete haplotypes were compared to a database of 9,821 mtDNA profiles from 59 modern populations from the Near East and South Eastern Europe and 2 Early Neolithic populations from Central Europe (LBK-AVK Neolithic, [24]) and North Eastern Iberia (Cardial/Epicardial Neolithic, [27]) (see Figure S1 and Table S4). Haplogroup affiliation was also considered in the haplotype search. The number and percentage of shared haplotypes between the PPNB population and the populations in the database plus the number and percentage of individuals from each population carrying PPNB are presented in Table S5. Figure 2 displays a contour map of the latter built using the same data in a subset of 51 populations. Two out of the 7 different complete PPNB haplotypes (16356C and 16293C, 28. 57% of studied samples) were not represented in any of the modern and ancient populations of the database. From the remaining haplotypes only 16224C 16311C, the basal node of haplogroup K, was shared with the other two ancient populations, displaying a frequency of 9. 52% in the Cardial/Epicardial dataset and of 23. 08% in the LBK-AVK. This haplotype is distributed nowadays both in South Eastern Europe and the Near East with an average frequency of 4%. However, some populations such as Ashkenazi Jews, Csángó and Cyprus exhibit frequencies of this haplotype higher than 10% (Table S5, Figure 2). The remaining haplotypes had a very limited geographic distribution, being only documented in 1 individual from Bulgaria (16311C-K); 3 individuals from Turkey, Qatar and Yemen (16223T-L3); 4 Irani, Karakalpak, Turkish and Bedouin individuals (16256T-H) and 3 Druze from Israel (16224C 16311C 16366T-K). MtDNA haplogroups could be assigned to 14 out of the 15 skeletons according to the HVS1 sequences obtained and on the diagnostic Single Nucleotide Positions (SNPs) typed following Phylotree rCRS oriented version 15 (Tables 1 and S6). Haplogroup K was the most prevalent, (N = 6,42. 8%) followed by R0 (N = 3,21. 42%) and H (N = 2,14. 28%). The observed haplogroup frequencies were compared to those of 59 modern populations from the Near East and South Eastern Europe and 2 Early Neolithic populations from Central Europe (LBK-AVK Neolithic, [24]) and North Eastern Iberia (Cardial/Epicardial Neolithic, [27]) (N = 11,610) (Table S7). Haplogroup K was present in almost all populations compared, and its mean frequency in South Eastern Europe and the Near East was around 7%. It reached its highest frequencies in certain populations that have experienced recent population bottlenecks, such as the Askhenazi Jews and the Csángó in Transylvania, Romania [33], [34] and also among Greek Cypriots. Moreover, it was also highly represented in both Cardial/Epicardial (15. 56%) and LBK-AVK (23. 08%) Early Neolithic datasets. Haplogroup R0 is especially prevalent in the Near East and North Africa with a mean frequency in both regions around 6%. The maximum frequencies of R0 were detected in South Arabian populations such as Bedouin, Oman and Saudi Arabia (Table S7). The rare European haplogroups U* and N* were also detected in 2 individuals in our ancient sample. The mean frequency of haplogroup U* is 2% in the Near East, 0. 9% in the Caucasus region and around 1% in Europe, whereas the N* mean frequency is less than 1% in all three datasets. However, both haplogroups reach peaks of frequency in certain populations, such as haplogroup U* in Crete. The case of N* is especially interesting, because apart from Bulgaria, Crete, Romania and Serbia it was only represented in Near Eastern populations (Iran, Jordan, Near Eastern Jews, Oman, Palestine, Saudi Arabia, Syria, Turkmenistan and United Arab Emirates). Moreover, this haplogroup was also detected in 4 Neolithic specimens from Catalonia, in North Eastern Spain, associated to the Cardial/Epicardial culture [27]. Carry- over contamination from these samples processed in the same laboratory can be ruled out, as results were validated in a second independent laboratory. Finally, the skeleton H8 belonged to the African L3 lineage, this being the most prevalent African haplogroup found in present-day Near Eastern populations. Principal Component Analysis with Hierarchical Clustering (PCA-HCA) was used to compare mean haplogroup frequencies of our dataset (Table S7) with the other populations of the database. Details about the method can be found in Table S8. The first six Principal Components (PCs), accounting for a 90. 6% of the variance, were selected for Hierarchical Clustering Analysis. Six clusters (1–6) were defined based on the topology of the hierarchical tree (Figure S2). The decomposition of the inertia computed on 6 axes supported this partition, indicating that with a division in 6 clusters up to an 80% of the data variation could be explained (Table S8). The main haplogroups contributing to the cluster separation were Asian (AS: test value = 12. 66; P = 0. 000), African (AF: test value = 8. 55; P = 0. 000), H (test value = 8. 96; P = 0. 000) and K (test value = 8. 01; P = 0. 000). The two biggest groups detected were Clusters 1 and 3, joining 43 of the 60 populations of the database. Cluster 1 mainly included European populations and it was distinguished by high frequencies of haplogroups H, U5, U4 and HV0 and by low frequencies of Asian and African types (Table 2). Near Eastern and some Caucasian datasets were grouped in Cluster 3. They were separated from European populations mainly by high frequencies of haplogroups J and T and low frequencies of H, HV0 and U5. Interestingly, LBK-AVK population was also included in this group. Its similarity with Caucasian populations like Georgia and Chechnya previously suggested by [24] was also evident in our analysis. Cluster 2 included our PPNB sample, grouped together with Ashkenazi Jews, Csángó, Cyprus and Cardial/Epicardial populations. High frequencies of haplogroups K and N* characterized this cluster (Table 2), pinpointing the genetic affinities between the PPNB and the Cardial/Epicardial Neolithic dataset already stressed by the qualitative haplogroup and haplotype analyses. Cluster 4 included populations from Africa or with a strong African component and it was defined by high frequencies of African haplogroups (L and U6) and low frequencies of haplogroup H. Western Asian populations were clearly separated from Near Eastern datasets in clusters 5 and 6. Both were distinguished by a high frequency of Asian haplogroups and a low frequency of European types. The inclusion of Romani population within cluster 5 is in agreement with its Asian origins [35]. The partition model proposed here supports the existence of geographic barriers for mitochondrial markers. Major geographic zones like Europe, the Near East and Eastern Asia are clearly distinguished. However, populations at boundary zones such as the Caucasus are clustered both with European and Near Eastern pools. The PCA-HCA for the two first PC factors, accounting respectively for 48. 32% and 19. 78% total genetic variation, is represented in Figure 3. On one hand, the first PC distinguished populations with and without Asian haplogroups, separating clusters 5 and 6 from 1,2, 3 and 4. On the other hand, the second PC separated those populations with African (Cluster 4) and non-African (Clusters 1,2 and 3) haplogroups. Cluster 3, containing Near Eastern and Caucasian populations, occupied an intermediate position in the plot. According to the two first PCs the PPNB population, included in Cluster 2, was equidistant to the centers of this cluster and Cluster 3 and close to modern populations from the Fertile Crescent, such as Jordan and Palestine. Affinities of the PPNB population with populations within Cluster 3 were due to high frequencies of haplogroup R0 in all of them. The Cardial/Epicardial Neolithic population, also member of Cluster 2, was in this case closer to Cluster 1 due to its moderate frequencies of haplogroups H and U5. Cluster 2 was clearly distinguished from the other 5 clusters by PC4, which summed up a 6. 64% of the global genetic variability (Table S8). The graphical plot of PC3 and PC4 separated populations by their frequencies of haplogroups HV, J and T (PC3) and K (PC4) (Figure S3). This graph situated the PPNB sample at the edge of PC4 axis, close to Cardial/Epicardial and Ashkenazi Jew populations. Pairwise FST genetic distances were computed between the PPNB and the other populations of the database (Table S9). Non-significant pairwise FST values were obtained between PPNB and Cyprus (FST = 0. 013; P = 0. 2734), Ashkenazi Jews (FST = 0. 028; P = 0. 1087), Csángó (FST = 0. 022; P = 0. 1087) and Khoremian (FST = 0. 0456; P = 0. 0805). These populations also exhibited the lowest FST values. The highest significant distances corresponded to Gilaki, Caucasian Jews and Mazandarian populations (FST>0. 2). When modern populations were grouped in geographic regions, the PPNB population was genetically closer to Near Eastern and Caucasian than to Southern European populations (Table 3). The Cardial/Epicardial and LBK-AVK populations showed low FST values with the modern Near Eastern pool, as previously stated [24], [27]. It is important to note, however, that the FST index between LBK-AVK and the pooled Southern European populations was lower than the one reported by [24]. FST distances between the PPNB and the modern populations were plotted in a contour map (Figure 4). The map showed minimum FST values in the Fertile Crescent area (Northern Egypt, Palestine, Jordan, Syria and Southern Anatolia) and Cyprus. From this region genetic distances gradually increased westwards across the Balkans, southwards to the Arabic Peninsula and eastwards through the northern Zagros to the Caspian Sea. Peaks of low distance were also detected in the Carpathian basin, Yemen and in North Uzbekistan, South from the Aral Sea. One of the inherent limitations of ancient DNA human studies is the possibility of contamination with exogenous DNA, a risk that is enhanced when human DNA is studied and a PCR approach is used. As a result a series of authentication criteria were proposed early at the beginning of the discipline [36], [37]. However, it has been recognized that on one hand, a complete level of authentication cannot be achieved in most of the cases and on the other, the strict application of all the criteria does not provide a 100% proof of the authenticity of the data [38]. The importance of the retrieved results as a potential comparative framework for other ancient DNA studies requires the reported data to be solid and unambiguous. As such, to guarantee the authenticity of our results we have used a combination of classical criteria of authenticity and a self-interpretative approach as suggested by [38]. These criteria include the replication of the results within the same or in a second laboratory, Real-Time PCR estimation of the number of DNA copies in the extracts, bacterial cloning of amplicons and a self-critical analysis of the obtained results. Trace contaminant DNA was detected through a detailed analysis of clone sequences. Phylogenetic sense observed between HVS1 mtDNA fragments and haplogroup specific SNPs at the mtDNA coding region provided further support to the authenticity of the obtained results. Sequence artifacts like chimerical haplotypes arisen by amplification of fragments of multiple origin (i. e contaminant, endogenous and damaged) could be ruled out through replication, as they occur at random and are not reproducible in different amplifications and extractions from the same or different skeletal samples [37]. Moreover, DNA content in the amplified extracts provided in all cases a number of starting copies higher than 1,000, thus making the possibility of displaying hybrid haplotypes highly improbable. The possibility of contamination between samples displaying the same haplotype (i. e. H4, H7, H28, H25; H3, R64-4II, R69 (2) and R65-14, R65-C8-SEB, R65-1S) could be also discarded as they were processed in different extraction and amplification batches and validated through independent replications, some of them conducted in two different laboratories. Even though the success recovery ratio is low (23. 8%), this study demonstrates that it is feasible to recover ancient DNA genetic information from temperate environments and suggests that other variables rather than the temperature operate in the DNA preservation through several millennia. Ancient DNA preservation in Near Eastern open-air sites has been previously stated [39]–[42]. The reported success ratios are variable, ranging from 4% [40] to 86% [42]. In the case of Tell Halula, the skeletons were located at opened pits under the main floor of the house. The pits were sealed using a cover made of mud brick of about 20 cm that in some cases was also plastered at the top [43]. This particular burial structure might have protected the human remains from DNA degradation. The absence of sample cleaning with water and the storage in freezers shortly after the excavation, should have also prevented skeletal remains from post-depositional degradation and contamination [39]. The recovery of insoluble collagen fractions (>30,000 Da) in the same remains is also an indicator of their good biomolecular preservation status [44], [45]. In recent years, the body of ancient DNA data of Neolithic populations has increased dramatically, providing a more accurate picture of local Neolithic dynamics. Some of these studies have also explored the Mesolithic genetic background, interpreting the results in terms of continuity or genetic break with the predecessor hunter-gatherer communities of the area [20], [23], [25], [28]. However, most of the attempts to estimate the Neolithic genetic input in those populations and/or to reconstruct the routes of dispersion of the first farmers into Europe have relied on extant data from modern Near Eastern populations [19], [24], [27], [29]–[31]. In the present research, ancient DNA results from the original human Near Eastern Neolithic communities are presented, to our knowledge, for the first time. The present study shows that even though the mitochondrial variability of the PPNB population is within the limits of modern Near Eastern, Caucasian and South Eastern European populations (Table 3), both haplotype and haplogroup PPNB frequencies clearly deviate from their modern successors (Figures 2 and 3, Tables S5 and S7). This indicates that the mitochondrial DNA make-up of modern Near Eastern populations may not reflect accurately the genetic picture of the area at the emergence of the Neolithic. All the detected haplotypes but one -the basal node of haplogroup K- have a null or limited distribution in the modern genetic pool, suggesting that a great bulk of ancient Neolithic lineages were not integrated into their succeeding populations or were erased by subsequent population movements in the region. This is in agreement with previous observations from other Early Neolithic populations [27], [46], and underlines the importance of genetic drift processes at the beginning of the Neolithic [16]. Nevertheless, the multi-population comparative analyses performed here also suggest that certain population isolates of Middle Eastern origin, like the Druze, could have retained an ancient Neolithic genetic legacy through cultural isolation and endogamous practices [47]. Another interesting case are the Ashkenazi Jews, who display a frequency of haplogroup K similar to the PPNB sample together with low non-significant pairwise Fst values, which taken together suggests an ancient Near Eastern origin. This observation clearly contradicts the results of a recent study, where a detailed phylogeographical analysis of mtDNA lineages has suggested a predominantly European origin for the Ashkenazi communities [48]. According to that work the majority of the Ashkenazi mtDNA lineages can be assigned to three major founders within haplogroup K (31% of their total lineages): K1a1b1a, K1a9 and K2a2. The absence of characteristic mutations within the control region in the PPNB K-haplotypes allow discarding them as members of either sub-clades K1a1b1a or K2a2, both representing a 79% of total Ashkenazi K lineages. However, without a high-resolution typing of the mtDNA coding region it cannot be excluded that the PPNB K lineages belong to the third sub-cluster K1a9 (20% of Askhenazi K lineages). Moreover, in the light of the evidence presented here of a loss of lineages in the Near East since Neolithic times, the absence of Ashkenazi mtDNA founder clades in the Near East should not be taken as a definitive argument for its absence in the past. The genotyping of the complete mtDNA in ancient Near Eastern populations would be required to fully answer this question and it will undoubtedly add resolution to the patterns detected in modern populations in this and other studies. Our PPNB population includes a high percentage (80%) of lineages with a Palaeolithic coalescence age (K, R0 and U*) and differs from the current populations from the same area, which exhibit a high frequency of mitochondrial haplogroups J, T1 and U3 (Table S7). The latter have been traditionally linked with the Neolithic expansion due to their younger coalescence age, diversity and geographic distribution [11], [12], [49]. In addition to the PPNB population, haplogroup T1 is also absent in other Early Neolithic populations analyzed so far [17], [22], [26], [30]. Haplogroup U3 has been found only in one LBK individual and it has been suggested that it could have been already part of the pre-Neolithic Central European mitochondrial background [19]. Haplogroup J is present in moderate frequencies in Central European LBK-AVK populations (11. 75%) and it has been proposed as part of the Central European “mitochondrial Neolithic package” [19]. However, it has also been described in one late hunter-gatherer specimen of Germany, raising the possibility of a pre-Neolithic origin [23]. Haplogroup J is present in low frequency (4%) in Cardial/Epicardial Neolithic samples of North Eastern Spain [27], [28], [31]. Absence of Mesolithic samples from the same region prevents making any inference about its emergence during the Mesolithic or the Neolithic. However, its absence in the PPNB genetic background reinforces the first hypothesis. These findings suggest that (1) late Neolithic or post-Neolithic demographic processes rather than the original Neolithic expansion might have been responsible for the current distribution of mitochondrial haplogroups J, T1 and U3 in Europe and the Near East and (2) lineages with Late Paleolithic coalescent times might have played an important role in the Neolithic expansive process. The first suggestion alerts against the use of modern Near Eastern populations as representative of the genetic stock of the first Neolithic farmers while the second will be explored in depth in the following section. The sharing of mitochondrial haplotypes and haplogroups between pre-pottery farmers from the Fertile Crescent and European Neolithic populations, suggests a genetic contribution of the first Neolithic communities in the European mitochondrial genetic pool. Haplogroup composition and PCA-HCA of the three ancient datasets compared here allow us to identify K and N*-derived haplogroups as potential Neolithic genetic contributors. Haplogroup K is present in all Early Neolithic datasets published so far with frequencies ranging from 7. 7 to 43% (Table S7, [19], [28], [31]). Moreover, it is absent in Central European and Northern Iberian Paleolithic/Mesolithic mitochondrial backgrounds [20], [23], [28]. The presence of “rare” paragroup N* in both Cardial and Epicardial samples from North Eastern Iberia and PPNB populations confirms the connection between both edges of the Neolithic expansion previously suggested in [27]. Haplogroup N1a, representing 12. 75% of LBK-AVK samples [19], [24], is not present in our PPNB sample, making it unlikely that this cluster was introduced from the earliest PPNB farmers of this region [23]. A more complex pattern for the LBK-AVK Neolithic expansion route, involving migration and admixture episodes with local hunter-gatherers in frontier zones (for example the predecessor populations of Starčevo-Criş-Körös cultures) should be considered in order to explain the available data for Neolithic populations of Central and Northern Europe. To solve this uncertainty, ancient DNA analysis from the Balkans region seems of vital importance. The signal of both rare N-derived haplogroups over the Neolithic genetic pool must have been erased by subsequent genetic drift events after the consolidation of Neolithic practices, as it has been suggested in other works [15], [27], [50]. In the absence of ancient genetic data from populations in the primary and secondary Neolithization areas, a detailed comparison of the genetic composition of the PPNB population with modern adjacent populations can shed light on possible routes of Neolithic expansion. As for modern Near Eastern populations, the frequency distribution of PPNB mitotypes in modern South Western European populations is limited (see Tables S5 and S7). However, strong genetic affinities at different levels of comparison could be detected with the islands of Cyprus and Crete (Figures 2,3, 4 and S2, Tables S5, S7 and S9), pointing out at a survival of ancient Neolithic genetic stock in these populations probably through endogamy and geographic isolation. The absence of an equivalent detectable genetic pattern in modern South-Western Anatolia suggests a primary role of pioneer seafaring colonization through Cyprus and the Aegean islands along the southern coast of Anatolia to the western coast of Greece. This observation is supported by three facts: An alternative scenario of land-mediated expansion through Western Anatolia would assume a survival of the genetic traits observed in the PPNB sample until the end of the period, when Middle-PPNB descendant populations would have expanded to secondary, adjacent areas of Neolithization around 7,500–7,000 years B. C. [57], [58]. This framework is not supported by the obtained data, but cannot be completely discarded as genetic drift or post-Neolithic genetic remodeling of the area might have erased ancient genetic signatures, as already stated from modern Near Eastern populations. Considering that the Neolithic expansion process was not uniform [59], the development of appropriate, spatially-explicit, model-based, statistical inference tools could be of great assistance in fully exploring the probabilities of these and other, competing demographic scenarios. In conclusion, the study of ancient DNA from the original geographic areas of Neolithic expansion performed here suggests a demic contribution of the first Near Eastern Neolithic in both main European routes of Neolithic expansion. Moreover, the population comparative analysis performed here points out at a leading role of seafaring colonization events in the first Neolithic expansions reaching the European continent. Further ancient DNA data from other primary and secondary areas of Neolithization and new data from frontier zones will be needed to add more resolution over the routes of expansion and the extent and nature of the genetic impact of the Neolithic over the European genetic pool. The studied material consisted of 63 ancient human skeletons from 3 different archaeological sites dating back to the PPNB time period (Table S10 and Figure 1). Tell Halula is located in the Middle Euphrates basin, 150 Km East of the city of Aleppo in the present territory of Syria. Excavations in the site, 8 hectares in area, have been in progress for the last 18 years by a Spanish Archaeological Mission in Syria. The excavations performed over an area of 2,500 m2 documented more than 40 occupation levels with thousands of stratigraphic units. A continuous occupation of the site can be assumed between 7,900–5,700 cal. B. C. , spanning from the PPNB to the Neolithic-Chalcolithic transition (Halaf and Obeid periods) [60]–[63]. PPNB occupation phases (I-XX) are located at the southern part of the Tell (sectors 2/4). Each phase is defined by successive human occupations followed by destruction/construction of habitation units. The houses were built one beside the other, oriented southward, and the deceased were buried by digging the graves in the floor of the house and covering them with a slide that allowed a clear association between the graves and the occupation floors. Most of the graves were located at the main entrance of each house, under a porch area. A total of 21 houses from PPNB levels have been unearthed to date, although only 14 of them have documented burial structures [64]. The skeletons analyzed in this paper belonged to occupational phases VIII-XIII on PPNB levels (7,500–7,300 cal. B. C.). The Tell Ramad archaeological site is located 20 Km south of Damascus on the slopes of Mount Hermon, in a basaltic plateau 830 m in height at the end of the river Wadi Sherkass, which flows in the Damascus basin. Human occupation was documented from early PPNB to ceramic Neolithic [65]. Radiocarbon dating of the site provided dates from 8,300 to 7,750 years B. P. for the PPNB levels (7,300–6,650 cal. B. C.) [66]. The burial model found at Tell Ramad is very similar to that in Tell Halula. The inhumations consisted of narrow tombs in the floor of the house, but evidence of common graves has also been documented. Dja' de El Mughara is located on the left bank of the Middle Euphrates, also in Syria. The excavations revealed three historical horizons corresponding to Early PPNB (9,400–8,700 B. P. , 8,700–8,250 cal. B. C.), Pre-Halaf ceramic Neolithic (7,700–7,400 B. P.) and Early Bronze Age (5,000 B. P.). The burial patterns found at the PPNB levels are very similar to those documented at Tell Halula and Tell Ramad with the graves located under the floor of the houses, but big collective funerary practices were also documented [67], [68]. Most samples from this site were collected by an experienced researcher in ancient DNA analysis (AP-P). The same person selected additional samples during the anthropological analyses. All the collected samples were neither washed nor treated after excavation. After collection, the samples were sent directly to the laboratory, where they were immediately studied. Contamination prevention measures were taken during all the selection processes, including the use of gloves and face shields. All the researchers involved in the handling of the samples during and after the excavation were typed for mtDNA (Table S11). Whenever possible two samples -preferably teeth- were taken from each individual. Clean and unbroken samples without visible fissures were selected, and then deposited in a sterile container until processed. The surface of all samples was removed with a sandblaster (Base 1 Plus, Dentalfarm) and subsequently UV-irradiated (254 nm) for 30 minutes on all sides. Cleaned samples were finally ground to a fine powder in a cryogenic impact grinder filled with liquid Nitrogen (Spex 6,700). Approximately 600 mg of the obtained powder was washed a minimum of 3 times with 8 ml 0. 5 M EDTA pH 8, and then incubated over-night at 37°C in 10 ml of lysis buffer solution (5 mM EDTA, 10 mM TRIS, 0. 5% SDS, 50 µg/ml proteinase K) in a hybridization oven. Tissue remains were removed by centrifugation and DNA was extracted from the supernatant with Phenol/Chloroform. The aqueous phase was concentrated by centrifugation dialysis using Centriplus 30,000 micro-concentrators (Millipore) and desalted with 15 ml sterile water (Braun) to a final volume of 300 µl. Extraction controls without powdered sample were processed in parallel to detect contamination during the extraction process. A region of 305 base pairs (bp) (np 16,095–16,399) of the mtDNA-HVS1 was amplified in the obtained extracts in two overlapping fragments. HVS1 fragment amplification was used as a screening method to detect the presence of amplifiable DNA in the studied samples. Samples were discarded if two consecutive amplifications produced no results. Two strategies were adopted for the HVS1 PCR amplification. In the laboratory at the Universitat de Barcelona, nested-PCR reactions using outer and inner primers (Table S12) were performed in a final volume of 25 µl with 5 µl of DNA extract, 1X Taq Expand High Fidelity PCR buffer (Roche), 2 mM MgCl2 (Roche), 0. 2 mM dNTP mix (Biotools), 0. 4 µM of each primer and 0. 06 U of Taq Expand High Fidelity (Roche). Nested amplification reactions were subjected to 30 cycles (first reaction) and 40 cycles (second reaction from the first amplified DNA solution) in a Perkin Elmer TC1 Thermocycler (94°C 60 s, 52°C 60 s and 72°C 60 s), with an initial denaturation step at 94 °C for 5 min and a final elongation step at 72 °C for 5 min. In the laboratory at Universidad Complutense de Madrid, one-round PCR reactions were set up in a final volume of 25 µl using the Multiplex PCR Kit (Qiagen) (5 µl of DNA extract, 1X Multiplex PCR Kit (Qiagen) and 0. 2 µM of each outer primer). This kit has proven to be extremely efficient for the amplification of ancient DNA [27], [69]. In this case, the cycling conditions using an Eppendorf Mastercycler consisted of 40 cycles of 30 s at 95°C, 90 s at 54°C and 60 s at 72°C, with a previous activation cycle of 15 min at 95°C and a final extension cycle of 10 min at 72°C. Amplicons were visualized in a 2% agarose gel stained with Ethidium Bromide and purification was performed directly from the amplification reaction using the Qiagen PCR purification Kit according to the manufacturer' s instructions. Sequencing reactions were performed with the Dye-Terminator Cycle Sequencing Reaction Kit vs 1. 2 (Applied Biosystems, Darmstadt, Germany) with one internal primer (L16125, H16259, L16257, H16370). Six microlitres of the PCR product were added to a final volume of 10 µl containing 3 µl of the kit and 16 pmol of the selected primer. Cycling sequencing was performed in an Eppendorf Mastercycler according to the supplier' s recommendations. Amplification products were analyzed on an automated sequencer ABI PRISM™ 310 (Applied Biosystems, Darmstadt, Germany). A detailed account of the extractions and amplifications performed can be seen in Table S2. MtDNA coding regions containing diagnostic SNPs were amplified at the Universidad Complutense de Madrid in monoplex reactions using primers described in Table S12 and the Multiplex PCR Kit (Qiagen) (5 µL of DNA, 1X Multiplex PCR Kit (Qiagen) and 0. 2 µM of each primer). The cycling conditions using an Eppendorf Mastercycler consisted on 40 cycles of 30 s at 94°C, 90 s at 50–54°C (see Table S12) and 60 s at 72°C, with a previous activation cycle of 15 min at 95°C and a final extension cycle of 10 min at 72°C. PCR products were purified and sequenced as it has been described above. Haplogroup diagnostic SNPs were typed at least in two separate extracts from the same skeleton in all the cases with the exception of skeleton H53 (Table S6). Only consistent HVS1 amplifications displaying the same mutation pattern between different extractions and PCRs were cloned using the pGEM-T Easy Vector System (Promega). PCR products were first incubated for 30 min with 0. 2 mM dATP, 1X PCR buffer, 2. 5 mM MgCl2 and 0. 1 U/µl Taq polymerase at 70°C in order to increase the ligation ratio. Three microlitres of the A-tailed products were ligated into pGEM-T Easy vector at 16°C overnight following manufacturer' s recommendations. Five microlitres of the ligation product were transformed into 100 µl of competent cells and the mixture directly plated on IPTG/X-Gal agar plates. Clones carrying PCR insert were selected by colony-PCR of white colonies (1X PCR buffer, 2 mM MgCl2,0. 2 mM dNTPs, 0. 4 µM each primer and 1. 5 U Taq polymerase, all from Biotools) using mitochondrial primers (Table S12). The cycling conditions in an Eppendorf Mastercycler were as follows: 94°C 10 min, followed by 30 cycles of 94°C 60 s, 52°C 60 s and 72°C 60 s, linked to a final extension step of 10 min at 72°C. Positive clones were grown in liquid Luria-Bertani broth and plasmidic DNA was purified using the Jetquick Plasmid Miniprep Spin Kit (Genycell, Granada, Spain). Cloned DNA was sequenced with universal primer SP6 or T7 as described above. Direct and clone sequences were aligned to the revised Cambridge Reference Sequence (rCRS, [70]) and differences were computed using the Mutation Surveyor software (Demo version 3. 24, SoftGenetics, LLC). Carry-over and cross-contamination were examined by comparing cloning results from the same extraction and amplification batch (Table S3). Consensus haplotypes were established from clone and direct sequences considering mutation reproducibility in different extractions/PCRs, concordance with SNP typing and potential contaminations, following the approach of [27]. Mitochondrial haplogroups were assigned to the amplified samples considering the information of both the HVS1 and the coding region SNPs according to the rCRS oriented version of PhyloTree Build 15 and Haplogrep [71], [72]. The number and type of miscoding lesions per sample were computed from the clone alignments manually in the PPNB sample excluding priming sites. The values were normalized by dividing for the number of clones per PCR and the number of sequenced base pairs. Mutations and insertions/deletions within poly-C tracts (positions 16,182–16,193) were not considered. To provide a frame of comparison for our results, miscoding lesion values were computed in the same way in the clone alignments of two datasets of Mesolithic and Early Neolithic temperate environments [27], [73]. A Taq-Man Real Time assay was used for the specific quantification of mtDNA HVS1 (np 16,103–16,233) in the obtained extracts using a Taq-Man-MGB probe 5′ - AATACTTGACCACCTGTAGTAC (np 16,138-16,159) and primers L16123 (forward) (5′ -ACTGCCAGCCACCATGAATATT, np 16,103–16,123) and H16209 (reverse) (5′ - TGGAGTTGCAGTTGATGTGTGA, np 16,209–16,233). PCR reactions were performed using TaqMan Universal PCR Master Mix (Applied Biosystems). The samples were loaded onto a standard 96-Well Reaction Plate (Applied Biosystems) and fluorescence detection was performed in a Sequence Detection System ABI Prism 7700 (Applied Biosystems). Four negative controls were included per plate. The DNA concentrations of the extracts were derived from comparison with serial dilutions of a known concentration of human mtDNA standard (103–109 copies equivalent to 3. 58×10−6 ng/µl and 3. 58 ng/µl). All extracts were quantified twice and the average values were considered. The following precautions and authenticity standards were observed for validating the obtained results: (1) Samples were collected on the field by staff trained in ancient DNA analysis. (2) Collected samples were unwashed to prevent pre-laboratory contamination. (3) All the analyses were performed in exclusive ancient DNA laboratories in which extraction, preparation of PCR reactions and post-PCR procedures were physically separated. (4) Access to extraction and PCR laboratories was restricted to two researchers (EF and CG), who wore clean-room protective clothes, gloves and facemasks. (5) The laboratories were routinely cleaned with bleach and UV-irradiated. (6) The samples and reagents were manipulated in laminar flow hoods, which were previously cleaned with bleach and irradiated with UV light (254 nm) for 30 minutes. (7) Exclusive material for ancient DNA analysis was employed in every experimental process. (8) Before the analysis, plastic material and pipettes were placed in the cabinet and UV-irradiated for 30 minutes. (9) All individuals were independently extracted at least twice in two independent laboratories except in two cases (see Tables S2, S3). (10) Each studied mtDNA fragment was amplified in separate laboratories at least twice. (11) Only extracts and amplicons from extraction and amplification groups providing negative results at the blanks were considered. (12) Reproducible direct sequences were cloned, and between 10 and 15 clones per amplicon were sequenced (Table S3). (13) The DNA amount in the DNA extracts was estimated by Real Time PCR (Table S1), providing in all cases a number of copies higher than 1,000. This result is high enough to guarantee sequence reproducibility [74]. (14) Obtained mtDNA sequences were compared to those from all the archaeologists (MM), anthropologists (AP-P, JA, IO) and researchers (EF, CG, MT, EP) involved in the retrieval or manipulation of the studied samples in order to detect pre-laboratory and laboratory contaminations. For additional security, staff working at the two laboratories involved in the analysis during this period was also typed (DT, JG, EA, AL, CB, JA) (Table S11). (15) Sequences deriving from the same and close extraction and amplification groups were compared to detect carry-over contaminations and non-consistent results were discarded. (16) Phylogenetic sense was observed between retrieved consensus mitochondrial haplotypes and SNP typing of mitochondrial haplogroups. These criteria not only meet but exceed in different aspects other ancient DNA reports from Neolithic populations [23]–[25], [30], [31]. A database of 9821 mtDNA-HVS1 haplotypes from 59 modern populations from the Near East and South Eastern Europe and 2 Early Neolithic datasets from Central Europe [24] and North Eastern Iberia [27], belonging respectively to LBK-AVK and Cardial/Epicardial Neolithic cultures, was constructed using published data. Sequence alignment tables were transformed into haplotypes using the program “Haplotyper” designed ad hoc (Python). Haplotypes were converted into sequences using Haplosearch [75] and used for calculations. An additional database of haplogroup frequencies was built using published haplogroup data of 11,610 individuals. The same populations used for the haplotype database were included when haplogroups where known. Haplogroup frequencies from other populations not including published haplotypes were also used. A description of the populations included in both databases is provided in Table S4. Geographic location of the modern populations of the database is shown in Figure S1. The 95% confident interval was calculated for the frequencies of the mitochondrial haplogroups found in the PPNB sample in the three ancient population datasets (PPNB, Cardial and LBK), in the three modern meta-populations (Near East, SW Europe, Africa and Caucasus) and in the pooled modern population using non-parametric bootstrap with replacement in SPSSvs21. 0 [76]. The number and percentage of shared haplotypes between our PPNB population and the other populations in the database, and the number and percentage of individuals in the database carrying PPNB haplotypes, were estimated using the Arlequin software, version 3. 5 [77]. Only information contained in the mtDNA fraction analyzed in our ancient population (np 16,126–16,369) was considered. Pairwise FST genetic distances [78], [79] were computed between our ancient dataset and the populations included in the haplotype database using the Arlequin software version 3. 5 [77]. Only the mtDNA fraction analyzed in our ancient population (np 16,126–16,369) was used for comparison. The significance of the genetic distances was tested by permuting the individuals between the populations 10,000 times. P values were adjusted post-hoc to correct for multiple comparisons with the Benjamini-Hochberg method [80] as suggested elsewhere [19] using the function p. adjust in R [81]. The percentage of individuals carrying PPNB haplotypes and the percentage of shared haplotypes and pairwise FST values calculated between the PPNB population and the other populations in the database were graphically plotted in a map using Surfer version 9 (Golden Software). Ethnic groups with disperse geographic location were not considered in the analysis (see Table S4). A PCA was performed using basal mtDNA haplogroup frequencies of the populations included in the database (see Table S7). Haplogroups with frequencies >1% in the studied regions were considered: H, HV, I, J, K, T, U*, U1, U2, U3, U4, U5, U6, U7, HV0 (including pre-V, V, HV0b, HV0c), W, X, N*, N1, N2, R0 (former pre-HV). Rare U and R European haplogroups were clustered into two groups: U+: U8, U9 and R+: R1, R2. African and Asian haplogroups were also grouped as follows: African haplogroups (AF): L0-L7, M1; Asian haplogroups (AS): A, B, C, D, E, F, G, M*, M3-48, N9, R5, R9, Q, Y, Z. HCA was performed over the six first principal components using Euclidean distances and Ward' s linkage algorithm [82]. Cluster partitioning was chosen looking at the shape of the obtained Hierarchical tree and examining the inertia between clusters/total inertia ratio. In the present study, a partition in 6 clusters was explored. An analysis of the variance was used to evaluate the distances between the clusters. The statistical program SPAD. N Ver. 5. 6 (Système Portable Pour L' Analyse de Donnés), DECISIA, France; [83] was used for both PCA and HCA analyses. The 15 mtDNA sequences reported in this paper have been deposited in Genbank with accession numbers KF601411- KF601425.

Since the original human expansions out of Africa 200,000 years ago, different prehistoric and historic migration events have taken place in Europe. Considering that the movement of the people implies a consequent movement of their genes, it is possible to estimate the impact of these migrations through the genetic analysis of human populations. Agricultural and husbandry practices originated 10,000 years ago in a region of the Near East known as the Fertile Crescent. According to the archaeological record this phenomenon, known as “Neolithic”, rapidly expanded from these territories into Europe. However, whether this diffusion was accompanied or not by human migrations is greatly debated. In the present work, mitochondrial DNA –a type of maternally inherited DNA located in the cell cytoplasm- from the first Near Eastern Neolithic populations was recovered and compared to available data from other Neolithic populations in Europe and also to modern populations from South Eastern Europe and the Near East. The obtained results show that substantial human migrations were involved in the Neolithic spread and suggest that the first Neolithic farmers entered Europe following a maritime route through Cyprus and the Aegean Islands.

Abstract Introduction Results Discussion Materials and Methods

haplotypes social sciences archaeology gene pool genetic polymorphism physical anthropology paleontology genetics biology and life sciences population genetics gene flow population biology evolutionary biology anthropology paleoanthropology evolutionary processes genetic drift

2014

Ancient DNA Analysis of 8000 B.C. Near Eastern Farmers Supports an Early Neolithic Pioneer Maritime Colonization of Mainland Europe through Cyprus and the Aegean Islands

Until recently, NADPH oxidase (NOX) enzymes were thought to be a property of multicellularity, where the reactive oxygen species (ROS) produced by NOX acts in signaling processes or in attacking invading microbes through oxidative damage. We demonstrate here that the unicellular yeast and opportunistic fungal pathogen Candida albicans is capable of a ROS burst using a member of the NOX enzyme family, which we identify as Fre8. C. albicans can exist in either a unicellular yeast-like budding form or as filamentous multicellular hyphae or pseudohyphae, and the ROS burst of Fre8 begins as cells transition to the hyphal state. Fre8 is induced during hyphal morphogenesis and specifically produces ROS at the growing tip of the polarized cell. The superoxide dismutase Sod5 is co-induced with Fre8 and our findings are consistent with a model in which extracellular Sod5 acts as partner for Fre8, converting Fre8-derived superoxide to the diffusible H2O2 molecule. Mutants of fre8Δ/Δ exhibit a morphogenesis defect in vitro and are specifically impaired in development or maintenance of elongated hyphae, a defect that is rescued by exogenous sources of H2O2. A fre8Δ/Δ deficiency in hyphal development was similarly observed in vivo during C. albicans invasion of the kidney in a mouse model for disseminated candidiasis. Moreover C. albicans fre8Δ/Δ mutants showed defects in a rat catheter model for biofilms. Together these studies demonstrate that like multicellular organisms, C. albicans expresses NOX to produce ROS and this ROS helps drive fungal morphogenesis in the animal host. Reactive oxygen species (ROS) including superoxide anion and hydrogen peroxide play diverse roles in biology. ROS can inflict severe oxidative damage to cellular components, but when carefully controlled, ROS can also be used to combat infection and act in cell signaling processes. A well-studied example of controlled ROS production involves NADPH oxidase (NOX) enzymes [1]. These heme and flavin containing enzymes use electrons from NADPH to reduce molecular oxygen to superoxide [1]. In macrophages and neutrophils, NOX enzymes generate bursts of superoxide in the extracellular milieu or phagolysosomal compartments to assault microbial pathogens. In non-immune cells, ROS from NOX enzymes are widely used in cell signaling pathways to promote growth, development and differentiation [1]. As membrane proteins, NOX enzymes can vectorially release superoxide inside the cell or extracellularly and in either case, the superoxide can react with neighboring superoxide dismutase (SOD) enzymes that disproportionate superoxide to oxygen and hydrogen peroxide. In fact, NOX enzymes often partner with SODs in signaling processes, whereby SOD converts the cell impermeable superoxide to the diffusible hydrogen peroxide signaling molecule [1–5]. NOX-SOD interactions are also prevalent during infection where the microbial pathogen uses its arsenal of extracellular SODs to combat the oxidative burst of host NOX enzymes [6]. The opportunistic fungal pathogen Candida albicans has evolved with a family of three extracellular SOD enzymes (Sod4, Sod5, Sod6) believed to protect the fungus from the attack of host NOX-derived superoxide [7,8]. We recently reported that these extracellular SODs represent a novel class of Cu-only SOD enzymes that are unique to the fungal kingdom and oomycetes [9,10]. Much of what is known about fungal Cu-only SODs has emerged from studies on C. albicans Sod5. Sod5 can react with superoxide at rates limited only by diffusion [9,10], and can effectively degrade superoxide radicals derived from macrophage and neutrophil NOX enzymes [11,12]. Curiously C. albicans Sod5 appears specific to the filamentous form of the fungus [7,13]. C. albicans is a polymorphic fungus that can transition from unicellular yeast-like form to pseudo hyphal and true hyphal filamentous states [14,15]; Sod5 is evidently absent in the yeast-form of C. albicans. The rationale for selective expression of Sod5 during morphogenesis was not clear, as both yeast and hyphal forms exist in the animal host, are subject to immune surveillance and are essential for virulence [15–17]. Moreover, SOD5 is induced in filamentous C. albicans in the absence of any insult from the host [7,13]. This raises the possibility that filamentous C. albicans witness a source of superoxide not seen in the yeast-like form. Certain multicellular fungi are capable of generating superoxide themselves using fungal NOX enzymes as part of signaling during differentiation [18–20]. However, unicellular yeasts were believed to not express NOX, as NOX was characterized as a property of multicellular differentiation [21,22]. This dogma of no NOX in unicellular fungi was recently challenged by the identification of Saccharomyces cerevisiae Yno1, a NOX that localizes to the endoplasmic reticulum and generates intracellular (not extracellular) superoxide [23]. Other than Yno1, there has been no evidence for NOX enzymes in evolutionarily related yeasts including C. albicans. Here we provide the first evidence for a NOX enzyme in the opportunistic fungal pathogen C. albicans. This NOX, known as Fre8, produces a burst of extracellular ROS in filamentous but not yeast-form cells. We demonstrate that Fre8-superoxide serves as substrate for Sod5, providing a rationale for inducing this extracellular SOD during morphogenesis. Strikingly, the ROS from Fre8 is concentrated at the growing tip of C. albicans hyphae and can promote formation and/or maintenance of elongated hyphae in vitro as well as in infected kidneys during a mouse model for disseminated candidiasis. Moreover, Fre8 enhances C. albicans survival in a rat venous catheter model of candidiasis. These studies show that host NOX is not the only source of ROS at the host-pathogen interface; C. albicans makes its own ROS for hyphal morphogenesis through Fre8. Previously, Schroter et al. demonstrated that C. albicans produces ROS during the transition from the yeast form to the hyphal state [24]. To probe this fungal ROS, we used luminol, a chemiluminescence probe typically used to measure ROS bursts in macrophages and neutrophils [11,12,25,26]. Consistent with findings by Schroter et al [24], C. albicans cells induced to form hyphae by serum treatment exhibited a burst in luminol chemiluminesence not seen in yeast-form cells (Fig 1A). We observed that this ROS is not unique to serum stimulation but is also seen when morphogenesis is induced by elevated amino acid concentrations in IMDM medium and alkaline pH conditions (Fig 1A) [14,27]. The luminol probe used in Fig 1A is not expected to penetrate the fungal cell wall, and should therefore only detect extracellular ROS. This notion of extracellular ROS was corroborated using the chemiluminescence probe lucigenin that cannot cross cell membranes, and is specific for superoxide compared to luminol which can detect both superoxide and hydrogen peroxide [25,28]. As seen in Fig 1B, cells induced to form hyphae exhibited a defined lucigenin signal within 15 min of morphogenesis. However, in WT cells the lucigenin signal often declined at later times points (60 min, Fig 1B), and we tested whether this reflected induction of extracellular SOD enzymes. Indeed the superoxide signal from lucigenin was enhanced in sod4Δ/Δ sod5Δ/Δ sod6Δ/Δ cells lacking all three extracellular SODs (60 min, Fig 1B) and the same was true with luminol chemiluminescence (Fig 2B). Of the three extracellular SODs, deletion of SOD5 alone was sufficient to enhance ROS during morphogenesis stimulated by serum (Fig 2A) or IMDM (Fig 2B), consistent with the notion that Sod5 is the major extracellular SOD induced during hyphal formation [7]. Compared to effects of sod5Δ/Δ mutations, there was no change in luminol ROS in sod1Δ/Δ mutants lacking the major intracellular Sod1 (Fig 2C). Together these studies demonstrate that cells undergoing morphogenesis produce a burst of extracellular ROS including superoxide that can serve as substrate for extracellular Sod5. In multicellular organisms, extracellular ROS is derived from NOX enzymes, although C. albicans was not previously known to express NOX. To address whether the ROS burst of morphogenesis was derived from a NOX enzyme, we used DPI (diphenylene iodonium), a classical inhibitor of NOX enzymes [29–31]. As seen in Fig 2D, there was a dose response inhibition of the ROS burst of C. albicans using DPI, with full inhibition at 0. 5 μM. DPI also eliminated the enhanced ROS of sod5Δ mutants (Fig 2C). These studies suggested that C. albicans expresses a NOX enzyme for extracellular ROS during morphogenesis. However, since DPI can also inhibit other flavin containing enzymes [32], we applied molecular genetic approaches to examine the source of the ROS burst. In the fungal kingdom, NOX enzymes are part of an expanded family of NADPH oxidoreductases that use electrons from NADPH to either reduce oxygen to superoxide (NOX enzymes) or reduce ferric or cupric metal ions (FRE enzymes) [33]. NOX and FREs are highly similar and it is difficult to predict functionality based on sequence analysis alone [23,33,34]. Yeasts are generally thought to only express FRE, not NOX [21,22] although as mentioned above, this dogma was challenged by identification of S. cerevisiae Yno1 as an endoplasmic reticulum NOX [23]. In C. albicans, there are at least 17 genes annotated as FREs [35,36], three of which are known cupric or ferric reductases (Fre1, Fre7, Fre10 (S1 Fig, [37–39]); the remainder have uncharacterized functions. By qRT-PCR, we identified a number that are induced during early stages of hyphal morphogenesis coincident with the ROS burst (S1 Fig). The C. albicans orthologue to S. cerevisiaeYNO1 [18] was not among the FREs induced with morphogenesis (S1 Fig). The most highly induced hyphal specific gene was FRE8, also known as CFL11 or C. albicans CR_06670W, orf19. 701 (S1 Fig). FRE8 was additionally reported as the most abundantly induced FRE during C. albicans invasion of the kidney [40]. We chose to focus on FRE8 as a potential NOX enzyme. Recombinant versions of codon optimized FRE8 were expressed in Pichia pastoris under control of the methanol inducible AOX2 promoter. An identical procedure has been used to express and analyze activity of mammalian NOX enzyme complexes [41]. In parallel, we expressed a second unknown member of the FRE family, namely FRP1, that is only moderately induced by hyphal stimulation (S1 Fig). P. pastoris cells expressing recombinant FRE8 and FRP1 were assessed for ferric reductase activity and ROS production. As seen in Fig 3A left, Pichia cells expressing FRP1 exhibited clear ferric reductase activity when FRP1 expression was induced with methanol. These cells however, exhibited no ROS that could be detected by luminol (Fig 3A right), indicating that FRP1 encodes a metalloreductase, and not a NOX enzyme. The opposite profile was obtained with Pichia cells expressing FRE8. These cells exhibited no methanol-inducible ferric reductase activity (Fig 3B left), but produced a strong methanol-inducible ROS signal (Fig 3B right). This ROS was eliminated by addition of 0. 5 μM of the NOX inhibitor DPI or by addition of exogenous SOD enzyme (bovine SOD1). By contrast the ferric reductase activity of recombinant Frp1 was not altered by exogenous SOD or by DPI (Fig 3A left). These results indicated that FRP1 encodes a metalloreductase while FRE8 encodes a NOX. Fre8 contains all the features predicted for a member of the NADPH family of oxidoreductases, including seven transmembrane domains and sequences for binding heme, FAD and NADPH (Fig 3C). To assess the function of Fre8 in vivo, we deleted both copies of genomic FRE8 in C. albicans and tested effects on metals and ROS formation. We observed no impact of fre8Δ/Δ mutations on accumulation of copper or iron (Fig 3D) or on whole cell ferric reductase activity (Fig 3E). However, the fre8Δ/Δ strain was completely defective in generating ROS, and this result held true regardless of the stimuli for morphogenesis including serum (Fig 4A), alkaline medium, IMDM and spider medium (Fig 4B). The heterozygous fre8Δ/+ strain retaining a single genomic copy of FRE8 exhibited roughly a 50% reduction in ROS generation and the same haploinsufficiency was seen with the fre8Δ/Δ strain complemented with a single copy of FRE8 (Fig 4A). As previously mentioned (Fig 2), the ROS burst of hyphal cells is enhanced in a sod5Δ/Δ mutant, and as shown in Fig 4C, this elevated ROS is Fre8-mediated, as no ROS is detected in a double sod5Δ/Δ fre8Δ/Δ mutant. The ROS burst emitted by C. albicans during the morphogenic switch is clearly Fre8-dependent. We observed that FRE8 mRNA is induced within 1 hour of serum treatment (Fig 5A), congruent with the ROS burst (Fig 1A) and the induction of SOD5 (Fig 5A). In C. albicans, morphogenesis involves complex signaling pathways that converge on the Efg1 and Cph1 transcription factors, and efg1Δ/Δ cph1Δ/Δ null cells are incapable of forming hyphae [14,42]. We observed that efg1Δ/Δ cph1Δ/Δ mutations block induction of FRE8 and SOD5 by serum (Fig 5A) and accordingly, the ROS burst is also eliminated (Fig 5B). SOD5 has previously been shown to fall under control of Efg1 [13]. SOD5 and FRE8 were not identified by ChIP as direct targets of Efg1 or Cph1 [43] [44]; both are subject to chromatin remodeling control by Hir1 that works in concert with Efg1 to control genes for morphogenesis [45]. Is hyphal formation required for ROS production? We uncoupled FRE8 expression from morphogenesis through ectopic expression of FRE8 under control of the repressible MET3 promoter in cph1Δ/Δ efg1Δ/Δ cells, fre8Δ/Δ cells and WT C. albicans. Strains were grown under conditions favoring yeast-only growth, and then FRE8 expression was de-repressed by removal of methionine. As seen in Fig 5C–5E there was a burst of ROS from all cells in which FRE8 expression was de-repressed. Importantly, these cells remained in the budding yeast-form, with no evidence of hyphal forms or germ tubes (Fig 5C–5E bottom). Thus the ROS burst is not dependent on morphogenesis, but rather the expression of FRE8, regardless of the morphogenic state. This study also demonstrated that Fre8-ROS is not sufficient to induce hyphal formation. We next examined localization of Fre8-dependent ROS using nitrobluetetrazolium (NBT) which has been used to localize NOX superoxide in multicellular fungi [46,47]. NBT is reduced by superoxide, forming a purple formazan precipitate. As seen in Fig 6, intense NBT staining was observed at the tip of elongating germ tubes in WT cells but not fre8Δ/Δ mutants (Fig 6A). This staining was discernable within 30 min of stimulating morphogenesis (Fig 6B). Thus, Fre8-dependent ROS is specific to the growing tip of the polarized cell. Why would C. albicans produce ROS during hyphal formation? We considered whether this ROS might act as a signal to modulate morphogenesis, as has been shown for ROS modulating development in differentiating fungi [18–20]. Although no detectable change in morphology could be noted in the NBT experiment of Fig 6, these experiments were conducted at early time points when cells were initially forming germ tubes. However, at later time points when cell formed elongated hyphae, a fre8Δ/Δ deficiency could be observed (Fig 7A). This fre8Δ/Δ defect was complemented in the FRE8 re-integrant (Fig 8A top) and appeared completely specific to later stages of hyphal development. After 3–4 hours when WT cells were assembling into elongated hyphae, fre8Δ/Δ cells accumulated abundant yeast-forms (Fig 7C). The experiments of Fig 7 employed cultures seeded at 6 x 107 cells/ml, a relatively high density where quorum sensing is prominent [48]. However, quorum sensing cannot explain the fre8Δ/Δ defect, as these mutants showed no increased sensitivity to the quorum sensing molecule farnesol (S2 Fig parts A, B), and high density cultures of fre8Δ/Δ cells exhibited WT-like quorum sensing properties [48] (S2 Fig part C and legend). Most importantly, the fre8Δ/Δ defect in hyphal development can also be seen in low density cultures not subject to quorum sensing, e. g. , when 4 x 106 cells/ml are stimulated with serum at 34°C (Fig 8B–8D). It is important to note that the requirement for Fre8 in hyphal formation is not absolute, and can be bypassed by potent stimuli for morphogenesis, e. g. , high levels of serum (Fig 8A bottom) or when low density cells are shifted to temperatures ≥37°C (S2 Fig Part C). We conclude that Fre8 can act as a modifier of hyphal development, but is not unconditionally essential for the process. Since Sod5 reacts with Fre8 superoxide to produce H2O2, we tested whether sod5Δ/Δ mutations likewise affect hyphal morphogenesis. As seen in Fig 9A and 9B, sod5Δ/Δ mutants exhibited a defect in hyphal development that was less pronounced than that of fre8Δ/Δ cells. Such an intermediate effect could be expected since Fre8 superoxide may also be converted to H2O2 through spontaneous disproportionation [49] or through extracellular Sod4 and Sod6. Even so, the parallel trends seen with sod5 and fre8 mutations would imply that H2O2 (and not superoxide) underlies the Fre8-defect. To more definitely test this, we addressed whether exogenous H2O2 could bypass the requirement for Fre8 in hyphal development. Previous studies have shown that mM concentrations of H2O2 can induce hyperpolarized buds in C. albicans [50,51], or pseudohyphae [52,53], neither of which resemble true C. albicans hyphae. Rather than using a single bolus of H2O2 as was previously done, we used glucose oxidase (GO) to continuously generate exogenous H2O2. As seen in Fig 9C, as little as 0. 1 mU of GO (generates 100 pmoles H2O2/min) was able to restore hyphal development to fre8Δ/Δ cells while heat-inactivated GO was without effect. It is important to note that the hyphae formed with GO treated fre8Δ/Δ cells were indistinguishable from that of WT cells (Fig 9D), unlike the elongated buds and pseudohyphae reported for C. albicans treated with mM H2O2 [50–53]. Together, our in vitro studies of Figs 7–9 support a model in which the H2O2 produced by Fre8 can act as a modifier of hyphal morphogenesis. It was important to examine the impact of Fre8 in vivo, as the animal host is the only natural environment for C. albicans. One model examined was the mouse model for disseminated candidiasis where kidney is the target organ. In late stages of infection, kidneys were harvested from mice infected with WT versus fre8Δ/Δ strains and subjected to histological analysis of invading fungi by PAS staining. As seen in Fig 10A top, WT C. albicans predominantly showed elongated hyphal filaments in the infected kidney. By comparison, the fre8Δ/Δ mutant from 4 independent mice produced a mixture of morphological forms with a much larger proportion of shorter filaments or yeast form cells in fre8Δ/Δ mutants compared to WT C. albicans (Fig 10A and 10B). These findings are similar to fre8Δ/Δ effects on morphology in vitro (Figs 7–9). Yet in spite of the changes in morphology observed in vivo, there was no overall impact on pathogenesis with fre8Δ/Δ mutants. fre8Δ/Δ mutants showed no deficiency in virulence (Fig 10C), and markers of host inflammation [54,55] were similar between SC5314 and fre8Δ/Δ infected kidneys (S3 Fig part A). Colony forming units (CFUs) in the kidney 48 hours post infection were modestly (2–3 fold) lower in fre8Δ/Δ mutations relative to WT (S3 Fig part B). All morphological forms are thought to contribute to infection and invasion [14], and our data with fre8Δ/Δ mutants supports this view. We additionally tested the fre8Δ/Δ mutant in a rodent model of catheter biofilms. C. albicans is capable of forming surface adherent aggregates of biofilms on either biological surfaces (e. g. , epithelial cells) or on medical implant devices, and such dense fungal communities are highly tolerant to antifungals [56]. One of the most common clinical biofilm infections involves venous catheter implants and a rodent vascular catheter model faithfully recapitulates the human disease [56]. In this model, WT SC5314 C. albicans forms robust biofilms within 24 hours of injection into the animal catheter (Fig 11A–11C). Over three independent trials, fre8Δ/Δ cells exhibited deficiencies whereby the biofilms were either sparse in number (Fig 11A) or undetected (Fig 11C), and when present, biofilms were often attenuated with few elongated hyphae (Fig 11B). This defect was partially reversed with the Fre8 re-integrant (Fig 11C), consistent with the haploinsufficiency and partial restoration of ROS formation in vitro (Fig 4A). The fre8Δ/Δ defect in biofilms in vivo may very well reflect changes in morphogenesis similar to what we observed in vitro (Figs 7–9). Yet host factors may also contribute. Neutrophils represent the primary leukocytes of Candida biofilms in catheters [57] and we tested whether fre8Δ/Δ cells were more sensitive to neutrophil killing. As seen in Fig 12A, the total mass of C. albicans biofilms in vitro in the absence of any host cells was unchanged in fre8Δ/Δ cells, indicating that there is no primary defect in biofilm formation including adherence. By comparison, fre8Δ/Δ biofilms exhibited a consistent increase in killing by neutrophils (Fig 12B). Thus, the fre8Δ/Δ defect with in vivo biofilms may not reflect a deficit in biofilm formation but rather increased clearance by host immune mechanisms including neutrophils. We tested whether this increased killing was unique to neutrophils or could be extended to other phagocytes, e. g. , macrophages. Our control for macrophage studies was the sod5Δ/Δ C. albicans strain that has been previously shown to be sensitive to macrophage killing due to an inability to degrade host superoxide [11]. We observe that both sod5Δ/Δ and fre8Δ/Δ mutants show statistically significant increases in killing by bone marrow derived macrophages (BMDM) (Fig 12C). The impact of Fre8 derived ROS appears to extend beyond the hyphal morphology effects seen in fungal-only cultures and interactions with host cells are also important. NOX enzymes have evolved to intentionally produce ROS, and until recently, were believed to be a property of multicellular differentiation [21,22]. The discovery of Yno1 in S. cerevisiae demonstrated that a unicellular fungi can produce ROS through NOX, although the ROS in this case was found to be intracellular [23,34]. Here we describe C. albicans Fre8 as the second example of NOX in an organism that can grow as a unicellular yeast and the first for this opportunistic fungal pathogen. Moreover, unlike S. cerevisiae Yno1, C. albicans Fre8 is capable of producing extracellular ROS, akin to NOX enzymes in multicellular organisms [1]. In animal cells, NOX enzymes can partner with extracellular SODs that convert the extracellular superoxide free radical to the diffusible H2O2 molecule [1–5]. Likewise C. albicans Fre8 appears to partner with extracellular Sod5, providing a rationale for expressing Sod5 only in hyphal cells [7]. These studies also provide a new twist to Sod5 function at the host-pathogen interface. While Sod5 can clearly react with superoxide from macrophage and neutrophil NOX enzymes [11,12], our studies here with C. albicans Fre8 indicate that the superoxide for Sod5 is not just coming from the host. It is conceivable that when hyphal cells interact with neutrophils or macrophages, that a “superoxide superstorm” ensues with ROS coming from both the sides of the host-pathogen axis, and with Sod5 operating in the middle. Why does C. albicans produce ROS during hyphal morphogenesis? Lessons may be taken from multicellular fungi or fruiting body fungi that use NOX derived ROS for morphogenesis and differentiation [18–20]. For example, ROS from these fungal NOX have been implicated in calcium signaling [58], MAP kinase and Rac1 GTPase signaling [46,59] and cell re-modeling involving cytoskeleton effects [23,60]. We show that Fre8 derived H2O2 can modulate morphogenesis. Work in other systems has shown that NOX derived H2O2 targets multiple redox sensitive molecules, including protein tyrosine phosphatases, casein kinases with peroxide sensitive degrons; even actin itself can be modulated by oxidation [1,5, 61,62]. Since Fre8 ROS is specifically seen at the growing tip of developing hyphae, the H2O2 produced may act locally on one or more redox sensitive targets that promote polarized growth. The precise mechanism of Fre8 control of hyphal biology is the subject of ongoing investigations. In addition to the morphology defects of fre8Δ/Δ mutants in vitro and in vivo, we observed that these cells are more sensitive to killing by neutrophils and macrophages in vitro. It is possible that the morphological changes in fre8Δ/Δ may somehow render these fungal cells more susceptible to attack by phagocytes. As an alternative possibility, the ROS from Fre8 may help condition cells for the oxidative attack by immune cells. It has been proposed that low dose exposures of C. albicans to H2O2 or to ROS from macrophages can induce adaptive mechanisms to guard against subsequent oxidative insults [63–66]. FRE8 ROS may promote such adaptation against neutrophil and macrophage attack. In future studies, it will be important to determine the impact of fre8Δ/Δ mutations in immunocompromised settings. Is Fre8 the only NOX of C. albicans? This organism has evolved with a very large family of 17 NOX/FRE enzymes, four of which are metalloreductases (including Frp1 characterized here), leaving 12 with unknown functions [35–39]. The extracellular ROS burst studied here is completely eliminated in fre8Δ/Δ cells suggesting that Fre8 is the only NOX for extracellular ROS in C. albicans at least under these in vitro conditions. During C. albicans invasion of the kidney, Fre8 is the most abundantly expressed member of the FRE/NOX family [40]. Even so, it is possible that other members of this family induced during fungal infection and hyphal morphogenesis such as Fre2 (S1 Fig) may similarly function in a NOX capacity, perhaps secondary to Fre8 [40]. C. albicans may also express NOX enzymes for intracellular ROS analogous to Yno1 of S. cerevisiae. With such a large family of NOX/FRE enzymes, we speculate that additional NOX enzymes will come to light as mediators of ROS signaling in C. albicans. Regardless, it will be of interest to integrate Fre8-ROS signaling into known pathways of hyphal regulation. Cultures of Candida albicans cells were typically maintained at 30°C in a yeast extract, peptone based medium (YPD) with 2% (wt/vol) glucose, conditions which support the budding yeast-form of the fungus. The C. albicans fre8Δ/Δ and sod5Δ/Δ strains used in these studies grew identical to WT SC5314 in the yeast-form (S4 Fig). To induce hyphal morphogenesis, yeast-form cells were harvested, starved for 30 min in sterile H2O at 30°C, followed by harvesting and induction of hyphal formation by incubating at 37°C (or 34°C, see below) in various media known to stimulate hyphal formation, including Iscove’s Modified Dulbecco’s Medium (IMDM; Gibco), alkaline YPD (50 mM glycine, pH 9. 5), spider medium (1% nutrient broth, 1% mannitol, 11. 5 mM potassium phosphate, pH 7. 2) or YPD with 5–20% fetal bovine serum (heat inactivated, Corning/Cellgro). Hyphal morphogenesis was stimulated in either “low density” (optical density, OD600 = 0. 1–0. 2) or “high density” (OD600 = 3. 0) conditions. In studies of hyphal morphology, yeast-form cells were cultured to OD600 ≈ 8. 0, conditions where all cells grew identically (S4 Fig), and were stimulated to form hyphae with YPD-serum. The level of serum used to investigate the fre8Δ/Δ defect ranged from 5–15% depending on the lot of serum, with low density cultures typically requiring less serum and temperatures of 34°C to demonstrate a dependence on serum for hyphal morphogenesis. Where indicated, cultures were supplemented with 0. 1–10 mUnits glucose oxidase (Type II Sigma#G6125) to bypass the fre8Δ/Δ defect in hyphal development. Experiments involving yeast-form cells expressing FRE8 under the MET3 promoter used a synthetic complete (SC) based medium containing 0. 67% yeast nitrogen base lacking cysteine and either containing or lacking 85. 6 mg/L methionine. Cell were seeded at OD600 = 0. 1 and grown for 1–7 hours. All C. albicans strains used in this study were isogenic to SC5314 or its derivative CA-IF100 (arg4Δ/arg4Δ, leu2Δ/leu2Δ: : cmLEU2, his1Δ/his1Δ: : cdHIS1, URA3/ura3Δ). The sod1Δ/Δ, sod5Δ/Δ and sod4Δ/Δ sod5Δ/Δ sod6Δ/Δ strains derived from CA-IF100 were kind gifts of Karl Kuchler as previously described [11]. The cph1Δ/Δ efg1Δ/Δ strain derived from SC5314 was a gift from Gerald Fink [42]. Mutations in FRE8 and SOD5 were introduced in SC5314 using the SAT1-flipper cassette method [67]. Deletion in a single FRE8 allele was achieved using plasmid pJGFRE8LKO, in which FRE8 regions -926 to -581 and +2402 to +2802 were inserted into the Kpn1 and XhoI and the NotI and SacI sites respectively of pSFS2 [67]. Following liberation of the cassette by KpnI and SacI digestion and transformation of SC5314 by electroporation, accurate deletion of a single FRE8 allele was verified by PCR, generating the fre8Δ/+ mutant strain CA-JG201. The second FRE8 allele was deleted similarly using a pSFS2 construct, pJGFRE8SKO, containing FRE8 sequences -587 to -3 and +2075 to +2427, creating the fre8Δ/fre8Δ strain CA-JG211. Homozygous sod5Δ/Δ mutations were introduced in either SC5314 (generating strain CA-JG201) or CA-JG211 (generating CA-JG221) using a construct containing SOD5–492 to +53 and +808 to +1253 inserted into the Kpn1 and XhoI and the Not I and SacI sites respectively of pSFS2 [67]. Deletion of both copies of sod5Δ/Δ in strain CA-JG201 and CAJG-221 was verified by PCR. A single copy of FRE8 was introduced into the fre8Δ/Δ strain CA-JG211 as follows: FRE8 sequences -926–+2802 were inserted into the KpnI and XhoI sites of pJGFRE8LKO. Integration into the FRE8 locus at position -926 to +2802 was achieved by transformation of the cassette liberated by digestion with KpnI and SacI generating the fre8Δ/Δ: FRE8 re-integrant strain CA-JG231. To create the construct for expressing FRE8 under control of the MET3 repressible promoter, C. albicans MET3 sequences -1643 to -1 were inserted into Sph1 and Nhe1 sites engineered at FRE8 position -1 in the pJGFRE8LKO re-integrant plasmid described above. Following digestion with Kpn1 and Sac1, the MET3-FRE8 containing cassette was used to transform the cph1Δ/Δ efg1Δ/Δ strain, the fre8Δ/Δ strain CA-JG211 and SC5314 by electroporation. Accurate integration at the FRE8 locus -926 to +2802 was verified by PCR. Expression of recombinant FRE8 and FRP1 in Pichia pastoris used the PichiaPink Strain 1: ade2 (Thermo Fisher Scientific). P. pastoris cells were maintained in YP-Gal medium (1% yeast extract, 2% peptone, 2% galactose). Protein expression experiments used a buffered YP medium (1% yeast extract, 2% peptone, 100 mM potassium phosphate, pH 6. 0,1. 34% yeast nitrogen base, 0. 00004% biotin) that was supplemented with either 0. 5% methanol to induce protein expression under the AOX2 promoter or with 1% glycerol for non-inducing conditions. The plasmid for expressing FRE8 or FRP1 under the P. pastoris AOX2 promoter represented a modified version of pPINK α-HC (Thermo Fisher Scientific) in which a 10X HIS tag was introduced downstream of the α-factor pre-sequence (plasmid pRPp718, kind gift of Ryan Peterson). Following the insertion of a Afe1 site downstream of the HIS tag, FRP1 sequences +1 to +1665 and FRE8 +1 to +2220 were inserted into the Afe1 and Fse1 sites of this expression plasmid, creating in-frame fusions to the N-terminus secretion sequence and HIS tag. The CTG codons from both genes were altered to TCG for optimal expression in P. pastoris; plasmids were linearized by digestion with Spe1 and integrated into the TRP2 locus by transformation. For luminol and lucigenin measurements of ROS, either yeast-form cells (grown in YPD to OD600 of 1. 0–2. 0) or cells induced to form hyphae as described above were used. Cells were harvested, washed, and suspended in a final OD600 of 0. 2 in Hanks buffered saline solution (HBSS) containing 0. 2 mM luminol (Cayman chemicals) and 0. 5 units/ml horseradish peroxidase. In studies with lucigenin, cells were first washed in 25 mM glycine pH 9. 5,0. 5% glucose prior to resuspending in 200 μl of the same alkaline buffer containing 5 μM lucigenin. Samples were analyzed for luminol or lucigenin chemiluminescence in 96 well plates using a BioTek Synergy HT plate reader. Analysis was carried out over 1. 5 hours at 37°C with a gain setting at 120–135 and integration time of 1. 0 second. Results were plotted according to relative luminescence units (RLU) per 0. 04 OD600 units of cells. With experiments involving DPI, 5 μl of DMSO containing the indicated amount of DPI (or no DPI as control) was added to the reaction at time zero. For qRT-PCR analysis of fungal-only cultures, 50 ml cultures of cells were induced to form hyphae for 1 hr by growth in YPD-10% FBS (as described above); these early hyphal cells or the control yeast-form were harvested, washed and RNA prepared by the hot acid phenol method [68]. cDNA was prepared using the Maxima H Minus First Strand cDNA Synthesis Kit (ThermoFisher Scientific) and qRT PCR carried out using iTaq Universal SYBR Green Supermix (Bio-Rad). Values were normalized to TUB2 and graphed according to the fold change in FRE8 and SOD5 expression in early hyphal versus yeast-form cells. Amplicons of ≈150 residues were prepared using primers as described in S1 Table. For analysis of inflammatory mRNA markers (TNF-α, IL-17a, IL-6), RNA from whole kidneys was extracted as previously described [69]. cDNA prepared from 5. 0 μg of RNA was diluted 1: 50 prior to PCR analysis as above. Values were normalized to ActB and graphed according to the fold change over uninfected controls. Primers for host mRNA analyses are listed in S1 Table. For ferric reductase and NOX activity analyses in P. pastori transformants, cells were grown overnight in 10 ml YP-Gal, washed twice in either glycerol or methanol containing buffered media (described above) and resuspended at an OD600 of 0. 1 in 15 mls of the same medium. Following growth at 30°C for 6 hrs, cells were harvested and washed in either HBSS for the luminol assay or 50 mM citrate, pH 6. 6,5% glucose for the ferric reductase assay. Cells were subjected to luminol chemilumiscence precisely as described above for C. albicans. Compared to C. albicans assays, the luminol substrate appears rapidly depleted in P. pastoris expressing high levels of FRE8. For the ferric reductase assay, cells at a OD600 of 0. 5 were incubated in 200 μl of a reaction containing 1 mM FeCl3 and 1 mM bathophenanthrolinedisulfonic acid (BPS) in 50 mM citrate, pH 6. 6,5% glucose. Absorbance at 515 or 520 nm was read in 96 well plates on a BioTek Synergy HT plate reader over 1. 5 hours at 30°C. Where designated, 0. 1 U of bovine Cu/Zn SOD1 or 50 nM DPI were added to the luminol or ferric reductase assay at t = 0. Ferric reductase measurements in C. albicans cells was conducted similarly, using cells induced to form hyphae in IMDM for 1 hr as described above and assayed for ferric reductase using the same conditions described for P. pastoris except C. albicans cells were assayed at OD600 of 0. 1. Total cellular accumulation of copper and iron was measured by inductively coupled plasma mass spectrometry (ICP-MS) using C. albicans cells induced to form hyphae for 1 hr in 10% FBS as described above. Cells were washed twice with 10 mM Tris, 1 mM EDTA, pH 8 and twice with MiliQ deionized water. Cell pellets containing 10. 0 OD600 units of cells were resuspended in 500 μl of 20% nitric acid and digested by incubation at 90°C overnight. Samples were diluted 10-fold in MiliQ deionized water and subjected to elemental analysis on a Agilent 7700x ICP-MS instrument. In vitro biofilms were grown in the wells of 96-well microtiter plates, as previously described [70]. Briefly, C. albicans resuspended in RPMI-MOPS at 1. 5 x 106 cells/ml (200μL/well) was added, and incubated for 24 hours at 37°C with 5% CO2. To assess biofilm burden an XTT (2,3-Bis- (2-Methoxy-4-Nitro-5-Sulfophenyl) -2H-Tetrazolium-5-Carboxanilide) assay was performed as an estimate of viable burden, as previously described [70]. For assays involving neutrophils, human neutrophils were collected as follows: Blood was obtained from volunteer donors with written informed consent through a protocol approved by the University of Wisconsin Internal Review Board (IRB). Primary human neutrophils were purified by negative antibody selection using the MACSxpress Neutrophil Isolation and MACSxpress Erythrocyte Depletion kits (Miltenyi Biotec Inc. , Auburn, CA), as previously described [71]. Experiments with neutrophils were performed in RPMI 1640 (without phenol red) supplemented with 2% heat-inactivated fetal bovine serum (FBS) and glutamine (0. 3 mg/ml). Incubations were at 37°C with 5% CO2. An adaptation of the XTT metabolic assay was used to estimate C. albicans viability following co-culture with neutrophils [71]. Following a 24 h incubation period, biofilms were washed with DPBS and neutrophils were added at 1. 5 x 106 cells/ml, which represented an effector: target of 1: 2. Following a 4 h incubation, 90 μL of 9: 1 XTT working solution (0. 75 mg/ml XTT in DPBS with 2% glucose: phenazine methosulfate 0. 32 mg/ml in ddH2O) was added to each well. After a 25 minutes incubation, samples were transferred to a Falcon 96 well U bottom plate and centrifuged at 1,200×g for three minutes to pellet cells. Supernatants (110 μl) were then transferred to a 96 well flat bottom plate for absorption reading at 492 nm. A neutrophil only control was used to subtract their contribution to the XTT values. To determine percent killing, values were compared to wells without neutrophils after subtraction of the baseline absorbance. Macrophage infection assays used bone-marrow derived macrophages (BMDM) isolated from the marrow of hind leg bones of 5- to 8-wk-old C57BL-6 female mice. For differentiation, cells were seeded in 100 mm treated cell culture dishes (Corning, Corning, NY) in Dulbecco’s Modified Eagle medium (DMEM; Corning) with 20% L-929 cell-conditioned medium, 10% FBS (Atlanta Biologicals, Flowery Branch, GA), 2mM Glutamax (Gibco, Gaithersburg MD), 1% nonessential amino acids (Cellgro, Manassas, VA), 1% HEPES buffer, 1% penicillin-streptomycin and 0. 1% 2-mercaptoethanol for 6–7 days at 37°C with 9. 5% CO2. 105 BMDM were seeded on 96 well plates and activated by incubating overnight using 100 U/ml of IFN-γ (Roche, Indianapolis, IN). C. albicans obtained from overnight cultures in YPD (OD600 = 8. 0) and starved in water as for hyphal morphogenesis studies (see above) were washed twice with PBS and incubated for 30 min at 37°C with Guinea pig complement (MP biomedicals, LLC, OH) for opsonization. The fungus was then added to macrophages at a multiplicity of infection (MOI) ratio of 1: 10 for 4 hours. After incubation, the media was removed and macrophages lysed in water. Fungal viability was assessed by the XTT assay according to Pierce et al [72]. The same XTT assay was used to determined fungal viability following farnesol treatment. For the murine model of disseminated candidiasis, ten male BALB/c mice (10 weeks old) per strain were inoculated with 2x105 C. albicans cells of WT SC5314, the fre8Δ/Δ strain or the fre8Δ/Δ strain complemented by FRE8 by lateral tail vein injection. Moribund mice were sacrificed by CO2 asphyxiation and immediately dissected for harvesting kidneys for histology (see below). Fungal burden and host inflammatory markers were analyzed following 48 hours of infection. The spleen and one kidney was processed for CFUs as previously described [69]. The other kidney was placed in 500 μL Trizol and frozen at -80°C for subsequent RNA analyses (see above). Mouse survival was plotted using a log rank test (Mantel Cox) to query any statistical difference. A jugular vein rat central venous catheter biofilm infection model was used as previously described [73]. Briefly, 24 h following surgical implantation of a jugular venous catheter, C. albicans at 106 cells/ml was instilled in the catheter lumen and flushed after 6 h. After 24 h biofilm growth period, catheters were harvested and fixed overnight (4% formaldehyde, 1% glutaraldehyde, in PBS). They were then washed with PBS, treated with 1% osmium tetroxide, and washed again. Samples were dehydrated through series of ethanol washes followed by critical point drying and mounted on aluminum stubs. Following sputter coating with platinum, samples were imaged in a scanning electron microscope (LEO 1530) at 3kV. For NBT staining and microscopic analyses of cell morphology, C. albicans cells were induced to form hyphae as described above using YPD containing 10% FBS (in the case of NBT staining). Cells were harvested, washed once with HBSS and resuspended in 1 ml HBSS containing 0. 05% nitroblue tetrazolium (NBT). Following an incubation for 30 min in the dark at 37°C, cells were washed 1X with HBSS, 1X with 70% Ethanol and resuspended in 200 μl 50% Glycerol/HBSS. Cells were visualized by light microscopy at 100x magnification on a Zeiss Axio ImagerA2 microscope. For analysis of hyphal morphogenesis, dark field microscopy of live C. albicans cells was accomplished using a Nikon Infinity 1 microscope at 40x magnification. Where indicated, enumeration of cells was carried with culture aliquots first passed through 26 gauge and 31 gauge needles to help break up dense aggregates and enhance visualization of individual cells. Passage through these needles did not affect integrity of the individual cells. To analyze C. albicans morphology in infected kidneys, freshly harvested kidneys from infected mice were flash frozen in Tissue Tek O. C. T. compound in a dry ice/ethanol bath and were sectioned to 20 μM thickness by cryotome. Tissue slices were adhered to Superfrost Plus Microscope Slides (Fisherbrand Cat. No. 12-550-15) and subjected to Periodic Acid Schiff (PAS) staining by treatment with 0. 5% periodic acid (Sigma) for 5 minutes, rinsing briefly with distilled water, then staining 5 minutes with Schiff’s Reagent (Sigma Aldrich). Following a 5 min rinse with water, the mounted tissue was dehydrated using successive 2 min treatments with 50%, 70%, 80%, and twice 95% and 100% ethanol, followed by three 2 min treatment with xylene isomer mixture (Sigma Aldrich) to remove residual ethanol. Cover slips were then mounted with Permount (Fisher) and slides then imaged on a microscope at 40X magnification. All experiments involving animals were approved by the Johns Hopkins University (protocols # MO16M168 and MO15H134) and University of Wisconsin (protocol # DA0031, MV1947) Institutional Animal Care and Use Committees according to guidelines established by the Animal Welfare Act, The Institute of Laboratory Animal Resources Guide for the Care and Use of Laboratory Animals, and the Public Health Service Policy. Experiments involving neutrophils were approved by IRB (protocol #2013 1758) and involved cells isolated from healthy human adult donors in which written informed consent was obtained at the time of blood draw, following the guidelines and approval of the University of Wisconsin-Madison Center for Health Sciences Human Subjects Committee.

We demonstrate here that the opportunistic human fungal pathogen Candida albicans uses a NADPH oxidase enzyme (NOX) and reactive oxygen species (ROS) to control morphogenesis in an animal host. C. albicans was not previously known to express NOX enzymes as these were thought to be a property of multicellular organisms, not unicellular yeasts. We describe here the identification of C. albicans Fre8 as the first NOX enzyme that can produce extracellular ROS in a unicellular yeast. C. albicans can exist as either a unicellular yeast or as multicellular elongated hyphae, and Fre8 is specially expressed during transition to the hyphal state where it works to produce ROS at the growing tip of the polarized cell. C. albicans cells lacking Fre8 exhibit a deficiency in elongated hyphae during fungal invasion of the kidney in a mouse model for systemic candidiasis. Moreover, Fre8 is required for fungal survival in a rodent model for catheter biofilms. These findings implicate a role for fungal derived ROS in controlling morphogenesis of this important fungal pathogen for public health.

Abstract Introduction Results Discussion Materials and methods

biofilms blood cells medicine and health sciences immune cells pathology and laboratory medicine chemical compounds oxides pathogens immunology microbiology animal models developmental biology fungi model organisms experimental organism systems kidneys morphogenesis fungal pathogens neutrophils research and analysis methods mycology white blood cells animal cells medical microbiology microbial pathogens chemistry mouse models yeast candida macrophages eukaryota cell biology anatomy biology and life sciences yeast and fungal models physical sciences cellular types renal system superoxides organisms candida albicans

2017

Candida albicans FRE8 encodes a member of the NADPH oxidase family that produces a burst of ROS during fungal morphogenesis

The Vibrio cholerae type VI secretion system (T6SS) assembles as a molecular syringe that injects toxic protein effectors into both eukaryotic and prokaryotic cells. We previously reported that the V. cholerae O37 serogroup strain V52 maintains a constitutively active T6SS to kill other Gram-negative bacteria while being immune to attack by kin bacteria. The pandemic O1 El Tor V. cholerae strain C6706 is T6SS-silent under laboratory conditions as it does not produce T6SS structural components and effectors, and fails to kill Escherichia coli prey. Yet, C6706 exhibits full resistance when approached by T6SS-active V52. These findings suggested that an active T6SS is not required for immunity against T6SS-mediated virulence. Here, we describe a dual expression profile of the T6SS immunity protein-encoding genes tsiV1, tsiV2, and tsiV3 that provides pandemic V. cholerae strains with T6SS immunity and allows T6SS-silent strains to maintain immunity against attacks by T6SS-active bacterial neighbors. The dual expression profile allows transcription of the three genes encoding immunity proteins independently of other T6SS proteins encoded within the same operon. One of these immunity proteins, TsiV2, protects against the T6SS effector VasX which is encoded immediately upstream of tsiV2. VasX is a secreted, lipid-binding protein that we previously characterized with respect to T6SS-mediated virulence towards the social amoeba Dictyostelium discoideum. Our data suggest the presence of an internal promoter in the open reading frame of vasX that drives expression of the downstream gene tsiV2. Furthermore, VasX is shown to act in conjunction with VasW, an accessory protein to VasX, to compromise the inner membrane of prokaryotic target cells. The dual regulatory profile of the T6SS immunity protein-encoding genes tsiV1, tsiV2, and tsiV3 permits V. cholerae to tightly control T6SS gene expression while maintaining immunity to T6SS activity. Vibrio cholerae is the etiological agent of the diarrheal disease cholera. This pathogen utilizes a wide array of virulence factors during host infection including the well-characterized cholera toxin (CT) and toxin-coregulated pilus (TCP). In addition, V. cholerae possesses numerous other virulence factors including the type VI secretion system (T6SS), a recently described mechanism used by numerous Gram-negative bacteria to export effectors across their cell envelopes. In contrast to TCP and CT, whose presence is restricted to a subset of virulent V. cholerae strains, the V. cholerae T6SS is highly conserved and is present in strains of all serogroups sequenced to date. Structurally, the T6SS resembles an inverted bacteriophage; it assembles in the V. cholerae cytoplasm and docks onto a baseplate complex situated in the bacterial envelope. Two alleles on the small and large chromosome encode hemolysin-coregulated protein (Hcp) [1], which polymerizes [2] and forms the inner tube of the T6SS injectosome and acts as a chaperone for T6SS effectors [3]. Contraction of the outer sheath (formed by VipA and VipB) around the formed Hcp nanotube leads to the ejection of the Hcp tube decorated with a VgrG trimer consisting of three different VgrG proteins: VgrG-1, VgrG-2, and VgrG-3 [4]–[10]. VgrG-1 and VgrG-3 carry enzymatic C-terminal extensions that crosslink actin or degrade the peptidoglycan layer, respectively, upon translocation into target cells [4], [6], [10]–[12]. ATP hydrolysis by the inner membrane protein VasK provides the energy for Hcp secretion [13]. A recent report suggests that proteins from the PAAR (proline-alanine-alanine-arginine) repeat superfamily form a sharp conical extension on the VgrG cap. This extension is able to load additional T6SS effectors besides the VgrGs, which are then delivered simultaneously into target cells in a single contraction-driven translocation event [14]. The O37 serogroup V. cholerae strain V52 [15] constitutively synthesizes T6SS structural proteins and effectors and actively engages in T6SS-mediated virulence under standard laboratory conditions [4], [5], [16]. Conversely, the pandemic-causing O1 serogroup strain C6706 possesses a full complement of T6SS-encoding genes, but does not express T6SS genes encoding structural apparatus components and effectors under laboratory conditions. Null-mutations in the genes encoding the quorum sensing regulator LuxO and the global regulator TsrA lead to production and secretion of the T6SS protein Hcp, and T6SS-mediated virulence in V. cholerae strain C6706, demonstrating that this pandemic strain employs a tightly controlled, fully active T6SS [17]. The V. cholerae T6SS is encoded by three separate gene clusters [5] (Figure 1A). One source of T6SS regulation, VasH (encoded by VCA0117), is encoded in all V. cholerae genomes sequenced so far and acts as an activator of the alternate sigma factor 54. A regulatory role for VasH in the expression of genes within the two auxiliary T6SS gene clusters encoding (i) Hcp-1 and VgrG-1, and (ii) VgrG-2 and VasX has been established in V. cholerae El Tor and classical O1, and non-O1/non-O139 strains [18]–[20]. T6SSs mediate toxicity towards both eukaryotes and prokaryotes [4], [5], [16], [21]–[32]. The V. cholerae T6SS mediates virulence towards murine macrophages and the amoeboid host Dictyostelium discoideum [4], [5], [16], [18], [21] as well as a variety of Gram-negative bacteria including Escherichia coli, Salmonella Typhimurium, and Citrobacter rodentium [16], [33]. Importantly, V52 bacteria do not kill each other, and some bacteria such as V. cholerae C6706 are resistant to killing by V52 [16]. Although all sequenced V. cholerae strains possess a full complement of T6SS-encoding genes, only some strains are known to express these genes under laboratory conditions. V52 has a constitutively active T6SS and produces hallmark T6SS proteins such as Hcp and VgrGs when grown in liquid culture [5]. In contrast, C6706 maintains a T6SS-silent state and does not produce Hcp and VgrGs under standard laboratory conditions [34]; yet C6706 is immune to killing by V52 [11], [16]. T6SS-mediated killing of prokaryotes has been reported in other bacterial species including Pseudomonas aeruginosa, Serratia marcescens, and Burkholderia thailandensis [23], [25], [26], [35], [36]. The P. aeruginosa T6SS encodes effectors as part of a toxin/antitoxin (TA) gene system in which the effector (toxin) is injected into the target cell where it degrades the peptidoglycan in the target cell unless inhibited by an immunity protein (antitoxin) [35]. TA systems are not restricted to T6SS gene clusters, and one other notable TA system subset are colicins—small secreted peptides produced by and toxic to certain strains of E. coli [37]. Colicins employ a diverse range of mechanisms to kill target cells, including degradation of RNA and DNA [38]–[44], inhibition of murein synthesis [45], and pore-formation in the inner membrane [46]–[49]. In the latter case, the pore in the inner membrane forms a voltage-dependent ion channel that dissipates membrane potential and halts cellular respiration [48]–[55]. Pore-formation in the inner membrane can occur only if colicin is presented to the target cell from the periplasmic face [56], [57]. Consequently, cytoplasmic production of colicins is not toxic to the producing cell [57]. We previously identified the protein VasX and characterized this T6SS effector with respect to its role in T6SS-mediated killing of D. discoideum [21]. In this study, we characterize the mechanism by which VasX acts as a bacterial effector. We also present evidence that VasW, the product of the gene directly upstream of vasX, is necessary for VasX-mediated bacterial killing. By screening a C6706 T6SS transposon library [58], we identified the VasX immunity protein-encoding gene VCA0021 located directly downstream of vasX, and two other T6SS-encoding genes, VCA0124 and VC1419, that encode additional immunity proteins. We demonstrate that all three immunity protein-encoding genes are controlled in a dual regulatory fashion, ensuring that even T6SS-silent strains such as C6706 are protected from T6SS-mediated killing by neighboring bacterial competitors (Figure 1B). The strains and plasmids used in this study are listed in Table 1 and Table 2, respectively. A streptomycin-resistant V. cholerae strain V52 (O37 serogroup) lacking hapA, rtxA, and hlyA genes was used as a T6SS-active strain in all experiments in this study unless otherwise noted in the figure legend. A streptomycin- and rifampicin-resistant V. cholerae strain C6706 (O1 serogroup) was used as a T6SS-silent strain in all experiments. DH5α λpir and SM10 λpir were used for cloning, and mating of pWM91-based plasmids, respectively. Unless stated otherwise, bacteria were grown in Luria-Bertani (LB) broth at 37°C with shaking (200 rpm). Antibiotic concentrations used were 100 µg·mL−1 ampicillin, 10 µg·mL−1 chloramphenicol, 100 µg·mL−1 streptomycin, 50 µg·mL−1 kanamycin, and 50 µg·mL−1 rifampicin. Arabinose was added at a final concentration of 0. 1% to induce expression from the PBAD promoter. In-frame deletion of T6SS-encoding genes was performed as described previously [59]. Construction of the vasX, vasH, and vasK knockout constructs was also previously described [5], [18], [21]. VasW, vgrG-3, tseL, tsiV1, tsiV2, and tsiV3 knockout constructs were created using the primers listed in Table 3. PCR products resulting from primer combinations A/B and C/D were stitched together by overlapping PCR. The resulting knockout construct was digested with BamHI and cloned into the suicide plasmid pWM91 [59]. V52ΔvgrG-3 [5] was used as the parent strain in which vasW was deleted to create the V52ΔvgrG-3ΔvasW double mutant. V52ΔvasH was used as the parent strain in which vasX was deleted to create the V52ΔvasHΔvasX double mutant. Construction of plasmid pBAD24-vasX was described previously [21]. VasX: : FLAG and vasW: : FLAG were PCR-amplified using the primers listed in Table 3 with Accuprime Pfx DNA polymerase (Invitrogen) and cloned into pCR2. 1 TOPO TA (Invitrogen) following 10 minute incubation at 72°C with TopTaq polymerase (Qiagen) to add 3′ adenine overhangs. VasX: : FLAG was excised using KpnI and XbaI and subcloned into pBAD24. VasW: : FLAG was excised using EcoRI and XbaI and subcloned into pBAD24 [60]. The VasX immunity protein-encoding gene tsiV2 was PCR amplified with a C-terminal 6×His tag using Phusion Taq (Thermo Fisher) and ligated into pJET1. 2/blunt (Thermo Fisher). The fragment was excised with KpnI and XbaI restriction enzymes, subcloned into pBAD33 [60], and transformed into E. coli Top10 (Invitrogen). A triparental mating technique [61] was used to introduce pBAD33-tsiV2: : 6×His into V. parahaemolyticus RIMD2210633. Top10/pBAD33-tsiV2: : 6×His (CmR), V. parahaemolyticus (RifR), and DH5α/pRK2013 (helper plasmid containing tra and mob genes) were mixed at a 1∶1∶1 ratio, spotted onto nonselective LB plates, and incubated for 6 hours at 37°C. The spot was harvested into 1 mL LB and pelleted at 13,000 rpm for 5 minutes. Cells were resuspended in 100 µL LB and plated on thiosulfate-citrate-bile salts-sucrose (TCBS) agar plates containing 50 µg·mL−1 rifampicin and 10 µg·mL−1 chloramphenicol to select for V. parahaemolyticus/pBAD33-tsiV2: : 6×His. Plates were incubated overnight at 37°C. Resulting colonies were restreaked onto TCBSrif50chlor10 to exclude false-positive transformants. Construction of the periplasmic expression vectors pBAD24-LS and pBAD24-SecP: : core was described previously [12]. The vasX gene was cloned in-frame with the sequence encoding a sec signal peptide (SecP) using XbaI and HindIII restriction sites. The resulting construct was transformed into E. coli Top10 (Invitrogen) for expression analysis and then into the indicated Vibrio strains. Plasmid pAH6 [62] was a generous gift from Dr. Jun Zhu (University of Pennsylvania School of Medicine) and contains a promoter-less lacZ gene. VasX (4-3258,4-1345,1575-3258,2208-3258), tseL, vgrG-3, and Phcp were amplified using Phusion Taq (Thermo Fisher), cloned into the blunt-ended cloning vector pJET1. 2/blunt (Thermo Fisher), and transformed into E. coli Top10 (Invitrogen). All vasX constructs, Phcp, and tseL were excised using XbaI and HindIII restriction sites and subcloned into pAH6. VgrG-3 was excised using the XbaI restriction site and subcloned into pAH6. The resulting constructs were transformed into E. coli Top10, followed by V. cholerae V52, V52ΔvasH, and C6706. Plasmids pBAD24-tsiV1: : FLAG, pBAD24-tsiV2: : FLAG, and pBAD24-tsiV3: : FLAG were created using the Gateway recombination system. Briefly, the LR recombination site, including the ccdB negative selection sequence, was amplified out of pBAD-DEST-49 using the primers listed in Table 2 (forward primer adds a KpnI site, reverse primer adds a stop codon, XbaI site, and FLAG epitope tag sequence). The PCR product was digested with XbaI and KpnI and ligated into the corresponding sites of pBAD24 to create pBAD24gw: : FLAG. The Gateway recombination reaction between pDONR221-tsiV2 (Harvard Institute of Proteomics) and pBAD24gw: : FLAG was performed according to the manufacturer' s protocol (Invitrogen, Carlsbad, CA). The resulting plasmids were transformed into E. coli Top10, followed by transformation into V. cholerae C6706ΔtsiV1, C6706ΔtsiV2, C6706ΔtsiV3, C6706ΔvasX, C6706ΔtseL, or C6706ΔvgrG-3. Bioinformatics were performed using BLASTn (nucleotide) and BLASTp (protein) algorithms (http: //blast. ncbi. nlm. nih. gov/Blast. cgi). Statistical significance was determined using the Student' s one-tailed, paired t-test with a significance cut-off of p<0. 05. Killing assays were performed as described previously [16]. Briefly, bacterial strains were grown overnight on selective LB plates and resuspended in LB broth (or LB 3% NaCl for V. parahaemolyticus). Predator and prey were mixed at a 10∶1 ratio and spotted onto prewarmed LB agar plates and incubated at 37°C for 4 hours. Spots were harvested, serially diluted, and 10 µl of each dilution was spotted onto a selective LB plate. Plates were incubated overnight at 37°C and surviving prey (CFU·mL−1) were enumerated. The competitive index was calculated by dividing prey CFUs after exposure to V52 by prey CFUs following exposure to V52ΔvasK. An overnight culture of C6706ΔtsiV2/pBAD24-vasX was back-diluted 1∶100 in selective LB broth in the presence or absence of 0. 1% arabinose to induce expression of vasX. Strains were grown in a 96 well plate on a Heidolph Titromax 1000 vibrating shaker at 900 rpm. OD600 readings were taken every 30 min for 8 h using a BioRad XMark microplate spectrophotometer. At the 7-hour time point, a cell lysate sample was collected from each sample, mixed with sodium dodecyl sulfate (SDS) protein sample buffer, and boiled for 10 min. The protein samples were subjected to SDS-PAGE followed by western blotting to detect VasX and DnaK. The OD600 was measured of overnight liquid cultures diluted 1/10 in Z buffer (0. 06 M Na2HPO4,0. 04 M NaH2PO4,0. 01 M KCl, 0. 001 M MgSO4, pH 7. 0) containing 2. 7% β-mercaptoethanol. To lyse bacteria, 10 µL of 0. 1% SDS and 20 µL of chloroform were added and the mixture was vortexed for 5 sec. Tubes were incubated in a 28°C waterbath for 10 min, followed by addition of 200 µL ortho-nitrophenyl-β-galactoside (ONPG) buffer (4 mg·mL−1 ONPG dissolved in Z buffer using a sonicating waterbath). The time required for development of a yellow color was recorded and the reaction was halted by the addition of 500 µL of 1 M Na2CO3. The OD420 of each sample was measured using a BioRad XMark microplate spectrophotometer. Miller units were calculated based on the equation (OD420/ (OD600·time·volume) ×103). Strains V52, C6706, and V52ΔvasH used in this assay each possess a chromosomal copy of lacZ. Strains O395w, NIH41w, MAK757w, C6709w, and N16961w contain a disrupted lacZ gene. Mutation of lacZ was accomplished using the suicide vector pJL1. E. coli SM10 λpir/pJL1 was mixed at a ∼1∶1 ratio with the recipient V. cholerae strains and incubated on a prewarmed LB agar plate for 6 hours at 37°C. Bacterial mixtures were harvested and resuspended in 1 mL LB and subjected to 4 serial dilutions; 200 µL of diluted sample were spread onto selective LB agar plates. Transconjugants were restreaked onto LB containing ampicillin and 40 µg·mL−1 5-bromo-4-chloro-3-indolyl-β-D-galactopyranoside (x-gal) to ensure disruption of lacZ resulted in white colony formation. Strains grown overnight were back-diluted 1∶100 in selective LB broth with 0. 1% arabinose. At 0,2, 4,6, and 8 h, a 200 µL sample was taken and serially diluted. 10 µL of each dilution was spotted onto selective LB plates and incubated overnight at 37°C. The BacLight Bacterial Membrane Potential Kit (Molecular Probes, Invitrogen) was used to determine whether VasX, when provided with a Sec signal peptide (SecP) to export SecP: : VasX to the periplasm, caused dissipation of the membrane potential. Overnight cultures of the indicated strains were back-diluted 1∶30 in selective LB with arabinose (0. 1%) and grown for 2 h at 37°C with shaking (225 rpm). Cells were diluted to ∼106 cells·mL−1 in filtered PBS. Carbonyl cyanide m-chlorophenyl hydrazone (CCCP, depolarizing control) and 3,3-diethyloxacarbocyanine iodide (DiOC2 (3) ) stain were added (where appropriate) to final concentrations of 5 µM and 15 µM, respectively, and incubated for 30 min in the dark. Stained cells were analyzed using a BD LSR II flow cytometer with a 488 nm laser and PE-Texas Red (601 long pass and 616/23 band pass filters) and FITC (502 long pass and 525/50 band pass filters) detectors. Forward and side scatter and fluorescence data were collected with logarithmic signal amplification. Red/green ratios were calculated based on double positive (Texas Red and FITC) cells by collecting 50,000 events per strain tested. Overnight bacterial cultures were diluted 1∶100 in LB containing appropriate antibiotics in the presence of 0. 1% arabinose. Cultures were grown under inducing conditions for 2 hours at 37°C with shaking. As a positive control for propidium iodide staining, dead cells were prepared by ethanol treatment: 1 mL of bacterial culture was centrifuged at 13,000 rpm for 3 minutes. The supernatant was aspirated and the pellet was resuspended in 1 mL of 70% ethanol and incubated for 20 minutes at room temperature. Cells were pelleted by centrifugation at 13,000 rpm for 3 minutes and the pellet was resuspended in 1 mL filtered PBS. ∼106 cells (living or dead) were diluted into 1 mL of filtered PBS. 1. 5 µL of 20 mM PI was added to each sample followed by incubation for 15 minutes in the dark. Stained cells were analyzed using a BD LSRFortessa Cell Analyzer with a 488 nm laser, and a PE (550 long pass and 575/26 band pass filters) detector. Forward and side scatter and fluorescence were collected with logarithmic signal amplification. The percentage of cells permeable to PI was recorded by collecting 50,000 events per strain tested. Overnight bacterial cultures were diluted 1∶100 in LB (+/−0. 1% arabinose) and grown to late-logarithmic phase. Bacterial pellet samples were resuspended in SDS protein sample buffer and boiled for 10 min. Culture supernatants were isolated, filter sterilized, and concentrated with 20% trichloroacetic acid. Supernatant proteins were resuspended in SDS protein sample buffer and boiled for 10 minutes. Samples were subjected to SDS-PAGE followed by western blotting with α-Hcp [5] and α-DnaK (loading and lysis controls, obtained from Stressgen) antibodies. V. cholerae V52 uses its constitutively active T6SS to kill other Gram-negative bacteria such as E. coli strain MG1655 [16]. Interestingly, two 7th pandemic El Tor strains C6706 and N16961 (with silent T6SSs), as well as V52 are resistant to this killing phenotype. C6706 and N16961 contain a full complement of T6SS-encoding genes; however, when spotted on nutrient agar plates, these strains do not produce structural components (e. g. , Hcp and VgrGs) required for the assembly of a functional T6SS. Therefore, C6706 and N16961 do not engage in T6SS-mediated killing of prokaryotes or eukaryotes under laboratory conditions [18], [34], [63]. We hypothesized that C6706 and N16961 (and other V. cholerae strains with silent T6SSs) encode immunity proteins within theT6SS gene clusters that are produced in a silent (T6SS-off) state. Dong et al. applied a high-throughput parallel sequencing technique to address this problem and identified the T6SS immunity protein-encoding genes, tsiV1 (VC1419), tsiV2 (VCA0021), and tsiV3 (VCA0124), along with their corresponding T6SS effector genes tseL (VC1418), vasX (VCA0020), and vgrG-3 (VCA0123), respectively [11]. We independently identified the same immunity genes by screening a subset of a transposon library of C6706 [58] to identify insertions within the three T6SS gene clusters that result in the loss of T6SS immunity against V52. Mutation of VCA0021 (tsiV2), VCA0124 (tsiV3), and VC1419 (tsiV1) [isolated as a polar mutation in VC1418 (tseL) ] resulted in sensitivity to killing by V52 (Figure S1). We noted that mutants with transposons in VCA0114, VCA0115, and VCA0121 also became susceptible to killing by V52. However, we decided not to analyze these three genes further because deletion of these genes in strain V52 [64], which employs a constitutively active T6SS, does not interfere with the protection from killing by V52 kin bacteria. Episomal expression of VCA0021 (tsiV2) and VCA0124 (tsiV3) protects the respective C6706 mutants from T6SS-mediated killing by V52 (Figures 2A and 2B). Complementation of the VC1418 null-mutation did not restore immunity (Figure S2). However, episomal expression of VC1419 in C6706ΔVC1418 restored immunity to killing by V52 (Figure S2), implying that VC1419 is the third immunity protein-encoding gene and confirming the findings by Dong et al. [11]. Subsequently, in-frame deletion of VC1419 rendered C6706 susceptible to killing by V52 and this could be complemented by expression of VC1419 in-trans (Figure 2B). The bacterial killing mechanism of VasX has not been characterized. Thus, we focused on the T6SS gene cluster encoding vasX and tsiV2 to understand the mode-of-action of VasX and the ability of C6706 to employ TsiV2 to achieve immunity to VasX-mediated toxicity in the T6SS-off state. Secondary structure predictions indicated that VasX has homology to colicins [64]. Because we previously observed that V52 kills E. coli MG1655 in a T6SS-dependent manner, we first tested whether VasX was required for this killing phenotype. As shown in Figure 3A, V52 with a disabled T6SS due to a vasK null-mutation (V52ΔvasK) was unable to kill E. coli; however, V52 with a native vasK but lacking vasX retained the ability to kill E. coli to the same extent as wild-type V52. Thus, we concluded that VasX is dispensable for V52' s ability to kill E. coli, probably because the other two T6SS effectors, VgrG-3 and TseL, compensate for the absence of VasX. To test whether VasX is important for killing E. coli in the absence of VgrG-3 and TseL, we performed a killing assay with a V52 predator that does not produce TseL and VgrG3, and relies on VasX for T6SS-mediated virulence. We observed that this predator strain was able to reduce the number of viable E. coli by ∼10-fold compared to fully attenuated V52ΔvasK or V52 missing all three T6SS effector genes (Figure S3). Although predator strains using VasX or VgrG-3 as the only T6SS toxin exhibited a similar Hcp secretion defect, their killing ability differed significantly (Figure S3), suggesting that VasX contributes to killing of E. coli. In conclusion, these data suggest that VasX is sufficient to kill E. coli when it is the sole effector utilized by V52, but VgrG-3 and TseL can compensate for the lack of VasX toxicity. We determined that V. cholerae C6706 lacking tsiV2, the gene encoding the immunity protein for VasX, became susceptible to killing by V52 (Figure 3A). This killing occurred in a VasX-dependent manner such that TsiV2 conferred immunity specifically to VasX (Figure 3A). We also tested whether VasX was involved in killing of other Vibrio species, namely Vibrio parahaemolyticus strain RIMD2210633 (hereafter referred to as V. parahaemolyticus RIMD). V. parahaemolyticus RIMD encodes a functional T6SS [65], [66]; however, the ability of this strain to engage in T6SS-dependent killing of bacterial neighbors has yet to be determined. As shown in Figure 3A, VasX is required for maximal killing of V. parahaemolyticus RIMD as a vasX mutant of V52, still capable of employing the T6SS effectors TseL and VgrG-3, is attenuated with regard to the killing of V. parahaemolyticus RIMD. As expected, episomal expression of vasX complemented the killing phenotype. Other Vibrio species including Vibrio alginolyticus and Vibrio fischeri were also susceptible to VasX-mediated killing (data not shown). Episomal expression of the immunity protein-encoding gene tsiV2 in V. parahaemolyticus RIMD provided intermediate protection against killing by V52 (Figure S4). In contrast, production of TsiV2 in MG1655 did not provide protection against VasX-mediated killing (data not shown). Taken together, these data indicate that although VasX is not required for V52 to kill E. coli MG1655 due to the compensatory effects of VgrG-3 and TseL, VasX is important for killing other Vibrio species. The V. cholerae VasX protein is important for T6SS-dependent killing of other bacteria [11] (Figure 3A) and has homology to colicins [64]. Colicins can kill target cells by degrading DNA and RNA, by forming pores in the inner membrane, or by inhibiting murein synthesis. Previous bioinformatic analysis suggested that VasX has three transmembrane domains in its C-terminus [21]; therefore, we hypothesized that the toxic activity of VasX involves damage to the inner membrane of target cells. To be toxic, pore-forming colicins must be presented to the target cell from the periplasmic face [56], [57]. Fittingly, we observed that VasX expression was not toxic towards V. cholerae C6706 lacking tsiV2 (Figure 3B). To ensure that VasX was produced, we collected a lysate sample from each sample presented in Figure 3B and performed a western blot with antibodies against VasX and DnaK (loading control) (Figure S5). As we detected VasX under inducing conditions in live bacteria, we reasoned that VasX is not toxic to the producing cell, because to be toxic it has to be presented to the inner membrane of target cells from the periplasmic space similar to pore-forming colicins. We then determined whether VasX is toxic when presented to bacteria from the periplasmic face. As shown in Figure 3C, providing VasX with a SecP secretion signal to target SecP: : VasX to the periplasm reduced the growth rate of C6706 lacking tsiV2. The observed toxicity caused by SecP: : VasX was not due to increased export of proteins to the periplasm as VgrG-3 (missing the peptidoglycan binding domain) [12] targeted to the periplasm (denoted as “SecP: : core”) was not toxic to the producing cell. We concluded from this data that VasX targeted to the periplasm of the producing cell is toxic to that cell. Knowing that periplasmic localization of VasX resulted in autotoxicity, we postulated that VasX inserts into the inner membrane of target cells similar to pore-forming colicins that share homology with VasX. To test whether production of SecP: : VasX compromised membrane integrity of the producing cell, we performed an SDS lysis assay [67] and determined that production of SecP: : VasX increased the sensitivity of a tsiV2 mutant of C6706 to SDS compared to the control strains (data not shown). This finding suggests that periplasmic VasX compromises membrane integrity. If VasX resembles a pore-forming colicin, insertion into the inner membrane would result in increased cellular permeability, ion leakage, and dissipation of the bacterial membrane potential. To test whether production of SecP: : VasX dissipates the membrane potential in the producing cell, we used the BacLight Bacterial Membrane Potential Kit (Molecular Probes, Invitrogen). In this experiment, the fluorescent dye DiOC2 (3) was used to stain all cells green. If the cell is actively respiring, the dye accumulates within the cell and shifts towards red emission. Stained cells were analyzed by flow cytometry using Texas Red and FITC filters and the red/green ratio was indicative of the strength of the membrane potential in analyzed cells. CCCP, which uncouples the proton gradient, served as a positive control for dissipation of membrane potential. Production of SecP: : VasX in a tsiV2 mutant of C6706 resulted in a lower red fluorescence compared to cells producing cytoplasmic VasX, SecP-tagged VgrG-3 core (SecP: : core), or empty vector control (Figure S6). When CCCP was added, the red fluorescence intensity decreased to levels comparable to cells producing SecP: : VasX, implying that SecP: : VasX disrupts the membrane potential to an extent similar to CCCP. Calculation of the red/green ratios indicated that the membrane potential for cells producing SecP: : VasX is significantly reduced compared to controls (Figure 4A). Furthermore, SecP: : VasX uncoupled the proton gradient to the same extent as the positive control CCCP (Figure 4A). Thus, only when VasX was presented from the periplasmic face, did it compromise the integrity of the inner membrane and dissipate the membrane potential (Figure 4A). Next, we tested whether SecP: : VasX made the tsiV2 null-mutant of C6706 permeable to propidium iodide (PI) – an intercalating agent bound to a fluorescent molecule that binds DNA. Normally, PI is used to assess bacterial cell viability because it is membrane impermeant and does not penetrate healthy cells. However, in the case of dead cells, or those with damaged membranes, PI enters the cell and binds DNA. We hypothesized that cells producing SecP: : VasX would be permeable to PI because they have damaged membranes. The tsiV2-mutant of C6706 producing SecP: : VasX or VasX was grown in liquid culture in the presence of arabinose and then incubated with PI. As a positive control for PI uptake, both strains were killed by incubation in ethanol following incubation with PI. Cells were analyzed by flow cytometry to assess the permeability of these cells at a population level. Both strains that were killed in ethanol exhibited increased permeability to PI (74% of the total population for cells producing VasX and 87% of the cell population for those producing SecP: : VasX) (Figure 4B). ∼30% of cells producing SecP: : VasX (without ethanol incubation) showed greater uptake of PI compared to those producing VasX (Figure 4B). This indicates that production of SecP: : VasX increases cellular permeability to PI. We thus conclude that VasX is a bacterial colicin-like effector that disrupts the inner membrane of target cells. The V. cholerae vasX gene is contained within a satellite T6SS-cluster along with hcp-2, vgrG-2, tsiV2, and vasW (Figure 1A). Speculating that the gene products act as a cooperative unit, we set out to determine whether VasW (encoded immediately upstream of vasX) was important for bacterial killing. We deleted vasW from the V52 chromosome and tested this mutant' s ability to kill V. parahaemolyticus RIMD. As shown in Figure 5A, deletion of vasW resulted in an attenuated phenotype towards V. parahaemolyticus RIMD, similar to the result observed when the V52 strain lacking vasX was used as the predator (Figure 5A). Episomal expression of vasX in the vasW null-mutant of V52 did not restore killing, demonstrating that the vasW mutation did not display a polar effect on vasX. Episomal expression of vasW restored killing of V. parahaemolyticus RIMD comparable to wild-type levels (Figure 5A). Based on these results, we hypothesized that VasW is important for VasX function. To determine whether the lack of bacterial killing by the vasW null-mutant of V52 resulted from impairment of VasX function, we used the vasW null-mutant V52ΔvasW as predator in a killing assay against a C6706 mutant that lacked the immunity protein-encoding gene tsiV2 and that had been shown to be sensitive to VasX-mediated killing by V52 [11]. As shown in Figure 5B, the killing ability of V52ΔvasW towards C6706ΔtsiV2 was attenuated and this phenotype could be complemented by expression of vasW from the plasmid pBAD24 (Figure 5B). Therefore, we conclude that V. cholerae VasW assists in VasX-mediated killing of other bacteria and this defect in killing could result from improper localization of VasX or the inability of VasX to be secreted. We previously determined that VasX depends on other T6SS proteins such as VgrG-2 and Hcp for secretion into culture supernatants [21]. Given that deletion of vasW results in similar phenotypes compared to V52ΔvasX, we asked whether VasW was required for V52 to secrete VasX. As shown in Figure 5C, the vasW null-mutant of V52 was unable to secrete VasX into culture supernatants, and episomal expression of vasW was able to restore VasX secretion. Because VasX secretion requires Hcp translocation [21], we tested whether Hcp could be found in culture supernatants of the vasW mutant. As expected, wild-type V52 secreted Hcp into culture supernatants, while V52 lacking vasK did not (Figure 5D). Hcp secretion was not affected by deletion of vasW in V52 (Figure 5D). We and others previously observed that V52 lacking both vasX and vgrG-3 is devoid of Hcp secretion, whereas V52 lacking either vgrG-3 or vasX retains the ability to secrete Hcp (Figure 5D and [11]). Because VasW appears to be important for VasX function, we tested whether a V52 double mutant lacking vgrG-3 and vasW was unable to secrete Hcp, similar to what we observed for a V52 mutant lacking vasX and vgrG-3. We observed that the absence of vgrG-3 and vasW abolished Hcp secretion (Figure 5D), suggesting that VasX and VasW act in a coordinated fashion. Taken together, these data indicate that VasW plays a crucial role in VasX secretion and in its function as a bacterial effector. Given that C6706 does not have an active T6SS under laboratory conditions (i. e. , does not produce VasX or Hcp), we hypothesized that expression of the immunity protein-encoding gene tsiV2 does not require the T6SS regulator VasH, an alternate sigma factor σ54 that works at the operon promoter (located in the 400 base pairs upstream of hcp-2) to drive expression of hcp-2, vgrG-2, vasX, and likely tsiV2 [18], [19]. This operon promoter is not active in C6706 as evidenced by a lack of Hcp production under standard laboratory conditions [34]. We speculated that a promoter exists within vasX capable of driving transcription of the downstream gene tsiV2, which would explain why wild-type C6706 cells are immune to killing by V52. To test this, we made transcriptional fusions of full-length vasX (4-3258, lacking the start codon) and different vasX truncations (nucleotides 4-1345,1575-3258, and 2208-3258) to promoterless lacZ in the plasmid pAH6, as outlined in Figure 6A. The hcp-2 promoter region (Phcp, basepairs −1 to −400 upstream of hcp-2) was also fused to lacZ as a control. Plasmids were transformed into V52, V52 lacking vasH, and C6706. The level of lacZ expression within cells was then determined by performing β-galactosidase assays. V52 permits constitutive expression of T6SS-encoding genes from the promoter upstream of hcp-2 as indicated by the significant amount of LacZ production driven by the operon promoter Phcp. Full length vasX, vasX (1575-3258), and vasX (2208-3258) drove expression of lacZ to levels significantly greater than the empty vector control by ∼5-fold, but less than Phcp (Figure 6B). Deletion of vasH in V52 prohibited transcription from the hcp-2 promoter, but not transcription from within full-length vasX (4-3258) and truncated vasX (1575-3258 and 2208-3258) (Figure 6C). When the reporter plasmids were transformed into the T6SS-silent strain C6706, we observed significant LacZ production in the presence of full-length vasX, vasX (1575-3258), and vasX (2208-3258), but not in the presence of Phcp or vasX (4-1345) (Figure 6D). These data suggest the existence of a promoter within the 3′-end of vasX that can drive expression of tsiV2 independently of VasH. Additional promoter dissection experiments demonstrated that the 3′-terminal 253-bp region of vasX (3006-3258) is sufficient to drive expression of a lacZ reporter (Figure S7). To test the biological role of the promoter within vasX, we performed two independent killing assays. First, we challenged V52 lacking vasH, vasX or both with wild-type V52 and isogenic mutants lacking vasK or vasX. When challenged with a wild-type V52 predator, V52 prey and V52 prey lacking vasH were resistant to killing; however, V52 prey that lacked vasX were more susceptible to V52 predation than V52 prey that lacked vasH (Figure 7A). The V52 ΔvasXΔvasH double mutant prey exhibited greater sensitivity to killing by V52 than V52 lacking either vasX or vasH (Figure 7A). Furthermore, a V52 predator lacking vasX was able to kill E. coli MG1655 but failed to kill V52 prey regardless of their respective mutations, suggesting that null mutations of vasX and vasH impaired immunity to VasX. In the second series of killing assays, we constructed a C6706 derivative containing an in-frame deletion mutation of vasX, and used this strain as prey in a killing assay with a V52 predator. We reasoned that in this strain, deletion of the putative tsiV2 promoter would render C6706 lacking vasX susceptible to killing by V52. C6706ΔvasX (with empty vector pBAD24) were killed by V52 (Figure S8). Importantly, trans-expression of vasX was unable to restore immunity to killing by V52 because expression from pBAD24 does not restore promoter function to drive expression of tsiV2 on the C6706 chromosome (Figure S8). Expression of tsiV2 from pBAD24 protected C6706ΔvasX from being killed by V52 (Figure S8). Taken together, these data indicate that the promoter within vasX can drive transcription of tsiV2 in a VasH-independent manner. To determine whether the promoter in vasX was functional specifically in V52 or in a variety of strains, we also transformed a pAH6 empty vector or pAH6-vasX (2208-3258) into V. cholerae O1 serogroup strains O395w, C6709w, NIH41w, MAK 757w, and N16961w. As shown in Figure 7B, V52, C6709w, NIH41w, MAK 757w, and N16961w had significantly higher LacZ activities compared to their empty vector controls. In contrast, O395w did not have higher levels of LacZ with pAH6-vasX (2208-3258) compared to the empty vector control (Figure 7B). These data indicate that the promoter found within vasX can elicit transcription of downstream genes in a variety of V. cholerae strains and that this phenomenon is not specific to V52. Based on the data presented in Figures 6 and 7, we tested whether tseL and vgrG-3 also contain internal promoters to drive VasH-independent transcription of tsiV1 and tsiV3, respectively. We made transcriptional fusions of tseL and vgrG-3 to promoterless lacZ in pAH6 and transformed the plasmids into V52, V52 lacking vasH, and C6706. Similar to what we observed with vasX, both tseL and vgrG-3 resulted in statistically significant production of β-galactosidase compared to the empty vector control. Production of β-galactosidase was independent of VasH (Figure S9), and the amount of β-galactosidase activity measured was higher for vgrG-3 than for either vasX or tseL (Figures 6 and S9). Last, we performed killing assays using C6706 lacking tseL or vgrG-3 as prey to determine if deletion of the effector gene, and thus the putative promoter for the downstream immunity protein-encoding genes, rendered these strains susceptible to killing by V52. As shown in Figure S2, C6706 lacking tseL or vgrG-3 (with empty vector pBAD24) were killed by V52. Episomal expression of tseL and vgrG-3 in their respective mutants did not restore immunity (Figure S2). However, immunity to V52 could be restored by complementing tseL and vgrG-3 null-mutations in C6706 with tsiV1 and tsiV3, respectively (Figure S2). Taken together, these data indicate that internal promoters located within tseL and vgrG-3 can drive transcription of tsiV1 and tsiV3 in a VasH-independent manner. TA systems are commonly employed by bacteria to ward off bacterial competitors [23], [33], [35], [37], [68]–[71]. In some bacteria, the presence of an antitoxin or immunity protein prevents self-intoxication and killing by toxins produced by sister cells. Similarly, it appears that T6SS immunity proteins protect against an oncoming attack by neighboring T6SS-active bacteria. By screening a C6706 T6SS transposon library for susceptibility to killing by V. cholerae V52 (Figure S1), we identified three T6SS immunity protein-encoding genes, namely tsiV1, tsiV2, and tsiV3 (Figure 1A). These data agree with recent findings of Dong et al. who also identified tsiV1, tsiV2, and tsiV3 as V. cholerae T6SS immunity protein-encoding genes using a Tn-seq approach [11]. Episomal expression of tsiV2 in the Vibrio species V. parahaemolyticus RIMD provided significant protection against killing by V52 (Figure S4). In contrast, episomal expression of tsiV2 in E. coli MG1655 did not provide protection against T6SS-mediated killing – even when VasX was the only bacterial effector employed (data not shown). Importantly, V. parahaemolyticus RIMD has a functional T6SS but does not possess homologs to VasX or TsiV2 according to BLASTn and BLASTp analyses (data not shown). MG1655 also does not possess genes similar to vasX and tsiV2 (data not shown). Therefore, we postulate that even though V. parahaemolyticus RIMD contains T6SS gene clusters, putative immunity proteins are not cross-protective against V52 effectors. Furthermore, we suggest that TsiV2 provides (intermediate) protection in V. parahaemolyticus RIMD, but not in E. coli MG1655, because TsiV2 requires other T6SS proteins or cofactors for proper function and/or localization within the prey cell that are produced by V. parahaemolyticus but not E. coli. In all TA systems, including those in which the toxin is a colicin, the presence of an immunity protein prevents toxicity. Commonly, the gene encoding the immunity protein is located directly adjacent to the effector gene [56], [72], [73]. Using β-galactosidase assays, we demonstrated that a promoter exists within the 1,051 nucleotides of the 3′end of vasX to drive expression of lacZ in V52, V52ΔvasH, C6706, C6709w, NIH41w, MAK 757w, and N16961w (Figures 6 and 7B). The promoter was determined to be within the 3′-terminal 253-bp region of vasX in a subsequent experiment with strain C6706 (Figure S7). Expression driven by promoters within vasX, vgrG-3, and tseL in the strains tested implies that C6706 produces immunity proteins under laboratory conditions to mediate protection against a T6SS-onslaught by bacterial neighbors (Figures 6,7, S2, S8, and S9). We are currently conducting experiments to identify the exact sequence that encodes these promoters. Constitutive expression of T6SS immunity protein-encoding genes could be advantageous for several reasons. The most obvious reason for maintaining basal expression of genes encoding immunity proteins is for protection against an attack by sister cells. Bacteria with activated T6SS gene clusters could readily kill sister cells that live within close quarters (i. e. , biofilms) and lack expression of immunity protein-encoding genes. Constitutive expression of immunity protein-encoding genes, independent of other genes encoding T6SS components, would also allow the bacterium to conserve energy as the T6SS secretion apparatus is a large, dynamic structure [7], [8], [33], the production of which likely requires vast energy expenditure. Previously it was shown that infections of infant mice with V. cholerae V52 unable to shut down its T6SS caused inflammation and actin cross-linking in the mouse gut [74]. By employing strict regulatory mechanisms for T6SS proteins while maintaining continuous expression of immunity protein-encoding genes (such is the case with C6706), V. cholerae could prevent identification by the host immune system and avoid a loss of viability at the time of T6SS activation [75]. Outside the human host, the T6SS-off state can be advantageous in mixed bacterial populations. Immunity protein-encoding genes only protect against T6SS effectors of the same species, sensitizing V. cholerae to T6SS effectors of other organisms including P. aeruginosa [33]. In a mixed population of V. cholerae and P. aeruginosa, P. aeruginosa employed its T6SS to kill V. cholerae only if V. cholerae used its T6SS to attack first. V. cholerae with an inactive T6SS did not induce a counter-attack from neighboring P. aeruginosa and the two bacterial species maintained a peaceful coexistence [33]. We speculate that maintaining constitutive expression of genes encoding immunity proteins independently from other T6SS-encoding genes provides the bacterium with a fitness advantage in the event it engages in T6SS dueling with another T6SS-active kin bacterium. The three immunity protein-encoding genes tsiV1, tsiV2, and tsiV3 are located directly downstream of their corresponding effector genes tseL, vasX, and vgrG-3, respectively. We previously demonstrated that VgrG-3 degrades the peptidoglycan of target cells [12], a phenotype that has also been described for T6SS bacterial effectors produced by P. aeruginosa [35]. TseL is a predicted class III lipase [11], however the role of TseL within the context of the V. cholerae T6SS remains to be determined. Here we showed that VasX is an effector that targets prokaryotes lacking the VasX immunity protein-encoding gene tsiV2 and that the accessory protein VasW is required for secretion of VasX. We previously demonstrated that V. cholerae V52 uses its T6SS to kill E. coli [16], and data presented here indicate that all three effectors – TseL, VgrG-3, and VasX are active against E. coli; however, the presence of TseL and VgrG-3 can compensate for the lack of VasX (Figures 3A and S3). A similar conclusion was reached by Dong et al. such that VasX is sufficient but not required for V52 to kill E. coli [11]. Interestingly, when other Vibrio species such as V. parahaemolyticus RIMD or a mutant of V. cholerae strain C6706 lacking tsiV2 were used as prey, VasX had a much stronger phenotype (Figure 3A). In this case, VasX was a more effective effector against V. parahaemolyticus RIMD and C6706, but less effective for killing E. coli. This could be due to factors such as a requirement for specific receptors on target cells or varying degrees of susceptibility to T6SS effectors. We determined that VasW, a protein whose gene lies directly upstream of vasX, is required for secretion of VasX and that V52 lacking vasW has a phenotype similar to V52 lacking vasX. We and others [11] demonstrated that a double mutant lacking vasX and vgrG-3 fails to secrete the T6SS substrate Hcp into culture supernatants. Dong et al. suggested that VgrG-3 and VasX interact to form part of the T6SS secretion apparatus [11]. Because V52 lacking both vgrG-3 and vasW also does not secrete Hcp (Figure 5D), we propose that VasW acts as an accessory protein responsible for mediating the interaction between VasX and structural proteins of the secretion apparatus. In light of the recently proposed mechanism for T6SS effector translocation [14], VasW could bind to the PAAR protein located at tip of the injectosome to prepare VasX for T6SS-mediated translocation. VasW-mediated recruitment of VasX to the T6SS injectosome could then function as a checkpoint for ejection of the inner Hcp tube decorated with a VgrG-3-containing tip. Previously, VasX was reported to have homology to colicins [64], secreted proteins produced by and toxic to certain strains of E. coli [37]. The data presented here suggest that similar to pore-forming colicins, VasX perturbs the inner membrane of target cells. Although we cannot conclude that VasX acts via pore-formation, our data indicate that VasX dissipates the target cell' s membrane potential and increases cellular permeability – both characteristics of pore-forming toxins [70], [76]–[78]. The observation that periplasmic localization of VasX is toxic to the producer cell also indicates similarities between VasX and pore-forming colicins. Cytoplasmic production of pore-forming colicins does not result in toxicity, because pore formation can occur only when the colicin is presented to the cell from the periplasmic face [56], [57], [67]. We also observed that production of VasX in C6706 that lacks the immunity protein-encoding gene tsiV2 was not toxic to the producing cell (Figure 3B), but providing VasX with the Sec signal peptide resulted in autotoxicity (Figure 3C). Along these lines, we previously demonstrated that VasX is a secreted protein that is present in the cytoplasm and membrane fractions of V52; however, VasX is undetectable in the periplasmic fraction by western blot analysis [21]. It appears that VasX either bypasses the periplasm, or is present transiently in small quantities en route out of the cell, as its presence in the periplasm could be toxic. Interestingly, cells producing SecP: : VasX initially increased in cell number, but then began to die. After 8 hours of induction the bacterial concentration was similar to the starting bacterial concentration (Figure 3C). It is unclear why we did not observe a more significant reduction in the number of surviving bacteria; however, a similar growth/toxicity phenotype referred to as “quasilysis” has been described for colicin lysis proteins that perforate colicin-producing cells and release colicin molecules into the extracellular milieu [56], [79]–[82]. We propose a model whereby VasX is injected into the periplasm of target cells along with T6SS structural proteins and other putative effectors. Upon arrival in the periplasm, VasX inserts into the inner membrane permeabilizing the target cell – an activity similar to quasilysis We previously determined that VasX has a role in T6SS-mediated virulence towards the amoeba D. discoideum and that VasX binds eukaryotic lipids [19]. The fact that VasX has the ability to target both prokaryotic and eukaryotic cells suggests that VasX targets a cellular structure common to both cell types such as the cytoplasmic membrane. A similar model has been proposed by the Mougous group regarding T6SS lipase effectors that target cytoplasmic membranes of both prokaryotes and eukaryotes [79]. In the case of D. discoideum, we propose that perturbation of the plasma membrane is responsible for the VasX-mediated phenotype observed in this eukaryotic host. Targeting the membrane rather than a single protein/receptor is an evolutionarily advantageous mechanism employed by bacterial effectors because developing resistance to this toxic mechanism would prove challenging for target cells.

Vibrio cholerae is the causative agent of the diarrheal disease cholera. This bacterium uses the type VI secretion system (T6SS) to kill other bacteria and host cells. The T6SS is a molecular syringe that Gram-negative bacteria use to inject toxic effectors into target cells in a contact-dependent manner. The V. cholerae T6SS secretes at least three distinct effectors, VasX, TseL, and VgrG-3 to confer antimicrobial activity. To protect itself from an oncoming attack by neighboring bacteria, V. cholerae produces three immunity proteins, TsiV1, TsiV2, and TsiV3 that specifically inactivate the activity of their respective effectors. We determined that the genes encoding TsiV1, TsiV2, and TsiV3 are controlled in a dual fashion that ensures expression of these genes at all times. This provides V. cholerae with constant protection from a T6SS attack by nearby close relatives. Thus, the T6SS gene cluster is a toxin/immunity system that can both kill and protect bacterial cells. Here, we characterize the mechanism of one T6SS effector, VasX, that disrupts the inner membrane of susceptible bacteria. The immunity protein TsiV2 protects prokaryotic cells against VasX-mediated toxicity.

Abstract Introduction Materials and Methods Results Discussion

2013

Dual Expression Profile of Type VI Secretion System Immunity Genes Protects Pandemic Vibrio cholerae

Interactions among microbial community members can lead to emergent properties, such as enhanced productivity, stability, and robustness. Iron-oxide mats in acidic (pH 2–4), high-temperature (> 65 °C) springs of Yellowstone National Park contain relatively simple microbial communities and are well-characterized geochemically. Consequently, these communities are excellent model systems for studying the metabolic activity of individual populations and key microbial interactions. The primary goals of the current study were to integrate data collected in situ with in silico calculations across process-scales encompassing enzymatic activity, cellular metabolism, community interactions, and ecosystem biogeochemistry, as well as to predict and quantify the functional limits of autotroph-heterotroph interactions. Metagenomic and transcriptomic data were used to reconstruct carbon and energy metabolisms of an important autotroph (Metallosphaera yellowstonensis) and heterotroph (Geoarchaeum sp. OSPB) from the studied Fe (III) -oxide mat communities. Standard and hybrid elementary flux mode and flux balance analyses of metabolic models predicted cellular- and community-level metabolic acclimations to simulated environmental stresses, respectively. In situ geochemical analyses, including oxygen depth-profiles, Fe (III) -oxide deposition rates, stable carbon isotopes and mat biomass concentrations, were combined with cellular models to explore autotroph-heterotroph interactions important to community structure-function. Integration of metabolic modeling with in situ measurements, including the relative population abundance of autotrophs to heterotrophs, demonstrated that Fe (III) -oxide mat communities operate at their maximum total community growth rate (i. e. sum of autotroph and heterotroph growth rates), as opposed to net community growth rate (i. e. total community growth rate subtracting autotroph consumed by heterotroph), as predicted from the maximum power principle. Integration of multiscale data with ecological theory provides a basis for predicting autotroph-heterotroph interactions and community-level cellular organization. Microorganisms are the largest component of the biosphere and drive biogeochemical cycles through metabolic activity [1]. Microorganisms commonly exist in biofilms or mats that contain numerous microenvironments due to the interplay between convection-, diffusion-, and chemical concentration gradients induced by microbial activity [2–4]. In addition, most natural microbial communities have diverse microbial populations and an array of nutrient and energy sources, which often precludes detailed analyses of microbial interactions linked to metabolic activity. Natural microbial communities containing well-characterized microbial populations and tractable nutrient inputs are excellent systems to elucidate the principles that organize microbial metabolism and interaction. The phylogenetic diversity of microorganisms within Fe (III) -oxide microbial mats of acid-sulfate-chloride springs in Yellowstone National Park (YNP) is limited due to high temperature (65–75°C) and low pH (~ 3) [5–7]. These biomineralizing communities are formed and inhabited by a limited number of distinct phylotypes, including crenarchaea from the order Sulfolobales (e. g. Metallosphaera yellowstonensis str. MK1) and candidate phylum Geoarchaeota (e. g. Geoarchaeum str. OSPB) (supplemental material) [6–12]. The aqueous and solid-phase geochemistry of two such environments in Beowulf and One Hundred Springs Plain (OSP) hot springs have been studied in detail [5,12–15], and provide bounding conditions and physicochemical context for modeling microbial community interactions. The primary electron donors that drive chemolithoautotrophy in Fe (III) -oxide microbial mats include Fe (II) (25–40 μM) and possibly reduced forms of sulfur (dissolved sulfide < 10 μM) and As (III) (25–30 μM) [9]. The oxidation of Fe (II) coupled with the reduction of oxygen provides energy necessary for the fixation of carbon dioxide by M. yellowstonensis, a major autotroph in these mats [6,8, 14,16,17]. The consumption of oxygen is diffusion-limited in Fe (III) -oxide microbial mats, and results in steep gradients in dissolved oxygen from 60 μM to below 1 μM over 0. 5 to 1 mm [14]. The steep oxygen concentration gradients and corresponding relative abundance of community members as a function of mat depth indicate microbial competition for this limiting electron acceptor [12]. Genomic and mRNA data indicate that predominant autotrophs (e. g. M. yellowstonensis) and heterotrophs (e. g. Geoarchaeum str. OSPB) in Fe (III) -oxide mats utilize oxygen as an electron acceptor [7,16,18]. Carbon dioxide fixation by autotrophs in the community contributes 42 to 99% of the total microbial biomass carbon in Fe (III) -oxide mats from Beowulf and OSP hot springs; the remaining carbon originates from exogenous sources that are produced independent of system electron donor and acceptor requirements [17]. Carbon dioxide fixation has been demonstrated in M. yellowstonensis, which is one of the primary autotrophs in the oxic zones of the Fe (III) -oxide mats [17,19]. Metagenome analysis has established Geoarchaeum str. OSPB as a primary aerobic heterotroph, which comprises 30 to 50% of the total microbial community in the oxic zones of Fe (III) -oxide mats found at OSP [7,12]. Chemolithoautotrophic metabolism and the subsequent transfer of nutrients and energy to heterotrophs (e. g. Geoarchaeum str. OSPB) is hypothesized to drive major autotroph-heterotroph interactions along with competition for the primary terminal electron acceptor, oxygen. Sulfolobales viruses are highly represented in the metagenome sequence of these Fe (III) -oxide mats [10], which suggests that viral predation of M. yellowstonensis and other Sulfolobales populations contributes to the turnover of autotrophic biomass and creates reduced carbon sources for heterotrophs. Genome-enabled stoichiometric modeling is a powerful approach in systems biology for examining metabolic acclimation to environmental stress across size scales from individual cells to communities of interacting populations [20–22]. A summary and graphical representation of stoichiometric modeling can be found in the supplemental information (Figure A in S13 File). Briefly, these approaches construct in silico representations of cellular metabolism inferred from genome sequence analysis [23]. The metabolic models define possible routes of electron transport, cellular energy production, carbon acquisition and central metabolism, and include details of the anabolic processes necessary to synthesize biomass. There are two major types of stoichiometric modeling: elementary flux mode analysis (EFMA) and flux balance analysis (FBA). EFMA identifies all distinct and indecomposable routes through a metabolic network; these routes are termed elementary flux modes (EFMs) [24]. EFMs, and non-negative linear combinations thereof, describe all possible physiologies independent of kinetic parameters, which makes EFMA well-suited for evaluating energetic efficiencies of different electron donors, acceptors, and nutrient sources involved in biomass production. FBA identifies optimal routes through a metabolic network that, for example, maximize growth rate for a given substrate uptake rate [25]. The use of optimization to identify these routes and rates for a given set of nutrient uptake rates is ideal for sampling a large metabolic space bounded by known but flexible kinetic parameters. Stoichiometric modeling has been used to predict optimal genotypes in engineered systems [20,26], interpret physiological behavior [21,27], and evaluate the transfer of mass and energy between distinct populations in a natural phototrophic microbial community [22] and others [28]. A primary goal in environmental microbiology is to understand and predict microbial behavior in communities, where interactions among different populations lead to emergent properties, such as enhanced productivity, stability, and robustness [29]. Consequently, the objectives of this study were to 1) construct metabolic network models for major autotroph and heterotroph populations present in oxic zones of high-temperature Fe (III) -oxide mats using metagenome sequence assemblies from representative sites; 2) analyze the use of electron donors and acceptors for the production of biomass and cellular energy under different nutrient limitations; 3) integrate individual population models to examine possible autotroph-heterotroph interactions that may be fundamental to the ecology of microbial communities (e. g. relative population abundances and oxygen competition); and 4) perform a sensitivity analysis based on parameters measured in situ to determine model limitations and identify future priorities for field measurements. The approach used here integrates data from the nanoscale of electron transport to microscale oxygen depth-profiles to hot spring-scale measurements of biotic Fe (III) -oxide deposition. Novel insights and governing principles of community structure and function were established through the integration of metagenome-enabled in silico approaches and in situ measurements. Genomic data were used to construct in silico representations of the electron transport network responsible for the oxidation of Fe (II) and sulfur species and the reduction of oxygen in M. yellowstonensis str. MK1 (Fig 1). Genomic and physiological evidence indicated that this chemolithoautotroph oxidizes Fe (II) using proteins encoded by the fox operon, which are also found in other Sulfolobales [16]. M. yellowstonensis can also oxidize a variety of sulfur species, including sulfide, elemental sulfur, sulfite, and thiosulfate, to reduce the quinone pool to quinol [16]. Electrons in the quinol pool can drive cellular energy production via oxidative phosphorylation through the reduction of oxygen using a high affinity heme copper oxidase. Alternatively, electrons in the quinol pool can reduce NAD+, enter central carbon metabolism, and be used to reduce inorganic carbon via the 3-hydroxypropionate / 4-hydroxybutyrate (3-HP / 4-HB) pathway [17]. These electron transport pathways are hypothesized to provide the majority of energy to the studied microbial communities. The modeled central carbon metabolism of M. yellowstonensis included the tricarboxylic acid cycle, gluconeogenesis, the pentose phosphate pathway via ribulose monophosphate [30], and the mevalonate pathway [31]. Electron donor and acceptor pathways for M. yellowstonensis were integrated into a cellular-level metabolism model to quantify the relationship between growth and the consumption of environmental resources (Fig 1, supplemental material). Biological systems often minimize their requirements for growth limiting resources, providing an ecologically relevant basis for the in silico prediction of metabolic phenotypes [32]. A total of 6,337 elementary flux modes (EFMs) were calculated that produce M. yellowstonensis biomass using the inorganic electron donors, Fe (II), sulfide, elemental sulfur, sulfite, or thiosulfate. Each EFM was plotted as a function of moles of electron donor and moles of electron acceptor required to produce one carbon mole (Cmole) of M. yellowstonensis biomass (Fig 2A). The moles of electron donor required to form a Cmole of biomass was lowest for sulfide and highest for Fe (II) (Fig 2A), which follows the available free energy predicted for the oxidation of these electron donors coupled to the reduction of oxygen [33]. The oxidation of Fe (II) also required the most moles of oxygen per Cmole of biomass produced of the electron donors evaluated. These relationships were consistent on both a moles of electron donor and moles of oxidized electrons basis (Figure B in S13 File, Table A in S13 File for relevant degrees of reduction). Genomic analysis of Geoarchaeum str. OSPB indicated that this organism is an organoheterotroph with the metabolic potential to utilize a wide variety of reduced carbon species, including biomass macromolecules (i. e. lipids, peptides, polysaccharides, and nucleic acids; see the supplemental material or [34] for a description of reactions involved). Details of Geoarchaeum biomass production from components of lysed Metallosphaera cells were elucidated in a prior report [34]. Briefly, the production of heterotroph biomass based on the consumption of autotroph biomass as a carbon source was evaluated with respect to the Cmoles of carbon substrate or moles of oxygen required to produce a Cmole of biomass (Table B in S13 File). The production of a Cmole of Geoarchaeum str. OSPB biomass required 2. 4 Cmoles of autotroph biomass to supply the cellular energy and structural components for growth under carbon-limited conditions (Fig 2B, point A). Production of a Cmole of Geoarchaeum str. OSPB biomass under oxygen-limited conditions requires 3. 8 Cmoles of autotroph biomass (Fig 2B, point B). The growth of autotroph required substantial inputs of oxygen and was governed largely by the degree of reduction and redox potential of the electron donor (Fig 2C, Table A in S13 File). The predicted growth of autotroph on elemental sulfur required 3. 5 moles of oxygen per Cmole of biomass produced while oxidation of Fe (II) required 38 moles oxygen per Cmole of biomass produced. Both predicted oxygen requirements are within the range of measured values for similar organisms growing on sulfur (3. 5 moles of oxygen per Cmole biomass, assuming 1. 8 * 10−13 grams per cell) and iron oxidation (35–104 moles oxygen per Cmole biomass) [35,36]. Different scenarios were used to evaluate heterotroph growth in this system: organic carbon was either made available, opportunistically, as a result of viral predation and subsequent lysis of the autotrophic cells, modeled as free monomer pools, or organic carbon was supplied as the result of a biological strategy where metabolites were excreted from the autotroph. Alternatively, organic carbon was provided from the surrounding landscape (hereafter called landscape carbon) which allowed for heterotrophic growth independent of the examined autotroph. The landscape carbon was modeled to have the same macromolecular composition as autotroph biomass for simplicity (supplemental material). The landscape carbon was produced from electron donors and acceptors external to the system boundaries and therefore did not contribute to oxygen consumption in the mat. Heterotroph growth on landscape carbon required 1. 6 moles oxygen per Cmole of biomass produced, which is 54 and 96% less than the oxygen required to support the production of autotroph biomass from either elemental sulfur or Fe (II), respectively (Fig 2C, Table B in S13 File). The lower oxygen requirement for organoheterotrophic growth was due to utilization of reduced carbon substrates as both electron donors and anabolic precursors. By comparison, autotrophic metabolism required substantially more energy to reduce carbon dioxide, which resulted in high oxygen requirements, especially when Fe (II) was the primary electron donor. Consequently, the oxygen requirements to produce heterotroph biomass increased sharply when community-level interactions were included. For example, Geoarchaeum str. OSPB requires 2. 4 Cmoles of autotroph biomass to produce a Cmole of heterotroph biomass under carbon-limited conditions (Fig 2B, Table B in S13 File). The total, aggregate oxygen requirement to produce heterotroph biomass from the biomass of an Fe (II) oxidizing autotroph included both the oxygen requirements of the heterotroph and the oxygen requirements to produce the autotroph biomass consumed, which resulted in 93 moles of oxygen being consumed per Cmole heterotroph biomass (Fig 2C). Oxygen is required to produce biomass and cellular energy for both the autotroph and heterotroph, which results in resource competition between these two trophic levels. Additionally, the heterotroph consumed autotroph biomass as an electron donor, creating a competitive interdependence. Mass balances on the autotroph and heterotroph within the oxic system boundaries (0. 7 mm depth, Fig 1C) were described as functions of growth rates, resource requirements, and biomass concentrations (Tables 1 and 2). The system was assumed to be at steady-state for the analyzed time scale of days to weeks, which resulted in a relationship between specific growth rates of the autotroph and heterotroph. This relationship and the flux of oxygen into the system were used to solve for steady-state total community growth rate, which sums both the autotroph and heterotroph, and net community growth rate, which subtracts autotroph carbon consumed by heterotroph from the total community growth rate, (Table 2). The rate of Fe (II) oxidation was then calculated using the autotroph growth rate. These relationships (Table 2) provided a mechanism to examine the effects of oxygen flux and the relative population abundance of autotroph and heterotroph on specific and community growth rates (Figs 3 and 4). Three oxygen fluxes into the mat were analyzed: 50,100, and 200% of the average in situ oxygen flux (420 nmol O2 cm-2 h-1) measured with microelectrodes [12]. In situ values of relative population abundances of autotroph to heterotroph (0. 3–0. 5) were determined based on metagenome analyses from Fe (III) -oxide mat samples collected from Beowulf and OSP springs (Table C and D in S13 File). Simulated total biomass concentrations were bounded by two constraints: 1) maximum specific growth rate of the autotroph and 2) oxygen flux into the mat. A microbial community composed of mostly autotroph was constrained to a minimum biomass concentration of 0. 04 mg cm-3 by the maximum measured specific growth rate for M. yellowstonensis in vitro (μ = 0. 1 h-1 [8,37]) and in situ oxygen flux into the mat (420 nmol O2 cm-2 h-1 [12,14]) (Figure C in S13 File). Increasing total biomass concentration decreased the specific growth rate for an autotroph dominated community and increased the feasible steady-state heterotroph abundance in the community at a given oxygen flux into the mat (Fig 3). Two scenarios were examined for heterotrophic growth: 1) a carbon-limited scenario, which was expected to occur at the top of the mat where oxygen is plentiful, and 2) an oxygen-limited scenario, which was expected at the bottom of the oxic zone (Fig 3). Heterotroph acclimation to intermediary oxygen availabilities would be bounded by these two scenarios. The observed in situ relative population abundance (0. 3–0. 5 autotroph per heterotroph) and oxygen flux was predicted to be feasible at steady-state with a total biomass concentration of ~0. 2 mg cm-3, thereby defining a minimum total biomass concentration for the studied microbial community (Fig 3). Increases in total biomass concentration decreased the specific growth rate for the heterotroph. The total community growth rate, which was the sum of autotroph and heterotroph produced, was constrained by competition for oxygen and the autotroph biomass requirement of the heterotroph. Increases in flux of oxygen into the mat increased total community growth rate supporting oxygen limitation of the modeled community. Increases in relative population abundance of autotroph to heterotroph decreased the total community growth rate, as it did with the specific rate of the autotroph (Figs 3 and 4A). This trend highlights the available mass and energy in autotroph biomass. This mass and energy can be utilized by heterotrophs with less oxygen required per biomass to further increase the total biomass using the same oxygen flux. Optimized carbon usage by the heterotroph, hypothesized to occur under carbon limitation, increased community growth rates by minimizing the autotroph biomass required to produce heterotroph. Conversely, optimized oxygen usage by the heterotroph increased carbon requirements for the heterotroph and decreased all steady-state growth rates by consuming more autotroph biomass (Figs 3 and 4). The net community growth rate, which quantified autotroph and heterotroph accumulation as opposed to production, was constrained by competition for oxygen and the heterotrophic requirement for autotroph biomass. The net and total specific rate of an autotroph only community was equal because no biomass was consumed (Fig 4). Net community growth rate increased with oxygen flux, as observed for specific rate, and relative population abundance of autotroph to heterotroph (Fig 4B). The maximum net community growth rate occurs in a community composed of mostly autotroph and is equal to the total growth rate of the autotroph. The observed oxygen flux into the mat (420 nmol O2 cm-2 h-1 [12,14]) and biomass concentrations above 0. 7 mg cm-3 resulted in negative net community growth rates, which are not sustainable in a steady-state community, due to the minimum metabolic activity to maintain an active population (Figure C in S13 File). The observed relative population abundances and oxygen flux predicted a maximum feasible, total biomass concentration of ~ 2. 0 mg cm-3, above which the net community growth rate was negative. A negative net community growth rate indicated that the total biomass concentration in the mat was not sustainable and would decrease for the oxygen fluxes simulated (Fig 4B). In situ measurements in natural ecosystems are complicated by numerous variables, such as seasonal variation in weather, wind, precipitation, temperature, and inherent heterogeneity. A sensitivity analysis was performed using sulfide and landscape carbon as alternative electron donors for the autotroph and heterotroph, respectively, using a multiscale, hybrid EFMA and FBA approach. The hybrid methodology first identified cellular-level, metabolic acclimation strategies to simulated environments, such as carbon- or oxygen-limited environments, using EFMA. The analysis hypothesized that organisms maximized the desired product, such as biomass or cellular energy, per limiting resource utilized (discussed above) (Fig 2). The same optima are calculable using FBA, but cellular-level EFMA provides additional insights, such as the number of reactions and distribution of suboptimal yields, which are difficult or prohibitive to determine using FBA [34]. The EFM resource requirements to produce biomass and associated maintenance energy were adjusted for specific growth rate using the relationship discussed in Pirt et al [38]. These cellular-level activities were represented by overall resource transformation reactions, quantified using the model exchange reactions, then incorporated into a model of community-level function, which was analyzed using FBA. FBA provides a convenient tool for assessing the impact of rate constraints and different modeling optimization criteria, such as maximizing net or total community growth rate, on the feasible range of metabolic activities. The hybrid FBA simulations evaluated three different electron donor scenarios (A, B, and C in Fig 5, Figure D in S13 File). Firstly, Fe (II) and autotroph biomass were the only electron donors for the autotroph and heterotroph, respectively (Scenario A), which addressed the capacity of the community to oxidize Fe (II) based on varying oxygen flux into the mat (Fig 5, Figure E in S13 File). The resource requirements for the heterotroph to produce biomass and cellular energy were bounded by results from the carbon- and oxygen-limited EFMA simulations to approximate the high and low oxygen regions in the mat, respectively. The growth rate constraints used for FBA are described in Figure D in S13 File and Table 1 and included relative population abundance of autotroph to heterotroph, total biomass concentration, and the maximum specific growth rates of autotroph and heterotroph. The maximum specific growth rates of both populations were set to 0. 1 h-1 based on in vitro studies of M. yellowstonensis [8,37]. Maximization of net and total community growth rate impacted the simulated rates but not the range of metabolic activities feasible by the system. Secondly, autotroph biomass and landscape carbon were both evaluated as possible carbon sources for the heterotroph (Scenario B). This scenario examined the effect of 42 to 99% of the biomass carbon in the system having been supplied by the Fe (II) -oxidizing autotroph with the balance from exogenous, landscape carbon (Fig 5, Figure F in S13 File) as determined by in situ isotopic analyses [17]. The biomass carbon origins were simulated by setting the relative population abundance of autotrophy-based populations (i. e. autotroph and autotroph consuming heterotroph) and the landscape carbon consuming heterotroph. Growth of the autotroph-consuming heterotroph resulted in a lower limit for iron oxidation below which sustaining the heterotrophic population would consume more autotroph than produced (Fig 5, Figure F in S13 File). Finally, sulfide was evaluated as an alternative electron donor for the autotroph (Scenario C). Elemental sulfur and sulfide can be present at low levels in Fe (II) oxidizing mats. The presented analysis only considered sulfide as it was predicted to have the largest impact (Fig 2A, Figure G in S13 File). Autotrophy was simulated to vary between oxidation of sulfide exclusively or Fe (II) exclusively with a total autotroph maximum specific growth rate of 0. 1 h-1; therefore, as Fe (II) oxidation rates increased, less oxygen could be directed toward sulfide oxidation (Fig 5, Figure G in S13 File). A scenario where heterotroph was able to consume Fe (II) and/or sulfide oxidizing autotroph but not landscape carbon, predicted a trend similar to scenario C. The three scenarios were compared to the in situ measurements of Fe (III) -oxide deposition and oxygen flux into the mat. The average in situ measurements for OSP Spring were within the predicted metabolic space for both scenarios B and C in a microbial mat that has a high biomass concentration (~ 2. 0 mg cm-3 as determined above). This overlap between in silico and in situ observation suggests that oxidation of landscape carbon or combinations of landscape carbon and reduced sulfur species could account for the additional aerobic activity in OSP Spring communities. The average in situ measurements for Beowulf spring were within the predicted metabolic space for scenario C for the high biomass concentration simulations, which indicated that the community would have to use landscape carbon, as well as Fe (II) and sulfide oxidation to account for the oxygen consumption. At lower biomass concentrations the metabolic space covered by each scenario decreased, excluding the in situ measurements of both Beowulf and OSP Springs, which indicated if these springs have such low biomass concentrations there must be additional oxygen consumption occurring, such as the oxidation of additional electron donors (see Discussion). Microbial processes span multiple spatial scales from metabolites, enzymes, individual cells, populations, and communities to microbial mats that can ultimately impact planetary biogeochemical cycling (Fig 1). Metagenomic data from Fe (III) -oxide mats in YNP provided a foundation for identifying enzymes responsible for electron transport in a dominant autotroph and heterotroph that occur within the oxic zones of Fe (III) -oxide microbial mats. Biochemical pathways for autotrophic and heterotrophic metabolism were integrated into stoichiometric models to quantify cellular-level resource requirements for biomass and cellular energy production (Fig 2). In situ measurements of relative population abundances (i. e. metagenomes) and fractions of autotrophic- versus landscape-based biomass carbon (i. e. stable carbon isotopes [17]) provided context for modeling microbial interactions within these communities. In addition, in situ measurements of oxygen flux into these mats using microelectrodes [12,14], and Fe (III) -oxide accretion rates from long-term temporal studies [12] (Fig 5) were integrated with the in silico models to quantify possible interactions between a dominant primary producer and secondary consumer, as well as the total contribution of these populations to the biogeochemical activity of the natural Fe (III) -oxide mats. Autotroph-heterotroph interaction modeling indicated that at least 98% of the measured oxygen flux in situ is consumed by the autotroph (e. g. M. yellowstonensis). This is due in part to the relatively low energy content of the electron donor (i. e. Fe (II) ) and the high energy demands of carbon dioxide fixation. Conversely, heterotrophic activity had little effect on predicted Fe (II) oxidation rates as a function of oxygen consumed (Scenario A and B in Fig 5), though it represented the major fraction of steady-state biomass. The relatively high energy content of autotroph biomass compared to Fe (II) meant heterotroph growth increased the total biomass produced at the cost of a decreased net community growth (Fig 4). The turnover of autotroph biomass is important to nutrient and energy flux through the mat community. It has been hypothesized that viral predation and subsequent cell-lysis represents a major mechanism of carbon cycling in microbial communities [39,40]. Numerous Sulfolobales viruses have been identified [41,42], and more specifically, the genome of M. yellowstonensis contains an extremely high abundance of transposases [16] and viral CRISPR spacers, which suggests that viruses are important in the life-cycle of these organisms. Assembled sequences of Sulfolobales viruses have been obtained from Fe (III) -oxide microbial mats [10], and the lysis of archaeal cells by viruses has been observed in situ using scanning electron microscopy (Inskeep, unpl). The turnover of autotroph biomass due to viral predation releases organic carbon enabling the production of heterotroph biomass, which leads to increased total community activity (Fig 4A). The predicted autotroph turnover is greater than 80% at the observed relative population abundances (Figure H in S13 File). To put this turnover value in perspective, it falls within the experimentally observed range of 26 to 1200% reported in other studied geothermal systems [43]. Cell lysis provides one mechanism of carbon and energy exchange; alternatives, such as metabolite exchange, could also contribute to these relative rates and abundances. Metabolite exchange is not expected to be the sole mechanism of exchange in these communities as it would necessitate greater than 80% of all autotrophically fixed carbon be secreted (Figure H in S13 File). The presented analysis does not exclude these mechanisms; indeed, a combination of lysis and metabolite exchange likely occur to varying degrees in all microbial communities. In addition to the lysis mechanism of carbon and energy transfer, simulations were run which considered only metabolite exchange between autotroph and heterotroph populations subject to the previously described community constraints. The metabolite exchange mechanism was simulated where the autotroph secreted monomer distributions consistent with the autotroph biomass composition to facilitate comparison with simulations of the lysis mechanism. The predicted maximum and minimum biomass concentrations were relatively unaffected (supplemental material). The simulated lysis mechanism required cellular energy expenditures by the autotroph to polymerize the monomer pools; this cellular energy was lost to the community when the organic material was transferred to the heterotroph. This energy expenditure was not required during simulated metabolite exchange mechanism conserving community energy and the oxygen used to produce it. However, the oxygen required for polymerizing monomer pools was very small relative to the oxygen required to fix DIC resulting in minimal differences in the predicted maximum and minimum biomass concentrations (Supplemental material). The relative population abundances observed in situ suggest that the Fe (III) -oxide mat communities have metabolic activity that maximize energy acquisition from the environment thereby making them more competitive than those that exhibit lower rates of energy acquisition; this theory is known as the maximum power principle [44]. The ecological benefit to maximizing energy acquisition may be the dilution of autotroph biomass, which increases community diversity and has been shown to promote resilience and stability of the entire community from phage and/or predatory bacteria [45,46]. Applying the maximum power principle to metagenome-derived models may provide theoretical context for community structure-function relationships in other systems. Analysis of the carrying capacity (i. e. maximum sustainable population numbers) of Fe (III) -oxide mats indicated that minimum and maximum total biomass concentrations exist where the specific growth rate of autotroph or oxygen flux into the mat become limiting, respectively. The minimum predicted biomass carrying capacity of 0. 2 mg cm-3 (Fig 3) corresponds remarkably well with the calculated biomass concentration in Beowulf Spring (0. 3 mg cm-3) (Table E in S13 File), which suggests that this microbial community is governed by the maximum specific growth rate of autotroph. Increases in heterotroph abundance would require an autotroph specific growth rate higher than the value observed in vitro [8]. This reduction in biomass concentration, and therefore autotroph limitation, is supported by high channel flow rate (20–30 cm s-1) and associated shear force in Beowulf Spring [12]. Conversely, the maximum predicted biomass carrying capacity of 2. 0 mg cm-3 (Fig 4B and Figure I in S13 File) corresponds well with the calculated biomass concentration of OSP Spring (2. 2 mg cm-3) (Table E in S13 File), which suggests that biomass production at this site is governed by the flux of oxygen. Indeed, in OSP Spring the channel flow rate is 10-fold lower than Beowulf (2–5 cm s-1), which results in less shear force and more limited oxygen transfer [12]. The presented modeling framework combined with ecological theory provides a powerful tool for understanding factors that control community structure and function, and can be applied to many natural and/or engineered microbial systems. A variety of additional oxygen consuming processes in the Fe (III) -oxide mats could explain the higher amounts of measured oxygen flux relative to measured Fe (III) -oxide deposition rates [14] (Fig 5). For example, at least three major oxygen consuming processes (i. e. , additional electron donors) could account for the higher measured oxygen flux than predicted for Fe (II) oxidation alone; these include the oxidation of reduced sulfur species, As (III), and reduced carbon from landscape carbon sources [14]. Small amounts of sulfide (< 5 μM) and occasional flocs of elemental sulfur can be present in Fe (III) -oxide depositing zones of acid-sulfate-chloride springs, such as Beowulf Spring, where high sulfide at discharge results in the deposition of elemental sulfur upstream of Fe (II) oxidation [5]. However, analysis of landscape carbon as an additional stimulant for heterotrophic activity could not account for the Fe (III) -oxide deposited for a given oxygen flux in the low biomass concentration simulations expected to be more representative of Beowulf Spring (Fig 5). Additionally, 25 to 30 μM As (III) is present in many acid-sulfate-chloride springs, and has been shown to be oxidized to As (V) by a bacterial autotroph population (i. e. Hydrogenobaculum spp.), which is commonly found in high abundance upstream of the examined mat community [47]. Elemental analysis of the poorly-crystalline Fe (III) -oxide phases indicates that significant amounts of As (V) are incorporated into the solid phase [5]. Any of these reduced species would result in consumption of additional oxygen if oxidized and warrant further study. The integration of in situ measurements and metagenome-based in silico analyses provide unprecedented approaches to understand microbial interactions and community function. The analyses applied in the presented study provide context for the relative population abundance of autotrophs and heterotrophs, the minimum and maximum biomass concentration of microbial mats, and the possible effects of additional electron donors. The multiscale analyses presented here suggests that microbial interactions contribute to emergent properties of complex organization, such as increased productivity. These principles are likely shared across microbial and macro ecology. In silico stoichiometric models were constructed for autotroph M. yellowstonensis str. MK1 (NCBI Taxon ID 671065, GOLD ID Gi04920) and heterotroph Geoarchaeum str. OSPB (NCBI Taxon ID 1448933, GOLD ID Gi0000638) in the following 5 step process. 1) Initial models were constructed using RAST [48–50]. 2) The elementally and electronically balanced reactions that represent carbon and energy metabolism were manually curated based on genome sequence (Joint Genome Institute—Integrated Microbial Genomes (JGI-IMG) [51]), prior genome analyses [6–8,10,16,17], literature surveys, and the MetaCyc [52,53] and KEGG databases [54]. 3) Reactions were assumed to close pathway gaps that would have resulted in auxotrophy or deviated from observed physiological behavior [17]. 4) The modeled macromolecular compositions of biomass for each population were 2% DNA, 11% lipid, 11% polysaccharide, 16% RNA, and 60% protein based on previous reports [55]. Genomes of each organism were mined for monomer distributions of DNA, RNA, and protein, based on the GC content, the ribosomal subunits, and the average amino acid distribution of all protein encoding genes, respectively (supplemental material, [51]). 5) Maintenance energy requirements were adjusted to calibrate the M. yellowstonensis and Geoarchaeum str. OSPB models to observed yields for Acidothiobacillus ferroxidans, a representative thermoacidophilic autotrophic Fe (II) oxidizing bacterium [56,57] and Alicyclobacillus acidocaldarius, a representative thermoacidophilic heterotrophic bacterium [58], respectively. Nongrowth associated maintenance energy (0. 11 mmol ATP g biomass-1 h-1 for both M. yellowstonensis and Geoarchaeum str. OSPB) was calculated for 65°C assuming 50 KJ mol ATP-1 (supplemental material) [59], which determined the growth associated maintenance energy (171 and 149 mmol ATP g biomass-1 for M. yellowstonensis and Geoarchaeum str. OSPB, respectively). EFMA-based cellular metabolisms (Fig 2A and 2B) were used as input reactions for FBA-based sensitivity analysis of electron donors available to the community. Briefly, the net exchange reactions from the EFMs were expressed on a specific biomass or specific cellular energy basis, such as iron per autotroph (mol Fe (II) per gram of autotroph), by normalizing to the biomass formed. The maximum specific growth rate (0. 1 h-1), relative population abundances (0. 3–0. 5 autotroph to heterotroph), and DIC incorporation fractionation were then used to establish constraints on the maximum and minimum rates and population abundances (See supplemental material for explicit constraints and optimization criteria for the different environmental electron donor scenarios). In silico stoichiometric models were constructed using CellNetAnalyzer version 2014. 1 [60], and exported to RegEFMTool version 2. 0 [61] to enumerate EFMs for each metabolic model. The macronutrients available to both populations were modeled as illustrated in Fig 1. Metabolic reconstructions for both populations, including reactions, genomic evidence, specific literature references, metabolites, stoichiometric balance, and gene regulatory rules can be found in the supplemental material. SBML files provided were exported and tested for import using CellNetAnalyzer version 2017. 3 and the CNA SBML parser. Community FBA was performed using functions from the COBRA Toolbox [62] and overall reactions obtained from EFMs deemed ecologically competitive based on resource utilization. Functions used for the community analysis are available in the supplemental material. All computations were processed on a machine with at most two Intel Xeon X5690 and 120 GB RAM.

Microbial communities often display emergent properties, such as enhanced productivity, stability, and robustness, compared to their component populations in isolation. However, determining the governing principles of these emergent properties can be elusive due to the complexities of interpreting and integrating genomic and geochemical data sets collected at largely different observational scales. Here, we use multiscale, metagenome-enabled modeling of an Fe (II) -oxidizing community to extract information regarding biomass productivity limitations, relative population abundance, total biomass concentration, and electron acceptor uptake rates. The systematic approach used herein is broadly applicable to any microbial community with modest activity and metagenomic data as well as provides a mechanism to characterize interaction motifs in communities that include uncultivated organisms.

Abstract Introduction Results Discussion Materials and methods

cell physiology ecology and environmental sciences chemical compounds electron donors oxygen cell metabolism sulfur oxygen metabolism sulfides ecological metrics chemistry oxidation biochemistry biomass cell biology ecology biology and life sciences chemical reactions physical sciences metabolism chemical elements

2018

Multiscale analysis of autotroph-heterotroph interactions in a high-temperature microbial community

Gene duplication is the predominant mechanism for the evolution of new genes. Major existing models of this process assume that duplicate genes are redundant; degenerative mutations in one copy can therefore accumulate close to neutrally, usually leading to loss from the genome. When gene products dimerize or interact with other molecules for their functions, however, degenerative mutations in one copy may produce repressor alleles that inhibit the function of the other and are therefore exposed to selection. Here, we describe the evolution of a duplicate repressor by simple degenerative mutations in the steroid hormone receptors (SRs), a biologically crucial vertebrate gene family. We isolated and characterized the SRs of the cephalochordate Branchiostoma floridae, which diverged from other chordates just after duplication of the ancestral SR. The B. floridae genome contains two SRs: BfER, an ortholog of the vertebrate estrogen receptors, and BfSR, an ortholog of the vertebrate receptors for androgens, progestins, and corticosteroids. BfSR is specifically activated by estrogens and recognizes estrogen response elements (EREs) in DNA; BfER does not activate transcription in response to steroid hormones but binds EREs, where it competitively represses BfSR. The two genes are partially coexpressed, particularly in ovary and testis, suggesting an ancient role in germ cell development. These results corroborate previous findings that the ancestral steroid receptor was estrogen-sensitive and indicate that, after duplication, BfSR retained the ancestral function, while BfER evolved the capacity to negatively regulate BfSR. Either of two historical mutations that occurred during BfER evolution is sufficient to generate a competitive repressor. Our findings suggest that after duplication of genes whose functions depend on specific molecular interactions, high-probability degenerative mutations can yield novel functions, which are then exposed to positive or negative selection; in either case, the probability of neofunctionalization relative to gene loss is increased compared to existing models. All major models of duplicate gene evolution to date assume that the two genes' products do not interact with each other physically or functionally. Mutations in one copy therefore have no effect on the functions of the other. In the classic model [7], [11], having two copies of a gene is phenotypically equivalent to having only one. After duplication, one copy drifts neutrally, free to amass mutations without the constraints of purifying selection (Figure 1A). In the vast majority of cases, degenerative mutations cause one copy to irreversibly lose its function and ultimately disappear from the genome, a process called nonfunctionalization. Rarely, however, mutations that yield a novel, beneficial function occur by chance; selection fixes these mutations and subsequently maintains both the original and “neofunctionalized” copies in the genome. The ultimate fate of a duplicate gene therefore depends on the outcome of a race between neo- and nonfunctionalization. Because gain-of-function mutations are very rare compared to those that compromise function [12]–[15], the vast majority of drifting duplicate genes will be lost, typically within a few million years, long before new functions are expected to evolve [10]. The plausibility of the classic model as a general explanation for the evolution of new functions has therefore been called into question[16], [17]. In the second model–Duplication, Degeneration and Complementation (DDC) [18]–duplication of a gene with multiple independent subfunctions, such as different expression domains controlled by separate regulatory elements, can be followed by degenerative mutations that knock out different subfunctions in each copy. Because the two copies complement each other but do not interact, these high-probability changes can occur neutrally in small populations, and purifying selection will subsequently conserve both copies and their remaining subfunctions (Figure 1B). The DDC model appears applicable in numerous case studies [18]–[21] and may explain in part why many duplicate genes bear the signature of continuing purifying selection after duplication [22]. Although subfunctionalization does not preclude subsequent evolution of new functions [23], [24], the DDC model per se does not explain how or why novel gene functions evolve after duplication, as has been observed in numerous examples [25]–[36]. In the third model, an increase in the dose of a gene' s product due to duplication yields an immediate selective advantage; thereafter, purifying selection purges degenerative mutations and conserves the ancestral function in both copies[37]–[39]. This model says nothing about how gene pairs with different functions evolved, and it does not explain the long-term retention of duplicate genes for which dosage is not likely to be limiting for fitness, such as many signaling molecules and enzymes in multicellular eukaryotes. Other models have also been suggested, largely variants of the three major models [16], [40]–[43]. Given the high probability of degenerative mutations relative to gain-of-function mutations, it remains unclear how large numbers of duplicate genes have evaded nonfunctionalization long enough to evolve new functions. A key omission from the major models is their assumption that the products of duplicate genes do not interact, either through direct physical contact or by competing for other molecular partners. Gene duplication is recognized as providing raw material for the evolution of elaborate molecular interaction networks [3], [44], but only limited research has addressed how interactions affect the evolutionary fate of duplicate genes. Most of this work has focused on the possibility that duplications may alter the stoichiometry of proteins in a complex, resulting in selection for or against the duplication per se [44]–[46]. But molecular interactions may affect the fate of gene duplicates much more directly, and two types are of particular interest (Figure 1C). First, many proteins function as homodimers. After duplication of such genes, their products will initially cross-dimerize [44]. Even if the duplicate genes are functionally redundant–in the sense that two copies produce the same phenotype as one–mutations that compromise one duplicate may interfere with the functionality of the other by tying up its products in non-functional dimers, just as null mutations at a single locus can produce dominant negative alleles (see refs. [8], [47]). Second, many gene products form functional complexes with other molecules, such as other proteins, DNA binding sites, or small-molecule ligands and substrates. After duplication, the two genes' products initially compete for the same binding partners [44]; mutations in one copy that compromise function but not binding will tie up partner molecules in nonfunctional complexes, reducing the activity of the other copy. As a result, degenerative mutations after duplication may not be phenotypically silent, even when the other copy retains the ancestral alleles. If repression of the ancestral function is deleterious, as seems likely in most cases, then purifying selection will tend to remove degenerative mutations from both copies of a gene. As a result, the temporal window before nonfunctionalization will be longer than expected under the neutral scenarios of the classic and DDC models, and the relative probability of neofunctionalization will increase. More rarely, the evolution of a repressor molecule may allow beneficial new modes of gene regulation; selection would favor such mutations, and a neofunctionalization will occur by high-probability degenerative mutations rather than the low-probability events required under the other models. Steroid hormone receptors (SRs) exemplify both types of functional interaction described above [48], [49], and their evolution has been the subject of considerable interest [50]–[57]. SRs are the major mediators of the effects of gonadal and adrenal steroid hormones on development, reproduction, behavior, and homeostasis throughout the vertebrates. These ligand-controlled transcription factors bind tightly as homodimers to specific DNA response elements in the control region of target genes. Jawed vertebrates have six SRs–two for estrogens (ERα and ERβ) and one each for testosterone and other androgens (AR), progestins (PR), glucocorticoids (GR), and mineralocorticoids (MR) –although some lineages, such as the teleosts, have additional duplicates of these genes [55], [57], [58]. Each hormone binds with high affinity and specificity to a receptor, triggering a change in the receptor' s conformation that allows it to attract coactivator proteins that facilitate transcription of the target gene. SRs have a modular structure, including a highly conserved DNA-binding domain (DBD) –which recognizes and binds to response elements–and a moderately conserved ligand-binding domain (LBD), which binds the hormone and contains the hormone-activated transcriptional activation function. This modular structure allows, for example, construction of chimeric proteins that combine the functions of one protein' s DBD with those of another protein' s LBD [59], and many mutations are known that independently compromise ligand binding, DNA binding, or transcriptional activation, without impairing the other functions [60]–[65]. The SR gene family evolved by duplication and acquisition of novel functions from a single ancestral SR gene (AncSR1). AncSR1 is as ancient as the protostome-deuterostome divergence, based on discovery of a steroid receptor gene in mollusks [27], [66]. Duplication of this ancestral gene produced two major subfamilies of vertebrate SRs[29]. One contains the estrogen receptors ERα and ERβ, which are activated by estrogens and bind to estrogen-response elements (EREs, inverted palindromes of AGGTCA) [67]. Members of the other subfamily–the ketosteroid receptors (kSRs), including AR, PR, GR, and MR–are activated by steroids with a keto group at the 3-position on the steroid backbone (in contrast to estrogens, which have a 3-hydroxyl) and bind to ketosteroid response elements (kSREs, inverted palindromes of AGAACA) [67]. Phylogenetic reconstruction, gene synthesis, and experimental characterization of AncSR1 has shown that it had the ligand- and DNA-specificity of the vertebrate estrogen receptors [29], [66], indicating that the gene duplicate leading to the kSRs evolved novel DNA- and ligand-binding functions. The evolutionary events by which this neofunctionalization process occurred are obscure. It was complete by the time of the vertebrate ancestor, some 470 million years ago, because agnathans, the most basal vertebrate taxon, contain one ortholog each for the ERα/ERβ, the GR/MR, and the AR/PR pairs of jawed vertebrates [29]. But little information is available on the interval between the AncSR1 duplication and the ancestral kSR, because SRs have not been characterized in taxa that branch off the animal phylogeny between the protostome-deuterostome ancestor and the ancestral vertebrate. Sequenced genomes are available from the echinoderm Strongylocentrotus purpuratus and the urochordate Ciona intestinalis, but all SRs were lost completely in these lineages [68], [69]. Cephalochordates therefore have the potential to provide key information about the early evolution of the kSR gene after duplication of AncSR1 [50]. These animals–commonly known as lancelets or amphioxus–represent the earliest-branching chordate taxon [70], [71]. In many gene families, cephalochordates contain a single gene orthologous to up to four paralogs in vertebrates [72], so they are a good candidate to have a single kSR ortholog. Further, the genome of the cephalochordate Branchiostoma floridae contains the set of cytochrome P450 (CYP) genes required for the biosynthesis of the sex steroids testosterone, progesterone, and estradiol, and the synthesis of these hormones in B. floridae has been experimentally established [73], [74]. To gain insight into the early functional diversification of the SRs, we therefore sought to isolate the SRs of B. floridae, characterize their functions, and analyze their evolution. Using a reciprocal BLAST search strategy, we identified two loci as steroid receptor orthologs in the completely sequenced genome of B. floridae. Coding sequences of both genes were determined using RACE (rapid amplification of cDNA ends) on RNA extracted from B. floridae adults, and full-length coding sequences were then isolated using the polymerase chain reaction. One of the B. floridae receptors has high amino acid identity to the human ERs, particularly in the DNA-binding domain, and much lower similarity to the AR, PR, GR, and MR. The other has approximately equal similarity to the ERs and the other SRs (Figure 2A, S1, S2). To determine orthology relationships, we phylogenetically analyzed an alignment of the two B. floridae SRs with the protein sequences of 140 other steroid and related receptors (Table S1). Both maximum likelihood and Bayesian analysis place the ER-like B. floridae receptor at the base of the chordate estrogen receptor clade and the other receptor at the base of the kSR clade (Figure 2B). We therefore named the former gene BfER and the latter BfSR. The affinity of BfER with other deuterostome and protostome ERs and of the BfSR with the vertebrate kSRs are both well-supported, with high posterior probabilities and likelihood ratios (Figure 2B). Alternative topologies that place the BfER and BfSR as cephalochordate-specific duplicates are very unlikely, with a cumulative posterior probability <0. 20 in the BMCMC analysis. Only the taxonomic relationships within the ERs are not well-resolved; we cannot rule out an alternative topology in which the protostome “ERs” are sister to a clade of all deuterostome ERs and kSRs, implying duplication of AncSR1 after the protostome-deuterostome divergence. The phylogeny we recovered is not an artifact of unincorporated heterotachy (site-specific changes in evolutionary rate), which under some conditions can confound ML and BMCMC methods using homogeneous models [75], [76]: when sequences were analyzed using a mixed branch-length method that incorporates heterotachy [77], the placement of BfER and BfSR did not change (See Figure S3). This phylogeny indicates that BfER is orthologous to the ERα and ERβ pair of humans, and BfSR is orthologous to the AR/PR/GR/MR cluster. The duplication of AncSR1 to produce these two major SR clades therefore occurred before the divergence of cephalochordates from the lineage leading to vertebrates. Subsequent duplications to produce the four kSRs and ERs of vertebrates occurred after that split, likely in the two vertebrate-specific genome duplications [29], [78]. This scenario is consistent with evidence of conserved synteny between the scaffold containing BfER and human chromosomes 6 and 14, where ERα and ERβ are located (Figures 2C, S4, see also ref. [79]). The short scaffold that contains BfSR does not contain signal of strong synteny to any human chromosomes (Figure S4). The phylogenetic placement of the B. floridae receptors close to the ER-like AncSR1 explains why both have substantial sequence similarity to the human ERs. The relatively long branch subtending the AR/PR/GR/MR clade indicates a period of relatively rapid sequence evolution in the ancestral kSR in the lineage leading to the vertebrates after the cephalochordates diverged and before subsequent gene duplications. To determine the molecular functions of the B. floridae steroid receptors, we assayed both full-length proteins and specific functional domains in a cell culture system. Surprisingly, we found that BfSR retains the ER-like functions of the ancestral receptor, but BfER does not. BfSR' s LBD activates transcription in the presence of nanomolar concentrations of estradiol and estrone, but is insensitive to ketosteroids, including a broad panel of androgens, progestins, and corticoids (Figures 3A, 3B). In the presence of estrogens, the full-length BfSR activates transcription of an ERE-driven reporter gene (Figures 4A, S5), but it does not activate genes driven by the SRE recognized by vertebrate kSRs (Figure 4B). These data indicate that the derived ligand-binding and DNA-recognition properties of the vertebrate kSRs evolved by neofunctionalization after the cephalochordate/vertebrate divergence. In contrast, the BfER LBD has lost the ancestral capacity to activate transcription in response to estrogens. Neither the BfER LBD nor the full-length BfER activates transcription in response to high doses of any of a large panel of vertebrate steroid hormones, including estradiol, estriol, and estrone (Figures 3A, S5). In the absence of hormone, full-length BfER (like other estrogen receptors) represses basal levels of ERE-reporter transcription, but hormone treatment does not alter this effect (Figure 4A). On an SRE-driven reporter, full-length BfER has no effect (Figure 4B). BfER does retain the ancestral DNA-recognition function: a fusion construct of the BfER-DBD with the constitutively-active NFkB activation domain yields robust transcription of an ERE-driven reporter (Figure 5A), but no activity on an SRE-driven reporter (Figure 5B). Further, BfER strongly binds EREs in an electrophoretic mobility shift assay (EMSA, Figure 5C). Because BfSR and BfER both interact with EREs, but only the former is activated by estrogens, we hypothesized that BfER might function as a competitive inhibitor of BfSR-activated transcription. To test this hypothesis, we cotransfected full-length BfER and BfSR in varying ratios with an ERE-driven reporter in the presence of estrogens. As predicted, BfER inhibited the transcriptional activity of BfSR in a manner dependent on the ratio of BfER to BfSR (Figure 4A). BfER' s inhibitory activity is mediated by competition for EREs: alanine replacement of Cys205, a residue in the DNA-binding domain crucial for response element recognition (Figure 5A, see also [64]), abolishes BfER' s capacity to repress BfSR-mediated transcription, even when transfected in very high ratios (Figure 5D, see also Figure 7C). In the EMSA, BfER binds very strongly to and competes for EREs as a homodimer; whether it can also heterodimerize with BfSR is unclear (Figure 5C). Taken together, these data indicate that competition for DNA is the primary mode by which BfER represses BfSR, although we cannot rule out a minor additional role for heterodimerization. The hypothesis that BfER inhibits BfSR-mediated transcription predicts that the two genes should be coexpressed in some but not all tissues in vivo. We characterized the expression of BfER and BfSR in cross-sections of adult B. floridae by in situ hybridization. In females, transcripts of both genes are abundant in the cytoplasm of oocytes at multiple stages of differentiation (Figure 6, A–F) in a pattern identical to Vasa, a germ cell marker [80]. In addition, the Bfer transcript, but not Bfsr, is weakly expressed in female gills. In males (Figure 6, G–L), Bfsr is expressed broadly throughout the testes, in cells apparently at all stages of spermatogenesis. Bfer, in contrast, is expressed solely in the germinal epithelium of the testis, in a narrow band of cells that are likely to be early spermatogonia based on the expression of Vasa and their basal location within the testis [81]–[83]. These results are consistent with a role for the BfER and BfSR in regulating gonadal function. Because BfER suppresses the transcriptional activity of BfSR and these genes are coexpressed, we conclude that BfER functions, at least in part, as a competitive inhibitor of BfSR activity in vivo. These results indicate that, after duplication of the estrogen-sensitive AncSR1, BfSR retained the ancestral function, whereas BfER evolved a new function as a repressor of the BfSR by losing its estrogen-activated transcriptional function but retaining its ability to compete for DNA response elements. To identify the mechanisms by which the ancestral ER was transformed into the repressor BfER, we compared the sequences of the BfER and BfSR LBDs to those of the human ERα and the reconstructed AncSR1 in light of existing structure-function information. Of the 18 residues that line the ligand-binding pocket of the human ERα [84], only three differ between BfSR and ERα. In contrast, 11 of 18 amino acids differ between the estrogen-insensitive BfER and the human ERα (Figure 7A). Examination of the crystal structures of human SRs predicts that two of the substitutions–R394C and F404L (numbered according to human the ERα sequence) –would have major effects on ligand binding and activation, because these residues, which are conserved in all vertebrate SRs, participate in a network of packing interactions and water-mediated hydrogen bonds that stabilize the receptor-ligand complex [85]. This network consists of a water molecule, Glu353, Arg394, Phe404, and the ligand' s A-ring. Replacing Phe404 with a less bulky leucine would weaken the network' s interaction with the ligand, and replacing the basic amino acid Arg394 with cysteine would abolish hydrogen bonds to both Phe404 and the water (Figure 7B). Both of these changes require a single nucleotide change and occurred in the BfER gene lineage, because the human ER-like states are also found in AncSR1 and all other steroid receptors sequenced to date, including BfSR. To test the hypothesis that one or both of these historical replacements could have converted an estrogen-sensitive transcriptional activator into a competitive inhibitor, we introduced these changes into BfSR by site-directed mutagenesis. As predicted, the double-mutant BfSR R394C/F404L loses the wildtype' s capacity to activate transcription in response to estrogens. Further, when co-transfected, the BfSR double-mutant repressed the activity of BfSR from an ERE in a concentration-dependent manner (Figure 7C). This repressive activity is dependent on ERE-binding and is not an artifact of co-transfection per se: unlike BfER and BfSR R394/F404L, co-transfecting increasing quantities of BfER-C205A, which does not compete for EREs, fails to repress BfSR-mediated reporter transcription (Figure 7C). Either mutation in isolation is sufficient to confer the full repressive phenotype on BfSR (Figure 7D). These data indicate that simple degenerative mutations, which abolish the protein' s activation by estrogens but do not interfere with its ancestral DNA-binding activity, are sufficient to confer on BfER its novel function as a competitive inhibitor. We found that BfSR, the B. floridae ortholog of the vertebrate kSRs, has molecular functions similar to those of the vertebrate estrogen receptor, whereas the cephalochordate ER ortholog is an estrogen-insensitive transcriptional repressor of BfSR-mediated transcription. The presence of estrogen-activated, ERE-recognizing receptors in both clades descending from AncSR1 corroborates previous findings that AncSR1 was ER-like in function [29], [66]. Our results make sense in light of the fact that cephalochordates and vertebrates diverged relatively recently after the duplication of AncSR1. In both taxa, one duplicate retained the ancestral function and the other evolved a new function, but the copy that experienced each fate differed between the two lineages. In the vertebrates, the ancestral kSR evolved novel ligand and response-element specificity, and ER retained the ancestral response to estrogen and EREs. In cephalochordates, the kSR ortholog retained the ancestral specificity for ligand and DNA, and the ER ortholog lost its estrogen-stimulated transcriptional activity, becoming a competitive repressor of BfSR on its DNA binding sites. BfER and BfSR have been recently recognized in the amphioxus genome sequence [50], [74]; this is the first characterization of their functions. Our experiments shed light on the functions of steroid hormones and their receptors in cephalochordates. B. floridae produces estradiol [73], so it is likely that BfSR functions, at least in part, as a classic estrogen receptor. Although B. floridae also produces progesterone and testosterone, these steroids do not activate either receptor in this species, suggesting that they may function primarily as intermediates in the synthesis of estrogens or in other signaling pathways [29]. BfSR and BfER are strongly coexpressed in some but not all tissues, so it is likely that BfER functions as a BfSR regulator, presumably allowing finer tissue-specific modulation of BfSR activity. These receptors may also have other functions, such as activation by unknown cephalochordate-specific hormones or post-transcriptional modification [86], [87]. The expression of both BfER and BfSR primarily in testis and ovary suggests a role for both in gonadal function, such as regulating germ cell development and maturation. In females, both receptors are expressed strongly in ovary. In males, BfSR is expressed throughout the testis, including the apical zone near the testicular lumen, where later-stage spermatocytes, spermatids, and sperm are found. BfER, in contrast, is expressed only in the basal cells of the germinal epithelium, where early spermatogonia are located and where the germ-cell marker Vasa is also expressed [81]. In vertebrates, ERs also play important roles in testicular development and spermatogenesis, suggesting a conserved ancestral function [88], [89]. Further work in vivo is required to elucidate the physiological and developmental roles of BfER, BfSR, and estrogens in cephalochordates. Our results demonstrate the strong effect that functional interactions can have on evolution after gene duplication and suggest the need to supplement existing models of this process. In both the classic and DDC models, degenerative mutations in one copy produce no change in phenotype. As a result, redundancy shields one copy from purifying selection, allowing high-probability degenerative mutations to accumulate neutrally; neofunctionalization must occur only through relatively low-probability gain-of-function mutations. Our results are not predicted by either model. Because steroid receptors function by interacting physically with other molecules, simple degenerative mutations in the BfER LBD produced a novel molecular phenotype–a protein that can no longer carry out the full functions of the ancestral gene but still competes for binding partners, thereby repressing the activity of its duplicate wherever the two are coexpressed. Evolutionary changes in the expression of BfSR and BfER also occurred, yielding partially overlapping expression domains. The order in which these coding and expression changes evolved, and their relative roles in the maintenance of the duplicated ancestral SR, cannot be resolved with current data. The production of repressive functions by degenerative mutations may strongly affect the dynamics of evolution after gene duplication, particularly the relative probability of neofunctionalization versus gene loss by nonfunctionalization. If the repressor allele is adaptive — allowing beneficial new modes of regulating the other copy' s activity–then it will be conserved, leading to long-term maintenance of both gene copies, as occurred with BfER and BfSR. This scenario points to a creative role for high-probability degenerative mutation not envisioned by any existing models. Because degenerative mutations are common, including them as a source of new functions would increase both the absolute and relative probability of neofunctionalization. We found that either of two historical point mutations in the BfER is sufficient to generate a repressor of BfSR; given the modular organization of steroid receptor domains, it is likely that there are many other potential mutations that could also have abolished estrogen-activation while leaving DNA-binding intact. If, on the other hand, repression of the ancestral function is deleterious–as will often be the case–then purifying selection will tend to purge degenerative mutations from both copies, leading to conservation of both duplicates with the ancestral function. In such cases, the accumulation of degenerative mutation would be much slower than under the neutral scenarios of the classic and DDC models. In turn, the rate of nonfunctionalization would be slowed, and the temporal window during which neofunctionalization can occur would be lengthened substantially. Evolution of beneficial duplicate repressors may be a widespread phenomenon. Numerous other members of the nuclear receptor (NR) superfamily have evolved by partial degeneration to function primarily as repressors of the transcriptional activity of paralogous NRs. Some have lost their capacity to activate transcription but retain their ability to bind DNA, so–like BfER–they compete for the binding sites of other receptors, whose activity they downregulate. Others have lost the capacity to bind DNA but retain the ability to form dimers with other NRs and thus inhibit their activity [90], [91]. Repressor duplicates have also evolved in beta-helix-loop-helix transcription factors: some duplicates have lost the canonical DNA-binding domain but still heterodimerize with and therefore silence their paralogs [47]. Similarly, in the family of transmembrane tumor necrosis factor receptors, which trigger apoptosis in response to extracellular ligands, primate-specific duplicates have evolved which bind ligand but have lost their capacity to interact with the intracellular factors that stimulate apoptosis. These “decoy receptors” compete for ligand and thereby repress the activity of their paralogs, preventing ligand-triggered apoptosis in cells that express receptors of both classes [92], [93]. As for the cases in which degenerative mutations are deleterious, it is not possible to directly determine the historical importance of selection against duplicate repressors in shaping genome evolution, but there are reasons to believe it may have played a significant role in delaying nonfunctionalization. Most protein families–enzymes, transcription factors, hormones, neurotransmitters, growth factors, to name a few–depend upon specific molecular interactions for their functions. In many of these families, separate domains, independent molecular surfaces, or even specific sets of residues mediate interactions with different partners or other aspects of function. It is therefore reasonable to expect that after duplication a large class of mutations would compromise specific aspects of function without abolishing all interactions, thereby producing competitive repressors. It is also likely that a nontrivial fraction of these alleles would be deleterious. If these assumptions are correct, then purifying selection would have played a role of general importance in purging degenerative mutations after gene duplication, maintaining duplicates for longer periods of time, and extending the temporal window during which neofunctionalization can occur. Indeed, it has been observed that duplicate genes involved in signal transduction have been preferentially retained in Arabidopsis [94]. This hypothesis also predicts that many genes after duplication will bear the mark of purifying selection–nonsynonymous-to-synonymous substitution rates considerably less than one–even in the absence of subfunctionalization. It has been observed that the majority of recent duplicate genes in Paramecium genomes are under strong purifying selection [46]. This signature is predicted to be particularly strong in duplicate genes that form homodimers or interact with partner molecules whose concentration is limiting for function, such as specific DNA binding sites. Not all degenerative mutations in genes whose functions depend on interactions are likely to produce repressor alleles. Those that radically reduce expression, impair protein folding or stability, or increase the tendency for a protein to aggregate are expected to yield alleles compromised in their ability to interact with any and all partners. These fully nonfunctional alleles would have no effect on their duplicate' s function and would be sheltered by redundancy as predicted by the classic and DDC models' neutral scenarios. Only those mutations that affect modular domains or molecular surfaces that control distinct subfunctions within the coding sequence have the potential to eliminate one aspect of a protein' s functions without abolishing its interactions with at least some molecular partners. It is therefore likely that purifying selection would be partially relaxed after duplication–in contrast to the situation for unduplicated genes, in which all mutations that compromise function, including those that also generate novel functions, would be exposed to the constraining influence of selection. Ohno' s idea that functional diversity can evolve by random exploration of sequence space after gene duplication may strain credulity in its original conception of a purely neutral setting and an inexorable tendency toward “entropic decay. ”[95] It becomes far more plausible, however, when selection can play an anti-entropic role, and when creativity can arise not only from rare mutational combinations but also from far more common ones. Steroid receptor orthologs were identified in the Branchiostoma floridae genome database (Joint Genome Institute, v. 1. 0) by tblastn search using as queries exons from the conserved DBD and LBD of human steroid receptors, as well as the inferred sequence of the ancestral steroid receptor [29], [66] and the ancestral corticosteroid receptor [96]. Recovered B. floridae sequences were used as queries to reciprocally search the human protein database using blastx, and those that recovered SR family members as best hits were retained. Using this technique, two B. floridae sequences were identified as likely SR orthologs. From the two partial SR sequences recovered, primers were designed for RACE (Rapid Amplification of cDNA Ends). Amphioxus RNA was extracted from gravid individuals collected in October 2003 (Gulf Specimen Marine Lab, Panacea, FL) using the RNeasy kit (Qiagen, Valencia, CA). The isolated RNA was reverse transcribed using Thermoscript (Invitrogen, Carlsbad, CA) and oligo dT primers or PowerScript reverse transcriptase (Clontech, Mountain View, CA) with gene specific primers. Both 5′ and 3′ sequences were obtained by RACE using the SmartRace method (Clontech) and Phusion polymerase (New England BioLabs, Ipswich, MA). Products were cloned into TOPO TA cloning vector (Invitrogen) and multiple clones were sequenced (accessions EU371730 and 371729). Numerous synonymous polymorphisms were found in both receptor genes. The two amphioxus SR sequences were aligned to a database containing 140 other SR and closely related nuclear receptor protein sequences, including the B. floridae estrogen related receptor ERR (Table S1). The alignable portions of the amino acid sequences (DBD, LBD, conserved parts of the hinge, and CTE) were aligned using MUSCLE software [97]. Phylogenies (Figures 2B, S6, S7 and S8) were inferred using both Bayesian Markov-Chain Monte Carlo (BMCMC) and maximum likelihood (ML) methods. For BMCMC analysis, we used MrBayes v. 3. 1 [98], using a search consisting of two independent runs of four chains each (one cold and three heated) of 4. 8 million generations each, integration over protein models, gamma-distributed among-site rate variation (prior on the alpha parameter uniform (0. 01 to 5), and uniform branch length priors (0,5]. Two million generations, a point well past stationarity (indicated as standard deviation of clade probabilities between runs of <0. 015), were discarded as burn-in. The Jones-Taylor-Thornton model had 100% posterior probability in BMCMC analysis; this model and gamma-distributed rate variation were therefore used for ML analysis using PHYML-aLRT software; support was calculated as the approximate likelihood ratio and the chi-square confidence estimate derived from that likelihood ratio [99], [100]. Mixed-model maximum likelihood analysis was conducted using our SAML software [77], which calculates likelihoods as the weighted sum over heterogeneous branch length sets, with the number of sets defined by the user and the weight and branch length vector for each set estimated by maximum likelihood using a simulated annealing algorithm. Because of computational demands, a reduced dataset of 33 SR sequences was analyzed. For each model, 1000 perturbations were examined at 1000 temperatures from 1. 0 to 0 with a setback of 10. For each proposal, the probability of topology rearrangement was 0. 3 (of which 40%, 40%, and 20% were tree bisection/reconnection, subtree pruning/regrafting, and nearest-neighbor interchanges, respectively), of branch-length changes was 0. 6, of changes in the alpha parameter of the gamma distribution was 0. 1, and of the weight for each branch length set was 0. 1. Analyses were conducted with models with from 1 to 6 branch length sets, and Akaike' s Information Criterion was used to choose among models. The best-fit model had five categories and very high support (Aikake weight>0. 999). Analysis with this model led to recovery of an ML phylogeny that also placed BfER sister to the clade of estrogen receptors and BfSR sister to the vertebrate AR/PR/GR/MR clade (Figure S3). Synteny relationships were investigated using the human genome (NCBI v. 36, obtained from Ensembl v. 41) and the B. floridae genome (DOE Joint Genome Institute v. 1. 0). To classify human proteins into paralogous groups, we used a single-linkage clustering algorithm based on reciprocal best hits in a BLASTp search of each protein in the human genome against all proteins in the human genome. To identify orthology relationships, BLASTp searches were performed between each human protein sequence and all proteins in the B. floridae genome, and between each B. floridae protein and all proteins in the human genome. Orthology relationships were inferred for reciprocal best hits between a human paralog group and a B. floridae protein. If multiple amphioxus proteins were reciprocal best hits for a human paralog group, the paralog group was split accordingly. A sliding window analysis was then performed between each human chromosome and all amphioxus scaffolds to group orthologs into conserved syntenic regions. A sliding window size of 100 genes was used. Figure 2C displays 19 ortholog pairs spanning ∼82 megabases of human chromosome 14 (77% of total chromosome length), ∼3 megabases of B. floridae scaffold 42 (100% of scaffold length), and 6 megabases of human chromosome 6 (3. 5% of chromosome length). Non-orthologous genes that fall between regions of conserved synteny are not shown. Full-length amphioxus SR cDNAs were amplified using specific forward and reverse primers designed at start and stop codons and were cloned into the mammalian expression vector pcDNA3 (Invitrogen, Carlsbad, CA). Fusion constructs were prepared by amplifying the DBD (the canonical zinc finger region plus the first 30 amino acids in the hinge) or the LBD (including the hinge and carboxy-terminal extension) and subcloning them into pCMV-AD (Stratagene, La Jolla, CA) or pSG5-DBD (gift of D. Furlow), respectively. Site-directed mutagenesis was performed using QuickChange II (Stratagene, La Jolla, CA). All clones were verified by sequencing. Reporter gene transcription assays were performed in Chinese Hamster Ovary (CHOK-1) cells grown to 90% confluence then harvested with trypsin (Invitrogen) and transferred to a 96-well plate containing phenol red-free αMEM supplemented with 10% dextran-charcoal-stripped fetal bovine serum and no antibiotics (Hyclone, Logan, UT). For LBD assays, cells were transfected using lipofectamine and Plus reagent (Invitrogen) with 1 ng receptor plasmid, 100 ng pFRluc reporter plasmid (Promega, Madison, WI), and 0. 1 ng pRLtk reporter plasmid for normalization in Optimem (Invitrogen). For DBD assay, cells were transfected in Optimem with 4 ng receptor plasmid, 2 ng of reporter pGL3-4 (EREc38) -luc (a firefly luciferase reporter driven by four estrogen response elements, a gift from C. Klinge) or SRE-luc (containing a TAT3 glucocorticoid response element and firefly luciferase ORF, a gift from B. Darimont), 0. 1 ng normalization reporter pRLtk (Promega), and 95 ng of pUC19 DNA as filler for transfection efficiency. Four hours later, the transfection mixture was removed and replaced with antibiotic-free αMEM with 10% fetal bovine serum; cells were incubated overnight. LBD assays were then incubated with hormones or vehicle control in triplicate at each dose for an additional 24 hours. Cells were lysed and assayed for luminescence using Dual-Glo Luciferase System (Promega) on a Perkin-Elmer Victor3 plate reader. To calculate normalized luciferase activity, firefly luciferase luminescence was divided by Renilla luciferase luminescence. Dose-response relationships were calculated using nonlinear regression using Prism software (Graphpad, San Diego, CA). Transcriptional activity of full-length receptors was assayed by transfecting CHO-K1 cells using reporter plasmids pGL3-4 (EREc38) -luc or SRE-luc. Full length human GR in pcDNA3 (gift from B. Darimont) was used as the positive control on SREluc and treated with cortisol at a concentration of 10−6 M. Co-transfection of BfER and BfSR full-length plasmids was performed using identical conditions to individual full-length transcriptional assays except that cells were transfected with varying concentrations of each receptor (2 ng, 4 ng, 8 ng, 16 ng, or 32 ng per well), and treated with estradiol at 10−6 M. In screens of full-length receptor sensitivity to a broad panel of hormones, 5 ng receptor and 50 ng reporter were transfected. All experiments were done in triplicate and repeated at least twice with the same results. CHO-K1 cells were transfected with plasmids containing full-length BfER alone (4 µg), BfER (1 µg) and BfSR (4 µg), or BfSR alone (4 µg) as described above, treated with 1 µM estradiol for 4 hours and harvested in TEGDK buffer (10 mM Tris-HCL, 1 mM EDTA, 0. 4 M KCL, 10% (vol/vol) glycerol, 1 mM dithiothreitol) with 1% protease inhibitor cocktail (Sigma-Aldrich, St. Louis, MO). Cells were lysed with four freeze-thaw cycles and centrifuged at 10,000×g for 20 min at 4 C. Protein was quantitated using the Bradford protein assay (Bio-Rad, Hercules, CA). Protein from BfER-transfected cells (2 µg), BfER+BFSR-transfected cells (5 µg), or BfSR-transfected cells (10 µg) was prepared and incubated with 10 ng biotinylated ERE- probe according to the manufacturer' s protocol (Panomics, Redwood City, CA). Reactions were separated on a 5% native polyacrylamide gel in 1× tris-borate EDTA buffer for 3 hours at 90 V at 4 C. The long run was required to separate the sizes of BfER/ERE and BfSR/ERE complexes; in this time, unbound labelled probe migrated off the end of the gel. Gel contents were transferred to Biodyne B membrane (Pall, Ann Arbor, MI) for 30 min at 300 mA. Chemiluminescent detection of biotinylated DNA was performed using the Panomics EMSA kit. Gravid amphioxuswere collected in March, 2007 (Gulf Specimen Marine Lab, Panacea, FL) and fixed in 4% paraformaldehyde, 0. 5 M NaCl, 1 mM MgSO4,2 mM EGTA and 0. 1 M MOPS. Fixed animals were dehydrated in a stepwise series of PBS∶methanol and stored in methanol at −20°C. Samples were re-hydrated in PBS, embedded in agar, and cross-sectioned in a cryostat at 16 mm. The germ-cell marker vasa was used as a positive control and to identify putative germ cells; vasa was identified in the B. floridae genome database using the human sequence as a query and amplified as described below. B. floridae cDNA fragments of BfER, BfSR, and BfVasa were amplified using the following primers: er-forward 5′- GCTAGTGCCTTTGACAAGTC -3′ and er-reverse 5′- CAGACACCTGGTCAGTGAG -3′ (corresponding to nucleotides 661–1501 of the BfER open reading frame), sr-forward 5′- AACTCATAGTGAGCCCCACC -3′ and sr-reverse 5′- CTGCAGAGTTCTCCACTGC -3′ (nucleotides 875–1724), and vasa-forward 5′- GAGCCAACGTGGCGAAGG -3′ and vasa-reverse 5′- CATCAGGGCCTCCTCTCTC -3′ (1–849 nt). The Vasa sequence has been deposited with accession EU371731. Amplicons were cloned into pCR4-TOPO vector (Invitrogen) and used to synthesize digoxigenin-labeled riboprobes (Boehringer-Mannheim). In situ hybridizations on cryo-sections were performed as described previously [101]. In situ hybridizations with different probes were performed on adjacent sections in alternate slides to facilitate comparison among the expression patterns of different genes. Negative control hybridizations were performed without riboprobe to identify any endogenous alkaline phosphatase activity.

Most genes evolved by duplication of more ancient genes. Under existing models of this process, mutations that compromise one copy have no effect on the other; as long as one copy remains intact, such “degenerative” mutations are shielded from selection. Because degenerative mutations are common, most duplicates are expected to be disabled before new functions can evolve. The great functional diversity of genes is therefore somewhat puzzling. Here, we reconstruct how simple degenerative mutations produced a new function in the steroid hormone receptors (SRs), a gene family crucial to reproduction and development. We characterized the two SRs of B. floridae, a cephalochordate that diverged from vertebrates ∼500 million years ago, just after the ancestral SR duplicated. One retained the ancestral gene' s estrogen receptor–like functions, while the other evolved a new function as a competitive repressor of the first. Either of two historical mutations is sufficient to recapitulate evolution of this function by disabling the receptor' s response to estrogen, but leaving its DNA-binding capacity intact. Our results suggest that, for the many genes that function by specifically interacting with other molecules, simple mutations can yield novel functions that, beneficial or deleterious, are exposed to selection.

Abstract Introduction Results Discussion Materials and Methods

evolutionary biology genetics and genomics

2008

Evolution of a New Function by Degenerative Mutation in Cephalochordate Steroid Receptors

Clinical examination of trachoma is used to justify intervention in trachoma-endemic regions. Currently, field graders are certified by determining their concordance with experienced graders using the kappa statistic. Unfortunately, trachoma grading can be highly variable and there are cases where even expert graders disagree (borderline/marginal cases). Prior work has shown that inclusion of borderline cases tends to reduce apparent agreement, as measured by kappa. Here, we confirm those results and assess performance of trainees on these borderline cases by calculating their reliability error, a measure derived from the decomposition of the Brier score. We trained 18 field graders using 200 conjunctival photographs from a community-randomized trial in Niger and assessed inter-grader agreement using kappa as well as reliability error. Three experienced graders scored each case for the presence or absence of trachomatous inflammation - follicular (TF) and trachomatous inflammation - intense (TI). A consensus grade for each case was defined as the one given by a majority of experienced graders. We classified cases into a unanimous subset if all 3 experienced graders gave the same grade. For both TF and TI grades, the mean kappa for trainees was higher on the unanimous subset; inclusion of borderline cases reduced apparent agreement by 15. 7% for TF and 12. 4% for TI. When we assessed the breakdown of the reliability error, we found that our trainees tended to over-call TF grades and under-call TI grades, especially in borderline cases. The kappa statistic is widely used for certifying trachoma field graders. Exclusion of borderline cases, which even experienced graders disagree on, increases apparent agreement with the kappa statistic. Graders may agree less when exposed to the full spectrum of disease. Reliability error allows for the assessment of these borderline cases and can be used to refine an individual trainee' s grading. The World Health Organization (WHO) recommends clinical examination of the upper tarsal conjunctiva of children for trachoma to determine when to start and stop mass antibiotic distributions, and when to declare elimination as a public health concern [1]–[4]. A considerable portion of the evidence justifying interventions is based on the clinical examination as primary or secondary outcomes [5]–[10]. Laboratory diagnostic tests for Chlamydia trachomatis, the causative agent of trachoma, are relatively expensive and rarely performed in trachoma-endemic areas, so the clinical examination will likely remain important in the future [11]. Clinical grades are assigned using the WHO' s simplified grading system, which has 2 grading classes instead of 4, as compared to its predecessor. The simplified grading system was developed for use by trained non-specialist personnel to obtain reliable information on trachoma in population-based surveys or for the simple assessment of the disease at the community level. Trachoma programs almost universally use the simplified system. While its predecessor is able to more finely discern disease activity, it requires more training to use accurately [12]. Agreement with experienced trachoma graders using a kappa statistic is the most common method currently used for certifying competence of field graders [13]–[15]. Unfortunately, clinical trachoma grading can be extremely variable. Even experienced graders disagree on the marginal cases [11]. It could be argued that little information is gained from these marginal cases; if 50% of experienced graders declare a case clinically active, then a trainee' s evaluation, whether positive or negative, reveals little. Statistics such as reliability error, a measure derived from the decomposition of the Brier score, can assess grading in these marginal cases. The Brier Score is the mean squared error of a set of predictions. It can be decomposed into three terms: reliability error, resolution, and uncertainty. Reliability error measures how often a set of predictions given the same forecast probability came true. Resolution measures whether different classifications of forecasts in fact had different outcomes, and uncertainty measures the variance of the outcomes, having nothing to do with the forecasts themselves. Decompositions of the Brier score have been used in meteorology to assess accuracy of weather forecasts [16], [17]. Here, we assessed trachoma grading agreement using photographs from a trachoma-endemic area of Niger, estimating inter-grader agreement using both the kappa statistic and reliability error. In addition to the 18 trainees, three experienced graders (TML, BDG, JDK) graded each of the 200 cases as either having TF or no TF and as having TI or no TI. Each was masked to the others' grades. A consensus grade was defined for each case as the one that at least 2 of the 3 experienced graders agreed upon. Cases for which all three experienced graders were in agreement were sub-classified as unanimous. Borderline or marginal cases are defined as those photos in the testing set where the three experienced graders did not unanimously agree on presence or absence of clinical activity. Kappa statistics on TF grades and, separately, TI grades, were calculated for each of the 18 trainees on the full set of 200 photographs by comparing the trainee' s grade with the consensus grade. Kappa statistics were then recalculated on the unanimous subset of cases only. Bootstrap 95% confidence intervals were determined by resampling trainees (n = 999). Equation 1 [16], [17]Brier score and reliability error for TF and TI were separately calculated for each trainee. (Equation 1). The Brier score can be decomposed into three component parts: reliability error, resolution, and uncertainty. Resolution and uncertainty were not analyzed in this study. Reliability error was calculated by placing the N = 200 cases into K = 4 mutually exclusive bins, representing forecast probabilities. Cases that the three experienced graders unanimously agreed were not TF were placed into the “0/3 TF activity” bin. Cases which one, two, or three experienced graders called as TF, were placed into the “1/3 TF activity”, “2/3 TF activity”, and “3/3 TF activity” bin, respectively. nk is the number of cases in the bin, fk is the forecast probability for that bin (either 0,1/3,2/3,3/3) and is the observed frequency of TF for the bin (ie proportion of cases in that bin the trainee graded as TF). Reliability error was sub-analyzed to reveal the proportion of cases trainees called as TF in each bin (term from Equation 1); mean values across all 18 trainees for each bin are reported here with bootstrapped 95% confidence intervals (n = 999). Calculations were repeated for the clinical grade of TI. To compare the effect of bin size, reliability error was recalculated for TF using K = 3 bins. Cases for which there was unanimous agreement amongst our 3 experienced graders as being not TF were put in the 0/2 TF activity bin. Cases for which there was unanimous agreement amongst experienced graders as TF were placed in the 2/2 TF activity bin. Cases which had any level of disagreement amongst experienced graders were placed in the 1/2 TF activity bin (equivalent to combining 1/3 TF and 2/3 TF activity bins into a single bin). Linear regression was used to assess the relationships of these measures with each other. All calculations were performed in Mathematica 9. 0 (Wolfram Research, Champaign, Illinois). Out of 200 cases, the three experienced graders all agreed 76 were not TF (39% of cases) and 80 were TF (40%), giving unanimous agreement to 156 cases (79%) while disagreeing on 44 cases (21%). When assessing inter-grader agreement on the full set of 200 cases, the mean kappa score of the 18 trainees for TF was 0. 774 (95% CI 0. 746 to 0. 800). Restricting the assessment to cases for which there was unanimous agreement amongst the 3 experienced graders (156 out of 200 cases), the mean kappa increased to 0. 896 (95% CI 0. 861 to 0. 926). The difference in mean kappa scores was 0. 122 (95% CI 0. 108 to 0. 136) higher when restricting analysis to the unanimous subset of cases. With TI grading of the full set of 200 cases, the three experienced graders all agreed 98 were not TI (49% of cases) and 51 were TI (25. 5%), giving unanimous agreement to 149 cases (74. 5%) and disagreeing on 51 cases (25. 5%). Mean kappa across 18 trainees for TI was 0. 707 (95% CI 0. 671 to 0. 744) on the full set of 200 cases. Restricting analysis to the unanimous subset (149 out of 200 cases), the trainees' mean kappa for TI increased to 0. 795 (95% CI 0. 756 to 0. 833). The difference in mean kappa scores was 0. 088 (95% CI 0. 070 to 0. 107) higher for the unanimous subset. Reliability error and Brier score for TF were calculated for the 18 trainees using the 200 cases placed into 4 bins. The three experienced graders unanimously agreed 78 cases were not TF (0/3 TF activity bin) and 80 cases were TF (3/3 TF activity bin). There were 18 cases which were called TF by only 1 experienced grader (1/3 TF activity bin) and 26 cases which were called TF by 2 experienced graders (2/3 TF activity bin). Mean reliability error for the 18 trainees on the full set of cases was 0. 013 (95% CI 0. 007 to 0. 021). Mean reliability error for the unanimous subset (i. e. the 156 cases in the 0/3 TF activity and 3/3 TF activity bins) was 0. 009 (95% CI 0. 004 to 0. 018). The difference in mean reliability error was 0. 004 (95% CI 0. 001 to 0. 006) higher for the full set as compared to the unanimous subset. Mean Brier score for TF on the full set of cases was 0. 089 (95% CI 0. 078 to 0. 101). Mean Brier Score the unanimous subset (i. e. 2 bins) was 0. 052 (95% CI 0. 038 to 0. 069). The mean Brier score was 0. 037 (95% CI 0. 033 to 0. 040) higher on the full set of cases for TF grading. Reliability error and Brier score for TI were calculated for the 18 trainees using the full 200 cases placed into 4 bins. The three experienced graders unanimously agreed 98 cases were not TI (0/3 TI bin) and 51 cases were TI (3/3 TI bin). There were 40 cases which were called TI by only 1 experienced grader (1/3 TI bin) and 11 cases which were called TI by 2 experienced graders (2/3 TI bin). Mean reliability error for TI on the full set of cases was 0. 034 (95% CI 0. 025 to 0. 045). Mean reliability error on just the unanimous subset (i. e. 2 bins) was 0. 025 (95% CI 0. 018 to 0. 035). The difference in mean reliability error was 0. 009 (95% CI 0. 005 0. 013) higher for the full set. Mean Brier score on the full set of cases was 0. 110 (95% CI 0. 098 to 0. 122). Mean Brier Score for the unanimous subset (i. e. the 0/3 and 3/3 bins) was 0. 087 (95% CI 0. 070 to 0. 104). The mean difference in Brier score was 0. 023 (95% CI 0. 019 to 0. 029) higher on the full set of cases for TI. The mean proportion of cases the 18 trainees scored as TF in the 0/3 TF activity bin, which contained the 76 cases that all three experienced graders scored as normal, was 6. 9% (95% CI 3. 9% to 10. 5%). Mean proportion of cases called TF in the 1/3 TF activity bin, containing the 18 cases which only 1 experienced grader called TF, was 50. 9% (95% CI 43. 8% to 57. 7%). Mean proportion of cases scored as TF in the 2/3 TF activity bin, containing the 26 cases that 2 experienced graders called TF, was 80. 1% (95% CI 75. 0% to 85. 7%). Mean proportion of cases called TF in the 3/3 bin, containing the 80 cases all experienced graders called TF, was 96. 5% (95% CI 95. 7% to 97. 4%). Similarly for TI, the mean proportion of cases the 18 trainees scored as TI in the 0/3 TI bin, which contained the 98 cases that all three experienced graders scored as normal, was 0. 6% (95% CI 0. 06% to 1. 24%). Mean proportion of cases called TI in the 1/3 TI bin, containing the 40 cases which only 1 experienced grader called TI, was 11. 8% (95% CI 6. 3% to 17. 8%). Mean proportion of cases scored as TI in the 2/3 TI bin, containing the 11 cases that 2 experienced graders called TI, was 46. 5% (95% CI 37. 4% to 55. 1%). Mean proportion of cases called TI in the 3/3 bin, containing the 51 cases all experienced graders called TI, 75. 8% (95% CI 71. 5% to 80. 3%). As an example, we report kappa scores and reliability errors for TF for 2 individual graders. Grader A had a kappa of 0. 736 and a reliability error of 0. 033. %TF in the 0/3 TF, 1/3 TF, 2/3 TF and 3/3 TF activity bins were 14. 5%, 72. 2%, 96. 1%, and 98. 8%, respectively. Grader B had a kappa of 0. 739 and a reliability error of 0. 005. %TF in the 0/3 TF, 1/3 TF, 2/3 TF and 3/3 TF activity bins were 5. 3%, 50%, 57. 7%, and 97. 5% respectively. Thus trainees with similar kappas may have different reliability scores. Reliability error for TF grades was recalculated by categorizing the 200 cases into 3 bins, instead of 4, by merging the 1/3 and 2/3 TF activity bins into a single 1/2 TF activity bin. The three experienced graders unanimously agreed 76 cases were not TF (0/2 TF bin) and 80 cases were TF (2/2 TF bin). There were 44 cases which the experienced graders disagreed on (1/2 TF bin). Mean reliability error for TF on the full set of cases in 3 bins was 0. 008 (95% CI 0. 004 to 0. 014). Mean reliability error on just the unanimous subset (i. e. 2 bins) was 0. 009 (95% CI 0. 004 to 0. 018). The difference in mean reliability error across all 18 trainees between the full set of cases (all 3 bins) and the unanimous subset (2 bins: 0/2 and 2/2 TF activity) was 0. 001 (95% CI −0. 002 to 0. 004). The mean proportion of cases the 18 trainees scored as TF in the 0/2 TF activity bin, which contained 76 cases that all three experienced graders scored as normal, was 6. 9% (95% CI 3. 9% to 10. 5%). The mean proportion of cases called TF in the 1/2 TF activity bin, containing 44 cases which the three experienced graders disagreed on, was 68. 2% (95% CI 62. 9% to 73. 1%). Mean proportion of cases called TF in the 2/2 bin, containing 80 cases all experienced graders called TF, was 96. 5% (95% CI 95. 7% to 97. 4%). Figure 1 depicts the relationships between our calculated measures on the full set of 200 cases. Figure 1A and 1B show a loose correlation between reliability error and kappa for TF (R2 = 0. 55) and a weak correlation for TI (R2 = 0. 36). Brier score and kappa are much more highly correlated, as Figure 1C and 1D show with R2 = 0. 92 for TF and R2 = 0. 94 for TI. As shown elsewhere with a different test set and trainees [19], we found higher agreement with a kappa statistic when analysis was limited to those cases with unanimous agreement amongst experienced graders. Removing the cases where experienced graders disagreed led to a 15. 7% increase in mean kappa scores for the 18 trainees for TF grades and 12. 4% increase for TI grades. Kappa has been the traditional method for assessing inter-grader agreement and for certifying trachoma field graders [13], [14], [20]. Proper training sets should contain the full spectrum of disease, presumably matching field conditions, not just easy-to-grade cases. Otherwise agreement cannot be expected to be as high as found during training and testing. When the kappa statistic is used to compare a trainee' s score with a gold-standard, it is essentially a scaled accuracy, with the relationship between kappa and accuracy perfectly linear when the prevalence of disease is 50%, and close to linear otherwise. Kappa requires marginal cases be classified as either having clinical activity or not. In contrast, reliability error treats cases as having a probability of possessing activity, which here we set equal to the proportion of 3 experienced graders scoring that case active. Reliability error assesses how close the proportion of positive observed outcomes, given a forecast probability, are to that forecast probability. For a trainee to have perfect (ie 0) reliability error on the 200 cases in this study, he/she must grade no cases in the 0/3 TF activity bin as TF (0 out of 76 cases), one-third in the 1/3 TF activity bin (6 out of 18 cases), two-thirds in the 2/3 TF activity bin (approximately 17 out of 26 cases), and all in the 3/3 TF activity bin (80 out of 80 cases). In contrast to kappa and accuracy, reliability error does not assess the individual grades a trainee gives for borderline cases in the 1/3 and 2/3 TF activity bins; rather, it assesses the proportion of cases scored active in those bins overall. Brier score is highly correlated to kappa (Figure 1 C, D), and thus provides little or no additional information. However reliability error, derived as a portion of the Brier score, does capture information not found in kappa—as evidenced by our finding that reliability error and kappa are not well-correlated (R2 = 0. 56 for TF and R2 = 0. 36 for TI). In contrast to kappa, reliability error can be constructive. We expect the proportion of TF or TI called in the 0/3,1/3,2/3 and 3/3 activity bins to be 0%, ∼33%, ∼66%, and 100% respectively. We found for TF, mean activity (across all 18 trainees) in those bins to be 6. 9%, 50. 9%, 80. 1%, and 96. 5%, respectively. For TI, the mean activity was 0. 6%, 11. 8%, 46. 5%, and 75. 8%. Thus, there was a tendency for our trainees to over-call TF and under-call TI, especially marginal cases (1/3 and 2/3 activity TF and TI bins). The proportion of activity called in each bin could be used, at the level of the trainee, to specifically refine scoring over portions of the disease spectrum, making reliability error a constructive measure. For example, two graders had nearly identical kappas (0. 736 and 0. 739), but reliability errors nearly 7-fold different (0. 033 vs 0. 005). One grader clearly over-called obviously normal cases (0/3 TF bin) as well as moderate cases of TF (1/3 and 2/3 TF bin). We used this information to remediate the grader' s tendency to over-call clinical activity. Furthermore, in contrast to kappa scores, reliability error scores are not necessarily subject to reduction by inclusion of borderline cases. Though we see a statistically significant difference in mean 4-binned reliability error between the full set of cases and the unanimous subset for both TF and TI, our trainees disproportionately over-called TF and under-called TI grades for borderline cases (the 1/3 and 2/3 activity bins). In our re-calculation with 3 bins for TF, there was no statistically significant difference in mean reliability error scores between the full set of cases and the unanimous subset. Though trainees tended to over-call TF in this recalculation, the borderline cases (1/2 TF activity bin) were not disproportionately over-called. Further studies must be done to determine an optimal number of bins to use when calculating reliability error for trachoma grades. Our study has limitations which may affect generalizability. We only analyzed cases from a specific hyper-endemic region in Niger. Other countries may have a different spectrum of disease. We had 3 experienced graders score the 200 cases; there may be variability among other experienced graders on these 200 cases. Additionally, using a larger number of experienced graders may allow for better resolution in categorizing cases as marginal. We used 4 bins to categorize cases, based on the proportion of the three graders that scored the case as having activity. A different binning procedure can demonstrate different results, as discussed previously in the 3-bin recalculation. Conjunctival photographs were used to train graders and perform this study. Field examination has several advantages over photo grading, including that the conjunctiva may be examined from multiple angles, is always in focus, and illumination can be adjusted. The conjunctiva is a three-dimensional structure, particularly when inflamed, whereas a photograph is a two-dimensional representation [15]. For the purposes of our study, however, testing 18 graders in the field on the same cases would not have been feasible. Lastly, this study looked at reliability of trachoma grading using the WHO' s simplified system currently used by most trachoma programs. We may have seen different results using the expanded classification system [12]. Because of its relatively low cost, trachoma control programs will likely continue using clinical examination to make treatment decisions. Thus proper training of field graders is important. To ensure high-quality grading, these graders should be trained on the full spectrum of disease that they are likely to encounter in the field. Using the kappa statistic to judge certification can be difficult to interpret, depending on how widely experienced graders disagree on cases in the test set, given that inclusion of marginal cases tends to deflate apparent agreement. If even experts disagree, a trainee' s answer may reveal little and lower their inter-grader agreement, as assessed by a kappa statistic. However, information can be learned about how a grader is assessing marginal cases by looking at the breakdown of their reliability error. Further studies can help determine if reliability error would also be an important metric to certify graders.

Trachoma is the leading infectious cause of blindness and the World Health Organization plans to eliminate it as a public health concern worldwide by the year 2020. This effort in large part involves mass oral antibiotic distributions to communities. A simplified trachoma grading scale is used to assess presence of active infection. Field workers must be properly trained and certified to perform these eye exams because their findings inform when to start and stop community-wide antibiotic treatments. Certification involves measuring agreement in trachoma grades between a trainee and an experienced grader on a test-set of trachoma photographs. Often, these test-sets have hard-to-grade cases of trachoma removed. We found that removing these borderline cases inflates agreement. Including these borderline cases in the test-set allows a more realistic estimate of agreement, but it is still difficult to assess a trainee' s grades for cases which even experts disagree on. We found that reliability error, a measure derived from the decomposition of the Brier score (the mean squared error of a set of forecasts), can be used to assess a trainee' s evaluation of these borderline cases.

Abstract Introduction Methods Results Discussion

eye infections public and occupational health bacterial diseases infectious diseases medicine and health sciences global health ophthalmology neglected tropical diseases tropical diseases trachoma

2014

Reliability of Trachoma Clinical Grading—Assessing Grading of Marginal Cases

Amyloidogenic neurodegenerative diseases are incurable conditions with high social impact that are typically caused by specific, largely disordered proteins. However, the underlying molecular mechanism remains elusive to established techniques. A favored hypothesis postulates that a critical conformational change in the monomer (an ideal therapeutic target) in these “neurotoxic proteins” triggers the pathogenic cascade. We use force spectroscopy and a novel methodology for unequivocal single-molecule identification to demonstrate a rich conformational polymorphism in the monomer of four representative neurotoxic proteins. This polymorphism strongly correlates with amyloidogenesis and neurotoxicity: it is absent in a fibrillization-incompetent mutant, favored by familial-disease mutations and diminished by a surprisingly promiscuous inhibitor of the critical monomeric β-conformational change, neurotoxicity, and neurodegeneration. Hence, we postulate that specific mechanostable conformers are the cause of these diseases, representing important new early-diagnostic and therapeutic targets. The demonstrated ability to inhibit the conformational heterogeneity of these proteins by a single pharmacological agent reveals common features in the monomer and suggests a common pathway to diagnose, prevent, halt, or reverse multiple neurodegenerative diseases. Amyloid-related neurodegenerative diseases are incurable disorders that are currently classified as conformational diseases (a subset of the proteinopathies). These diseases are thought to be primarily caused by an unknown conformational change in specific proteins, termed “neurotoxic proteins” (NPs), which confers them a cytotoxic gain of function. This initial transition in the soluble monomer triggers a series of events that lead to the formation of small soluble oligomers and protofibrils, and ultimately the generation of highly ordered fibrillar insoluble aggregates. More than 20 of these NPs have been found to be intrinsically disordered proteins (IDPs). Despite the fact that the monomers of NPs have different primary and, when folded, tertiary structures, their oligomers and aggregates share common structural properties [1], [2], including recognition by a conformational antibody [3]. Whereas the nature of the cytotoxic species and the molecular mechanisms involved in their misfolding and aggregation remain largely unknown, there is growing evidence that conformational changes preceding aggregation are fundamental in the molecular pathogenesis of these diseases [1]. Furthermore, as the formation of monomeric misfolding intermediates is regarded as the earliest event in the pathogenic cascade, this process represents an ideal therapeutic target [4]. Remarkably, it was recently demonstrated that the key conformational change that endows amyloidogenesis occurs in the monomer, triggering aggregation and fibrillogenesis as well as cytotoxicity [5]–[7]. Thus, to understand the acquisition of toxic properties by NPs, the three-dimensional structure of this toxic monomeric conformer must be characterized. However, the propensity of monomers to aggregate in vitro hampers their structural analysis using classical high resolution techniques such as X-ray crystallography or solution nuclear magnetic resonance (NMR). Nevertheless, it was shown that some NPs are rich in α-helical [8], [9] or disordered [10], [11] segments. It must be noted that the oligomeric state of α-synuclein, the NP associated to Parkinson disease, remains controversial. Thus, while some reports validate it as a natively unfolded monomer [12], other studies have shown, interestingly, that its potential native state is a dynamic α-helical tetramer [13], [14] that undergoes unfolding/folding cycles associated to its physiological function to promote SNARE complex formation in neurons [15]. Furthermore, it was suggested that many alternative monomeric conformations may coexist in a dynamic equilibrium that progressively shifts toward conformers enriched in β-structure, which are thought to be the first toxic species [2]. However, the inherent ensemble averaging of these techniques prevents the full complexity of monomer conformational equilibria from being elucidated and the fast fluctuations involved from being resolved. Moreover, solid-state NMR and X-ray crystallography can only provide structural information about the final species (i. e. , amyloid fibers [16]). Bulk-based, theoretical, and single-molecule studies support the existence of conformational heterogeneity in intrinsically disordered NPs [1], [17], [18]. However, the co-existence of several species in solution at equilibrium, some of which may be scarcely populated, makes analyzing IDPs a particularly challenging task using established biophysical techniques [19]. By contrast, single-molecule techniques provide a unique opportunity to tackle this problem being able to reveal monomeric conformers [18]. Some of these techniques have already been used to monitor the activity and conformation of proteins with conformational plasticity [20], resolving fast fluctuations (20–300 ns) at the monomeric level [21]. Of the available techniques, nanomanipulation has two additional advantages. First, it allows the forces involved in stabilization of the intramolecular interactions to be directly measured (mechanical stability); and second, it enables the mechanical stability of a protein to be related with its conformation. In particular, atomic force microscopy (AFM) -based single-molecule force spectroscopy (SMFS, for short) shows remarkable potential to analyze the conformational equilibrium of monomers, and even to detect low-abundant conformers [18]. In principle, measuring the mechanostability of NPs using SMFS should be feasible, as the toxic gain of function of NPs parallels the acquisition of β-structure [1], [2], a secondary structure that is usually detectable and is often more mechanostable (M) than α-helical or disordered (random coil [RC]) conformations [22]. Furthermore, determining the mechanostability of NPs is particularly relevant, given their processing by chaperones and chaperone-related proteases [23], which unfold their substrates mechanically using AAA+ ATPase molecular motors [24], [25]. Previous SMFS studies of NPs [26]–[29] have been hindered by three serious problems. Firstly, standard polyproteins were used as single-molecule markers fused to the NP, and since the NP is placed in series with the marker repeats, the SMFS recordings obtained are typically contaminated in the proximal region (Figure 1), thereby compromising the reliability of the data obtained. Secondly, the amyloidogenic behavior of the NP fused to the polyprotein was not demonstrated. Finally, there is no convincing evidence demonstrating that the selected data originated from bona fide intramolecular interactions (rather than intermolecular interactions). In the present study, these limitations have been overcome to reveal the existence of a rich conformational polymorphism in the monomer of NPs and to demonstrate its link to amyloidogenesis and neurotoxicity (Figure S1). To circumvent the first of the aforementioned limitations, we recently developed a novel strategy that uses a new vector to mechanically protect the NP as a guest inside a carrier protein of well-known mechanical properties (plasmid for force spectroscopy-2 [pFS-2]) (Figure 1A) [30]. Accordingly, unlike previous works [26], by means of this approach (termed the “carrier-guest” strategy) the SMFS data obtained from the NP are always read after the unfolding of the carrier, far from the proximal region to the AFM substrate, which is known to be typically noisy and thus unreliable (Figure 1B; Text S1). We used SMFS to analyze the nanomechanics of NPs expressed as fusion polyproteins in pFS-2. Using the so-called length-clamp mode of this technique, two basic parameters are directly obtained (see Text S1): the mechanical stability of the resistance barriers (F, measured directly by the height of the force peaks in a force-extension recording), and the length released after unfolding (increase in contour length (ΔLc) as measured by fitting the force-extension recording to the worm-like chain [WLC] model of polymer elasticity). We followed highly stringent criteria when selecting single-molecule recordings using the carrier-guest mechanical protection strategy whereby we only selected those recordings in which the ΔLc of the carrier (alone or including a non-mechanostable fraction from the NP) was followed, although not necessarily consecutively [29], by force peaks with ΔLc values that accounted for the total expected length released by the carrier-guest protein (Figure 1B; Text S1). This strategy was employed to study four model NPs: three human NPs (expanded poly-glutamines -polyQ-, β-amyloid1–42 -Aβ42-, and α-synuclein) and a neurotoxic-like protein that has been extensively used as a model for human prions (Sup35NM, a yeast prion). These NPs are amyloidogenic IDPs [18], and they are thought to cause the most representative neurodegenerative diseases: polyQ (familial disorders such as Huntington' s disease), Alzheimer' s, Parkinson' s, and prion diseases, respectively [1], [18]. All these NPs, including Aβ [31], are found intracellularly, and they are therefore susceptible to be stretched by the protein-processing machinery [23]. To test whether the properties of these engineered NPs and carriers are preserved in the carrier-guest proteins (the second limitation of the former SMFS studies listed above), we performed standard structural, calorimetric, and fibrillogenic experiments on the isolated carrier-guest proteins (Figures S2, S3, S4, S5, S6, S7, S8, S9; Table S1; Text S1). Whereas the conformational stability of the carrier proteins is reduced when they contain the guest NP, and in some preparations some molecules even lose their folded tertiary structure, the carriers essentially maintain their structure and mechanical properties (Table S2), which enables the unambiguous observation of the conformational polymorphism of the monomeric NPs. Furthermore, regarding the possible presence of spurious interactions, in these studies we found no evidence of contacts or other interactions between the carrier and guest proteins beyond the covalent linkage of the polypeptide chain. This rules out the possibility of artifactual contacts, and hence a possible gain of structure, induced in the NPs when nested inside the carriers or the other way around. To the best of our knowledge, the pFS-2 vector is the first strategy capable of successfully avoiding the contamination of SMFS data with non-specific interactions from the proximal region of the AFM, which is particularly useful for studying IDPs [30]. Indeed, this approach guarantees that the selected data originate from the stretching of single NP molecules (Figure 1). However, in principle, the stretching of a single NP monomer does not exclude possible interactions of the NP with the carrier, the AFM elements (substrate or tip) or other surrounding NPs (intermolecular interactions from oligomers and/or fibers). Such potential interactions (the third limitation of those SMFS studies listed above) must be ruled out before force events can be unequivocally attributed to intramolecular interactions in the NP monomers. By using specific controls, we provide compelling evidence that in our experimental conditions, the stringently selected SMFS data (we sampled about 100 molecules of each NP with each set of recordings resulting from 50,000–100,000 pulling attempts) exclusively arose from monomeric NPs and, therefore, represent bona fide intramolecular interactions (Text S1). Finally, by restricting the degrees of freedom of the hosted NPs, the pFS-2 strategy has the additional advantage of dramatically slowing down fibrillogenesis (Figure S6D; Text S1), thereby minimizing the formation of aggregates, during the nanomechanical analysis of the monomeric NPs. PolyQ diseases are unique among amyloidogenic neurodegenerative diseases in that they are both genetically determined and their NPs, polyQ expansions, have pathological thresholds (21–30 and 36–40 glutamine residues in spinocerebellar ataxia type 6 [SCA6] and Huntington disease, respectively). Interestingly, there is a positive correlation between the number of glutamine repeats and the severity of the disease, the age of onset, and the oligomerization kinetics [4], [32]. Thus, in the simplest hypothesis, the disease determinants are expected to correlate with the number of glutamine residues. Using SMFS we looked for possible mechanical differences in the monomeric polyQ tracts of three lengths that are considered to be sub-, near- (except for SCA6) and super-threshold for triggering polyQ diseases: Q19, Q35, and Q62 (Figure 2A). We found that the subthreshold tract showed no force peaks; we term these conformers “non-mechanostable” (NM) (i. e. , there were no detectable SMFS conformations; the force detection limit of SMFS was taken as twice the force error, 20 pN, at a pulling speed of 0. 4 nm/ms) (Figure 2A, orange bars; this representation is followed also in Figures 3–5; Text S1). Similar mechanical behavior was observed for non-amyloidogenic/non-neurotoxic IDPs [33], [34], including VAMP2 (Figure 2B) [35]. Accordingly, neither Q19 nor VAMP2 constructs produced amyloidogenesis (Figures 2, S2C, S2D, S3C, and S3D). By contrast, Q35, which showed moderate amyloidogenesis (Figure S2C and S2D), exhibited not only numerous NM conformations (F≤20 pN) but also several force peaks with a variety of ΔLc values; we call these conformers “mechanostable” (M) (F>20 pN) (Figure 2A, red bars). ΔLc and F were found to be uncorrelated (Figure S10A) as recently reported for the NP tau [29]. Interestingly, this incipient conformational polymorphism included conformers with extremely high mechanical stability (F≥400 pN). Given that the mechanical stability (upon N-C stretching) of eukaryotic proteins studied to date is below 400 pN [22], we have operationally defined an additional subset in the latter class based on an arbitrary force value of 400 pN. This subset of M conformers that accounts for events with F≥400 pN (which are rare) is termed hyper-mechanostable (hM). Based on our current knowledge of NPs [1] and recent studies of polyQ tracts [5], it is likely that NM and M conformers correspond to RC and β-stranded structures, respectively (see Text S1), as has been previously assumed [26], [27]. Together, these results indicate that polyQ tracts of 35 glutamines can experience a structural transition that permits conformational fluctuations. Significantly, Q62 exhibited even greater conformational polymorphism that also included a higher number of hM conformers (Figure 2A). Thus, expanded polyQ tracts (Q35 and Q62) allow the appearance of conformations that can be detected by SMFS; these mechanical conformations include hM conformers. This polymorphism (estimated as the frequency of M conformers, including hM ones) is positively correlated with the polyQ length, which in turn correlates with amyloidogenesis. Interestingly, a structural correlation with the pathological threshold for polyQs has been previously reported [5]. The SMFS analysis of the three NPs used as models of sporadic neurodegenerative diseases (Aβ42, α-synuclein, and Sup35NM) [1], [2] showed that the wild-type (wt) form of each protein exhibited a rich conformational polymorphism, which also included rare hM conformers (except for Aβ42) (Figures 3–5). The differences in the occurrences of stable conformers in NPs relative to non-amyloidogenic IDPs (VAMP2, Q19, and F19S/L34P Aβ42, see below) are statistically significant (Table 1). In the wt form of the four model NPs, the frequency of M conformers ranged from 0% in Q19 (Q21 is the minimum pathological threshold detected in polyQ diseases, specifically in SCA6 [2]) to 64% in Sup35NM, which for each NP was positively correlated with amyloid fiber density (Figures 2–5, S2D, S3D, S4D, S5D, S6D; Table S1). The frequency of hM conformers for wt NPs ranged from 0% to 8% (Figures 2–5). These rare events are associated with extremely high rupture forces (up to ∼800 pN), greater than those reported for the most mechanostable proteins described to date (i. e. , bacterial cohesin modules [36]) and just below the force range required to break a covalent bond [37]. Table 1 summarizes the frequency for each conformer type. Frequently, more than one force peak was found for each NP molecule (Figure S11), suggesting the presence of more than one structured region in the M conformers. It should be noted that the number (and thus the frequency) of M conformers may in reality be greater than that estimated due to the existence of additional SMFS curves that could correspond to incomplete recordings, i. e. , compatible with the unfolding of a carrier bearing a mechanostable element that was not fully observed probably due to the premature detachment of the molecule. Although mechanostability is a stochastic property, these premature detachment events should often be lower than the forces needed to unravel the undetected NP (putative M conformers). A subset of these recordings contained putative hM events (Figure S12). Although these putative events were not included in the sample size, n, it is interesting to note that there is a good correlation between the putative and detected hM events among the different NPs and conditions tested (Table S3). Taken together, SMFS analysis revealed that whereas VAMP2, a non-neurotoxic IDP, always shows a low (undetectable) resistance to mechanical unfolding, all the NPs studied behaved similarly, displaying a high degree of mechanical conformational polymorphism that often included a few hM conformers. We postulate that this behavior is associated with amyloidogenesis and neurotoxicity, which, in agreement with previous results [5], strongly suggests that a critical conformational change in the monomer leads to the generation of the first toxic NP species. Hence, we have tested this working hypothesis by studying four conditions in which the disease and/or amyloidogenesis are affected in two ways: diminished with a peptide that blocks the oligomerization process, an aggregation-defective mutant, or a peptide that inhibits the β-conformational change; and intensified by familial-disease mutations. According to our hypothesis, in these conditions the conformational polymorphism (and hM frequency) should be unaltered, abolished, decreased, or increased, respectively. We first studied the effect of blocking oligomerization on the conformational polymorphism of monomeric NPs. Specifically, we used the SV111 peptide, which inhibits Aβ42 oligomerization and fibrillogenesis by binding to a specific β-hairpin structure [38], but should not affect the conformational fluctuations of the unbound monomers. Notably, incubation of Aβ42 with SV111 (100 µM [38]) yielded SMFS results similar to those from the untreated Aβ42, including the number of events per molecule (Figures 3 and S11; Table 1), whereas fibrillogenesis was inhibited (Figures 3 and S4D; Table S1). Since this peptide effectively inhibits oligomerization, this experiment also served as a control to rule out a significant contribution of intermolecular interactions to our SMFS data, strongly suggesting that the behavior observed originated from bona fide intramolecular interactions within the NP monomer (Text S1). We then analyzed a double mutant of Aβ42 known to suppress fibrillogenesis (F19S/L34P) [39]. As expected, the results obtained (Figures 3 and S4) were similar to those obtained with Q19 or VAMP2 (Figures 2, S2, and S3). This experiment also served as a control to exclude potential interactions between the NPs and the AFM elements as a source of the M events (Text S1). We next studied the effect of additional representative familial neurodegenerative disease mutations on the conformational polymorphism, using dominant forms responsible for the early onset of the disease and for the acceleration of oligomerization kinetics [1], [2]. In addition to Q35 and Q62, we also studied the arctic (Arc: E22G) mutation in Aβ42 that has been implicated in familial Alzheimer' s disease, as well as the A30P and A53T mutations in α-synuclein that are involved in familial Parkinsonism [1], [2]. Like the results obtained with Q35 and Q62, we found an increase in the frequency of M conformers, including hM ones, for all the mutants analyzed (Figures 3 and 4; Table 1). Similarly, the number of mechanostable regions (i. e. , force peaks) per molecule increased for each NP mutant studied (Figure S11). Finally, two of the familial-disease mutations also increased the number of putative hM recordings (Q62 in polyQ tracts and A53T in α-synuclein) (Figure S12). Taken together, these results indicate that familial-disease mutations increase the propensity to form M conformers, including hM forms, an effect that is correlated with aggregation and fibrillogenesis, at least for polyQ tracts. This increased conformational polymorphism may also explain the dominant character of these NP mutations. We finally analyzed how the conformational polymorphism of NPs was affected by PolyQ-binding peptide 1 (QBP1), an amyloidosis inhibitor known to block the critical β-conformational change of expanded polyQs at the monomer level. QBP1 binds RC conformers [5], [40], suppressing oligomerization, amyloidogenesis, cytotoxicity [5], and neurodegeneration [41]. Specifically, we used the minimal active core of this peptide (QBP1-M8) [42]. Incubation of Q62 with QBP1 completely abolished the formation of fibers (Figures 2 and S2; Table S1), and effectively reduced both the frequency of detected and putative M conformer formation, including hM forms (Figures 2, S12B; Tables 1 and S3). The number of mechanostable events per molecule (Figure S11A) was also reduced. In what was originally intended as a negative control experiment, we also tested the effect of this peptide on the other three NPs, selecting two well-characterized familial-disease mutations (Arc Aβ42 [Figure 3] and A53T α-synuclein [Figure 4]) and the yeast prion Sup35NM (Figure 5). Surprisingly, QBP1 strongly diminished the formation of M conformers (detected and putative) of both A53T α-synuclein and Sup35NM (Figures 4,5, and S12; Table S3). The subset of hM conformers (detected and putative) showed a significant decrease as well (Figure S12; Table S3). Aggregation and fibrillization were also reduced or suppressed by QBP1 in these two NPs (Table S1). It should be noted that the effect of the peptide on Sup35NM was incomplete as some fibers were still formed (Figures 5 and S6D), while its effect on A53T α-synuclein was not as significant, given the small number of fibers formed by this mutant in the absence of the inhibitor (Figures 4 and S5D). Interestingly, in the specific conditions used in our experiments, QBP1 appeared to have no significant effect on Arc Aβ42, either on the conformational polymorphism or its aggregation and fibrillization capabilities (Figures 3, S4C, and S4D; Tables 1 and S1). Furthermore, the frequency of hM events in the untreated Arc Aβ42 was so low (the lowest of the four NPs) that any apparent effect could not be considered significant. To rule out any effect of QBP1 on Aβ42, higher concentrations of QBP1 (from 100 µM to 1 mM) were tested on the isolated Aβ42 and no effect in fiber formation was observed (Figure S4D). These results point to the existence of common characteristics at the monomer level in at least three of the NPs studied (expanded polyQs, α-synuclein, and Sup35NM), which appear to be recognized by QBP1. In spite of the similarities found for all the NPs studied, the inability of QBP1 to recognize Aβ42 indicates that there may be some variation in nascent amyloidogenesis. It should also be noted that structural characterization as well as aggregation and fibrillogenesis experiments (Figures S2, S5, and S6) suggest that the inhibitory effect of QBP1 is stronger on expanded polyQ tracts than on A53T α-synuclein or Sup35NM, while it seems not to affect Arc Aβ42 (at the same peptide concentration) (Figure S4), in accordance to its mentioned effect on the conformational polymorphism. Given that QBP1 blocks the formation of β-structures [5], it is likely that at least some M conformers contain such β-structures, as discussed below. These results extend previous observations of common molecular mechanisms in both oligomerization and fibrillogenesis on the basis of the recognition of similar oligomeric structures in all NPs tested by the A11 conformational antibody [3] and in the existence of a similar though somewhat variable cross-β spine structure discovered in these fibers [16], [43]. Thus, our results point to the existence of additional common molecular mechanisms upstream in amyloidogenesis acting in the initial stages of this process at the monomer level. Single-molecule techniques are ideal to analyze the conformational polymorphism of NPs and SMFS in particular is a highly promising technique that has already been used to this end [26]–[29]. However, the disordered nature of NPs and their tendency to oligomerize pose significant challenges to this kind of analysis. We have successfully developed a strategy to unequivocally analyze at the single-molecule level the conformational polymorphism of IDPs, and in particular NPs. Accordingly, we have sampled the conformational space with about 100 molecules from each of the four most representative NPs and their variants: polyQ tracts, Aβ42, α-synuclein and Sup35NM (see Table 1). We demonstrate that proteins (neurotoxic or not) that do not form amyloid fibers (i. e. , non-pathological polyQ tracts such as Q19, VAMP2, and the F19S/L34P Aβ42 mutant) show no mechanical conformational polymorphism, as monitored by SMFS. By contrast, proteins with propensity to form amyloid fibers (expanded polyQs, Sup35NM, and wt and familial-disease mutations for Aβ42 and α-synuclein) exhibit a rich conformational polymorphism that includes rare hM conformers. As mentioned, it is likely that NM and M conformers correspond to RC and β-stranded structures, respectively. In the case of prions, conformers separated by a high-energy barrier and stabilized by strong non-covalent forces (presumably high-density hydrogen bonding), have recently been proposed to account for the heritability of conformations [44]. It is tempting to speculate that these structures may contain β-strands in a shear (rather than zipper) configuration, as this is the most mechanostable structure observed experimentally to date [36], [45]. These hM conformers were already suggested in previous SMFS studies of polyQs [28], although the authors assumed they involved collapsed rather than β-stranded structures, even for non-pathological tracts. It is important to emphasize that our SMFS results correlate closely with our structural and aggregation/fibrillogenesis controls (Table S1). However, our results are not directly comparable to those of previous studies, which did not include unequivocal single-molecule markers, structural controls, or controls for intermolecular interactions. Thus, while our findings are in line with previous results from tau [29], except that these authors did not report NM conformers, they diverge significantly from those reported for α-synuclein [26], [27], possibly owing to the larger sample size used for tau as compared with α-synuclein (214–453 versus 34–63, respectively). Neurotoxic and neurotoxic-like proteins involved in sporadic neurodegenerative diseases (Aβ42, α-synuclein, and Sup35NM) [1], [2] exist as a broad ensemble of NM and M conformers, including a few hM forms. In contrast, Q19 (which fails to trigger disease) does not show M conformers. These findings and previous observations [41] strengthen the connection between M conformers, amyloidogenesis, and disease. Indeed, mutations that induce an early disease onset and accelerate amyloidogenesis [1], [2] increase the frequency of M (and hM) conformers. Conversely, treatment with the QBP1 inhibitor peptide strongly reduced the frequency of M (and hM) conformers in A53T α-synuclein and Sup35NM. Intriguingly, this treatment did not affect the conformational polymorphism of Arc Aβ42. Moreover, incubation of Aβ42 with SV111 yielded a similar conformational polymorphism as for the untreated Aβ42, although it did inhibit aggregation and fibrillogenesis, in agreement with previous results [38]. Overall, we observed an excellent correlation between the propensity to develop a neurodegenerative disease, previously demonstrated by others [1], [2], and conformational polymorphism, including hM conformers, of the relevant NPs. We show that this polymorphism can be altered in both directions, towards more “structured” conformers by pathological mutations or towards the NM conformation, by the QBP1 inhibitor and by non-fibrillogenic mutations. The dramatic reduction of the frequency of M conformers, including hM forms, produced by QBP1 suggests that hM conformers and/or their precursors from the M pool are good candidates to represent the primary cause of neurodegenerative diseases (i. e. , the hypothetical initial misfolded intermediate). Since QBP1 inhibits the formation of these conformers and exerts anti-neurodegenerative effect [41], we conclude that hM conformers, and/or their precursors, are likely on-pathway intermediates in amyloidogenesis and/or cytotoxicity. The discovery of these common features offers hope that a single therapeutic agent may be able to prevent, halt, or reverse the development of many neurodegenerative diseases arising from these different proteins. Prior to our work, it was known that common molecular mechanisms existed in the later stages of amyloidogenesis, including oligomerization [3] and amyloid fiber formation [43]. Also, prion-like behavior has been recently shown for polyQ tracts, Aβ and α-synuclein [46]. Our identification of a common pattern of conformational propensities among the monomers of the four most representative NPs further extends this unifying view to the earliest stages of the amyloidogenesis process. Furthermore, the inhibition of fibrillogenesis by the same peptide in three out of those four NPs suggest that amyloidogenesis shows common molecular features from the very beginning of the process, although some diversity is also likely [16]. In general, our results strongly support the conformational change hypothesis as the mechanism to explain the initiation of amyloidogenesis [47], wherein amyloidogenesis is thought to be preceded and triggered by a conformational change in the monomer precursor [4]. In the particular case of polyQs (the clearest-cut system), the nature of the conformational changes that occur during polyQ amyloidogenesis have remained controversial [2], [4], [48]. In fact, other models for polyQ diseases have received greater attention in the literature, such as the “polymerization hypothesis” (i. e. , the conformational change is induced upon monomer polymerization into fibers) or recruitment mechanisms like the “linear lattice” model [2], [48]. Our data strongly suggest that it is the monomer that undergoes the critical conformational change that results in the generation of the first toxic species [5]. Furthermore, our data provide strong evidence supporting the so-called “structural threshold hypothesis, ” which states that above the pathological threshold polyQ tracts undergo a structural transition from RC to β-sheet structure [49] and indicating that it occurs in the monomer. Our single-molecule approach also provides an explanation for the failure to detect this change using NMR spectroscopy in a previous study [50]. M conformers represent just 7. 5% or less of species, a fraction that is difficult to detect even if all these conformers were from a unique species (which seems not to be the case according to our ΔLc data). Finally, we note that in the aforementioned study the Glutathione S-transferase-Qn (GST-Qn) fusion proteins used did not show evidence of aggregation. We have adopted a new strategy to unequivocally identify and quantify the conformational polymorphism of NPs and their pathological and non-pathological variants. This approach permits—for the first time to our knowledge—rare hM conformers to be unequivocally detected. We postulate that these conformers, or their precursors, may mediate the primary event in the pathological cascade leading to amyloid fiber formation and concomitant neurodegenerative disease. On the basis of these findings, we propose the self-explanatory term “proteinoscleroses” to collectively refer to amyloid-related neurodegenerative diseases. In this context, hM conformers (or their earliest precursors, if we assume a sequential pathway) constitute ideal pharmacological targets and biomarkers of the propensity to develop such diseases. Moreover, the conformational inhibitor QBP1 appears to target the pathway that leads to hM conformer formation in expanded polyQs, α-synuclein, and Sup35NM prion. Whereas there is evidence for rapid interconversion of conformers in NPs [21], we propose that once formed, the hM conformers are likely to be kinetically trapped (because of the high energy unfolding barriers) and might mechanically jam the unfoldases of the protein-processing machinery of the cell as postulated previously [51]. Such an effect would slow down their processing, which would unbalance the cell proteostasis by increasing the cytosolic concentration of hM conformers. This imbalance could affect distinct cellular processes and ultimately, it may lead to the formation of oligomers and amyloid fibers (Figure S1). There are three findings that support this hypothesis: (1) amyloid-related neurodegenerative diseases have a late onset, developing as aging impairs the efficiency of the protein-processing machinery [52], [53]; (2) the proteasome is transiently impaired in vivo by the expression of the N-terminus of pathologic huntingtin [54]; and (3) the AAA+ ATPases from the degradative machines have recently been shown to unfold their substrates mechanically using relatively low forces (with a pulling geometry apparently similar to that of SMFS experiments and a low calculated loading rate of 0. 15 pN/s) [24], [25]. Alternatively, toxic NP conformers may sequester other IDPs involved in essential cell functions [55] or affect the normal lysosomal autophagy process [56]. It should also be noted that the causal relationships between misfolding and oxidative stress remain a matter of debate [57]. In spite of this, it must be noted that, alternatively, hM conformers may represent dead-end species, i. e. , by-products, off-pathway of amyloidogenesis and cytotoxicity; in this case the toxic species would be part of the remaining species of the M pool. In conclusion, our findings show a clear correlation in all NPs studied between conformational polymorphism (including the presence of hM conformers) and neurotoxicity. They strongly indicate that hM events represent highly collapsed or structured conformers in the fibrillization pathway and that these conformers, or their precursors, may be the primary cause of amyloid-related neurodegenerative diseases. Our approach opens the door to understanding the molecular mechanism of amyloidogenesis and developing a single therapeutic strategy that may treat distinct neurodegenerative diseases using a single drug. Our novel approach also offers a new means of studying other proteins involved in conformational diseases and should allow the dissection (by using the appropriate conditions) of the toxicity and fibrillogenic pathways involved in amyloidogenesis. Polyprotein engineering was performed using a new vector recently described by our group (pFS-2) (Table S4) [30]. Detailed methods on polyprotein sample preparation, AFM-SMFS, circular dichroism (CD), differential scanning calorimetry (DSC), NMR, turbidimetry, Congo Red binding assay, transmission electron microscopy (TEM), and imaging AFM experiments can be found in Text S1.

Neurodegenerative diseases like Alzheimer' s or Parkinson' s are currently incurable. They are caused by different proteins that, under certain circumstances, aggregate and become toxic as we grow older, but the molecular events underlying this process remain unclear. The lack of a well-defined structure, and the tendency of these “neurotoxic proteins” to aggregate make them difficult to study using conventional techniques. Here, we use an established single-molecule manipulation technique combined with a new protein-engineering strategy to show that all these proteins can adopt a rich collection of structures (conformers) that includes a high proportion of mechanostable conformers, which are associated with toxicity and disease. We also find that a known drug can block the formation of these mechanostable structures in different neurotoxic proteins. We suggest that the most mechanostable conformers, or their precursors, may trigger the pathogenic cascade that results in toxicity. We thus propose that these mechanostable structures are ideal targets for early diagnosis, prevention, and treatment of these fatal diseases.

Abstract Introduction Results Discussion Methods

medicine protein interactions drugs and devices chemical biology biomechanics protein folding protein structure proteins chemistry biology recombinant proteins biophysics physics biochemistry protein chemistry

2012

Common Features at the Start of the Neurodegeneration Cascade

An outstanding question in human genetics has been the degree to which adaptation occurs from standing genetic variation or from de novo mutations. Here, we combine several common statistics used to detect selection in an Approximate Bayesian Computation (ABC) framework, with the goal of discriminating between models of selection and providing estimates of the age of selected alleles and the selection coefficients acting on them. We use simulations to assess the power and accuracy of our method and apply it to seven of the strongest sweeps currently known in humans. We identify two genes, ASPM and PSCA, that are most likely affected by selection on standing variation; and we find three genes, ADH1B, LCT, and EDAR, in which the adaptive alleles seem to have swept from a new mutation. We also confirm evidence of selection for one further gene, TRPV6. In one gene, G6PD, neither neutral models nor models of selective sweeps fit the data, presumably because this locus has been subject to balancing selection. Detection of selected regions has been a major goal in population genetics in recent years [9]–[13]. Rather than working with the full data, all of these studies simplified their data by using various statistics designed to detect the signal of selection (see e. g. [12], [14]). These statistics may be classified in different categories, based on the information they exploit. First, functional differences between different codon positions, and the substitution rates of synonymous and non-synonymous sites were used by [15], [16]. Another approach relies on finding related populations, where selection acts on only one of them. This leads to locus-specific high population differentiation, which may be detected by statistics such as FST [17] or XP-EHH [18]. A third category of statistics is based on the length of haplotypes associated with a given allele. Haplotypes associated with the selected allele will on average be younger than haplotypes carrying the derived allele, and there will therefore be fewer recombination events that break up the haplotypes. Statistics such as EHH [9] and iHS [19] were developed to detect this pattern. Finally, the site frequency spectrum (SFS) can also be used to detect departures from neutrality and hence selection. SFS based statistics usually compare various estimators of the population mutation rate θ. The first and perhaps most well-known statistic in this category is Tajima' s D [20], but the statistic can be generalized [21], [22], and other statistics such as Fay and Wu' s H [23] belong to the same family. In this study, we are interested in distinguishing the SDN and SSV models of evolution for a single putatively adaptive mutation. Barrett & Schluter [24] identify three possible ways of identifying SSV: i) the selected allele may occur in an ancestral population, ii) an allele is shown to be older than the environment it is adaptive in and iii) the signature of selection at linked loci, the selected sweep, is different between SSV and SDN. Our approach is based on differences in the genetic signature of selection, but when possible, we will compare to inferences based on i) and ii). To understand the difference between the SSV model and the SDN model, it is important to realize that all the information regarding selection, and mode of selection, is captured by the allele frequency trajectory through time. In other words, the full allele frequency path through time would be a sufficient statistic for the selection coefficient, if it was known. As selection acts only to change the allele frequency in the selected site, and does not act directly on adjacent sites, the effects of the selection on linkage disequilibrium, haplotype patterns, allele frequencies in linked sites, etc. , are only through the effects caused by the change in allele frequency of the selected allele (hitch-hiking effects). This observation is the foundation for standard population genetic theory on selective sweeps (e. g. , [27]–[28]) and forms the basis for several simulation methods, in which the path of the selected mutation is first simulated and then neutral simulations are performed conditional on the allele frequency path [28]. Such simulation methods would be invalid if the allele frequency path did not contain all information regarding the selection coefficient acting on the selected mutation. Similarly, if the path of an allele is the same under the SSV and the SDN model, no additional genomic data could help us distinguish between the two models. Armed with this insight, we can further explore the differences between the two models. Figure 1a, 1b depict the trajectory, the number of copies of the selected allele through time for an SSV and SDN model. Looking backward in time, the adaptive alleles are selected at first in both models, and during this stage the two models do not differ at all. In the SSV model, however, the mutation stops being advantageous at some point in the past. Backwards from this time point, the mutation in the SSV model acts as a neutral allele, whereas the mutation in the SDN model is under selection. As selection is the same in the phase when both alleles are selected, the difference between the models is during the phase in which selection is acting on the mutation in the SDN model but not in the SSV model. How big is this difference? It depends on two parameters: the selective advantage of the mutation under the SDN model, and the frequency of the mutation at the time when selection first start acting in the SSV model. A good measure of the difference might be the allele age distribution at this point, which is plotted in Figure 1c and 1d for a mutation at a frequency of 1% and 5%, respectively. Unfortunately, it turns out that the difference it is rather small: While the allele age of a mutation at a low frequency does depend on the selection coefficient, the difference is very small if selection is weak. Clearly, it will be much easier to distinguish between the two models if selection is strong and if the frequency of the mutation is initially high in the SSV model. However, we cannot observe the trajectory directly, but only the diversity at linked site. It has been shown that the genetic signature of sweeps from standing variation differs in three important aspects from the signature of sweeps from new mutations [29]: at the same selection coefficient, the signal of selection from standing variation is 1) weaker and 2) affecting a narrower region. As a third difference, we expect an increased variance in both allele age and trajectory. Under the SSV model, the selected allele may be present on several haplotypes when selection starts, and these haplotypes will be affected equally strongly by selection. Thus, there will be more variation compared to SDN, and the, loss-of-diversity signal of selection will be weaker. The fact that the signal of selection affects a narrower region is due to the fact that the selected allele is older in the SSV model, and hence recombination had more time to break it up (Figure 1a, 1b). The increase in variance is evident from the large variance in the neutral phase of the allele trajectory in Figure 1b, and the wider distribution of the allele age of neutral alleles in Figure 1c and 1d. In Figure 1e and 1f we give the expected distribution of Fay and Wu' s H [23] and EHH [9], two statistics used to detect selection, and where we show that the signal is indeed expected to be weaker and affecting a narrower region under the SSV model. The objective of this paper is to develop and explore a statistical method for distinguishing between SSV and SDN models, and for providing associated estimates of relevant parameters. However, the method we develop is not intended as a new method for performing scans for selection in genome-wide data or for quantification of genome-wide levels of selection. For computational reasons, other methods might be more suitable for such genome-wide analyses. We focus on illustrating the method on a few loci previously hypothesized to be under selection in humans, but the method could as well be applied to other human loci or data from other species. To exploit the characteristics of selective sweeps discussed in the previous section, we combine different statistics and calculate them for different genomic regions. Using combinations of statistics to improve inference is not a new concept, and has been applied previously (e. g. [30]). Here, we choose an Approximate Bayesian Computation (ABC) framework for combining statistics [31], [32]. ABC has the advantage that it extends naturally to allow both model choice and parameter estimates under a given model. ABC was developed to estimate parameters of complex models in manageable computer time, and has been widely used in population genetics, most frequently to infer parameters for complex models of demographic history [32]–[37]. Several implementations of the ABC algorithm have recently been published [38]–[40], and in the past few years, various variations of the algorithm have been developed [41]–[43]. ABC is a rejection sampling algorithm used to calculate the posterior distribution of a parameter under a given model, used frequently when the likelihood cannot be calculated analytically. In ABC inference, a large number of data sets are simulated using parameters randomly drawn from a prior distribution. If a simulation does not match the observed data, it is rejected, otherwise it is retained. However, if the data is complex, the probability of a match is prohibitively low, and two important approximation steps are used: First, the data is transformed into a set of summary statistics. If these statistics are sufficient (i. e. retain all the information present in the data), this step is exact. However, in many cases, including this study, no sufficient statistics are known, and this step results in a first approximation step. In many cases, however, this transformation will still result in very low acceptance probabilities. Therefore, the condition of an exact match is relaxed. Specifically, the summary statistics based on the simulations (S) are compared to the summary statistics of observed data (S*). Using some distance measure δ, simulations are retained if |δ (S, S*) |<ε for an arbitrarily small distance ε. Frequently, some post-sampling adjustment is used in an attempt to correct for the error introduced in the second approximation step, and posterior distributions are estimated from the parameters of the retained simulations. In this study, we propose to use ABC to distinguish between a selective sweep from a new mutation and a selective sweep from standing variation. We use simulations to determine which parts of the parameter space the method has power to make this distinction, and aim to estimate parameters under both models. We then apply our method to seven genes that were previously reported to be under selection. We first wanted to assess how accurately we can estimate the selection coefficient and the age of the selected mutation from the SSV and SDN models. For this purpose, we performed ABC inference on simulated data sets with known parameter values. Results for a case of moderately strong selection (α = 400) are given in Figure 2, with α being the population scaled selection coefficient α = 4Ns. As can be seen from the figure, the mode is an accurate estimator of the true value for both models. However, in the SSV case the posterior distribution is much broader than under the SDN model, and the 95% confidence interval extends to the edges of the prior, indicating low accuracy in the estimate. For the initial frequency parameter, f1, the posterior differs only marginally from the prior, and therefore this parameter cannot be reliably estimated. We aim to identify parameter regions where we can distinguish between the SSV and the SDN model. As a control, we also consider a model of neutral evolution (NT), where an allele increases to high frequency solely due to genetic drift. In particular, we are interested in three parameters that are expected to have a strong influence on model choice accuracy: the selection parameter α, the frequency of the mutation when it became selective advantageous, f1, and the current frequency of the selected allele fcur. In Figure 3 and Figure S1, we explore the accuracy of our model choice procedure in three series as a function of α, f1, and fcur. . We find (Figure 3a) that in cases where α<100; the method cannot reliably distinguish between selection and a neutral model. This is not surprising, as for such values of α, standard neutrality tests have little or no power to detect selection [44], [45]. For selection coefficients of α = 100 and α = 200 the neutral model has a very low posterior probability and would be rejected, but we still do not have sufficient power to distinguish the signals from SSV from SDN. Only under strong selection (α = 1,000) do we have reasonable power to distinguish between SSV and SDN. Thus, we find that there is a parameter range of α between 100 and 500, in which selection can be reliably detected, but the two models of selection are statistically indistinguishable. In the second series (Figure 3b), we vary the initial allele frequency (f1). We find that simulations under the SSV model, with f1 = 1%, are identified as SDN models, but that the accuracy in model choice increases with f1. For larger values of f1, we can detect selection when selection is strong (α = 1,000). For high initial allele frequency (f1 = 20%) we correctly infer the true mode of selection even when α is 200. This suggests that the ability to distinguish the two models increases with f1. Furthermore, we also find a negative relationship between the estimated value of f1 for a data set and the posterior probability of the SDN model (Figures S2, S3 and S4): As we would expect, the larger the estimate of f1, the lower is the posterior probability of the SDN model, and we find a strong negative correlation (R2 = 0. 51) between these two quantities based on 1,000 simulations. In the third series (Figure 3c), we investigate the effect of the current allele frequency fcur on the model comparison. For simulations under the SSV model, we find that the accuracy strongly decreases with fcur. For fcur = 0. 2, we classify slightly less than half of the data sets correctly. This is in contrast to simulations under the SDN model, where the power to correctly classify simulated data sets gradually increases with fcur. Thus, in studies aimed at detecting selection on standing variation, the false positive rate should depend only slightly on fcur, but the false negative rate is expected to increases drastically when fcur is low. Figure 4 illustrates how the selection parameter (α) and the initial allele frequency (f1) affect the accuracy of model choice between the SSV, SDN and NT models for three values of fcur (fcur = 0. 95, fcur = 0. 8 and fcurr = 0. 5). As in Figure 3, the number of correctly assigned data sets increases with α, f1 and fcur. Under the SSV model, the gradient with which the power declines is strongest when fcur is large (95%, Figure 4a), and becomes less pronounced for smaller fcur (see Figure 4c and 4e). For fcur = 95% (Figure 4a), there is a region with f1>0. 05 and α>1,000 where there clearly is very high power to infer the correct model. On the other hand, for α<200 or f1<0. 03, we make incorrect inferences more than half of the time, indicating that in these regions of the parameter space, the signal of the sweep is too weak to discriminate between the SSV and SDN models. While that global pattern is the same for fcur = 0. 8 and fcur = 0. 5 (Figure 4c, 4e), the distinction between regions where we can and cannot assign simulated data sets correctly is less pronounced. Quite surprisingly, however, we find that for fcur = 0. 8, the number of correctly assigned data sets increases when selection is low. The same trend holds for fcur = 0. 5 (Figure 4e), however here the influence of selection is even weaker, and inference becomes quite ambiguous, with posterior probabilities ranging from 60% to 80% in the entire parameter space. In contrast, the pattern is much simpler for simulations under the SDN model (Figure 4b, 4d and 4f), where the probability to correctly identify the model increases with decreasing fcur. When fcur is set to 0. 95, we need a selection coefficient of α = 1,500 to make confident inferences. For fcur of 0. 8 and 0. 5, this value decreases to 900 and 300, respectively. In summary, a high current allele frequency increases the power to distinguish between SSV from SDN (Figure 3c, Figure 4). The frequency with which the SDN model is correctly inferred increases slightly with decreasing fcur, presumably because the selected phase makes up a larger proportion of the trajectory. We illustrate our model choice procedure by analyzing seven genes that have previously been identified as candidates for being under selection. These genes are ADH1B, ASPM, EDAR, G6PD, LCT, PSCA and TRPV6. The genes were selected using the following set of criteria: i) there is evidence for selection from a previous study, ii) a putative causal mutation has been identified and iii) the putative causal site has reached a high frequency in at least one population, but has not yet reached fixation. In addition, we also analyzed four regions that were noncoding and presumably neutral. We retrieved polymorphism data from the 1000 Genomes Project low coverage data [46] using tabix [47]. Ancestral genotypes were inferred by comparison to the homologous chimpanzee allele. If a signal of selection was present in more than one population, we used data for the population where the selected site was most frequent, to facilitate inference. Model choice and parameter estimation were performed using the procedures described in the methods section. In contrast to the inference on simulated data sets, here we explicitly model varying recombination rates and the complex demographic history of the human population. Results for the sample genes are given in Figure 5 and Figure S4, as well as Table 1. For six of the seven genes analyzed, the neutral scenario was strongly rejected with a posterior probability of less than 1%, and we can confirm the prior evidence that these genes are under selection. Three of those genes, ADH1B, EDAR and LCT, were found to be under selection from a new mutation and one gene, TRPV6 could not be assigned with any significant probability to either model. Two genes, ASPM and PSCA, were found to be under selection from standing variation. Finally, none of the three models provided a good fit to observed data in the G6PD gene, suggesting that neither of the models is appropriate for this gene. In the following paragraphs, we will discuss each gene in some detail, and give estimates for selection coefficient and time when appropriate. All estimates are given with a point estimate for the mode, and the lower and upper bound of a 95% Highest Posterior Density interval in brackets. Estimates in years were made assuming a generation time of 25 years. The distribution of summary statistics in Figure S5 illustrates the impact of choice of summary statistic for model inference [92]. Very high values of EHH are clearly indicative of the SDN model, at both a 10 kb and 20 kb distance. Both the SSV and SDN models are associated with low IHS values, whereas the neutral regions have IHS values closer to zero. Tajima' s D and Fay and Wu' s H are both very informative for model comparison, with SDN genes having very low D values, SSV genes having D values close to zero and neutral regions having positive D values. The main exception is the LCT gene, however, which we inferred to be selected from a de novo mutation, but which has a high D. The signal for SDN apparently comes more from the high EHH and low IHS values. In general, our results are highly concordant with previous studies of these genes. Our estimates tend to gene, G6PD, we could not make any inferences, because we could not reproduce the observed pattern of diversity using simulations of positive directional selection. G6PD shows an extremely narrow region of reduced diversity, surrounded by a region of high diversity. This may be due to balancing selection between malaria resistance and reduced efficiency oxygen transport introducing a signal that cannot be reproduced by our simple model of directional selection. Alternatively, the X-linked mode of inheritance of this locus is not concordant with the assumptions of our model. This also highlights one of the dangers of ABC: It is crucial that the models investigated are able to reproduce the data observed; otherwise false inferences may be drawn. This danger inherent to any ABC approach is also highlighted by the fact that misidentification of the selected site will bias model choice results towards SSV (Figure S6). This can be explained by the fact that even if the neutral site is closely linked to the selected site, it is likely to “escape” the sweep by recombining away from the selected haplotype, thus giving the signal of selection from standing variation. Similarly, analyzing data simulated under a population bottleneck under a constant size model will bias the results towards stronger selection and SDN (Figure S7), presumably due to the younger age of mutations being taken as evidence of strong selection. We have shown that it is much more difficult to estimate the model parameters α, and t1 from the SSV model than from the SDN model. This is unsurprising, as the SSV model has been shown to have a higher variance in allele age, which results in a higher expected variance for most summary statistics [8]. We further show that there is not enough information to estimate the initial frequency of the sweep f1. This is unsurprising, as the exact position of the switching point has likely only a minor effect on the data, especially as the effect of selection on the trajectory is weak when the allele frequency of the beneficial allele is low [27]. We further notice that the accuracy of our model choice procedure decreases when the signal of selection is weak. Consistent with previous findings, selection is very hard to detect if α is below about 100 [5], [11], [45]. This is also the point where our method gains power to distinguish between SVN from SDN. The initial frequency required to detect standing variation is moderate at around 3% for weak selection and 2% for stronger selection. However, selection has to be rather strong, at around α = 1,000 and initial frequencies have to be above 5% to allow accurate inference. Presumably, this is because below this threshold, the stochasticity of the trajectory is very large even under selection, and the difference between the two scenarios is small (see also Figure 1c). These findings are not particularly surprising, as selection scans based on summary statistics have been shown in general to have low power under these conditions [45]. These findings certainly limit the scope of our approach. Could we do better with a different strategy? As discussed in the introduction, the ABC approach simplifies data in two ways. First, instead of using the full data, we use an array of summary statistics. Second, we substitute an exact match between observation and simulations with an approximate match, depending on “close” simulations. Regarding the use of summary statistics, we note that summary statistics have been widely used to detect selection from genetic data [9]–[11], [14], [19], and currently provide the only way to detect selection from DNA sequence data. No full likelihood based method is available to detect selection from DNA sequence data that could be adapted to distinguish between the two sweep models entertained here. The second simplification step is based on the number of simulations performed and the tolerance interval and is imposed by computational constraints. We examine the effect of different numbers of simulations and tolerance cutoffs on our results by calculating relative error rates of the posterior mean and the false negative rate of the model choice. We show in Table S1 that increasing the number of simulation by a large amount or changing the rejection parameter does not significantly improve our results, indicating that we do not lose a lot of information at this stage. This shows that the ABC approach reliably estimates the posterior based on the summary statistics, and as such use all the information available in these statistics. Statistics such as EHH, iHS, Tajima' s D, etc, do not contain information that will allow us to provide more reliable estimates. It the light of this, it may appear disappointing that our method does not provide more accurate parameter estimates and more power to distinguish between models. However, it is important to realize that as previously argued, all information regarding selection is in the frequency path of the selected allele [26], [27]. For relatively small selection coefficients and/or small initial frequencies of the selected allele, the paths are very similar for the SSV and SDN models. Even if a full likelihood method could be developed, it is unlikely that it had much more power to distinguish between models. A further simplification in our method is the restriction to a single population. Population differentiation measures, such as FST, are one of the most successful ways to detect sweeps from standing variation [24], [86], and the inclusion of more realistic models of demography may improve our accuracy. Such models, however, require an additional estimation of multi-population demographic history, which greatly increases the complexity of the model. While we applied our method only to human candidate loci, it should be possible to easily translate it to other species. In particular, as our simulation results suggest that we have more power to distinguish SDN and SSV if selection is strong, species with large population sizes, such as e. g. Drosophila or many microorganisms may be very promising targets for a similar study. Another possible target might be species with very strong artificial selection, such as domesticated animals or plants, where we may gain valuable insights on the domestication history of these species. Of course, our approach could also be combined with ancient DNA (e. g. [83]), which could provide much narrower confidence intervals on time estimates and also help improve estimates of selection coefficients. The two selection models we consider here, the SSV and the SDN models, are nested models, Setting f1 = 1/2N in the SSV model recovers the SDN model. To facilitate Bayesian model choice we assign positive probability to f1 = 1/2N, and base our inferences on a choice between f1 = 1/2N and f1 ∼ U (0,0. 2) (See Methods). ABC based model choice has recently been criticized and been shown to be biased in some cases where the statistics used are not sufficient [93], [94]. While some of the specific issues raised by [91] are not applicable in our setting because we consider nested models, we do not base our inference on sufficient statistics and the statistical properties of our model choice procedure are, therefore, largely unknown. To address this issue, and in general to validate our approach, we use a method introduced in [95]. We show in Figure S8 that our estimated probabilities only show bias for very small values of the Bayes factor, where there appears to be a bias towards inference of the SDN model for simulations generated under the SSV model with very low values of f1. In order to keep our problem simple, we condition on two important parameters: We assume that the exact site under selection is known from extraneous information, and we furthermore assume that the allele frequency fcur of that site at the time of sampling, tcur = 0 is known. The interpretation of the parameters is depicted graphically in Figure 1a. Unless noted otherwise, we assume a panmictic diploid population of size N with an additive selection model where the ancestral homozygous, heterozygous and derived homozygous genotypes have fitness 1,1+ s/2 and 1 + s, respectively. However, the methodology applied here can easily be adapted to more complex scenarios, e. g. models involving multiple populations, more sophisticated demographic models, and other models of selection. For most simulated data sets, we will report the population scaled mutation rate α = 4Ns, as the shape of the allele frequency trajectory depends only on that compound parameter [25]. However, for most of the genes we analyze previous estimates were made directly on s rather than the compound parameter. To facilitate comparisons, we report s for the genes we analyzed. We use a standard ABC approach [31], [32], using a post-sampling adjustment in the form of a GLM [96]. We used the ABCToolbox package [40], for specifying priors, rejection sampling and post-sampling adjustment. Unless specified otherwise, we perform 105 simulations per model, and retained the 100 (0. 1%) simulations with associated Euclidean distance between observed and simulated summary statistics closest to zero. To assess how the number of simulations and acceptance rates influence our results, we analyze 10,000 random data sets with up to 107 simulations and varying acceptance rates. We show that these parameters have very little impact on the relative error for both the model choice and parameter estimates in Table S1. All our data sets used for both the ABC inference and the assessment of our procedure are simulated using a modified version of the coalescent simulator mbs [102]. Mbs allows simulation of genetic data sets with a single selected site using the structured coalescent [26]. mbs first simulates the allele frequency trajectory of the site under selection, and then generates a data set conditional on that trajectory. We simulate allele frequency trajectories using Euler' s method on the unscaled backwards diffusion equation with selection (eq 7. 1 in [103]). This equation makes it very easy to incorporate population size changes by just changing the variance term. To simulate sweeps from standing variation, we set the selection coefficient (s) to zero the first time the trajectory reaches f1. To analyze simulated data sets, we generally simulate a 100 kb region with a recombination rate of 1. 5 cM/Mb. For the human genes, we simulate the gene and a 50 kb flanking region on both sides, resulting in regions that are usually between 100 kb and 150 kb wide. Recombination rates and hotspots are modeled by using the HapMap recombination map [104] in the application to selected genes. For all simulated data sets we assume a constant-sized population. For the analysis of human genes, we use the population history estimated by [105]. Specific regions and details of the used regions are given in Table S2. To ensure that our method does not suffer from a high false positive rate, we also analyze regions 5 Mb downstream from the candidate genes, as they are presumably neutral. For three of genes (ASPM, G6PD, and PSCA), no data was available for these downstream regions, so we analyzed the remaining loci. Candidate loci for selection were chosen using the following criteria: i) they were required to have a derived allele frequency between 0. 7 and 0. 9 and ii) to be as closely to 5 Mb away from the actual candidate locus in the upstream gene as possible. We estimate parameters from our models using the standard ABC procedure described above. The parameters we estimate are the mutation rate μ, the age of the sweep t1 the selection coefficient s and, only under the SSV model, the initial frequency f1 for the SSV model. In particular we want to determine if our posteriors are unbiased, and if we were able to get reasonable confidence in our estimates. To do this, we simulate data sets with fixed parameters and plotted the average posterior distribution for all parameters in Figure 2. For model choice, our main goal is to calculate the relative probabilities of the models given the data, i. e. Pr (SSV | data), Pr (SDN | data) and Pr (NT | data), which we calculate using the marginal densities as proposed by [96]. To identify parameter regions where there is power to distinguish between the models, we simulate 1,000 data sets each under 30 different scenarios in three series, corresponding to three parameters of interest: The strength of selection α, the frequency when the mutation became selective advantageous f1 and fcur, the frequency at which the mutation is observed. To test the algorithm for approximating Bayes factors, we also use a simulation approach. The estimator of the posterior probability from k simulations, , should have the property where m is a model indicator functions for a specific model. Also, for a particular draw from the posterior, m (0), we expect, if the simulation algorithm works properly. In other words, , should asymptotically equal, i. e. if = c, we expect a proportion c of simulations to have been obtained from model m. Equivalently, for an estimated log Bayes factor, log10 = c, we expect a proportion 10c/ (1+10c) of draws to be from model m. This prediction is tested in Figure S8, based on 10,000 random data sets from both the SDN and SSV model.

Considerable effort has been devoted to detecting genes that are under natural selection, and hundreds of such genes have been identified in previous studies. Here, we present a method for extending these studies by inferring parameters, such as selection coefficients and the time when a selected variant arose. Of particular interest is the question whether the selective pressure was already present when the selected variant was first introduced into a population. In this case, the variant would be selected right after it originated in the population, a process we call selection from a de novo mutation. We contrast this with selection from standing variation, where the selected variant predates the selective pressure. We present a method to distinguish these two scenarios, test its accuracy, and apply it to seven human genes. We find three genes, ADH1B, EDAR, and LCT, that were presumably selected from a de novo mutation and two other genes, ASPM and PSCA, which we infer to be under selection from standing variation.

Abstract Introduction Results Discussion Methods

natural selection human genetics genetics population genetics biology evolutionary biology evolutionary processes genetics and genomics evolutionary genetics evolutionary theory

2012

Distinguishing between Selective Sweeps from Standing Variation and from a De Novo Mutation

CMT4J is a severe form of Charcot-Marie-Tooth neuropathy caused by mutation of the phosphoinositide phosphatase FIG4/SAC3. Affected individuals are compound heterozygotes carrying the missense allele FIG4-I41T in combination with a null allele. Analysis using the yeast two-hybrid system demonstrated that the I41T mutation impairs interaction of FIG4 with the scaffold protein VAC14. The critical role of this interaction was confirmed by the demonstration of loss of FIG4 protein in VAC14 null mice. We developed a mouse model of CMT4J by expressing a Fig4-I41T cDNA transgene on the Fig4 null background. Expression of the mutant transcript at a level 5× higher than endogenous Fig4 completely rescued lethality, whereas 2× expression gave only partial rescue, providing a model of the human disease. The level of FIG4-I41T protein in transgenic tissues is only 2% of that predicted by the transcript level, as a consequence of the protein instability caused by impaired interaction of the mutant protein with VAC14. Analysis of patient fibroblasts demonstrated a comparably low level of mutant I41T protein. The abundance of FIG4-I41T protein in cultured cells is increased by treatment with the proteasome inhibitor MG-132. The data demonstrate that FIG4-I41T is a hypomorphic allele encoding a protein that is unstable in vivo. Expression of FIG4-I41T protein at 10% of normal level is sufficient for long-term survival, suggesting that patients with CMT4J could be treated by increased production or stabilization of the mutant protein. The transgenic model will be useful for testing in vivo interventions to increase the abundance of the mutant protein. The lipid phosphatase FIG4/SAC3 is broadly expressed in eukaryotic cells from yeast to mammals. Mutations of FIG4 are responsible for Charcot-Marie-Tooth Disease type 4J (OMIM 611228), an atypical, autosomal recessive form of CMT with severe motor dysfunction and rapid progression [1], [2]. FIG4 phosphatase activity specifically removes the 5-phosphate from the inositol ring of PI (3,5) P2, a membrane-bound phospholipid that acts as a molecular signal for trafficking and fusion of intracellular vesicles. In yeast, Fig4p is localized to the vacuole membrane in a protein complex that regulates the synthesis and turnover of PI (3,5) P2 [3]–[5]. In mammalian cells, the PI (3,5) P2 biosynthetic complex is localized in the endosomal/lysosomal vesicle system [6]. Deficiency of mammalian FIG4 or VAC14 leads to accumulation of cytoplasmic vacuoles in tissues and in cultured fibroblasts and neurons [1], [3], [7], [8]. We previously identified a spontaneous null mutant of mouse Fig4 caused by a transposon insertion [1]. The most striking phenotypes of the Fig4 null mice are spongiform degeneration of the brain and loss of neurons from the dorsal root ganglia, resulting in a severe movement disorder and lethality between 1 and 2 months of age (see video supplement to [1]). At the cellular level, Fig4 null fibroblasts exhibit reduced levels of PI (3,5) P2 [1], [9], [10]. In the CNS, astrocytes and neurons exhibit accumulation of p62, ubiquinated protein and other autophagic components in cytoplasmic inclusion bodies [11]. These abnormalities demonstrate that PI (3,5) P2 is required for completion of basal autophagy, and indicate that there is a defect in resolution of autolysosomes in deficient cells [12]. The biosynthetic complex that regulates PI (3,5) P2 contains two major proteins in addition to FIG4, the 5-kinase FAB1/PIKfyve, which phosphorylates position 5 of the inositol ring in PI3P, and the scaffold protein VAC14, composed of multiple heat-repeat structural domains [3]. Stable localization on the yeast vacuolar membrane requires interaction between Fig4p, Fab1p and Vac14p, and loss of one protein results in mislocalization of the other two [3], [9]. In the mouse, the phenotype of the spontaneous Vac14 mutation L156R mimics the Fig4 null phenotype, with neurodegeneration, cellular vacuolization and defective autophagy [3], [11]. Vac14-L156R is located in a heat repeat domain and the mutation reduces binding affinity for FAB1, thereby disrupting the PI (3,5) P2 biosynthetic complex [3]. These mutants in yeast and mouse demonstrate the importance of the stable complex between FIG4, FAB1 and VAC14. Patients with CMT4J are compound heterozygotes at the FIG4 locus, carrying the shared missense mutation I41T, on a common haplotype, in combination with a unique or “private” null allele [1]. The frequency of the I41T allele is less than 1/500 in the Northern European population [1]. The corresponding yeast mutant, I59T, retains partial function in a yeast assay for correction of the vacuole phenotype [1], [13]. Disease onset in CMT4J patients with the genotype FIG4I41T/− may occur in childhood or adult life. The rapid decline of motor function in adult onset patients resembles the course of ALS, and deleterious mutations of FIG4 have also been identified in patients with ALS [13]. In order to generate a mouse model of human CMT4J, we have expressed a Fig4 cDNA construct containing the I41T mutation in transgenic mice. Here we report the dose-dependent rescue of the Fig4 null phenotype by the Fig4-I41T transgene. We also demonstrate that the pathogenic mechanism of the I41T allele is based on defective interaction with the scaffold protein VAC14, resulting in destabilization of the FIG4 protein in vivo. While this work was in progress, Shisheva and colleagues reported related work indicating that the short half-life of a GFP-FIG4 fusion protein in cultured cells is increased by over-expression of myc-VAC14, and that the I41T mutation prevents this increase [14]. The authors proposed that VAC14 has a novel regulatory role in the turnover of FIG4 protein [14]. Using the yeast two hybrid system, we demonstrate here that the direct interaction between VAC14 and FIG4 is impaired by the I41T mutation. The reduced interaction results in greatly reduced abundance of FIG4-I41T protein in patient fibroblasts. We further demonstrate that wildtype FIG4 requires VAC14 for stability in vivo. Finally, we find that overexpression of mutant I41T protein can compensate for its reduced binding affinity and rescue the mouse model of CMT4J. This work extends the previous observations and advances our understanding of the pathogenic mechanism of the FIG4-I41T mutation. The corresponding yeast mutation, Fig4-I59T, results in impaired vacuole morphology and defective regulation of PI (3,5) P2 [1], [13]. We tested the interaction of Fig4p-I59T with the Fig4p binding partners Vac14p and Fab1p using a directed yeast two hybrid assay. Fig4p was fused to the DNA binding domain of GAL4 and Vac14p was fused to the transcription activation domain. The I59T mutant did not support growth under stringent selection in the presence of 3AT (Figure 1A). This result demonstrates reduced binding of the Fig4p-I59T mutant to Vac14p. This result was confirmed in a co-immunoprecipitaton assay using myc-tagged Fig4p and GFP-tagged Vac14p [15]. Wildtype and mutant yeast proteins were expressed at similar levels, but co-precipitation of Vac14p was reduced by approximately 75% for Fig4p-I59T (Figure 1B). Co-immunoprecipitation of GFP-labeled Fab1p was reduced to a similar extent by the Fig4p-I59T mutation (Figure 1C). The latter could be an indirect effect of impaired interaction with Vac14p. To confirm the effect of the I41T mutation in the context of the mammalian proteins, we tested the interactions of human FIG4 and human VAC14 in the yeast two hybrid system. The I41T mutation significantly impaired the interaction between the two human proteins, preventing growth on both of the selective media (Figure 1D). These experiments demonstrate that the isoleucine-to-threonine substitution reduces the direct interaction of FIG4 with VAC14. To investigate the in vivo dependence of FIG4 on interaction with the VAC14 scaffold protein, we examined FIG4 levels in tissues from a Vac14 null mouse [7]. The absence of VAC14 in the null mouse was confirmed by Western blot (Figure 2A). To detect FIG4 protein we generated a monoclonal antibody to a bacterially-expressed 220 amino acid fragment from the C-terminus of FIG4 (Materials and Methods). The monoclonal antibody recognizes a single protein of ∼100 kDa in homogenates of mouse tissues, consistent with the calculated molecular weight of 103 kDa for the 907 amino acid FIG4 protein (Figure S1). Remarkably, the abundance of FIG4 protein was greatly reduced in the Vac14 null mouse (Figure 2B), although the level of Fig4 mRNA was normal (Figure 2C, 2D). Although FIG4 protein was undetectable in the tissue extract, a very low level of protein could be detected in cultured fibroblasts by Western blot (Figure S2). The data demonstrate that wildtype mammalian FIG4 protein is dependent on interaction with VAC14 for stability. To investigate the in vivo function of the mutant protein, we generated a transgene construct containing the mouse Fig4-I41T cDNA under the direction of the ubiquitously expressed chicken β-actin promoter (Figure 3A). Expression of Fig4-I41T on a wildtype background did not cause any visible abnormality in two independent transgenic lines, which exhibited normal fertility and life span. The transgene copy-number was measured by analysis of genomic DNA by qPCR and demonstrated a transgene copy number of 4 copies in line Tg705 and 2 copies in line Tg721 (Figure S3). Brain RNA was prepared from mice carrying the transgene on a wildtype genetic background. qRT-PCR reactions were carried out with Taqman primers in exon 3 and 4 and the product was detected with a probe spanning the junction between exon 3 and exon 4. Both endogenous transcripts and transgene-derived transcripts are detected by this assay. The abundance of Fig4 transcripts was compared with Tata binding protein (Tbp) transcripts as an internal control (see Materials and Methods). The ratio of Fig4 transcripts to Tbp transcripts in Fig4+/+, Tg705 brain was 3× higher than in nontransgenic Fig4+/+ mice (Figure 3B). Subtracting the 1× contribution from the endogenous Fig4 alleles, the contribution of the 705 transgene is 2× the level of endogenous expression. In the brain of the higher expressing line Tg721, Fig4 transcript was 6× higher than in Fig4+/+ mice, indicating that the transgene transcript is expressed at 5× endogenous expression (Figure 3B). In order to generate a model of CMT4J expressing I41T in the absence of wildtype Fig4 protein, the FIG4-I41T transgenic mice were crossed with heterozygous mice carrying the Fig4 null (plt) allele. Fig4-null mice carrying the I41T transgene were generated in the expected Mendelian proportions from standard two-generation crosses with both transgenic lines. When extracts from brain and kidney of Fig4−/−, Tg705 and Fig4−/−, Tg721 mice were examined by Western blotting, the abundance of FIG4-I41T protein was substantially lower than in wildtype tissues (Figure 3C, 3D). With longer exposure, a low level of FIG4-I41T protein could be detected in the Tg705 line, and a higher level in the Tg721 line, consistent with their relative transcript levels (Figure 3E). Comparison with a 1∶10 dilution of wildtype brain indicated that the level of FIG4-I41T protein in line Tg721 is approximately 10% of wildtype (Figure 3F). The low abundance of I41T protein was confirmed with a second antibody, a rabbit polyclonal antibody generated to the same C-terminal antigen (Figure S4). The abundance of VAC14 protein is normal in transgenic and Fig4 null mice (Figure 3G, 3H). Thus low VAC14 protein is not responsible for the low level of FIG4-I41T protein in the transgenic lines. The data are consistent with the evidence above from VAC14 null mice, and indicate that the FIG4-I41T protein is destabilized in vivo by its reduced affinity for VAC14. Inheritance of the low expressing transgene Tg705 increased survival of Fig4 null mice from 1–2 months to 3–6 months (Figure 4). These mice provide a model of human CMT4J, as described below. In the high-expressing Tg721 line, lethality was completely corrected (Figure 4). The oldest cohort of Fig4−/−, Tg721 mice have survived more than 28 months with no visible abnormalities. The data demonstrate that a relatively low level of Fig4-I41T protein can rescue the lethality of Fig4 null mice. Fig4 null mice exhibit a reproducible pattern of spongiform degeneration in the brain and extensive loss of neurons from peripheral ganglia (Figure 5, top panel). Neurons in layers 4 and 5 of the cortex, the deep cerebellar nuclei, and the dorsal root ganglia (DRG), are severely affected, with accumulation of vacuoles that fill the cytoplasm [1]. In null mice carrying the Tg721 transgene, these abnormalities are almost completely eliminated. Spongiform degeneration of the brain is minimal and DRG neurons are intact at P90 (Figure 5). In line Tg705, an intermediate level of degeneration is visible at P90. In both transgenic lines degeneration of the cerebellar nuclei is visible, indicating that this region is extremely sensitive to PI (3,5) P2 levels (Figure S5). Overall, neurodegeneration is rescued in a dose-dependent manner. Enlarged lateral ventricles and hydrocephalus are seen in both transgenic lines (Figure 5, left panels). High pressure hydrocephalus is indicated by the compression of the cerebellum and hippocampus observed in all of the Fig4−/−, Tg705 mice near the end of their lifespan (Figure 5, bottom panel). This is also indicated by the domed appearance of the head and the expulsion of cerebrospinal fluid during dissection. The hydrocephalus in these mice is very similar to that in the L156R (ingls) mutant of Vac14 [3]. Spongiform neurodegeneration in Fig4 null brain is accompanied by accumulation of p62 and other autophagy intermediates, predominantly in activated astrocytes [11]. In brain from transgenic mice, there is an intermediate level of accumulation of the autophagy markers p62 and LAMP-2 (Figure 6A). Accumulation of the astrocyte protein GFAP is also corrected in a dose-dependent manner (Figure 6A). Reduction of astrocytosis in Tg721 mice is indicated by the decreased number of GFAP positive cells (Figure 6B). Accumulated p62 and lysosomal membrane protein LAMP-2 are localized in astrocytes of the trangenic mice (Figure 6B), as previously shown in null mice [1]. In the high expressing Tg721 line, astrocytosis and GFAP accumulation are almost completely corrected (Figure 6B). The defective myelination of peripheral nerves characteristic of Charcot-Marie-Tooth disease is also seen in Fig4 null mice ([1] and Figure 7A, top panel, arrows). In the two transgenic lines, sciatic nerve myelination was comparable to wildtype (Figure 7A). To quantitate axonal myelination we calculated the g-ratio, the inner axon diameter divided by the diameter of the nerve fiber (Figure 7B). At postnatal day 21, the g-ratio of 0. 5 for wildtype sciatic nerve is increased to 0. 7 for Fig4 null sciatic nerve by the thinning of the myelin sheath. In the transgenic mice the g-values were restored to wildtype (Figure 7B). Slowed nerve conduction secondary to defective myelination is another characteristic of Charcot-Marie-Tooth disease that is reproduced in Fig4 null mice [1]. In sciatic nerve of null mice at 1 month of age, the conduction velocity was 50% of the wildtype value [1]. In contrast, conduction velocities in sciatic and sural nerves from the two transgenic lines, measured at 4 months and 14 months of age, varied between 65% and 80% of wildtype (Figure 7C, 7D), demonstrating partial rescue of the defect. The amplitude of the compound action potential in sciatic nerve was was also restored in the transgenic mice (data not shown). Fig4-null mice have diluted pigmentation due to reduced numbers of melanosomes in the mature hair follicle and clumping of melanosomes within the hair shaft [1]. Pigmentation is partially rescued in the higher expressing line Tg721, but not in line Tg705 (Figure S6). In view of the role of Fig4 in autophagy, it is interesting that autophagy components may play a role in melanosome biogenesis [16]. In order to determine whether the low abundance of the FIG4-I41T protein in transgenic mice was representative of patient tissues, we examined fibroblasts from a CMT4J patient who is a compound heterozygote for I41T in exon 2 and the null allele R183X in exon 6 [2]. Exon 2 was amplified from patient and control fibroblast RNA (Figure 8A). The sequence of the control RNA contains a T nucleotide at position 122, encoding the wildtype isoleucine allele. The patient RNA contains a C nucleotide encoding threonine (Figure 8B). The absence of the wildtype nucleotide in patient RNA indicates that the R183X transcript is not stably expressed, probably due to degradation by nonsense-mediated decay [17], consistent with its location at a distance from the C-terminal exon 23. Western blot analysis of fibroblast extracts demonstrated that FIG4-I41T protein in patient fibroblasts is dramatically reduced in comparison with wildtype (Figure 8C). The very low abundance of FIG4-I41T protein, well below the expected 50% of normal, indicates that the mutant protein is unstable in patient tissues, as it is in the transgenic mice. The level of endogenous FIG4 in HEK293 cells was increased by transfection of VAC14 (Figure 8D, lane 1 and 2). The level of tranfected wildtype mouse FIG4 in HEK cells was also increased by co-transfection of VAC14 (Figure 8D, lane 3 and 4). The abundance of FIG4-I41T after transfection was much lower than for the wildtype protein, and the effect of co-transfected VAC14 was much smaller (Figure 8D, lane 5 and 6). These experiments demonstrate that wildtype FIG4 can be stabilized by coexpression of VAC14 in cultured cells, and that mutant FIG4-I41T is less effectively stabilized. The level of FIG4 protein in primary fibroblasts from Tg721 transgenic mice is lower than in fibroblasts from wildtype mice (Figure 8D), but is sufficient to prevent vacuolization (data not shown). It was recently reported that the proteasome inhibitor MG-132 increases the level of endogenous FIG4 protein in COS7 cells, indicating that there is turnover of FIG4 in the proteasome [14]. To determine whether the mutant FIG4-I41T protein could be stabilized, we cultured primary fibroblasts for 8 hrs with 10 uM MG-132. The level of FIG4-I41T protein was significantly increased by this treatment in fibroblasts from Tg721 transgenic mice (Figure 8E). In control fibroblasts the level of wildtype FIG4 was also increased by culture with MG-132 (Figure 8F). The MW of protein produced in the presence of MG-132 was higher than in untreated cells (Figure 8E, 8F). Since MG-132 does not prevent ubiquitination of substrates, the higher MW may be a consequence of ubiquitination of the FIG4 protein. We have demonstrated that the FIG4-I41T protein is unstable in vivo in cultured cells and in transgenic mice, and that the amount of protein in cells from CMT4J patients is extremely low. FIG4 protein expression equivalent to 10% of wildtype levels is sufficient to prevent neurodegeneration and completely rescue lethality in transgenic line Tg721. In contrast, transgenic line Tg705, with lower expression of FIG4-I41T, provides an animal model of the human disorder with neurodegeneration. These key observations suggest that increasing the expression of the FIG4-I41T allele in CMT4J patients, or stabilizing the protein, would be therapeutic. The low level of FIG4 protein in rescued mice and in patient fibroblasts appears to be a consequence of the direct effect of the I41T mutation on interaction with the scaffold protein VAC14. In yeast the co-localization of Fig4p, Fab1p and Vac14p on the vacuolar membrane requires the presence of all three proteins, and loss of one protein prevents localization of the other two [3]–[5]. Similarly, in mammalian cells expressing an shRNA to downregulate Vac14 expression, a small reduction in endogenous FIG4 protein was reported [14]. We confirmed the importance of the FIG4-VAC14 interaction for stability of wildtype FIG4 with the demonstration that FIG4 protein is drastically reduced in mice homozygous for a null allele of VAC14. This experiment clearly demonstrates the dependence of wildtype FIG4 protein on VAC14 for in vivo stability. Because of its reduced affinity for VAC14, FIG4-I41T is a hypomorphic allele encoding an unstable protein, resulting in a very low steady-state level of protein in vivo. The VAC14 protein is composed of heat-repeat domains [3] and is thought to function as a scaffold for the PI (3,5) P2 biosynthetic complex. In another example of the importance of interactions between proteins in this complex, the missense mutation of Vac14 in the ingls mouse which reduces the affinity of VAC14 for FAB1 and results in a neurodegenerative disease that closely resembles the Fig4 null mice [3]. In the reciprocal experiment, VAC14 protein was not reduced in Fig4 null mice, demonstrating the greater intrinsic stability of the mammalian scaffold protein, and/or its stabilization by interaction with other components of the complex. This model of pathogenesis is consistent with the structure of the FIG4 protein, which was predicted by superimposition with the crystal structure of Sac1p, a closely related lipid phosphatase [18]. The I41T mutation is located near the surface of the non-catalytic domain, in a hydrophobic pocket between two β-sheets (Figure 9). The mutation was predicted to affect protein-protein interaction by destabilizing the SacN domain [18], consistent with our observations. The I41T mutation is located at a distance from the catalytic domain, and appears not to affect the enzymatic activity [14]. Based on our observations, increased expression or stabilization of the FIG4-I41T protein in patients with CMT4J should be therapeutic and could achieve complete rescue of this progressive neurodegenerative disorder. Inhibition of proteasome degradation with MG132 resulted in increased levels of FIG4-I41T protein in cultured fibroblasts. This non-specific agent increased the abundance of approximately 200 proteins in cultured fibroblasts [19]. The proteasome inhibitor Velcade (bortezemib) has been approved for treatment of multiple myeloma, and the widely used drug disulfiram (Antabuse) was recently shown to have activity as a proteasome inhibitor [20]. Pharmacological interventions like these might increase FIG4 concentration to a level sufficient to protect against loss of motor function in CMT4J patients with genotype FIG4I41T/−. Histone deacetylase inhibitors such as sodium butyrate and tricostatin A that increase the expression of other neurological disease genes such as SMN should also be evaluated in CMT4J fibroblasts. The correlated stepwise rescue of autophagy, gliosis, neurodegeneration and lethality in the two transgenic lines supports our proposed model of pathogenesis in which accumulation of autophagy intermediates leads to neuronal damage and then to gliosis, neural cell death and lethality [11], [12]. PI (3,5) P2 appears to be required at a step subsequent to the formation of the autolysosome [11], and may be involved in the regeneration of lysosomes from autolysosomes [21]. A newly recognized function of PI (3,5) P2 is activation of the lysosomal calcium channel TRPML1, which is mutated in the neurodegenerative disorder mucolipidosis type IV [22]. Transfection of TRPML1 into Vac14 null fibroblasts rescued the vacuolization caused by deficiency of PI (3,5) P2 [22]. Increased lysosomal ion concentrations resulting from PI (3,5) P2 deficiency in FIG4 mutant cells could contribute to vacuolization via osmotic retention of water. Low molecular weight activators of TRPML1 are under development and might provide another route to treatment of PI (3,5) P2 deficiency. The response of various tissues to rescue of FIG4 deficiency in transgenic mice is summarized in Table 1. The dramatically improved survival of cortical and dorsal root neurons, and the complete rescue of lethality, indicate that a small amount of I41T protein is sufficient for normal function in most affected cells. Fifteen CMT4J pedigrees segregating the I41T allele have been identified to date ([1], [2] Nicholson et al, unpublished data). In all of the unrelated families the I41T allele is inherited on the same chromosome haplotype, indicating inheritance of a shared founder mutation. In spite of their shared FIG4I41T/− genotype, the clinical course in CMT4J patients is highly variable. Age of onset ranges from an early childhood form with developmental delays that resembles Dejerine-Sotas syndrome to an adult onset form with rapid progression that may be triggered by trauma [1], [2]. Background genetic variation affecting the level of expression of VAC14 or other proteins in the PI (3,5) P2 biosynthetic complex could contribute to the clinical differences in patients with identical FIG4 genotype. The Tg705 model of CMT4J, which survives for 3 to 6 months and then develops severe disease, will be useful for testing therapies designed to increase the in vivo level of FIG4-I41T protein. This study was carried out in strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. The protocol was approved by the University Committee on the Use and Care of Animals (UCUCA) of the University of Michigan (Protocol No. 08629). The Fig4 null mutant “pale tremor” (plt) arose spontaneously in a mixed strain background that included C57BL/6J, C3H, 129 and SJL. This stock was crossed for one generation to strain C57BL/6J. Heterozygous plt/+ offspring were crossed to strain CAST/Ei to produce an F2 generation for genetic mapping [1]. plt/+ F2 mice were intercrossed to initiate a recombinant inbred line that is maintained by brother×sister breeding and is now at generation F12. The genetic background of this line, designated CB. plt, is approximately 50% CAST/Ei and 25% C57BL/6J, with smaller contributions from strains C3H, 129 and SJL. The mouse Fig4 transcript was amplified from brain RNA isolated from strain C57BL/6J using two primers containing SphI sites: F, ACG CAT GCT ATG CTA TGT GTC TGG TGT GCT GGA GGT CTG and R, TCG CAT GCA GTC CTT TAC CCA TGA GCT GCA TC. The product was digested with SphI and subcloned into the corresponding site of the vector pCAG3z [23], [24]. The I41T mutation was incorporated into the Fig4 cDNA clone by site-directed mutagenesis and the construct was completely sequenced. Plasmid DNA was isolated with the Qiagen MaxiPrep Kit and digested to generate a linear fragment containing the cDNA sequence with the promoter and polyadenylation site. Transgenic mice carrying the Fig4-I41T cDNA construct were generated by microinjection of (C57BL/6J X SJL) F2 mouse oocytes at the Transgenic Animal Model Core at the University of Michigan (www. med. umich. edu/tamc). Transgenes were maintained in heterozygous state during breeding to prevent unequal recombination between multi-copy inserts. The yeast two-hybrid test and immunoprecipitation analysis were carried out as previously described [3]. Human and yeast VAC14 were subcloned into the ClaI–BglII and XmaI–SalI sites of pGAD, respectively. Human and yeast wild type and mutant FIG4 were subcloned into the XmaI–SalI and BamHI–PstI sites of pGBD, respectively. pGAD and pGBD plasmids were cotransformed into the yeast strain PJ69-4A. Transformants were initially plated onto SC-LEU-TRP media for plasmid selection and replica-plated onto selective plates with SC-LEU-TRP-ADE-HIS+3AT media or SC-LEU-TRP-ADE-HIS media and grown at 24°C for 4 to 14 days for colony formation. The yeast strain PJ69-4A and the pGAD and pGBD vectors have been described [25]. The Venus-derivative of GFP was previously described [15]. Selective media contained 3-amino-1,2, 4-triazole (3AT), an inhibitor of His3p. Immunocytochemistry of p62, LAMP2 and GFAP on fresh frozen cryosections of mouse brain and Western blotting of brain extracts were carried out as previously described [11]. For immunoblotting of FIG4, tissues were homogenized in 0. 25 M sucrose, 0. 05 M Tris, pH 7. 5 and the 21,000× g soluble fraction was analyzed. Immunofluorescence images were captured on an Olympus BX51 microscope equipped with epifluorescence and processed and merged with Adobe Photoshop software. A rabbit polyclonal antiserum and a mouse monoclonal antibody were generated to a C-terminal fragment of FIG4 containing the final 220 amino acids, residues 688 to 907, which is encoded by exons 18 to 23. A 660 bp cDNA fragment was amplified from a full length Fig4 cDNA derived from strain C57BL/6J and cloned into the XhoI and EcoRI sites of the expression vector pRSETA (Invitrogen Corporation, Calrsbad, CA, USA), adding a polyhistidine tag (6× His) to the N-terminus. Recombinant protein was expressed in E. coli and purified from inclusion bodies using Novagen BugBuster Protein Extraction Reagent (EMD Chemicals, Gibbstown, NJ, USA). Rabbit polyclonal antiserum to the purified protein was generated and affinity-purified by Pocono Rabbit Farm and Laboratory, Inc. (Canadensis, PA). The purified polyclonal antiserum was used at 1∶50 dilution for western blots. The monoclonal antibody was generated using the same protein fragment at the UC Davis/NIH NeuroMab Facility (Clone N202/7). The monoclonal was purified from cell culture and used at 1∶200 for western blots. A broad crossreacting band of 50 to 55 kDa is ocassionally observed on Western blots of control and mutant tissues but there is no consistent association with genotype. Brain and spinal cord were fixed for 24 hours at 4°C in phosphate buffered 10% formalin and then in 70% ethanol for an additional 24 hours at 4°C. Paraffin embedding, decalcification of spine, and H&E staining were carried out at Histoserv Inc (Maryland). Images were obtained with an Olympus BX51 microscope and DP50 camera. Image capture settings were identical for sections from −/− and −/−, Tg705 mice that exhibited immunostaining for autophagy markers. To detect the outlines of brain sections from −/−, Tg721 and +/+ mice that lacked immunostaining, a longer image capture time was used. Five ug aliquots of total RNA from whole brain was treated with DNAse I (Invtrogen) and cDNA was prepared using the SuperScript First Strand Synthesis System for RT-PCR (Invitrogen). The Fig4 transcript was quantitated using TaqMan gene expression probe Mm00506074_m1 (ABI) which spans the junction between exon 3 and exon 4. As an internal control, the TATA binding protein (Tbp) transcript was quantitated using TaqMan probe Mm00446971_m1. Fluorescence was measured on a Step One Real Time PCR System (ABI) at the Microarray Core at the University of Michigan. A linear relationship between copy number and CT, the threshold cycle for detection of fluorescence, was observed (Figure S3). The mean CT value was determined from quadruplicate assays of each sample. The value of ΔCT was calculated by subtracting the CT for Tbp from the CT for Fig4. The ratio of Fig4 transcripts to Tbp transcripts was calculated as 2−ΔCT. Transgene copy number was determined using the TaqMan probe described above. A standard curve was generated using pCAG3Z-Fig4 diluted into wildtype mouse DNA for copy-number standards of 0,0. 5,1, 2,4, 8, and 16 copies per haploid genome using 2. 83 pg/ug for one haploid copy equivalent for this 3 kb transgene. Genomic DNA (50 ng) from Tg705/+ mice, Tg721/+ mice, and standards were assayed in quadruplicate. The Taqman probe includes the junction between exon 3 and exon 4, which are separated by 4 kb in genomic DNA. No fluorescent signal was obtained from 8 replicates of wildtype genomic DNA. R2 for the standard curve was 0. 999. Observed copy numbers for both transgenic lines are within the linear range of the assay. HEK293 cells (passage 15) were maintained in DMEM/F-12 supplemented with 10% fetal bovine serum, penicillin, streptomycin and amphotericin. Subconfluent cells in 10 cm dishes were transfected with 6 ug plasmid DNA and 18 uL FuGene (Roche) for 24 hours in complete culture media. Cells were lysed in RIPA buffer and lysates were spun at 15,000× g for 10 minutes. Supernatant protein was quantitated using the BCA kit (Pierce). Mouse fibroblasts were isolated from P0 mouse tail biopsy by digestion with collagenase type 2 (Worthington labs) and cultured in RPMI 1640 supplemented with 15% fetal bovine serum and containing penicillin, streptomycin and amphotericin. Experiments were carried out at passage 3 to 5. Treatment with 10 uM MG-132 (Sigma) was carried out in complete media for 8 hours. Primary cultured fibroblasts from CMT4J patient 2 were previously described [8]; passage 4 cells were kindly provided by Dr. Jun Li, Department of Neurology, Vanderbilt University. Nerve conduction velocity (NCV) was measured as previously described [26]. Mice were anesthetized and maintained at a 34°C core temperature with a heating lamp. Sural sensory NCV was determined by recording at the dorsum of the foot and antidromically stimulating with supramaximal stimulation at the ankle. NCV was calculated by dividing the distance by the take-off latency of the sensory nerve action potential. Sciatic-tibial motor NCV was determined by recording at the dorsum of the foot and orthodromically stimulating with supramaximal stimulation first at the ankle, then at the sciatic notch. Latencies were measured in each case from the initial onset of the compound muscle action potential. The sciatic-tibial motor NCV was calculated by subtracting the measured ankle distance from the measured notch distance. The resultant distance was then divided by the difference in the ankle and notch latencies for a final nerve conduction velocity. Mice were anesthetized with ketamine and xylazine and perfused transcardially with 3% paraformaldehyde (Electron Microscopy Sciences) and 2. 5% glutaraldehyde (Ted Pella, Inc.). Sciatic nerves were dissected and post-fixed for several hours in perfusion solution at 4°C. Tissues were incubated in a 1% solution of OsO4 and embedded in epoxy resin. Ultrathin (75 nm) sections were cut and visualized with a Philips CM-100 TEM. Images were analyzed at 2600-fold magnification using Image-J software for quantification of g-ratio (defined as the diameter of an axon within its myelin sheath divided by the diameter of the axon outside of its myelin sheath). Two measurements were made per axon in order to account for elongated or irregular shape, and the mean was used. Only myelinated axons were used in for the calculation of g-ratio. Data were compared using Microsoft Excel. Accession numbers for the genes referred to in this paper are: human FIG4 (Entrez Gene ID 9896), mouse Fig4 (Entrez Gene ID 103199), yeast Fig4p (Entrez Gene ID 855392), mouse Vac14 (Entrez Gene ID 234729), yeast Vac14p (Entrez Gene ID 851102), mouse Fab1/PIKfyve (Entrez Gene ID 18711), yeast Fab1p (Entrez Gene ID 850574).

Charcot-Marie-Tooth disease type 4J is a severe neurological disorder with childhood or adult onset and progression to loss of mobility and death. Patients inherit a mutation that changes amino acid residue 41 of the FIG4 protein from isoleucine to threonine. We report that this mutation destabilizes the FIG4 protein by blocking its interaction with a stabilizing protein partner. We developed a mouse model of CMT4J and found that a low level of expression of the mutant protein, 10% of wildtype level, is sufficient to prevent lethality. This work provides the scientific basis for development of a directed treatment for this rare, lethal disorder.

Abstract Introduction Results Discussion Materials and Methods

medicine model organisms neurological disorders neurology genetics biology neuroscience genetics and genomics

2011

Pathogenic Mechanism of the FIG4 Mutation Responsible for Charcot-Marie-Tooth Disease CMT4J

It has long been thought that iminosugar antiviral activity is a function of inhibition of endoplasmic reticulum-resident α-glucosidases, and on this basis, many iminosugars have been investigated as therapeutic agents for treatment of infection by a diverse spectrum of viruses, including dengue virus (DENV). However, iminosugars are glycomimetics possessing a nitrogen atom in place of the endocyclic oxygen atom, and the ubiquity of glycans in host metabolism suggests that multiple pathways can be targeted via iminosugar treatment. Successful treatment of patients with glycolipid processing defects using iminosugars highlights the clinical exploitation of iminosugar inhibition of enzymes other than ER α-glucosidases. Evidence correlating antiviral activity with successful inhibition of ER glucosidases together with the exclusion of alternative mechanisms of action of iminosugars in the context of DENV infection is limited. Celgosivir, a bicyclic iminosugar evaluated in phase Ib clinical trials as a therapeutic for the treatment of DENV infection, was confirmed to be antiviral in a lethal mouse model of antibody-enhanced DENV infection. In this study we provide the first evidence of the antiviral activity of celgosivir in primary human macrophages in vitro, in which it inhibits DENV secretion with an EC50 of 5 μM. We further demonstrate that monocyclic glucose-mimicking iminosugars inhibit isolated glycoprotein and glycolipid processing enzymes and that this inhibition also occurs in primary cells treated with these drugs. By comparison to bicyclic glucose-mimicking iminosugars which inhibit glycoprotein processing but do not inhibit glycolipid processing and galactose-mimicking iminosugars which do not inhibit glycoprotein processing but do inhibit glycolipid processing, we demonstrate that inhibition of endoplasmic reticulum-resident α-glucosidases, not glycolipid processing, is responsible for iminosugar antiviral activity against DENV. Our data suggest that inhibition of ER α-glucosidases prevents release of virus and is the primary antiviral mechanism of action of iminosugars against DENV. Iminosugars are considered to be promising candidates for broad-spectrum antiviral activity because of their presumed mechanism of action as glycoprotein processing inhibitors [1]. 1-Deoxynojirimycin (DNJ) iminosugar derivatives possess glucose stereochemistry and inhibit infectious virus production in vitro of viruses including dengue virus (DENV) [2–7], hepatitis B virus (HBV) [8,9], hepatitis C virus (HCV) [10], human immunodeficiency virus (HIV) [11], and influenza A virus [12]. Bicyclic iminosugars possessing glucostereochemistry, such as castanospermine, also inhibit infectious virus production in vitro [11,13–15]. Antiviral efficacy of both bicylic and monocyclic iminosugars has been further demonstrated in vivo, particularly against DENV infection, with protection in lethal mouse models conferred by post-exposure therapeutic administration of N-butyl-DNJ (NB-DNJ, Miglustat, Zavesca) [5], N- (9-methoxynonyl) -DNJ (UV4, MON-DNJ) [16], and 6-O-butanoyl-castanospermine (BuCAST, celgosivir) [17,18]. These promising in vitro and in vivo results have led to clinical trials of both MON-DNJ and celgosivir as dengue therapeutics. The antiviral activity of iminosugars is presumed to be a function of inhibition of endoplasmic reticulum (ER) -resident α-glucosidases I and II; thus, it has been hypothesized that antiviral efficacy of this class of drugs may be refractory to development of drug-resistant mutants because a host protein, rather than a viral protein, is the target of drug activity. Therefore treatment should be relatively refractory to escape and be effective against many viruses possessing N-linked glycoproteins [1,19]. DENV possesses four N-linked glycoproteins: envelope, pre-membrane, non-structural protein 1, and non-structural protein 4B [20]. Previous work has demonstrated that iminosugar treatment reduces association of DENV glycoproteins with the ER-resident glycoprotein chaperone calnexin, and as a result, secretion of the three DENV glycoproteins studied (envelope, pre-membrane, and non-structural protein 1) was reduced [6]. Furthermore, siRNA-mediated reduction of calnexin, calreticulin, or human immunoglobulin heavy chain binding protein (BiP), which is another glycoprotein chaperone, led to decreased production of infectious DENV [21]. Generation of free oligosaccharides (FOS), which are markers of inhibition of ER α-glucosidase activity [22], has been correlated with iminosugar antiviral activity for two DNJ-derived iminosugars [23]. Although these reports present strong circumstantial evidence for dependence of iminosugar antiviral activity on inhibition of ER α-glucosidase activity, the ubiquity of d-glucose in metabolism suggests that other pathways may be equally affected by iminosugar treatment. Indeed, NB-DNJ has been approved for clinical use since 2002 as a treatment for Gaucher’s Disease [24]–a lysosomal storage disorder. In this context, NB-DNJ is used as an inhibitor of glucosylceramide synthase (GCS) to reduce production of glycosphingolipids (GSLs) that accumulate to the detriment of normal cellular function [25,26]. It has been suggested that neutral GSLs are of importance for binding of DENV to both mammalian and mosquito cell surfaces [27], and large-scale rearrangements of the endoplasmic reticulum and associated membranes during the course of DENV infection suggest that significant rearrangement of the cellular (glyco) lipid pool is necessary to support DENV infection and replication [28,29]. Thus, iminosugar antiviral activity may be dependent upon perturbation of host glycolipid processing rather than, or concomitant with, host glycoprotein processing. A clear understanding of the role of interference with glycolipid processing in iminosugar antiviral activity is essential for defining the mechanism (s) of action of this class of drug against DENV. In this study, we delineate the relative antiviral contribution of glycoprotein and glycolipid modulation by iminosugars. In order to determine the role of glycolipid inhibition in iminosugar antiviral effect, we utilized a small panel of the most commonly investigated DNJ iminosugars and an equivalent panel of structurally similar 1-deoxygalactonojirimycin (DGJ) iminosugars (Fig 1). DGJ analogues are identical to their DNJ counterparts with the exception of the stereochemistry of the C4-hydroxyl group to achieve galactose, rather than glucose, stereochemistry of the sugar head-group. By comparison of in vitro and cellular enzyme inhibition profiles of DNJ and DGJ iminosugars, we determine the roles of iminosugar inhibition of glycolipid and glycoprotein processing on DENV antiviral activity. For in vitro infection assays, DENV2 strain 16681 (a gift from E. Gould, Centre for Ecology and Hydrology, Oxford, UK) was propagated in C6/36 Aedes albopictus cell line (US Armed Forces Research Institute of Medical Sciences, Thailand (AFRIMS) ), collected from supernatant, and concentrated by precipitation with 10% weight per unit volume (w/v) poly (ethyleneglycol) Mr 6,000 (Sigma), 0. 6% sodium chloride (Sigma) overnight at 4°C. Following precipitation, virus was centrifuged at 2830 x g for 45 minutes at 4°C, resuspended in Leibovitz’s L15 + 10% HI-FBS, and stored at –80°C until use. For in vivo experiments, mouse-adapted DENV2 strain D2S10 was amplified in C6/36 cells as described previously [30]. The iminosugar compounds tested herein were synthesized according to previously reported protocols or obtained from industrial sources. Monocylic iminosugars include d-DNJ (solubilised in water), NB-DNJ (solubilised in PBS, gift from Oxford GlycoSciences Ltd.), NN-DNJ (solubilised in 83% dimethyl sulfoxide (DMSO) ), d-DGJ (solubilised in water), NB-DGJ (solubilised in 83% DMSO, Toronto Research Chemicals) and NN-DGJ (solubilised in 83% DMSO, Toronto Research Chemicals). Bicyclic iminosugars include castanospermine (solubilised in water, Cambridge Biosciences Ltd) and celgosivir (solubilised in water, gift from Subhash Vasudevan). Inhibition of glucosidases and glycosidases was detected as previously described [32]. Briefly, assays were performed at the optimum pH of each enzyme, and appropriate disaccharide was provided as substrate for glucosidase release assays. Release of d-glucose was measured using a Glucose B-test (Wako Diagnostics). Appropriate p-nitrophenyl glycoside was added as substrate for additional glycosidase assays, and assays were stopped by adding 400 mM Na2CO3. Release of p-nitrophenol was measured spectrophotometrically at 400 nm. Recombinant Mus musculus ER alpha-glucosidase II inhibition was measured by incubating 20 nM enzyme with various concentrations of inhibitors in 100 mM potassium phosphate pH 7. 2 at 37°C for 5 minutes. Activity was measured by addition of 0. 25 mM 4-methylumbelliferyl α-d-glucopyranoside (4MUG) (Sigma Aldrich) in black, non-binding surface-treated microplates (Corning). Fluorescence upon hydrolysis of the 4MUG was measured at λex of 355 nm and λem of 460 nm in a M5 multi-mode plate reader (Molecular Devices) at 37°C, measuring fluorescence every minute for a total of 30 minutes. The initial velocity was fit in the linear range of measurements, and activity was normalized to the uninhibited control in order to plot activity against the inhibitor concentration. A sigmoidal, four-parameter function with a variable slope was used to determine the EC50 values in Prism (GraphPad). Cytotoxicity of compounds was assessed by a cell proliferation assay for metabolic activity using a CellTiter 96 AQueous One Solution Cell Proliferation Assay (Promega) as per the manufacturer’s instructions. Briefly, 20 μl of a solution containing tetrazolium compound [3- (4,5-dimethyl-2-yl) -5- (3-carboxymethoxyphenyl) -2- (4-sulfophenyl) -2H-tetrazolium; MTS] and electron coupling reagent (phenazine ethosulfate) was added to cells covered or suspended in 100 μl of culture medium in a 96-well plate. Samples were incubated for approximately 4 hours (37°C, 5% CO2), and the absorbance at 490 nm (A490) was measured on a SpectraMax M5 microplate reader (Molecular Devices). Absorbance was normalised to untreated controls whereby cytotoxicity resulted in decreased A490. MDMΦ were infected with DENV2 16681 diluted to a multiplicity of infection (MOI) of 1 in X-VIVO10 without supplements for 90 minutes (20°C, with rocking). Upon removal of virus, fresh MDMΦ growth medium without IL-4 but containing serial dilutions of drug or control was placed on the cells, and cells were incubated for 48 hours (37°C, 5% CO2). For collection of infectious virus, supernatant was harvested and centrifuged for 5 minutes (room temperature, 400 x g) to pellet any cells/debris, and supernatants were aliquoted and stored at -80°C until analysis by plaque assay. In the case of FOS and GSL assays where MDMΦ were uninfected, treatments were essentially as above except that no virus was added and cells were washed three times with sterile PBS after 24 or 48 hours as per sample time point. On the third wash, cells were removed from tissue culture plastic by mechanical disruption (scraping) and transferred to microcentrifuge tubes for centrifugation (room temperature, 4000 x g, 5 minutes). Cells were lysed by three cycles of freeze-thawing (alternating room temperature and –80°C) in deionised H2O and stored at –80°C for subsequent protein, FOS, or GSL assay. Virus titres for samples from human MDMΦ and C6/36 infections were obtained by LLC-MK2 plaque assays as per previous descriptions [33]. Briefly, LLC-MK2s were seeded in confluent monolayers, allowed to adhere, and washed once with Hank’s balanced salt solution (HBSS, Gibco). Log10 serial dilution of viral supernatants was conducted in DMEM10 (Gibco), and virus was incubated on cells for 90 minutes (20°C, with rocking). Upon removal of virus, first-overlay containing nutrients and low-melting point tissue culture-grade agarose [33] was added and allowed to solidify at 20°C before incubation for 5 days (37°C, 5% CO2). A second-overlay containing nutrients, low-melting point tissue culture-grade agarose, and neutral red stain [33] was added and allowed to solidify at 20°C before further incubation for 1 day (37°C, 5% CO2) after which plaques were counted by eye. The limit of detection for this assay is 33 plaque forming units per ml (pfu/ml). Virus titers of samples isolated from infected mice were obtained by BHK-21 plaque assays in accordance with previous protocols [15,34]. BHK-21 cells were seeded in 12-well plates (3 x 105 cells/well) in α-MEM medium containing 5% HI-FBS for 3 hours (37°C, 5% CO2). Medium was removed, log10 serial dilutions of virus in α-MEM with 2% HI-FBS were added, and cells were incubated for an additional 2 hours (37°C, 5% CO2). Following incubation, 1. 5 ml of α-MEM containing 5% (v/v) HI-FBS and 1% (w/v) low-melting point tissue culture grade agarose (Sigma) was added to each well and allowed to solidify at room temperature before further incubation (37°C, 5% CO2) for 5 days. Plaques were visualised after fixation in 10% formaldehyde (1 hour, room temperature) and removal of the agarose plug by staining briefly (approximately 30 seconds) with 1% (w/v) crystal violet (Sigma) in 20% (v/v) ethanol (EtOH, Sigma). Concentrations were normalised to tissue mass (in grams) or serum volume (in ml) as appropriate. MDMΦ FOS were detected as previously described by Alonzi et al. [22]. Cells were cultured in 6-well plates as described above, and protein lysates were prepared according to the freeze/thaw protocol. Following cell lysis, samples were subjected to mixed-bed ion exchange and then lyophilised. FOS were labelled with 2-aminobenzoic acid (2-AA) and purified using a DPA-6S column (Sigma). Unconjugated 2-AA was removed by phase splitting with ethyl acetate, and samples were lyophilised and resuspended in water and then purified using a con-A column. Glycans were separated by normal phase-high performance liquid chromatography (NP-HPLC), and peak area was used to assess molar quantity in comparison to standards of known identity and quantity. FOS generation was normalised to protein concentration based on a modified Bradford assay as previously described [35]. Briefly, cell lysates were added in a 1: 1 volumetric ratio to 1x Bradford Quick Start Reagent (Bio-Rad) and incubated at 20°C for 5 minutes. Absorbance values (Axxx) at 595 nm and 420 nm were measured, and the ratio of A595 to A420 was used to determine protein concentration in comparison to a serially diluted standard of bovine serum albumin. MDMΦ glycosphingolipids (GSLs) were detected using NP-HPLC as previously described [36]. Briefly, cells were cultured in TC-25 flasks at 1. 5 x 106 cells/ml as described above and treated with appropriate dilutions of iminosugars for 48 hours at 37°C, 5% CO2. Following incubation, supernatants were removed and cells were washed twice in pre-warmed PBS before mechanical removal by scraping. Flasks were washed once with warm PBS, and this wash was combined with the scraped cell fraction and pelleted by centrifugation (room temperature, 4000 x g, 5 minutes). The MDMΦ pellet was washed once more in PBS before resuspension in 300 μl of water. Three cycles of freeze/thawing were executed to lyse samples, and an aliquot was removed for modified Bradford protein assay for normalisation. The remaining fraction was used to extract glycolipids by adding 400 μl chloroform and 800 μl methanol (MeOH) to 300 μl of aqueous lysate for lipid extraction using a modified Svennerholm and Fredman method [37]. Extracted GSLs were hydrolysed overnight by incubation with ceramide glycanase (37°C, 50 mM sodium acetate buffer, pH 5. 0, containing 1 mg/ml sodium taurodeoxycholate). Oligosaccharides released from GSLs were brought to 30 μl with water and labelled with 2-AA as described for FOS. These labelled oligosaccharides were analysed by NP-HPLC as described above. DENV RNA in cell culture supernatants was isolated according to the manufacturer’s protocol for Direct-zol RNA MiniPrep Kit (Zymo Research) and assayed by reverse transcription-real time polymerase chain reaction (qRT-PCR) on an Applied Biosystems 7500 real-time PCR system (Life Technologies). Thermo-Start DNA Taq polymerase was used to enable Taq-mediated release of fluorescently labelled dyes from a DENV2 NS5-specific probe adapted from a previously published protocol [38]. Probe and primer sequence and concentrations were as previously described, with the reaction mixture prepared according to the manufacturer’s instructions for Verso 1-Step RT-PCR Kit with Thermo-Start Taq (Life Technologies). Thermal cycling was adapted to match enzyme components as follows. Synthesis of complementary DNA (cDNA) was performed for 30 minutes at 50°C, followed by a 15-minute activation of Thermo-Start Taq polymerase at 95°C. PCR thermocycling with fluorescence detection was executed for 45 cycles of 95°C for 15 seconds followed by 60°C for 60 seconds and a fluorescence read step. Samples were read in technical duplicate and compared to a standard curve generated from high-titer viral RNA isolated from C6/36-grown DENV2. 95% confidence intervals were determined based on biological and technical variation and graphed using Prism 6 (GraphPad Software, Inc.). For animal experiments, multiplex qRT-PCR was conducted on all tissue isolates and plasma as described for in vitro experiments. Rodent glyceraldehyde 3-phosphate dehydrogenase (GAPDH) control (Life Technologies #4308313) containing a VIC reporter dye was attached to a TAMRA (tetramethylrhodamine) quencher, and experimental DENV NS5-specific target was labelled with FAM (6-fluorescein amidite) reporter dyes with a TAMRA quencher. GAPDH primers were used at 10 μM and probe at 20 μM. In the case of DENV qRT-PCR, quantitation using a DENV2-specific primer/probe set directed to NS5 was conducted using all primers and probe at 10 μM final concentration [38]. Thermocycling was conducted as for in vitro experiments. Viral RNA from tissues was quantified by ΔΔCt method, whereas plasma samples were quantified relative to a standard curve as for in vitro experiments. Samples were read in technical duplicate and 95% confidence intervals were determined based on biological and technical variation and graphed using Prism 6 (GraphPad Software, Inc.). All experimental procedures were pre-approved by the UC Berkeley Animal Care and Use Committee and were performed according to the guidelines of the UC Berkeley Animal Care and Use Committee, the United States Public Health Service, and the USDA Animal Welfare Act. Mice of the 129/Sv background deficient in interferon (IFN) -α/β and -γ receptors (AG129 mice) were infected with DENV under antibody-enhanced conditions as previously described [5,39]. Mice were primed 24 hours prior to infection with a 5 μg intraperitoneal (i. p.) injection of a pan-flavivirus monoclonal antibody against E protein, 4G2. At t = 0,105 pfu of DENV2 D2S10 was injected intravenously. Beginning immediately following infection, mice were treated with 33. 3 mg/kg celgosivir three times daily (t. i. d.) via oral gavage (p. o.). Treatment was continued for 80 hours, at which point mice were anaesthetised by isoflurane inhalation and euthanized after cardiac puncture. Cardiac bleeds (approximately 1 ml) were collected in citrate anti-coagulant tubes (Becton Dickinson) and stored on ice prior to isolation of PBMCs and plasma. To separate plasma, whole blood was centrifuged (18,000 x g, 30 minutes, 4°C), and supernatant was collected. Cells were resuspended in 0. 14 M ammonium chloride (Sigma), 17 mM Tris (MP Biomedicals), pH 7. 2 in H2O and incubated at 37°C for 5 minutes to lyse red blood cells (RBCs). RBC lysis was repeated three times, and cells were washed in serum-free α-MEM (Gibco) prior to resuspension and storage in RNAlater (Qiagen). Spleen, liver, small intestine, kidney, and lymph node tissues were collected into tubes containing zirconia/silica beads (1. 0 mm diameter) and protease inhibitor cocktail (Roche) in α-MEM. Tissues were homogenised using a Mini-Beadbeater-8 (BioSpec Products) and stored immediately at -80°C until future use. PBMCs and other tissue samples were thawed on ice in the presence of 40 mM dithiothreitol (DTT) and mixed with three volumes of buffer RLT (Qiagen) containing 40 mM DTT prior to isolation of cellular RNA by RNeasy Mini Kit (Qiagen) as per the manufacturer’s instructions. Plasma RNA (viral RNA) was isolated by QIAamp Viral RNA Mini Kit (Qiagen) in accordance with the manufacturer’s instructions. Antiviral EC50s in cell culture were calculated using a 4-parameter logistic model to obtain 95% confidence intervals in GraphPad Prism 6 (GraphPad Software, Inc.). The same protocol was applied to GSL assays. Statistically significant differences (α<0. 05) in viral load for animal studies were assessed by non-parametric Mann-Whitney testing with Bonferroni correction for multiple comparisons. Cells of the myeloid lineage, including tissue-resident macrophages and dendritic cells as well as circulating monocytes, are instrumental in supporting early DENV infection and directing host immune responses [40–42]. We have previously established a monocyte-derived macrophage (MDMΦ) model using blood from dengue-naïve human donors [31] and demonstrated that iminosugars such as NB-DNJ and NN-DNJ can reduce production of infectious DENV in this model [5]. In order to dissect how these iminosugars reduce infectious virus, we first compared the antiviral activity of DNJ-derived and DGJ-derived compounds. MDMΦs were infected with DENV serotype 2 at an MOI of 1 for 90 minutes prior to treatment with drug. Upon removal of virus, a titration of iminosugar was added to the cells. At 48 hours post-infection, supernatants were collected and plaqued on LLC-MK2 cells. The relative titer of treated samples is presented as a percentage of the infectious titer of samples not treated with iminosugar (untreated) because of the variation observed in the titer of untreated samples from different donors (see S1 Table). IL-4-treated primary human macrophages produce an average of [2. 58 +/- 2. 07] x 104 pfu/ml (mean +/- SD) of infectious DENV 48 hours after infection with MOI 1 DENV2, strain 16681, covering a >40-fold range from 2. 3–97. 5 x 104 pfu/ml. DNJ-derived compounds successfully reduced production of infectious virus with EC50s between 1. 2 and 10. 6 μM, as we have previously reported [5]; however, there was no detectable reduction of infectious virus with any DGJ-derived compound up to the maximum non-toxic dose tested (Fig 2A). Cytotoxicity as assessed by MTS assay was not detected up to 100 μM for NB-DGJ, but for NN-DGJ the maximum non-toxic dose was determined to be 31. 6 μM by MTS assay. Although celgosivir has been tested previously in cell culture and animal models, the drug’s efficacy against DENV has not been assessed in primary human cells cultured in vitro. Furthermore, the antiviral properties of the pro-drug in comparison to castanospermine have not been addressed. In order to compare the optimization efforts of monocyclic and bicyclic iminosugars, we performed a titration of NN-DNJ and celgosivir as well as the “parent” compounds DNJ and castanospermine in our human MDMΦ model. Comparison of the two optimized candidates in MDMΦs reveals similar antiviral efficacy (EC50 NN-DNJ = 1. 25 μM; EC50 celgosivir = 5. 17 μM), and both drugs are at least 7-fold more potent than the parent compounds (Fig 2B). A summary of in vitro antiviral properties for all iminosugars tested is presented in Table 1. Given the diversity of functions of glycans in vivo and the rationale of iminosugar use as monosaccharide mimetics, such drugs might inhibit a considerable number of enzymes involved with glycan processing. Indeed, the differential antiviral effects of gluco- and galactostereochemistry-possessing iminosugars indicate that differential inhibition of molecular targets occurs in response to small, defined changes in stereochemistry. With alteration of one stereocenter, differences in inhibition of glucosidase and glycosidase enzymes specific for various sugar stereochemistries are likely. Such differences were hypothesized to account for the differential antiviral efficacy observed in cell culture. In order to address this possibility, a panel of isolated glycan processing enzymes were treated with the iminosugars used in this study to determine targets of inhibition (Table 2). Unsurprisingly, DNJ and derivatives thereof demonstrated high nanomolar to low micromolar in vitro inhibition of α-glucosidases, whereas DGJ-derived iminosugars only inhibited these enzymes at concentrations approaching 1 mM or not at all. Castanospermine and the pro-drug, celgosivir, were similarly capable of inhibiting the α-glucosidases; however, inhibition of mouse ER α-glucosidase II required considerably higher concentrations of drug than for the DNJs. Stereochemistry is similarly important for inhibition of α-galactosidase, as this enzyme is inhibited at low nanomolar concentrations by the DGJ iminosugars with identical stereochemistry to the natural substrate galactose, but absence of or very weak inhibition was detected for the glucose-mimetics. Of particular interest is the trend observed for inhibition of glucosyltransferase isolated from a human promyelocytic leukemia cell line (HL60) where iminosugars with increasing alkyl chain length, again mimicking the natural substrate of the enzyme glucosylceramide synthase, were more potent inhibitors of the enzyme. A similar trend can be observed for both DNJ and DGJ iminosugars, suggesting that this effect is not specific to sugar stereochemistry. The in vitro testing of isolated enzymes indicates that inhibition of either α-glucosidases or glucosyltransferases is likely the essential property of iminosugars required for antiviral efficacy. Furthermore, the exclusive ability of glucose-mimicking iminosugars to potently inhibit α-glucosidases provides considerable support for the hypothesis that α-glucosidase inhibition is the single class of target enzymes responsible for antiviral activity against DENV; however, it is equally possible that only DGJ-derived iminosugars fail to access glycolipid processing enzymes in cells. Thus, we next assessed whether these iminosugars could achieve inhibition of glycolipid processing in our human MDMΦ model. In order to address this question, MDMΦs were treated for 48 hours with the maximal non-toxic dose (MNTD) of drug used in antiviral studies as in Table 1. Whole cell lysates were then collected and assayed for monosialodihexosylganglioside (GM3) levels. Gangliosides are reliable markers of iminosugar-mediated alteration of glycolipid processing [25,43], and the relative simplicity of GM3, possessing one β-d-galactose and one β-d-glucose head group, facilitates analysis by normal phase high performance liquid chromatography (NP-HPLC). As shown in Table 3, DNJ and DGJ derivatives both effectively lowered GM3 levels normalized to total protein concentration by at least 90 percent. In contrast, bicyclic glucose iminosugars failed to inhibit GM3 synthesis. In addition to inhibition at maximal dosing, the dose-response relationship of iminosugar to GM3 levels was investigated to determine whether this correlated with antiviral efficacy. The response of GM3 production in MDMΦs to NB-DNJ was therefore investigated in three donors over a titration of 1 to 100 μM of drug. Fig 3 shows that treatment resulted in a dose-dependent decrease in GM3 production. The EC50 of NB-DNJ with respect to reduction of GM3 is approximately 10-fold lower than the antiviral EC50 (95% CI for GM3 EC50 = 0. 80–1. 37 μM; 95% CI for antiviral EC50 = 7. 77–14. 5 μM). Although 90–95% inhibition of GM3 levels could be reached with 10 μM treatment of NB-DNJ, increasing concentrations up to 100 μM did not lead to further inhibition of GM3 levels, suggesting that this is the maximal effect possible for iminosugars. In summary, these data indicate that all of the monocyclic iminosugars tested could effectively inhibit glycolipid processing to equivalent levels at the concentrations tested in the antiviral assays presented in Fig 2A. Whereas both DNJ- and DGJ-derived iminosugars can effectively inhibit glycolipid processing in cell culture, the inhibition data from isolated enzymes suggests that only glucose-mimic iminosugars are likely to inhibit glycoprotein processing. Generation of FOS can be used as a marker of successful inhibition of ER-resident α-glucosidases [22], and this technique has previously been used to demonstrate inhibition of these glycoprotein processing enzymes by DNJ-iminosugars in a human lung cancer-derived cell line (A549) [23]. In order to address whether the iminosugars used in this study can also productively inhibit ER α-glucosidases in a primary human cell, more relevant for DENV infection, we treated our MDMΦ model with the MNTD of all iminosugars as well as a titration of celgosivir or NB-DNJ for 48 hours in the absence of DENV infection. As a consequence of inhibition of ER α-glucosidase II, a monoglucosylated glycoprotein is produced. After trimming of the glycan precursor by mannosidases and recognition of the glycoprotein as terminally misfolded, a Glc1Man4GlcNAc1 free oligosaccharide species is cleaved from the peptide during ER-associated degradation. In the case of inhibition of ER α-glucosidase I, a similar process produces a Glc3Man5GlcNAc1 species. Thus, the presence of each species of FOS can be correlated with successful inhibition of the respective cellular α-glucosidase [22]. Nearly all species of FOS were undetectable in the absence of iminosugar treatment (Fig 4A); however, addition of 100 μM NB-DNJ led to accumulation of peaks representative of inhibition of both α-glucosidases (Fig 4B). Similar inhibition of both α-glucosidases was also noted for NN-DNJ at the MNTD; however, all DGJ-derivatives tested failed to produce any detectable monoglucosylated or triglucosylated oligosaccharide species, which could be a product of inhibited glycoprotein processing at the MNTD (Fig 4C and 4D). The major peaks observed in the NP-HPLC trace for DGJ-treated samples, which occur at 23. 3 and 32. 6 minutes, represent FOS generated by inhibition of lysosomal β-N-acetylhexosaminidases [44]. These enzymes are responsible for trimming terminal glycans from molecules possessing terminal N-acetyl hexosamines (e. g. GlcNAc) and are frequently implicated in disorders of glycolipid processing [45]. Analysis of the dose-response relationship of iminosugars and FOS generation may provide further insights into the mechanism of action of the drugs. Addition of 1 μM NB-DNJ or celgosivir led to accumulation of Glc1Man4GlcNAc1 but not Glc3Man5GlcNAc1 in all donors tested (Fig 4E and 4F). With increasing concentrations of celgosivir, further accumulation of monoglucosylated FOS was observed, with maximal levels achieved at 3 and 10 μM drug. Whereas triglucosylated FOS were produced in relatively limited amounts at treatment below 10 μM, increasing accumulation of Glc3Man5GlcNAc1 was noted with 31. 6 and 100 μM celgosivir treatment. A concomitant decrease in Glc1Man4GlcNAc1 accumulation was reliably detected, with the more prolific accumulation of triglucosylated FOS at the highest dose tested. Reduced accumulation of monoglucosylated FOS is presumed to be a function of substrate limitation as α-glucosidase I activity occurs upstream of α-glucosidase II activity; thus, inhibition of α-glucosidase I limits the available pool of substrate for α-glucosidase II. A similar trend to celgosivir was observed for NB-DNJ, except that the difference in inhibition of α-glucosidase I and α-glucosidase II was considerably less pronounced for NB-DNJ (Fig 4F). These differences in specificity may be a useful starting point for understanding glucosidase-specific structure-activity relationships of iminosugars. With a growing understanding of the specificity of different drugs in this class, it may be possible to ascertain the relative importance of inhibition of each α-glucosidase. Thus, although glycolipid inhibition fails to differentiate antiviral from non-antiviral iminosugars, glycoprotein inhibition correlates with antiviral activity. Strict dependence of DENV glycoproteins on α-glucosidase trimming and subsequent interactions of the glycoproteins with the ER chaperones calnexin and/or calreticulin would explain the historical data as well as our observations of the antiviral efficacy of iminosugars against DENV. The sum of these historical data and the data presented above still fails to demonstrate whether the glycoproteins, particularly E, are grossly misfolded and targeted for degradation or whether partially misfolded glycoproteins escape the folding control of the ER, and the virus that subsequently buds from the cell is less infectious due to this partial misfolding, as seen for example in the case of iminosugar-treated HIV-secreting cells [46]. In order to address this question, we infected MDMΦs with DENV as in Fig 2 and treated them with a serial dilution of NB-DNJ or celgosivir. Supernatants were collected at 48 hours post-infection for coordinated assay of infectious virus by plaque assay and total virus production by qRT-PCR for DENV RNA. For both drugs and in every donor tested, the drop in infectious viral titer correlated with a drop in total virus secreted (Fig 5). To ensure that the reduction in total virus secretion observed at 48 hours was not related to re-infection of MDMΦs following a round of DENV replication within the cells (i. e. related to decreased infectivity), total DENV secretion levels at 24 hours were assayed and compared to those observed at 48 hours post-infection. Infectious DENV is only detectable in our MDMΦ model with incubation of greater than 12 hours as demonstrated in other cell systems [47,48]; thus, assays conducted at 24 hours are likely to be exclusively the product of events occurring only as a result of a single round of infection. In summary, the direct correlation of effects on total virus and infectious virus further supports the prevailing hypothesis that glucostereochemistry iminosugars, as a class, induce gross misfolding of DENV glycoproteins leading to ER-associated degradation and reduced secretion of virus. Previous work has demonstrated that orally administered iminosugars can rescue mice infected with a lethal dose of DENV in models of severe disease [16–18,49]. Dosing of the iminosugars used in these studies has varied from once daily (QD) to thrice daily (t. i. d.) regimes with more frequent dosing generally associated with greater success. Although celgosivir has demonstrated efficacy at twice daily (b. i. d.) intraperitoneal (i. p.) dosing of 50 mg/kg/dose, recent human trials were conducted with an oral (p. o.) regime of a 400 mg loading dose followed by 200 mg/dose b. i. d. These trials failed to demonstrate reduction in viral load or fever burden in patients with dengue infection [50,51]. To address these inconsistencies and assess whether more frequent dosing with less drug might be more efficacious, we treated animals with 33 mg/kg/dose of celgosivir t. i. d. p. o. and assessed levels of viral RNA in circulation (Fig 6A) and in various tissues (S1 Fig) as well as infectious virus in circulation (Fig 6B). AG129 mice were administered 5 μg of 4G2 antibody i. p. 24 hours prior to infection with 105 pfu of DENV-2 D2S10 to mimic ADE infection [5,39]. Drug was administered immediately following infection and every 8 hours thereafter until mice were sacrificed at 80 hours post-infection. In this mouse model of DENV disease, mortality occurs at day 3. 5 to 4 (84–96 hrs) post-infection; thus, samples were collected just prior to sacrifice at a point when differences in variables anticipated to be relevant to clinical outcomes were predicted to be most pronounced. Serum was isolated from whole blood and used to determine the viral load (circulating viral RNA, Fig 6A) and the level of infectious virus (Fig 6B). Using this more frequent, oral treatment regime for celgosivir, there was a statistically significant reduction in viral RNA load (α<0. 05), and this trend appeared to extend to infectious virus; however, results for the plaque assay were not statistically significant (α = 0. 100) due to the considerable variation in the assay. Additional tissues including kidney, parenteral lymph nodes, liver, and small intestine showed similar trends in RNA levels; however, in the case of the spleen, celgosivir enhanced viral RNA levels by approximately two orders of magnitude (S1 Fig). Overall, these results indicate that more frequent oral dosing with less drug administered per dose may be a viable antiviral treatment strategy for celgosivir. As celgosivir does not inhibit the glycolipid biosynthesis pathway (Table 3), any antiviral effect observed here is due to ER α-glucosidase inhibition. In this work, we tested a panel of iminosugars covering a range of specificities of enzyme inhibition profiles to address the role of glycoprotein and glycolipid inhibition on DENV antiviral activity. Coordinated analysis of isolated enzyme inhibition profiles, primary human cell culture antiviral, glycoprotein, and glycolipid inhibition profiles, and animal model efficacy allows us to present the most complete deconstruction of iminosugar antiviral properties necessary for efficacy against DENV to date. Prior to this publication, a large body of work existed in support of the hypothesis that iminosugar antiviral activity against DENV is a function of inhibition of endoplasmic reticulum resident α-glucosidases; however, exclusion of alternative mechanisms of action related to inhibition of additional sugar processing enzymes had not been undertaken. Knowledge of the precise mechanism (s) of action of iminosugars as DENV therapeutics is essential as these drugs enter and progress through clinical trials. The possibility that some level of antiviral activity is a function of modulation of alternative targets including host glycolipid processing and/or viral ion channel inhibition bears precedent [52–55]. There is some debate as to the existence of an ion channel in DENV [56,57] although the best available evidence suggests that there is no equivalent molecule to the HCV ion channel p7 [56]. Iminosugars are also known to inhibit glycan processing enzymes in addition to ER α-glucosidases. Reports of inhibition of intestinal disaccharidases [58] were confirmed for both NB-DNJ and celgosivir (Table 2); however, we cannot conceive of a mechanism whereby this effect on enzymes located on the brush border of the small intestine could accomplish antiviral activity in the experimental setting investigated here, or indeed when iminosugars are administered systemically. To our current knowledge, there is no evidence for an antiviral mechanism mediated by inhibition of additional human glycosidases. In contrast, the role of glycolipids in viral genesis provides support for the hypothesis that this “off-target” effect may play a role in iminosugar antiviral activity. The ability of iminosugars to inhibit glycolipid processing as in the clinical use of NB-DNJ for Gaucher’s disease and Niemann-Pick type C disease [24,25,59] provides cause for concern that the promiscuity of iminosugars for multiple glycan processing enzymes has clouded our understanding of antiviral drug design and optimization. Our results demonstrate that production of infectious DENV in a primary human macrophage model was controlled in a dose-dependent manner with DNJ-derivative iminosugar treatment, as previously reported [5], and variable alkylation of the ring nitrogen enhanced antiviral activity. A similar inhibition profile was noted in MDMΦs for NN-DNJ and celgosivir, although the improvement of NN-DNJ in comparison to the parent compound DNJ (246-fold) was considerably more pronounced than the improvement of celgosivir in comparison to the parent castanospermine (7-fold). In a BHK-21 (hamster kidney) cell culture model, celgosivir exhibited ~100-fold enhancement of antiviral efficacy over castanospermine [17]; however, this difference in enhancement may be due to variable efficiency of cleavage of the butanoyl pro-drug moiety in different cell culture models. Strikingly, equivalent galactostereochemistry analogues of the DNJ derivatives all failed to impact infectious virus production in MDMΦs suggesting a clear importance for the stereochemistry of the iminosugar head group. How does the difference in glucose and galactose stereochemistry confer such a striking difference to antiviral efficacy of iminosugars? Although it remains a technical challenge to produce sufficient quantities of pure human ER α-glucosidase to test this enzyme in vitro, our panel of closely related enzymes helps to answer the question of iminosugar specificity. Inhibition of isolated enzymes occurs in line with anticipated headgroup stereospecificity (e. g. galactose analogues inhibit galactosidases and not glucosidases), but this specificity does not extend to glycolipid processing enzymes such as the glucosyltransferase of HL60 cells, where factors relating to alkyl chain structure and length are more likely to be relevant. It is worth noting that the inhibition of enzymes in vitro is less potent for celgosivir than for castanospermine, a result that is likely to be the product of interference of the butanoyl group with binding of the inhibitor. In cell culture and animal models, this group is cleaved so that the more potent inhibitor, castanospermine, is in abundance. Trends observed for isolated enzymes are replicated in cell culture; thus, the question of antiviral efficacy cannot be explained by differential uptake, a phenomenon much more likely to be mediated by differential alkylation. Rather, this work demonstrates that in the presence of maximal glycolipid processing inhibition (~95% reduction of GM3), DGJ-derived iminosugars fail to induce any accumulation of FOS related to ER α-glucosidase inhibition. Although equivalent glycolipid processing effects are observed for DNJ-derivatives, the cellular inhibition of ER α-glucosidases correlates remarkably well with observed antiviral efficacy of iminosugars. Whereas GM3 accumulation occurs at approximately 10-fold lower concentrations, α-glucosidase inhibition maps very well to antiviral efficacy. In particular, inhibition of α-glucosidase II appears to be of specific importance as can be observed by comparison of celgosivir and NB-DNJ titration curves which reveals a more potent inhibition of α-glucosidase II by the more potent antiviral, celgosivir. This trend can be extended even in the case where α-glucosidase I inhibition occurs at lower concentrations than α-glucosidase II inhibition. In the case of DNJ based iminosugars CM-10-18 and CM-9-78, α-glucosidase I inhibition occurs at 3- to 5-fold lower concentrations than α-glucosidase II concentrations. For these molecules, α-glucosidase II inhibition occurs with an EC50 of 2. 6 and 1. 6 μM for CM-9-78 and CM-10-18, respectively. Antiviral EC50s in the same A549 cell line are 1. 5 and 1. 1 μM while α-glucosidase I inhibition EC50s of 0. 42 and 0. 46 μM are noted, respectively [23]. Thus, we propose that inhibition of α-glucosidase II is the crucial property of iminosugars for antiviral activity against DENV. It has been suggested that iminosugars reduce infectivity of HIV [46,60] but prevent secretion of HBV [61]. In our hands, it appears that DENV secretion is blocked and the virus that does escape is similar in infectivity to virus produced in the absence of iminosugars, thus the effect that iminosugars have on DENV appears to be more similar to that of HBV rather than HIV. Previously published studies indicate that iminosugar treatment during the course of DENV infection reduces association of viral glycoproteins with the ER chaperones calnexin and calreticulin [6], and our ability to detect FOS that have been cleaved from terminally misfolded glycoproteins suggests that DENV glycoproteins are not retained by ER chaperones and are instead grossly misfolded and targeted for ER-associated degradation when cells are treated with iminosugars. Lack of antiviral efficacy observed in the first human studies of iminosugars as DENV therapeutics is a challenge for the field [50,62]. Use of the AG129 ADE model of DENV infection is well established in antiviral drug testing, so we sought to address the events that occur with drug administration immediately prior to the normal time of death to capture any virus-related phenotype that may be responsible for protection with iminosugar treatment. This experiment demonstrated that more frequent dosing of celgosivir at lower doses than have been previously tested are capable of reducing levels of tissue-resident and circulating viral RNA. It should be noted that a pronounced level of circulating viral RNA and infectious virions could still be detected in animals treated with celgosivir. This suggests that complete elimination of virus by iminosugar treatment is not essential for survival, but reduced viral load and any concomitant changes in response to pathogen clearance are sufficient to prevent mortality. Furthermore, any viral RNA measured by qRT-PCR in the mouse model that is not accounted for by infectious virus theoretically represents viral RNA encapsulated in non-infectious virions. As such, comparison of the in vitro macrophage data in Fig 5 and in vivo mouse data in Fig 6 suggests that in the mouse model there is an accumulation of viral RNA encapsulated in non-infectious virions that is not observed in primary macrophage cell culture. The extension of these observations in a mouse model to the treatment of human infection requires further investigation, but this is consistent with studies demonstrating that viral load correlates with disease severity [51]. Finally, these data further support the hypothesis that glucosidase inhibition is sufficient for anti-DENV activity. In summary, this study provides confirmation of the long-standing dogma that iminosugar inhibition of DENV is accomplished via inhibition of ER α-glucosidases. Our data are in agreement with the mounting bank of evidence that α-glucosidase II is the glucosidase responsible for anti-DENV activity, and we suggest that any further optimisation of iminosugar candidates focuses on enhanced inhibition of this enzyme. For monocyclic iminosugars, the necessary specificity for glucose-processing enzymes is provided by the DNJ headgroup. Alkylation or modification of the ring nitrogen on the other hand constitute amenable targets for medicinal chemistry efforts. Modulation of the alkyl group in particular is promising as we have demonstrated that glycolipid inhibition, which may be altered or abolished with changes in the alkyl group is of no consequence to antiviral activity against DENV. At present, high-resolution structural data for the complete α-glucosidase II enzyme is unavailable; however, a crystal structure of a lectin receptor domain has been solved [63]. Availability of more complete structural data for this enzyme would enable structure-activity relationship investigations of iminosugars and may provide insights to the differential inhibition observed in this study. Looking to the future, these data provide a rationale for further optimization of iminosugar structure and treatment regimes as dengue virus therapeutics.

Current treatment of dengue virus infection is supportive; however, iminosugars have been widely investigated as an antiviral strategy. The means by which these molecules are thought to exert their antiviral effects is through inhibition of host-resident glycoprotein processing enzymes, the endoplasmic reticulum-resident α-glucosidases, but many iminosugars are also capable of inhibiting host glycolipid processing and are utilized clinically for the treatment of lysosomal storage disorders such as Gaucher’s and Niemann-Pick type C diseases. The work presented here is the first to conclusively differentiate the antiviral properties of these two major mechanisms of action of iminosugars, and our data support the long-standing hypothesis that inhibition of glycoprotein processing is the essential antiviral property of iminosugars in the case of dengue virus infection. These results indicate that further development of iminosugars as dengue antivirals should focus on optimization of glycoprotein inhibition efficacy with reduction or elimination of glycolipid modulating properties to minimize off-target effects. These results are supported by the in vitro and in vivo efficacy of the bicyclic iminosugar, celgosivir, which we demonstrate to lack capacity for inhibition of glycosphingolipid processing.

Abstract Introduction Methods Results Discussion

glycolipids animal models of disease endoplasmic reticulum cell processes enzymology biological cultures microbiology animal models model organisms cell cultures glycoproteins enzyme inhibitors cellular structures and organelles research and analysis methods animal models of infection animal studies chemistry mouse models biochemistry stereochemistry cell biology secretory pathway biology and life sciences physical sciences glycobiology

2016

Iminosugars Inhibit Dengue Virus Production via Inhibition of ER Alpha-Glucosidases—Not Glycolipid Processing Enzymes

Recent studies have revealed an important role for hormones in plant immunity. We are now beginning to understand the contribution of crosstalk among different hormone signaling networks to the outcome of plant–pathogen interactions. Cytokinins are plant hormones that regulate development and responses to the environment. Cytokinin signaling involves a phosphorelay circuitry similar to two-component systems used by bacteria and fungi to perceive and react to various environmental stimuli. In this study, we asked whether cytokinin and components of cytokinin signaling contribute to plant immunity. We demonstrate that cytokinin levels in Arabidopsis are important in determining the amplitude of immune responses, ultimately influencing the outcome of plant–pathogen interactions. We show that high concentrations of cytokinin lead to increased defense responses to a virulent oomycete pathogen, through a process that is dependent on salicylic acid (SA) accumulation and activation of defense gene expression. Surprisingly, treatment with lower concentrations of cytokinin results in increased susceptibility. These functions for cytokinin in plant immunity require a host phosphorelay system and are mediated in part by type-A response regulators, which act as negative regulators of basal and pathogen-induced SA–dependent gene expression. Our results support a model in which cytokinin up-regulates plant immunity via an elevation of SA–dependent defense responses and in which SA in turn feedback-inhibits cytokinin signaling. The crosstalk between cytokinin and SA signaling networks may help plants fine-tune defense responses against pathogens. The first layer of active plant immunity begins with the recognition of microbial molecules, followed by activation of an effective defense response [1]. Non-adapted pathogens are halted by this defense response, whereas adapted pathogens are able to overcome these defense responses via deployment of virulence factors, eventually leading to manipulation of the host biology and culminating in pathogen growth and reproduction. The plant hormones salicylic acid (SA), jasmonic acid and ethylene have long been implicated in defense responses [2] and recent studies have also uncovered a role in plant defense for several other hormones [3], [4], but the extent of crosstalk among the hormonal networks in plant defense is only now beginning to be understood. Cytokinins are a group of N6-substituted adenine derivatives that regulate many plant developmental processes and responses to the environment [5]. Cytokinin perception and signaling is carried out by two-component element proteins [6], analogous to two-component signaling systems present in bacteria and fungi (Figure 1A). In Arabidopsis, binding of cytokinin to sensor histidine kinases (AHKs) receptors initiates a phosphotransfer cascade that culminates in the phosphorylation of response regulator proteins (ARRs), which are responsible for the regulation of cytokinin outputs. ARRs fall into two main groups [6]: type-A ARRs contain short C-terminal extensions and act as negative regulators of cytokinin responses [7]–[10]; type-B ARRs contain extended C-termini that include a DNA binding domain and directly mediate the transcription of cytokinin-responsive genes and positively regulate cytokinin signaling [7], [11]–[13]. Several lines of evidence support a role for cytokinins in plant-pathogen interactions. For example, application of cytokinin results in decreased replication of White Clover Mosaic Potexvirus and in the induction of defense gene expression in bean plants [14]. Treatment of Arabidopsis plants or plant cell cultures with cytokinin also up-regulates stress- and defense-related genes [15], [16] and promotes resistance to the bacterial pathogen Pseudomonas syringae pv. tomato DC3000 (Pst DC3000) in a process involving a type-B ARR and SA signaling [17]. Conversely, increases in cytokinin content are associated with the formation of ‘green-islands’, photosynthetically active leaf tissue supporting a region of pathogen growth and surrounded by senescent tissue [18]. Increased cytokinin content is associated with increased pathogen growth in several plant species [19], [20]. Finally, many fungal and bacterial pathogens can produce cytokinins [21], presumably used to manipulate host cell physiology to the pathogen' s benefit. These examples suggest that levels of pathogen- or host-derived cytokinins can alter host responses to pathogens and influence the outcome of plant-pathogen interactions. Here, we report that exogenous cytokinin alters immune responses of Arabidopsis to a pathogenic isolate of a biotrophic oomycete pathogen. We show that while high concentrations of cytokinin lead to decreased susceptibility through a process that requires SA accumulation and activation of SA-dependent defense gene expression, treatment with lower concentrations of cytokinin results in increased pathogen growth. We also demonstrate that SA negatively regulates cytokinin signaling, which may act to fine-tune this process. These functions for cytokinin in plant defense require an intact host cytokinin phosphorelay system, and are mediated in part by type-A ARRs, which act as negative regulators of defense responses. We examined the expression of Arabidopsis genes encoding elements involved in cytokinin signal transduction (Figure 1A) in response to pathogen treatment. Publicly available microarray data deposited at AtGenExpress (http: //www. uni-tuebingen. de/plantphys/AFGN/atgenex. htm) were analyzed using the e-northern tool of the Bio-Array Resource for Arabidopsis Functional Genomics (http: //bar. utoronto. ca/) [22] (Figure 1B). Among the genes encoding two-component elements, the expression levels of type-A ARR genes were most affected by pathogen treatment. This is similar to what is observed after treatment of plants with exogenous cytokinin [16], [23]. While the expression of genes encoding two-component elements was clearly altered by pathogen treatment, there was no direct correlation between the pattern of gene expression and pathogen lifestyle with respect to biotrophic (G. orontii), hemibiotrophic (P. syringae ES2346, P. infestans) or necrotrophic (B. cinerea) pathogens, or elicitors derived from biotrophic pathogens. These results are consistent with a role for two-component elements in the response to a variety of plant pathogens. Due to contrasting reports regarding the role of cytokinins during plant immune system responses [18], we examined the effect of a range of concentrations of exogenous cytokinin on the responses of wild-type Arabidopsis plants (accession Col-0) to the virulent oomycete Hyaloperonospora arabidopsidis isolate Noco2 (Hpa Noco2). Hpa Noco2 is a well-adapted obligate biotrophic pathogen of Arabidopsis that is able to overcome defense responses of wild-type plants and establish an intimate relationship with its host. Two-week-old plants were treated with increasing concentrations of the cytokinin benzyl adenine (BA) or a vehicle control (DMSO) 48 hours prior to pathogen treatment (Figure 1C). We observed distinctive effects of cytokinin on the susceptibility of wild-type plants to Hpa Noco2 at the different concentrations tested. Treatment with low concentrations of exogenous cytokinin led to enhanced susceptibility to Hpa Noco2, indicating that cytokinin-dependent processes contribute to the susceptibility to this pathogen (Figure 1C). In contrast, treatment with higher levels of cytokinin (>10 µM) led to decreased susceptibility to Hpa Noco2, indicating a threshold above which the action of cytokinin has a negative impact on the susceptibility of Arabidopsis to Hpa Noco2 (Figure 1C). We further investigated the effect of cytokinin on plant immunity. Pre-treatment of wild-type plants with high concentrations of cytokinin led to decreased susceptibility to Hpa Noco2 (Figure 2A). The added cytokinin induced a cytokinin response as shown by the up-regulation of the cytokinin-inducible gene ARR7 (Figure 2D). The effect of high levels of cytokinin on the growth of Hpa Noco2 was not due to off-target effects of BA or direct effects on Hpa Noco2 growth as it was abrogated by disruption of the AHK2 and AHK3 cytokinin receptors (Figure 2B). Comparable levels of cytokinin have been shown to elicit biologically relevant levels of cytokinin signaling in other assays for cytokinin responses [24], [25]. Two non-mutually exclusive hypotheses that could account for this effect of cytokinin on Hpa Noco2 susceptibility are: 1) changes in host metabolism that could result in poor pathogen growth; or 2) increased activation of defense responses. Because SA plays a significant role in plant immunity [26], [27], we tested if this response to high concentrations of cytokinin was a result of activation of SA-mediated responses by examining the eds16 mutant in which the ISOCHORISMATE SYNTHASE 1 (ICS1) gene required for SA biosynthesis [28], [29] is mutated. eds16 plants displayed a substantially reduced response to high concentrations (100 µM) of cytokinin (Figure 2C) as compared to wild-type plants. This indicates that the effect of high concentrations of cytokinin is largely dependent on SA biosynthesis, which is consistent with a role for cytokinin upstream of SA during activation of defense responses by Hpa Noco2. Interestingly, there is a slight decrease in pathogen growth in the eds16 plants at lower cytokinin concentrations (1 µM) that is not observed in wild-type plants, suggesting that eds16 plants are hypersensitive to cytokinin (Figure 2C; see also below). To further understand the relationship of high concentrations of cytokinin and plant immunity, we analyzed the expression of SA-responsive genes in response to cytokinin treatment and inoculation with Hpa Noco2 (Figure 2E, 2F). While much defense transcriptional reprogramming generally occurs early after pathogen recognition [30], [31], we chose to look at gene expression changes three days post inoculation (dpi) when an estimated 40% of plant mesophyll cells are in contact with Hpa hyphae and/or haustoria [32]. The SA-responsive genes tested were marginally up-regulated by cytokinin treatment alone in wild-type plants (Figure 2E). This increase in expression in response to cytokinin was partially dependent on SA biosynthesis as it was generally diminished in eds16 plants. As expected, the defense genes examined were markedly up-regulated by inoculation with Hpa Noco2 in the wild-type, but to a reduced extent in eds16 plants (Figure 2F). While cytokinin treatment alone only led to a slight induction in defense gene expression, pre-treatment with cytokinin followed by Hpa Noco2 inoculation led to a further enhancement of the expression of the defense genes tested (Figure 2F). These results suggest that cytokinin acts by priming defense-gene expression in Arabidopsis. This potentiation of defense gene expression by cytokinin was partially dependent on SA as revealed by gene expression analysis of similarly treated eds16 plants (Figure 2E, 2F). Treatment of wild-type plants with low concentrations of cytokinin (100 nM), which are sufficient to induce expression of the cytokinin-regulated gene ARR7 (Figure 3C), results in increased susceptibility to Hpa Noco2 (Figure 1C, Figure 3A). We examined if the effect of low concentrations of cytokinin on susceptibility to Hpa Noco2 was dependent on endogenous SA, as was observed for higher levels of cytokinin. eds16 plants did not show an increase in pathogen growth after cytokinin treatment (Figure 3B), suggesting that basal levels of SA may be required for the promotion of susceptibility by low levels of cytokinin. Alternatively, the hyper-susceptible phenotype of eds16 plants [28] may preclude quantification of marginal increases in the growth of pathogens. We examined the effect of SA on cytokinin responsiveness by examining the expression of cytokinin-regulated genes [16] in untreated eds16 plants. The basal level of expression of genes positively regulated by cytokinin (ARR7, ATST4B, and to a lesser extent CKX4; [16]) were significantly elevated in eds16 plants relative to the wild-type. Conversely, the expression of XTR7 (XYLOGLUCAN ENDOTRANSGLUCOSYLASE 7), which is negatively regulated by cytokinin, was further down-regulated in eds16 relative to the wild-type (Figure 3D). We also analyzed the response of eds16 plants to cytokinin using a primary root elongation assay [10], [33]. Wild-type seedlings showed inhibition of root elongation by BA concentrations above 50 nM, while ahk2,4 control plants were largely resistant to BA. eds16 plants displayed a significant and reproducible hypersensitivity to cytokinin at lower concentrations of cytokinin (Figure 3E). Together, these results suggest that SA negatively regulates cytokinin responsiveness, consistent with the hypersensitivity of eds16 mutants to cytokinin with regard to pathogen growth (Figure 2C). To further address the mechanism of cytokinin action in plant immunity, we examined the requirement for a host phosphorelay mechanism in the susceptibility of Arabidopsis to Hpa Noco2. Consistent with a role for cytokinin in plant immunity, disruption of the cytokinin receptors (AHK2, AHK3 or AHK4) resulted in enhanced susceptibility to Hpa Noco2 (Figure 4A). The ahk3,4 and ahk2,4 double mutants, but not the ahk2,3 double mutant, displayed an additive increase in susceptibility, indicating that the cytokinin receptors play partially redundant roles in defense responses to this pathogen, similar to their overlapping roles in other cytokinin-regulated physiological processes [34]–[36]. The triple receptor mutant was not used in this study due to its small, stunted phenotype, which precludes us from drawing any meaningful conclusions from pathogen assays in this background. Unlike the other two-component elements, type-A ARRs are negative regulators of cytokinin signaling [7]–[10]. There are ten genes encoding type-A ARRs in Arabidopsis. Due to partial redundancy in this gene family, increased sensitivity to cytokinin is apparent only in quadruple and higher order multiple mutants. The type-A arr multiple mutants arr5,6, 8,9 and arr3,4, 5,6, 8,9 showed decreased susceptibility to Hpa Noco2 (Figure 4B) as compared to wild-type plants. Similar to their roles in cytokinin signaling, the respective single mutations had no measurable effect on susceptibility to Hpa Noco2 (data not shown). Interestingly, the arr3,4, 5,6 mutant, which has an equivalent hypersensitivity to cytokinin as the arr5,6, 8,9 mutant in several response assays [10], did not exhibit any difference in susceptibility to Hpa Noco2 compared to wild-type plants, suggesting combinatorial specificity in this response (Figure 4B). Together, these results indicate that cytokinin signaling components play partially overlapping roles in plant immunity; the cytokinin receptors exert a mainly positive role, while the type-A ARRs have a negative regulatory effect. To further explore the role of type-A ARRs in plant immunity, we examined the effect of overexpression of type-A ARRs on the susceptibility to Hpa Noco2. Consistent with the decreased susceptibility phenotype observed in the loss-of-function type-A arr multiple mutants, transgenic lines overexpressing type-A ARR genes under the control of the constitutive CaMV 35S promoter [33] showed enhanced susceptibility (Figure 4C). This suggests that susceptibility to Hpa Noco2 is correlated to the level of type-A ARRs. Phosphorylation of type-A ARRs on a conserved residue (Asp87) in the receiver domain is required for type-A ARR activation and function in cytokinin signaling [33]. Therefore, we tested whether this phosphorylation is required for the enhanced susceptibility to Hpa Noco2 seen in transgenic lines overexpressing type-A ARRs. Transgenic lines overexpressing phospho-mimic (ARR5D87E) and phospho-deficient (ARR5D87A) forms of ARR5 were tested for their susceptibility to Hpa Noco2. These lines have been characterized and shown to express similar protein levels, and the ARR5D87E and ARR5D87A proteins have been shown to retain their ability to interact with two-component elements in a yeast two-hybrid assay, indicating proper folding [33]. Similar to lines overexpressing wild-type type-A ARRs, overexpression of ARR5D87E also led to enhanced susceptibility to Hpa Noco2 (Figure 4C). Conversely, plants overexpressing ARR5D87A were not statistically significantly different from wild-type plants in their susceptibility to Hpa Noco2 (Figure 4C). Moreover, the expression levels of the defense genes tested were reduced in both unchallenged (Figure 4D) and Hpa-induced (Figure 4E) ARR overexpressing lines in comparison to wild-type plants. These results indicate that it is the phosphorylated form of type-A ARRs that play a negative role in regulating defense responses in both unchallenged plants and in response to Hpa Noco2. The potential role of type-A ARRs in basal defense gene expression and Hpa Noco2-triggered responses led us to investigate transcriptional reprogramming in response to Hpa Noco2 in the type-A arr3,4, 5,6, 8,9 multiple mutant. Wild-type and arr3,4, 5,6, 8,9 mutant plants were treated with either water or inoculated with Hpa Noco2 and tissue harvested three days after treatment. RNA from replicate samples from independent experiments was prepared and gene expression analyzed using ATH1 Affymetrix microarrays. Samples were normalized to the water-treated wild-type samples. The expression levels of 1583 genes were significantly altered in wild-type plants in response to inoculation with Hpa Noco2 (Figure 5A). Transcriptome changes were similar in wild-type and arr3,4, 5,6, 8,9 mutant plants in response to Hpa Noco2, both in amplitude and in the set of genes regulated (Table S1). However, 292 of these regulated genes were expressed at levels 20–50% higher in water-treated arr3,4, 5,6, 8,9 mutant plants as compared to water-treated wild-type plants; hence they are under negative control by type-A ARRs. Representatives selected from the most markedly de-repressed cluster (Figure 5A, red asterisk) include several genes involved in SA-mediated defense signaling (e. g. FRK1, PAD4, FMO1 and WRKY18), SA biosynthesis (ICS1), and SA-mediated defense markers (e. g. PR5). Conversely, a subset of genes known to be down-regulated by SA, such as PDF1. 2, displayed reduced basal expression in arr3,4, 5,6, 8,9 plants (Figure 5B). We confirmed these results for a subset of genes in an independent experiment using qRT-PCR (Figure 5D). These results suggest that type-A arr3,4, 5,6, 8,9 mutant plants are primed for defense responses, exhibiting a slight elevation of SA-dependent defense gene expression even in the absence of applied biotic stress. Previously described cytokinin-responsive genes [16] were also differentially regulated by Hpa Noco2 in wild-type plants (Figure 5C). The overlap between the suites of cytokinin- and Hpa Noco2-regulated genes supports a function for cytokinin in plant immunity and suggests a role for cytokinin-regulated processes in the pathogenicity of Hpa Noco2. The altered expression of both suites of genes in arr3,4, 5,6, 8,9 mutants indicates that these processes converge at the level of type-A ARR function. Among the genes induced by Hpa Noco2 and de-repressed in arr3,4, 5,6, 8,9 mutant plants is ICS1, which encodes an enzyme involved in SA biosynthesis. We hypothesized that the altered expression of SA-dependent genes observed in arr3,4, 5,6, 8,9 plants is a direct result of altered regulation of ICS1 and SA metabolism, and subsequent activation of SA-dependent defense responses. We examined SA accumulation in the wild-type and arr3,4, 5,6, 8,9, and the contribution of Hpa Noco2 and high concentrations of cytokinin (100 µM BA) to this response. SA levels in wild-type plants, regardless of treatment, remained at or below levels of detection of our assay. These results are similar to published results of SA levels in Arabidopsis plants after infection with virulent isolates of Hpa [37] and reflect the relatively weaker defense responses elicited by virulent pathogens and the nature of the Hpa-Arabidopsis interaction, in which a limited number of plant cells are in contact with the pathogen at early stages of infection. The SA levels of arr3,4, 5,6, 8,9 plants treated with DMSO, cytokinin or Hpa Noco2 were also at or below the detection limits of our SA assay (Figure 6A). In contrast arr3,4, 5,6, 8,9 mutant plants pre-treated with cytokinin and subsequently challenged with Hpa Noco2 showed a significant and reproducible increase in SA levels, well above the detection limits of our assay (Figure 6A). These results suggest that the increased defense responses observed in arr3,4, 5,6, 8,9 mutant plants are due to increased SA content. The results of our SA assays led us to examine the expression of ICS1 and the defense marker PR1 in these plants. As expected, ICS1 expression was elevated synergistically by Hpa Noco2 and cytokinin treatment in both genotypes. arr3,4, 5,6, 8,9 plants showed even further up-regulation of ICS1, which could account for elevated SA levels observed in these plants (Figure 6B). Surprisingly, levels of PR1 were equally high in the arr3,4, 5,6, 8,9 mutant treated with cytokinin, or with cytokinin and Hpa Noco2, even though levels of SA and ICS1 differed (Figure 6C). These results indicate that in the absence of functional type-A ARRs, cytokinin can bypass the requirement for recognition of Hpa Noco2 on the activation of defense responses, suggesting a role for type-A ARRs in the suppression of defense responses downstream of SA accumulation. Consistent with increased defense gene expression and SA content, arr3,4, 5,6, 8,9 plants treated with cytokinin also exhibited increased resistance to Hpa Noco2 (Figure 6D). To better score susceptibility, plants were stained with lactophenol-trypan blue at 4 dpi. At this point during infection, wild-type plants pre-treated with DMSO showed widespread hyphal growth and sporulation, while wild-type plants pre-treated with cytokinin had not yet produced sporangiophores and displayed diminished hyphal growth. DMSO-treated arr3,4, 5,6, 8,9 plants showed decreased susceptibility compared to similarly-treated wild-type plants, and this was even more apparent in arr3,4, 5,6, 8,9 plants pre-treated with cytokinin, which showed substantially reduced hyphal growth (Figure 6D). We examined the influence of the plant hormone cytokinin on the immune responses of Arabidopsis plants in response to the biotrophic oomycete Hpa Noco2. The susceptibility of wild-type plants was increased after treatment with low concentrations of the cytokinin BA (<1 µM) and decreased with higher concentrations (>10 µm). This bell-shaped response is reminiscent of other physiological responses regulated by cytokinin, such as shoot initiation in vitro [38], [39] and the induction of ethylene biosynthesis [40]. In particular, this finding is similar to the effect of exogenous cytokinin on the response of wheat to powdery mildew (Erysiphe graminis), in which a complex dose response curve of pathogen growth was obtained in response to exogenous zeatin [20]. While multiple processes such as cytokinin uptake, degradation and conjugation likely contribute to the complexity of this response, our findings highlight the importance of hormone concentrations during the responses of plants to pathogens. All molecules with cytokinin activity are recognized in Arabidopsis by the three cytokinin receptors, AHK2, AHK3 and AHK4 [34], [35] that have varying affinities for different cytokinins [41]–[43]. Different cytokinins elicit different levels of cytokinin signaling upon binding to the cytokinin receptors [44]. It is thus possible that contrasting reports on the roles of cytokinin in susceptibility to pathogens might reflect the levels of signaling elicited by different cytokinins during plant-pathogen interactions and their different effects on pathogen growth, which would be similar to the effect of different levels of cytokinin on the susceptibility of Arabidopsis to Hpa Noco2 observed in this study. Treatment with lower concentrations of cytokinin resulted in a significant increase in Hpa Noco2 growth on wild-type plants. The mechanisms involved in this increased susceptibility may involve several physiological processes that are regulated by cytokinins, such as sink-source relationships, delay of senescence and/or nutrient acquisition [5], many of which likely affect to the ability of pathogens to grow optimally. Several plant pathogens produce cytokinins in order to manipulate plant physiology and development, thereby promoting optimal conditions for completion of their life cycle [21]. The role of lower concentrations of cytokinin for the susceptibility of Arabidopsis plants to Hpa Noco2 raises the question of whether Hpa-derived cytokinins could be contributing to the growth of this pathogen. Analysis of the Hpa genome does not reveal any isopentenyl transferases genes predicted to synthesize cytokinins, as found in plants and some plant pathogens [45]. Genes encoding tRNA isopentenyl transferases involved in a secondary cytokinin biosynthetic pathway are present in the Hpa genome, as they are in most genomes, but given the debatable role of tRNA-derived cytokinins in plant physiology [45] these are unlikely to contribute in a substantial way to the production of active cytokinins. Treatment of Arabidopsis with high levels of cytokinin led to an enhancement of defense responses, characterized by a decrease in susceptibility to Hpa Noco2. This effect of cytokinin was mostly abolished in eds16 plants, demonstrating that cytokinin acts primarily upstream of SA production in plant immune responses against Hpa Noco2. Treatment of plants with high concentrations of cytokinin led to a subtle increase in defense gene expression, which was further enhanced after treatment with Hpa Noco2. Consistent with our observations, a similar effect of comparably higher concentrations of cytokinin was observed in the induction of resistance and enhancement of defense gene expression to a pathogenic strain of Pst DC3000 in Arabidopsis [17], and a comparable effect of cytokinin on defense gene activation was reported in tobacco plants after wounding, also accompanied by increased SA levels [46]. This potentiation of defense gene expression by pre-treatment with cytokinin observed in our results indicates that cytokinin may act by priming the defense responses of Arabidopsis plants to Hpa Noco2. While the molecular mechanisms of priming are not well understood, it is hypothesized that priming may pre-activate defense signaling, but not defense responses, allowing plants to respond more rapidly to biotic and abiotic stresses [47] without the energy costs associated with pre-activation of full defense responses [48]. Given the role of cytokinins in carbon partitioning and energy allocation [49], [50], it is possible that cytokinin signaling might play a role in regulating the levels of energy that can be allocated into defense responses. The effect of high cytokinin concentrations on the susceptibility to Hpa Noco2 required the AHK2 and AHK3 cytokinin receptors, indicating that a cytokinin phosphorelay system is required for responses to Hpa Noco2. Similar to other processes regulated by cytokinin, the individual receptors contribute differently to this phenotype [34]–[36]. Type-A response regulators are negative regulators of cytokinin signaling [7]–[10]. We observed that type-A arr 3,4, 5,6, 8,9 multiple mutant plants exhibited decreased susceptibility to Hpa Noco2. While type-A ARRs exert mostly overlapping roles in cytokinin signaling, the combinatorial specificity observed in the responses of two different quadruple mutants to Hpa Noco2 may suggest distinct roles for individual type-A ARRs in regulation of plant immunity. Transgenic lines overexpressing type-A ARRs display decreased defense responses and allowed for increased pathogen growth; hence type-A ARRs are also negative regulators of plant immunity. Consistent with this conclusion, we note that type-A ARRs must function to regulate basal responses in uninfected plants, as unchallenged arr3,4, 5,6, 8,9 plants display elevated basal expression of several SA-regulated genes and as we observed a converse effect on SA-dependent signaling when type-A ARRs are overexpressed. Overexpression of ARR5D87A, which cannot be phosphorylated, did not lead to increased susceptibility. This result indicates that it is primarily the phosphorylated state of type-A ARRs that is active in the negative regulation of SA-dependent defense responses and that a complete phosphorelay cascade, initiated at the level of cytokinin receptors and culminating in type-A ARR phosphorylation and activation, is required in this process. This type-A ARR function is promoted by cytokinin and occurs downstream of SA; in the absence of functional type-A ARRs, defense gene expression, but not SA accumulation, is elevated following cytokinin treatment (Figure 7). While the exact mechanisms by which type-A ARRs function are still unknown, phosphorylation of their receiver domain has been shown to stabilize a subset of type-A ARR proteins [33], and likely to lead to phospho-specific interactions with target proteins, which in turn mediates cytokinin outputs [33]. A similar mechanism of response regulator action is employed in two-component systems in yeast [51]. Importantly, a type-B ARR transcription factor has also been shown to trigger enhancement of defense responses to the bacterial pathogen Pst DC3000. [17]. In this model, treatment of plants with comparably high concentrations of the cytokinin trans-zeatin leads the TGA3 b-zip transcription factor to associate with and recruit the type-B transcription factor ARR2 to specific cis-elements within the promoter of the PR1 gene, thus activating defense responses [17]. It is known that the phosphorelay cascade that is initiated after cytokinin perception promotes type-B ARR phosphorylation and activation, culminating in the transcription of cytokinin-regulated genes, which include type-A ARRs. In the context of plant immunity, high concentrations of cytokinin may lead not only to activation ARR2 and its association with TGA3 on the PR1 promoter, but also to the transcription of type-A ARR genes and their activation by phosphorylation, which might then counteract defense responses. In addition to cytokinin up-regulating SA-dependent responses, our results suggest that SA negatively regulates cytokinin signaling. Similarly to type-A arr mutants, eds16 plants showed hypersensitivity to low concentrations of cytokinin. In Hpa Noco2 susceptibility assays, eds16 plants also displayed hypersensitivity to high concentrations of cytokinin as compared to wild-type plants. Taken together, these results point to a possible feedback loop of SA on cytokinin signaling that would work to fine-tune the level of defense responses to pathogens. A possible trade-off between cytokinin-regulated and SA-dependent defense responses may have broad agricultural implications. Some species of plants, such as tomato, soybeans and particularly rice, have naturally high basal levels of SA [52]. If in these crop species SA negatively influences cytokinin-regulated processes, which include nutrient allocation and yield, manipulating this hormonal crosstalk may lead to increased crop productivity. Our results reveal a complex crosstalk between cytokinin and SA in plant immunity, in a mechanism involving two-component signaling elements and which incorporates regulation in part by type-A ARRs. Moreover, we show that cytokinin levels are important in determining the amplitude of plant immunity, ultimately influencing the outcome of plant-pathogen interactions. As the network of plant hormone interactions in plant immunity is further dissected, it is becoming clear that a for a complete appreciation of the role of plant hormones in this process, the levels of hormonal signaling will also have to be considered. The Col-0 accession was used as the wild-type in this study. The ahk T-DNA knockout mutants used in this study have been described in [53]. Type-A arr T-DNA knockout mutants (arr3,4, 5,6; arr5,6, 8,9 and arr, 3,4, 5,6, 8,9) and ARR-overexpressing transgenic lines have been described [10], [33]. eds16 plants have been previously described [28]. All mutants and transgenic lines described above are in the Col-0 accession. All plants were grown on soil (Metro 360) in growth chambers (Percival Scientific) under short days (8∶16 hour light∶dark, 22°C). Hpa Noco 2 was propagated on the susceptible Col-0 accession. Hpa spores (5×104/ml) were sprayed onto two-week-old plants using a pressurized sprayer (Preval). Inoculated plants were kept in growth chambers (Percival Scientific) (19°C, 8∶16 hour light∶dark) and covered with a transparent plastic dome to maintain high humidity. For Hpa assays, two-week old plants were inoculated as described above. One day after the first appearance of sporangiophores (5–6 dpi) the first pair of true leaves was collected from three individual plants, and added to a previously weighed 1. 5 ml microcentrifuge tube containing 300 µl of sterile water, for a total of six leaves per sample, and weighed again to determine fresh weight. Spores were counted using a hemacytometer. Spore counts from at least four samples per genotype were determined. Plants were harvested at 4 dpi and stained with a 3∶1 ethanol: lacto-phenol trypan blue solution (1∶1∶1∶1 phenol∶ lactic acid∶ water∶ glycerol and 0. 05% trypan blue (Sigma-Aldrich) ), at 95°C, for 5 min, and moved to room temperature for 10 min. Excess staining was removed with chloral hydrate (Sigma-Aldrich). Samples were moved to 50% glycerol for storage and mounting. Pictures were taken with an Olympus SZX9 stereomicroscope. Cytokinin (benzyl adenine, or BA) (Sigma-Aldrich) was sprayed onto two-week-old plants, using a Preval sprayer. BA solutions were prepared from a stock in DMSO, diluted into an aqueous solution to the required BA concentration plus 0. 002% Silwet L-77 (Lehle Seeds). Control plants were sprayed with the corresponding amounts of DMSO plus 0. 002% Silwet L-77. Total RNA was extracted using RNeasy Plant kit (QIAGEN), according to the manufacturer' s instructions. Quality and integrity of RNA were assessed by gel electrophoresis and A260/A280 and A260/A230 ratios. RNA samples of good quality were treated with DNAse-free Turbo (Ambion) and then checked for absence of genomic DNA by qRT-PCR using primers for At5g65080, (At5g65080 For 5′-TTTTTTGCCCCCTTCGAATC-3′, At5g65080 Rev 5′-ATCTTCCGCCAC-CACATTGTAC-3′). cDNA synthesis was performed using Superscript III (Invitrogen) and oligo-d (T) primers according to the manufacturer' s instructions. cDNA was checked for full extension by qRT-PCR using primers for 3 amplicons 1 kB apart within At1g13320 (At1g13320a, At1g13320b, At1g13320c); primers used are as follows: At1g13320a For 5′-TAGATCGCTCGGAACTTGGAAA-3′; At1g13320a Rev 5′-GGAGTGATTTGAGTTTTGGTGAGG-3′; At1g13320b For 5′-AACTAGGACGGATCTGGTGCCT-3′; At1g13320b Rev 5′-ATAATGAGGCA-GAAGTTCGGATAGC-3′; At1g13320c For 5′-AAATTTAACGTGGCCAAAA-TGATGC-3′; At1g13320c Rev 5′-ACCAAGCGGTTGTGGAGAAC-3′. cDNAs with Ct ratios of At1g13320a/At1g13320b and At1g13320b/At1g13320c below 1. 5 Cts were considered suitable for qRT-PCR. qRT-PCR reactions were performed using ExTaq SYBR Green (Takara) on a Bio-Rad Opticon2 machine using the following thermocycler program: (1) 2 min at 95°C; (2) 15 s at 95°C; (3) 15 s at 60°C; (4) 15 s at 72°C; (5) optical read, repeat 34 cycles of steps 2 through 5, followed by a final analysis of product melting temperature to confirm the PCR product. β-TUBULIN 4 (At5g44340) was used as housekeeping gene in all reactions. Gene-specific primers are as follows: ATST4B (At1g13420) For 5′-AGCCTCGTGTGCAAA-TCAAGAGAC-3′, Rev 5′-ACTCCTTCCGACAAGCT-TCCTGTT-3′; ARR7 (At1g19050) For 5′-ACTGTAGAGAGTGGAACTAGGGCT-3′, Rev 5′-AGTCCTGGCATTGAGTAATCCGTC-3′; ICS1 (At1g74710) For 5′-TGCATCCAACTCCAGCTGTTTGTG-3′; Rev 5′-AGCTGATCTGATCCCGA-CTGCAAA-3′; PR1 (At2g14610) For 5′-ACACGTGCAATGGAGTTTGTGGTC-3′; Rev 5′-TACACCTCACTTTGGCACATCCGA-3′; FRK1 (At2g19190) For 5′-AGCTTCTCTGTTGAAGGAAGCGGT-3′; Rev 5′-TTGAGCTTGCAATAGC-AGGTTGGC-3′; XTR7 (At4g14130) For 5′-AGCTCAATGCTTATGGCAGGAGGA-3′; Rev 5′-TTGCATTCTGGAGGGAAT-CCACGA-3′; ACD6 (At4g14400) For 5′-GTGACGTTTG-CTGCAGGCTTTACA-3′, Rev 5′-AGTTGGGTTAGTGGC-CAAAGTTGC-3′; CKX4 (At4g29740) For 5′-CACCCACAAGGGTGAAATGGTCTC-3′, Rev 5′-TGCGACTCTTGTTTGATCGGAGAG-3′; WRKY18 (At4g31800) For 5′-TGGGTCAAGCACAGTGAC-TTTGGA-3′; Rev 5′-GCAGCAGCAAGAGC-AGCTGTAAAT-3′; β-TUBULIN 4 (At5g44340) For 5′-AGAGGTTGACGAGCAA-GATGA-3′, Rev 5′-AACAATGAAAGTAGACGCCA-3′; PDF1. 2 (At5g44420) For 5′-GCTTCCATCATCACCCTTATCTTC-3′; Rev 5′-ACATGGGACGTAACAGATACA-CTTGTGT-3′. The relative expression of specific genes and 95% confidence intervals were determined using REST 2008; [54] (http: //rest-2008. gene-quantification. info). At least three biological replicates of each experiment were obtained and qRT-PCR performed as described above. Arabidopsis seeds were grown on vertical plates containing MS medium (1× MS salts, 0. 05% MES buffer, and 1% sucrose, pH 5. 8), with 0. 6% phytagel (Sigma-Aldrich) supplemented with a dose range of BA or 0. 1% (v/v) DMSO vehicle control for 10 days. Primary root lengths at days 4 and 9 were marked on the plates. The plates were scanned at 10 days, and root growth between days 4 and 9 was measured using NIH Image J version 1. 43u (National Institutes of Health, Bethesda, MD). Two-week old plants (wild-type and arr3,4, 5,6, 8,9) grown under short days (8∶16 hour light∶dark cycle, 22°C) were sprayed with distilled water (control) or Hpa Noco2 as described above. Plants were kept at 18°C and 8∶16 hour light∶dark cycle. Tissue was harvested three days after treatment. Two independent biological replicates of the experiment were obtained. Total RNA was extracted using RNeasy Plant Kit (QIAGEN). 30 µg of total RNA were converted into cRNA and hybrized to ATH1 chips (Affymetrix) according to the manufacturer' s instructions. Data were RMA-transformed and analyzed using Genespring software version GX 10 (Agilent). Raw values were filtered to a minimum expression of 20th percentile and statistical analysis was performed with two-way ANOVA (α≤0. 05) using Benjamini-Hochberg multiple testing correction. For interpretation of data, wild-type water-treated samples were used as a baseline (control) for comparison to the other samples. Two-week-old seedlings were pre-treated with either DMSO or cytokinin BA and subsequently inoculated with either water or Hpa Noco2 as in Figure 1. Tissue was harvested at 3 dpi. Total SA measurements, including free SA and SA glucoside (SAG), were performed as described [55]. Briefly, frozen samples were ground and tissue homogenized in 200 µl 0. 1 M acetate buffer pH 5. 6. Samples were then centrifuged for 15 min at 16,000 g at 4°C. 100 µl of supernatant was transferred to a new tube for free SA measurement, and 10 µl were incubated with 1 µl 0. 5 U/µl β-glucosidase (Sigma-Aldrich) for 90 min at 37°C for total SA measurement. After incubation, 60 µl of LB, 20 µl of plant extract (treated or not with β-glucosidase), and 50 µl of Acinetobacter sp. ADPWH-lux (OD = 0. 4) were added to each well of a black 96-well plate. The plate was incubated at 37°C for 60 min and luminescence was read with a Spectra Max M5 (Molecular Devices) microplate reader. For the standard curve, 1 µl of known amounts of SA stock (from 0 to 1000 µg/ml) was diluted 10-fold in eds16 plant extract, and 5 µl of each standard were added to the wells of the plate containing 60 µl of LB, and 50 µl of Acinetobacter sp. ADPWH-lux (OD = 0. 4). SA standards were read in parallel with the experimental samples. SA standard values were analyzed with linear regression for calculations of SA amounts. Results are depicted by gram of fresh weight.

Plant hormones play an important role in many aspects of a plant' s life cycle, from the regulation of development to responses to constantly changing environmental conditions. In the past decade, the importance of hormones in plant immunity against a variety of pathogens has been uncovered. In this manuscript, we demonstrate that in the model plant species Arabidopsis thaliana components of the signaling system of the plant hormone cytokinin also mediate plant immunity. We demonstrate that this involves the type-A class of Arabidopsis response regulators in a process that occurs downstream of the plant defense hormone salicylic acid and involves a host two-component phosphorelay. Moreover, we show that the levels of cytokinin are important in determining the amplitude of plant immunity, ultimately influencing the outcome of plant–pathogen interactions. Finally, our results indicate that salicylic acid negatively regulates cytokinin signaling, which may serve to fine-tune the effects of cytokinin in plant immunity. Given the high energy costs of defense responses and the role of cytokinins in carbon partitioning and energy allocation, we hypothesize that the mechanisms uncovered here may help regulate the levels of energy that can be allocated into defense responses, an important aspect in the biology of plants.

Abstract Introduction Results Discussion Methods

plant science plant biology biology

2012

Two-Component Elements Mediate Interactions between Cytokinin and Salicylic Acid in Plant Immunity

Activated protein kinase R (PKR) plays a vital role in antiviral defense primarily by inhibiting protein synthesis and augmenting interferon responses. Many viral proteins have adopted unique strategies to counteract the deleterious effects of PKR. The NSs (Non-structural s) protein which is encoded by Rift Valley fever virus (RVFV) promotes early PKR proteasomal degradation through a previously undefined mechanism. In this study, we demonstrate that NSs carries out this activity by assembling the SCF (SKP1-CUL1-F-box) FBXW11 E3 ligase. NSs binds to the F-box protein, FBXW11, via the six amino acid sequence DDGFVE called the degron sequence and recruits PKR through an alternate binding site to the SCFFBXW11 E3 ligase. We further show that disrupting the assembly of the SCFFBXW11-NSs E3 ligase with MLN4924 (a small molecule inhibitor of SCF E3 ligase activity) or NSs degron viral mutants or siRNA knockdown of FBXW11 can block PKR degradation. Surprisingly, under these conditions when PKR degradation was blocked, NSs was essential and sufficient to activate PKR causing potent inhibition of RVFV infection by suppressing viral protein synthesis. These antiviral effects were antagonized by the loss of PKR expression or with a NSs deleted mutant virus. Therefore, early PKR activation by disassembly of SCFFBXW11-NSs E3 ligase is sufficient to inhibit RVFV infection. Furthermore, FBXW11 and BTRC are the two homologues of the βTrCP (Beta-transducin repeat containing protein) gene that were previously described to be functionally redundant. However, in RVFV infection, among the two homologues of βTrCP, FBXW11 plays a dominant role in PKR degradation and is the limiting factor in the assembly of the SCFFBXW11 complex. Thus, FBXW11 serves as a master regulator of RVFV infection by promoting PKR degradation. Overall these findings provide new insights into NSs regulation of PKR activity and offer potential opportunities for therapeutic intervention of RVFV infection. Activated double-stranded (ds) RNA-dependent protein kinase (PKR or EIF2AK2) plays a vital role in antiviral defense primarily by inhibiting protein synthesis and augmenting interferon responses [1]. PKR is a serine/threonine kinase that is maintained as an inactive monomer and undergoes activation in response to dsRNA and/or cellular stress signals, primarily resulting from viral infection. Activated PKR undergoes auto-phosphorylation and inhibits protein synthesis by phosphorylation of the eIF2-α (eukaryotic translation initiation factor 2 subunit alpha or EIF2A). The importance of PKR in innate antiviral responses is suggested by the existence of a multitude of viral regulators of PKR action. Proteasomal degradation of PKR is one of the numerous strategies used by viruses to impair PKR function. Most of the targeted protein ubiquitination and subsequent proteasomal degradation are achieved by Cullin-RING E3 ligases (CRL [2,3]). The CRLs are modular assemblies centered on one of the several cullin scaffolds CUL1, CUL2, CUL3, CUL4A, CUL4B, CUL5, CUL7 and CUL9 forming the corresponding CRL1, CRL2, CRL3, CRL4A, CRL4B, CRL5, CRL7 and CRL9 ligases. The C terminal domain (CTD) of the cullin module contains an embedded RING finger protein (RBX1 or RBX2) that serves as the site for E2 binding and ubiquitin transfer activity. The amino terminus has an adaptor protein that binds to substrate receptors to recruit specific target proteins destined for ubiquitination. The SCF E3 ligase or CRL1, consisting of SKP1, CUL1, and an F-box protein (SCF), is the founding member of the CRLs. The substrate specificity is determined by the adaptor protein SKP1 and any of the 72 F-box substrate receptors bound to the N-terminus of CUL1 module. The enzymatic activity of CRLs is dependent on cullin modification by the covalent attachment of NEDD8, a 9-kDa ubiquitin-like molecule, which requires the activity of the NEDD8 activating enzyme, NAE1[3]. Recently, a specific NAE1 small molecule inhibitor, MLN4924, has been developed and is currently in clinical trials for cancer therapy [4]. MLN4924 inhibition of cullin NEDDylation blocks CRL activity resulting in the accumulation of CRL targets. Rift Valley fever virus (RVFV) is a member of the genus Phlebovirus within the family Bunyaviridae. This virus causes a severe disease, Rift Valley fever (RVF), that affects humans and livestock throughout Africa and the Arabian Peninsula. RVF occurs in large epidemics affecting public health and agriculture resulting in significant economic losses. Currently, there are no FDA approved drugs or vaccines to treat RVF. Understanding the molecular mechanisms of RVFV infection may lead to the development of new therapeutics. The RVFV genome is composed of three RNA segments, the S-, M- and L-segments, which code for the viral structural proteins, nucleocapsid (N) protein that encapsidates the viral RNA to form ribonuceloprotein (RNPs) complexes, the envelope glycoproteins (Gn and Gc), and the RNA-dependent RNA polymerase (RdRp), respectively. Virus entry into cells is mediated by the binding of the envelope glycoproteins (Gn/Gc) to an unknown cell surface receptor which mediates virus endocytosis. Acidification of the virus-containing endocytotic vesicle promotes virus-host membrane fusion and results in the release of the RNPs and RdRp into the cytoplasm, where transcription and replication of the viral genome occurs [5]. The glycoproteins Gn and Gc are produced by a glycoprotein precursor (GPC) that is co-translationally cleaved by an unknown protease. Gn and Gc form heteromeric complex which localize in steady-state at the Golgi apparatus where multiple interactions of the glycoproteins with RNPs and RdRp are believed to cause a change in membrane curvature leading to virus budding into the Golgi lumen [6–8]. Release of the virus-filled vesicles from the Golgi and their subsequent fusion with cell plasma membrane results in the release of mature virions into the extracellular medium. Thus, glycoprotein (Gn/Gc) expression on the cell surface reflects successful completion of the virus life cycle that spans approximately 10–12 h [9]. The viral genome also encodes three nonstructural proteins, NSs, NSm (NSm2) and the less characterized 78-kD NSm1. NSs is considered the major virulence factor of RVFV [10]. For instance, the C13 strain of RVFV, which has an in-frame deletion of the majority of the NSs gene, is avirulent in mice and sheep [11,12]. NSs effectively blocks antiviral responses by inhibiting both interferon synthesis and further downstream transcription of interferon-induced genes. NSs also promotes PKR proteasomal degradation during early stages of virus infection [13,14]. Recent evidence suggests that PKR degradation is required for efficient viral replication [15]; however, the molecular mechanisms regulating the NSs-mediated PKR proteasomal degradation have not been well-characterized. We sought to investigate the molecular mechanism of PKR degradation and found that early PKR degradation is essential for RVFV infection. NSs targets PKR degradation by assembling (SKP1-CUL1-F-box) SCFFBXW11 E3 ligase via its binding to F-box protein FBXW11. A degron sequence, DDGFVE263, in NSs protein regulates the NSs-FBXW11 interaction and the SCF E3 ligase assembly. Disrupting the assembly of the SCFFBXW11-NSs with the small molecule MLN4924 or RVFV encoding NSs degron mutants or siRNA knockdown of FBXW11 successfully blocks PKR degradation. Under these disruptive conditions, NSs expression was sufficient for early PKR activation, causing potent inhibition of RVFV infection by suppressing protein synthesis. Furthermore, FBXW11 is the limiting factor in SCFFBXW11-NSs assembly and thereby serves as a master regulator of RVFV infection. Overall, these findings illustrate that inactivation of SCFFBXW11-NSs is sufficient to induce early PKR activation and inhibition of RVFV infection. To explore the mechanism of NSs mediated PKR proteasomal degradation, we first examined how small molecule inhibitors targeting different stages of the ubiquitin (UB) -proteasome pathway (UPP), including MG132 (proteasome inhibitor) and MLN4924 (CRL E3 ligase inhibitor), regulated RVFV infection. Both MG132 and MLN4924 inhibited RVFV infection, but to gain mechanistic insights into this inhibition, we focused on MLN4924 since it targets a specific class of E3 ligases unlike MG132, which globally inhibits the UPP. As shown in Fig 1A, MLN4924 at 1 μM inhibits the viral antigen expression of both the virulent strain ZH501 and the vaccine strain MP-12 based on immunofluorescence analysis (IFA) of HeLa cells. The latter were infected with the corresponding viruses for 24h at an MOI (multiplicity of infection) = 1. The potency of MLN4924 was determined by a 10-point concentration response curve in RVFV-infected HeLa cells (Fig 1B and 1C). Viral infections were measured by high-content image based analysis (HCA) using previously described methods [16,17]. Briefly, images of immunofluorescently labeled viral antigen-expressing cells were acquired by automated high-speed microscopy and subjected to image analysis software to enumerate the viral antigen expressing cells. The assay conditions were optimized such that greater than 50% of cells expressed viral antigen after multiple rounds of infection at a minimal MOI (further details in Methods section). HeLa cells treated with increasing concentrations of MLN4924 were infected with either ZH501 (Fig 1B) or MP-12 (Fig 1C) in a 96-well plate at an MOI = 1. After 24h, cells were subjected to IFA of glycoprotein (G) or nucleocapsid (N) protein expression. Viral infection was measured by HCA of N or G expression and was also compared to viral titers determined by plaque assay of the cell supernatants (Fig 1B and 1C). As shown in Fig 1B and 1C, irrespective of the method used to measure viral infection, MLN4924 showed a dose dependent inhibition of RVFV infection. Furthermore IC50 values (of 10 nM—150 nM) fall within the range of values expected for MLN4924’s activity on CRL in cell culture assays [4] and there was no significant decrease in cell number when compared to mock-treated cells for the duration of the infection (24h). The IC50 values by plaque assay were lower than the IC50 values by HCA due to the differences in the parameters measured, but nevertheless followed the same trend of inhibition. For example when the percentage of G expressing cells decreased by two fold, the corresponding virus titers dropped by a log value (Fig 1B and 1C). For rest of the study, HCA was used for measuring RVFV infection. Under the conditions of infection described for HCA, RVFV infection showed a linear increase in percentage of antigen expressing cells with increasing MOI (S1A and S1B Fig). A similar trend in the increase of viral titers was observed when viral infection was measured by plaque assay; however, HCA is relatively more sensitive to small changes in MOI and is a more quantitative and robust assay compared to the plaque assay (S1A and S1B Fig). The MLN4924 mediated RVFV infection inhibition was also observed in multiple primary and cancer cell lines of human or mouse origin (Fig 1D). MLN4924 antiviral activity was specific to RVFV and was not observed with other RNA viruses including Venezuelan equine encephalomyelitis virus (VEEV), Lake Victoria Marburg virus (MARV), or Lassa virus (LASV) within the families Togaviridae, Filoviridae, and Arenaviridae, respectively, when viral infections were measured either by HCA (Fig 1E) or plaque assay (Fig 1F). MLN4924 inhibits CRL activity by blocking conjugation of NEDD8 to the Cullin subunit. Further genetic evidence of NEDDylation’s role in RVFV infection was determined by overexpression of a point mutant, UBC12 (C111S), in the E2 NEDD8-conjugating enzyme, which has a dominant-negative effect on NEDD8 conjugation [18]. UBC12 (C111S) expression, and not the empty vector, inhibited RVFV infection specifically thus providing genetic proof for the role of NEDDylation (S1C Fig). We next examined the stage of the viral life cycle that is critically dependent on CRLs by time of compound addition (ToA) studies (S1D Fig) and by viral RNA and gene expression kinetics during one viral replication cycle, which is ~12h (S1E–S1G Fig). HeLa cells were infected with a high MOI (MOI = 10) to achieve an infection rate >80% during one cycle of viral replication. Infection rates in ToA were evaluated by HCA of G expression in non-permeabilized cells in order to detect virus egress. ToA studies showed that MLN4924 treatment could be postponed until 5h post infection (PI [S1D Fig]). The kinetics of early N and late G gene expression in permeabilized cells showed that late gene expression was severely inhibited, while early gene expression was normal until 6h PI (S1E Fig). The inhibition in viral RNA expression levels also began to appear from 5-7h PI (S1F Fig) and compound addition could be postponed until 5h PI (S1G Fig). Collectively, these data suggest that MLN4924 does not target viral entry, rather it inhibits viral RNA and gene expression beginning at 5-6h PI as late gene expression was more severely inhibited than the early gene. Since MNL4924 regulated CRL E3 ligase activity, we hypothesized that MLN4924 treatment of RVFV infected cells might block NSs-mediated PKR proteasomal degradation resulting in accumulation of activated and phosphorylated PKR (p-PKR), which then might phosphorylate eIF2-α (p-eIF2-α) to suppress further viral protein synthesis. Thus, we examined the kinetics of PKR, p-PKR and p-eIF2-α expression during the course of a single cycle of infection (i. e. , 12 h) in DMSO (vehicle control) or MLN4924 treated HeLa cells (Fig 2A). Due to a lack of suitable antibodies to the NSs protein, recombinant MP-12 (rMP-12) virus containing the NSs gene engineered with a V5 tag at the C terminus i. e. , rMP-12-NSs-V5, was used in place of wildtype MP-12. As shown in Fig 2A, HeLa cells infected with rMP-12-NSs-V5 and treated with MLN4924 failed to clear PKR and led to a robust and simultaneous induction of p-PKR and p-eIF2-α beginning at 6h PI. Induction of p-PKR coincided with NSs expression. As expected, phosphorylation of eIF2-α suppressed further increases in NSs levels or synthesis of the late viral G protein, whose expression normally follows NSs. MLN4924 treatment also restored p62 expression levels, which under normal infection conditions would be targeted by NSs for proteasomal degradation [19]. As a control for MLN4924 activity, the lysates were also probed with anti-NEDD8 antibodies to determine the status of cullin NEDDylation. As expected, control treated cells showed a strong signal at a molecular weight that corresponded to that of cullin proteins, while no bands were detected with the MLN4924 treated lysates (Fig 2A). Collectively, these data show that MLN4924 prevented PKR degradation and instead activated PKR. The above data suggested that NSs probably blocked PKR activation by depleting PKR from cells. If that was the case, NSs deficient virus might induce PKR activation. Furthermore, if MLN4924’s antiviral activity was primarily due to PKR activation, loss of PKR expression should restore RVFV infection. Both these hypotheses were tested by measuring infection rates of wildtype or NSs-deficient RVFV mutants treated with DMSO (vehicle control) or MLN4924 under normal or low PKR siRNA knockdown conditions during one cycle of virus replication (Fig 2B). Two different NSs deficient mutants were used: Clone-13, which is a naturally attenuated isolate of RVFV encoding a non-functional NSs gene due to a large in-frame deletion, and rMP-12ΔNSs: : Luci virus, in which the NSs gene of MP-12 is replaced with a luciferase (Luci) gene by reverse genetics [20]. We confirmed that PKR expression levels were indeed reduced in HeLa cells transfected with siRNA targeting PKR (inset in Fig 2B). MLN4924 treatment inhibited MP-12 and ZH501 infection and as predicted this inhibition was completely rescued in cells with low levels of PKR expression (Fig 2B). Therefore PKR activation was responsible for MLN4924 antiviral activity. But surprisingly, NSs deficient viruses (clone 13 and rMP-12ΔNSs: : Luci) were insensitive to MLN4924 treatment. In addition, replication of these viruses did not increase under low PKR expression levels. One explanation was that NSs was essential for early PKR activation; alternatively, it was also possible that the NSs deficient viruses had delayed replication kinetics that may have further delayed PKR activation. To address these questions, we examined the kinetics of viral gene expression and the PKR activation pathway by Western blot analysis during one cycle of virus replication with the NSs deficient virus (rMP-12ΔNSs: : Luci) under mock or MLN4924 treatment conditions (Fig 2C). This virus showed the same kinetics of G expression as wildtype rMP-12, with an increase in G expression becoming apparent by 8 h PI in both viruses (compare G expression kinetics in Fig 2A and 2C). These data suggested that at least during the early stages of the first cycle of viral replication there is no significant difference in the replication kinetics between NSs deficient virus when compared to rMP-12 virus. In agreement with previous observations (Fig 2B), the NSs deficient virus was insensitive to MLN4924 treatment as no decrease in G expression levels during MLN4924 treatment was observed. Moreover, MLN4924 did not induce early PKR activation. However, the NSs deficient virus did induce late PKR activation at 10–12 h PI, by which time most of the viral proteins were expressed and the virus egress had already been initiated. Thus, NSs is essential for early PKR activation upon MLN4924 treatment in RVFV infected cells. We next examined if NSs is sufficient for PKR activation by transient transfection of NSs into HeLa cells that were treated with DMSO (vehicle control) or MLN4924 (Fig 2D). Western blot analysis showed that NSs expression alone was sufficient to induce PKR activation (based on increase in p-PKR and p-eIF2-α expression levels) by MLN4924 treatment. Collectively, these data suggest that NSs has dual antagonistic functions, promoting PKR degradation, but also activating PKR under conditions when NSs-mediated PKR degradation is blocked. We next sought out to identify the specific CRL that regulated PKR degradation by NSs. We hypothesized that NSs targeted PKR degradation by assembling CRL E3 ligase. Thus, we evaluated the binding of NSs to various cullin family members by co-immunoprecipitation (co-IP) assay (Fig 3A). Lysates from 293T cells that were transfected with vector alone or myc-tagged CUL1-3, -4A, -4B, -5 and -7 or HA-tagged CUL9 expression vectors were combined individually with rMP-12-NSs-V5 infected cell lysates and immunoprecipitated with anti-myc or anti-HA antibodies to determine NSs binding to the corresponding CRL. NSs bound to CUL1, but not with the other cullin family members (CUL2, -3, -4A, -4B, -5, -7 or -9). Furthermore, overexpression of dominant negative CUL1, but not other cullin family members, activated PKR in RVFV-infected cells (S2 Fig), thus providing functional significance to the NSs-CUL1 interaction. CUL1 forms a molecular scaffold to organize CRL1 by forming two distinct modules. The carboxyl (C) terminal domain binds to RBX1, which recruits the ubiquitin-loaded E2 enzymes for catalysis while the amino (N) terminus of CUL1 binds to the adaptor protein SKP1, which recruits the substrate via the substrate recognizing F-box protein (Fig 3C). Using previously characterized CUL1 mutants (see Methods section) that selectively lost their ability to interact with SKP1 or RBX1; we tested for NSs binding by co-IP assay. As shown in Fig 3B, NSs bound to wt-CUL1 (lane 1) and to the CUL1 mutants that have retained their capacity to interact with SKP1 (lanes 2,5, 6,7 and 8); in contrast, NSs binding to CUL1 was severely reduced with the mutants CUL1 ΔN53 and CUL1 Y42A/M43A, which cannot bind SKP1 (lanes 3 and 4). Taken together, these data suggest that NSs binds to the CRL1/SCF complex via the classical substrate binding motif in the CUL1 subunit, and thus this interaction is most likely facilitated by an F-box (see model in Fig 3C). We speculated that if NSs-PKR were recruited by an F-box protein, then depletion of the specific F-box protein would prevent PKR degradation, activate PKR, and inhibit RVFV infection. Thus, we used an siRNA screening assay targeting 70 out of the 72 F-box genes characterized in the human genome to determine the specific F-box gene (s) involved in RVFV infection [21,22]. CDRT1 and FBXO49 were not classified as F-box genes during the time this study was initiated and therefore were not included in our siRNA screening assay (refer to gene list in S3 Fig). For this assay, a pool of 4 different siRNAs was used to knockdown each of the 70 F-box genes in triplicate. As shown in Fig 4A, siRNA knockdown of four genes resulted in 30–50% inhibition of ZH501 infection in HeLa cells (marked in orange). Further validation of the hits was achieved by de-convoluting the pool of four siRNAs into individual siRNAs and testing their antiviral activity, cellular cytotoxicity and their ability to decrease target mRNA levels. A hit was considered genuine if two or more siRNAs showed antiviral activity without cytotoxic effects (<15% decrease in cell number). Among the four hits, siRNA’s targeting FBXW11 alone passed these criteria (Fig 4B and 4C). As shown in Fig 4B, 3 out of the 4 individual siRNAs targeting FBXW11 showed greater than 70% knockdown efficiency in mRNA levels and also inhibited viral infection of both MP-12 and ZH501 strains by greater than 50%. The fourth siRNA was able to knockdown FBXW11 mRNA by approximately 70%, however, it did not inhibit viral infection. These data suggest that a higher level (greater than 70%) of FBXW11 mRNA knockdown is required to inhibit RVFV infection. These data were also validated with three additional siRNAs targeting different regions of the mRNA (S4A Fig). Furthermore, antiviral activity due to loss of FBXW11 expression was specific to RVFV infection and was effective in multiple cell lines including 293T and HSAEC (human small airway epithelial cells) primary cells, thus following the same pattern of antiviral activity displayed by MLN4924 treatment (S4B Fig). FBXW11 and BTRC are two distinct homologues of the βTrCP gene and many studies suggest that they are functionally redundant [23]. Thus we determined, whether BTRC partially compensated for loss of FBXW11 expression in RVFV infection. As shown in Fig 4C, combined knockdown of both BTRC and FBXW11 was more potent in suppressing RVFV infection than knockdown of FBXW11 alone. Surprisingly loss of BTRC expression had no effect on RVFV infection and was compensated by FBXW11. The observed antiviral effects were not due to cytotoxicity or non-specific effects of siRNA knockdown (Fig 4D). Thus, FBXW11 plays a dominant role among the two homologues in RVFV infection. We then tested if knockdown of βTrCP inhibited RVFV infection through the same mechanism as global inactivation of CRL by MLN4924 treatment. This included loss of PKR degradation, PKR activation that resulted in increased expression levels of p-PKR and p-eIF2-α, followed by inhibition of viral gene expression within 8 h of RVFV infection. This pattern of antiviral gene expression was evaluated by Western blot analysis. As shown in Fig 4E, loss of FBXW11 expression alone (lane 5) or FBXW11 and BTRC combined together (lane 7) showed the same gene expression pattern as inactivating CRLs by MLN4924 treatment (lane 3). However the phenotype of FBXW11 knockdown was milder when compared to combined knockdown of BTRC and FBXW11. FBXW11 played a dominant role and loss of FBXW11 expression recapitulated loss of global activity by CRLs during RVFV infection. Based on these data, we predicted that FBXW11 might be the key factor that recruited NSs-PKR to the SCF complex, and thus examined the NSs-FBXW11 interaction. As shown in Fig 4F (top panel), the βTrCP protein contains the N terminal “D” or protein dimerization domain and the F-box motif that regulates recruitment to the SCF complex via binding to SKP1. The seven WD40 repeat motifs in the C terminus together form a beta-propeller like structure and regulate substrate binding. We characterized the NSs-βTrCP interaction by co-IP assay. As shown in Fig 4F (bottom panel), NSs binds to both homologues, BTRC and FBXW11, of βTrCP. Furthermore, WD40 repeats play an essential role in this interaction. As a control, the CUL1 binding to βTrCP mutants was analyzed. As reported previously, F-box mutants (ΔF and ΔNF deletion mutants) reduced the CUL1-βTrCP interaction. Overall, it appears that NSs targets PKR degradation by assembling SCFFBXW11. If SCFFBXW11 targets PKR degradation, then loss of expression of key components of this complex should inhibit RVFV infection by the same mechanism as global inactivation of CRL by MLN4924 treatment. We therefore examined the status of the PKR activation pathway and viral gene expression by Western blot analysis of mock or rMP-12-NSs-V5 infected HeLa cells in which the individual components of the SCFFBXW11 were knocked down by siRNA, either singly or in combination. As shown in Fig 5A, knockdown of FBXW11 in combination with CUL1 and SKP1 (lane 4) or with CUL1 alone (lane 5) resulted in the same gene expression signature as MLN4924-treated cells (lane 7). Knockdown of FBXW11 (lane 3) had a milder effect; whereas, knockdown of CUL1 (lane 1) or SKP1 (lane 2) alone had no effect and was identical to control siRNA treated cells. SiRNA knockdown reduces, but does not completely eliminate, the protein in the cells. Therefore, low residual expression of CUL1 or SKP1, in the presence of endogenous expression levels of FBXW11 was probably sufficient to rapidly degrade PKR. Overall, these data suggest that SCFFBXW11 is the primary complex that regulates PKR degradation, and failure to destroy PKR leads to a concomitant increase in p-PKR and p-eIF2-α levels, which suppresses viral protein synthesis. If this is the case, then depleting PKR in the cell should rescue the viral inhibition resulting from knockdown of the SCFFBXW11 complex. We approached this question by using a HeLa cell line that stably expresses a doxycycline-inducible PKR shRNA. We first confirmed by Western blot that doxycycline treatment decreased PKR expression in these cells (Fig 5B, inset). Next we examined the effects of knocking down SKP1, CUL1, FBXW11 alone or in combination, in the presence or absence of PKR, on infection levels of RVFV ZH501 in HeLa cells. As shown in Fig 5B, under normal PKR expression conditions, at saturating infection rates, depletion of CUL1 had no effect on RVFV infection, while SKP1or FBXW11 depletion resulted in greater than 40% or 80% reduction in infection. Moreover, knockdown of FBXW11 in combination with CUL1 and SKP1, individually or in combination, resulted in nearly complete inhibition of RVFV infection, similar to that observed when cells were treated with MLN4924. The infection inhibition by loss of key components of SCFFBXW11 was completely reversed by loss of PKR expression as expected. These data support the idea that SCFFBXW11 is the primary complex that regulates PKR degradation, and failure to destroy PKR acts to suppress viral protein synthesis and thus viral infection. To further examine the role of NSs in the degradation of PKR via SCFFBXW11 and in the activation of PKR, we repeated the experiment using RVFV Clone13, which encodes a non-functional NSs gene. As shown in Fig 5C, knockdown of any of the components of the SCFFBXW11 complex did not inhibit viral infection, which was similar to that observed with MLN4924 treatment. These data strongly suggest NSs regulated PKR activation in conditions when PKR degradation was blocked by inactivation of SCFFBXW11. We speculated that PKR is recruited to the SCFFBXW11 complex for degradation by the NSs protein. Thus, we examined the binding between PKR and the SCFFBXW11 complex in the presence and absence of NSs protein. As shown in Fig 5D, PKR bound to CUL1 or FBXW11 in the presence of NSs (lanes 1 and 3), but no binding of PKR to CUL1 or FBXW11 was detected in the absence of NSs (lanes 2 and 4). Furthermore, FBXW11 appears to serve as the NSs-PKR adaptor of the SCF complex and a limiting factor for this interaction as well since in the absence of exogenously expressed FBXW11, there was substantially lower amounts of NSs and no PKR detected in the CUL1-bound fraction (lanes 5 and 6). However, similar levels of NSs or PKR were bound to FBXW11, even in the absence of exogenous CUL1 expression (lanes 7 and 8). Taken together, these data provide strong evidence that NSs-FBXW11 serves as the PKR receptor of the SCFFBXW11 complex as shown in the model presented in Fig 5E. Furthermore, the protein binding data, combined with siRNA knockdown results, suggests that FBXW11 serves as the limiting factor in the assembly of SCFFBXW11 complex. Numerous studies have shown that βTrCP recognizes a 6–9 linear amino acid sequence also known as degron (or destruction) sequence that starts with two acidic amino acid residues followed by a small amino acid (usually glycine) and another acid moiety in the last position. These acidic amino acids and glycine were shown to play an essential role in βTrCP-substrate interaction. We found one degron-like sequence, DDGFVE263, in the C-terminal acid-rich region of the NSs protein (Fig 6A). Using a reverse genetics system, we generated several NSs degron mutants and subsequently rescued the mutant viruses. We speculated that rMP-12 encoding NSs mutants that cannot bind to FBXW11 will be non-infectious in HeLa cells due to PKR activity (Fig 4E). We also included in this analysis the rMP-12-NSs-R173A-V5 mutant virus, which was previously described as incapable of binding to PKR [15]. As shown in Fig 6B, NSs-GE and NSs-DGE mutants in lanes 5 and 6, respectively, showed the same gene expression profile as global inactivation of CRLs (lane 3) with potent inhibition of late viral gene expression. On the other hand, the NSs-D259A mutant was similar to wildtype virus. This mutation was probably compensated by the surrounding acidic amino acid residues. Most importantly, NSs-R173A induced PKR activity, as reported previously [19], but was not as pronounced as NSs-GE or DGE mutants and did not inhibit viral gene expression. We next examined if the GE and DGE mutants indeed inhibited infection due to PKR activation by testing virus infectivity in PKR knockdown cells. A HeLa cell line that stably expresses doxycycline inducible PKR-shRNA was used to evaluate the replication of rMP-12 expressing NSs-GE or DGE mutants under normal or low PKR expression conditions. We first examined NSs gene expression by IFA at 24h PI. As shown in Fig 6C, NSs-GE or DGE mutant viruses showed severe inhibition in NSs gene expression, but in cells treated with doxycycline in which PKR expression is reduced, the NSs expression was restored. The NSs-GE and DGE mutants formed filamentous structures in the nucleus identical to the wildtype NSs (see inset). This pattern of NSs gene expression was also similar to MLN4924 treated and wildtype rMP-12 virus infected cells. In contrast the NSs-R173A mutant, formed mosaic structures in the nucleus, as previously reported [15], and did not show any difference in NSs gene expression under normal or PKR depleted conditions. These data suggested that NSs-GE or DGE mutations, unlike NSs-R173A, did not undergo a conformational change and retained most of the properties of the wildtype NSs in structure and function. We also examined infection by these mutant viruses by enumerating the percentage of cells that expressed G on the cell surface (a reflection of successful virus egress) using HCA after multiple rounds of infection (Fig 6D). As expected, rMP-12-NSs-GE or DGE expressing mutant viruses showed potent inhibition of infection under normal PKR expression levels, but their infection was completely restored under low PKR expression levels similar to the observations made with MLN4924 treatment of wildtype rMP-12 infected cells. On the other hand, the R173A mutant replicated efficiently irrespective of PKR levels. We next examined the kinetics of PKR activation to understand the NSs mutant’s role in activating PKR. As shown in Fig 6E, rMP12-NSs-GE or DGE mutants, similar to MLN4924 treated and wildtype virus infected cells, failed to promote PKR degradation and instead caused a steady increase in PKR activity based on p-PKR levels starting from 6 to 18h PI. Although the NSs-R173A mutant blocked PKR degradation, it induced low levels of PKR activation and for a short period of time (6-8h PI). The low p-PKR expression levels were probably not sufficient to severely inhibit NSs-R173A mutant replication. Furthermore, similar to earlier observations (Fig 2C) a functional NSs gene was essential for early PKR activation since induction of p-PKR activity was not observed until the very late stages of infection (10h PI) in the rMP-12ΔNSs: : Luci infected cells. Taken together, these data suggest that NSs activates PKR when PKR degradation is blocked. In addition, NSs-R173A is not as effective as the wildtype NSs gene in activating PKR. We next examined if NSs-GE or NSs-DGE mutant viruses indeed blocked NSs binding to FBXW11. Only under very low conditions of PKR expression did the NSs-GE or NSs-DGE mutants express virus abundantly (Fig 6C). Therefore, HeLa cells stably expressing doxycycline inducible PKR shRNA were used for co-IP assay. Lysates of 293T cells that were transfected with HA-FBXW11 and myc-CUL1 expression plasmids were combined with the lysates of HeLa cells infected with wildtype rMP-12 or rMP-12 mutant viruses expressing NSs -R173A, -D259A, -GE, or -DGE mutants that were either mock treated or treated with doxycycline for 48h prior to start of infection. Cell lysates were immunoprecipated with the V5 antibody to detect NSs binding to FBXW11 and CUL1 (Fig 6F). The NSs mutants, NSs-GE and NSs-DGE expressed abundantly only in HeLa cells that had low PKR expression and did not bind to FBXW11 or CUL1 protein, unlike the wildtype or NSs-D259A or NSs-R173A mutant proteins. Furthermore, NSs-R173A did not interact with PKR similar to previously reported observations [19], while NSs-DGE or NSS-GE mutants bound PKR to the same extent as wildtype NSs. Overall, these data suggest that “DDGFVE” is the degron sequence that regulates the NSs-FBXW11 interaction. NSs-GE and DGE mutants that bind to PKR, but cannot bind to FBXW11, failed to assemble SCFFBXW11 to promote PKR degradation. Meanwhile, other unknown NSs activities in the cell (details in Discussion) stimulated PKR activity that induced phosphorylation of eIF2-α, leading to global translational suppression including inhibition of viral protein synthesis. PKR degradation is one of the numerous strategies that viruses such as RVFV and poliovirus have evolved to counter the PKR antiviral activities, although the specific mechanisms have not been clearly defined previously [13,14,24]. In this study, we show that NSs of RVFV regulates PKR degradation by recruiting PKR to the SCFFBXW11 complex (see model in Fig 7). In spite of our best efforts, we could not demonstrate ubiquitinated PKR. This could be due to a combination of low endogenous expression levels of PKR and high proteasome activity. Early degradation of PKR by NSs was absolutely essential during RVFV infection. Failure to do so resulted in PKR activation and impaired late viral protein production. MLN4924 that blocked PKR degradation by inactivating SCF/CRL activity or rMP-12 NSs-GE and NSs-DGE mutants that cannot assemble SCFFBXW11-NSs E3 ligase, induced robust PKR activation and potent inhibition of late viral protein synthesis (model in Fig 7). These inhibitory effects of PKR could be complemented by loss of PKR expression. This is similar to the antagonism of PKR activity by the E3L protein of poxviruses [25]. Under normal infection conditions, E3L blocks PKR access to double stranded (ds) -RNA, thereby preventing PKR activation. A E3L deletion mutant vaccinia virus (ΔE3L) shows potent activation of PKR activity and inhibition of late viral protein production. Loss of PKR expression rescued infection by an E3L deficient virus. However, unlike poxviruses, NSs and not dsRNA, is the trigger for early PKR activation during RVFV infection. We show that NSs has dual antagonistic functions. NSs activates PKR but only under conditions when NSs mediated PKR degradation is blocked. An NSs deficient virus (rMP-12 ΔNSs: : Luci) that cannot degrade PKR did not induce early PKR activation (Figs 2B, 2C and 6E). This lack of early PKR activation was not due to a delay in virus replication kinetics, since no differences in the viral gene expression kinetics were observed with the NSs deficient virus when compared to a wildtype virus during the first cycle of replication (Fig 2C). This also excludes dsRNA as a possible trigger since viral replication kinetics during the first cycle of infection was not different between NSs deficient and wildtype viruses. Furthermore, unlike positive-stranded RNA or DNA viruses, dsRNA accumulation was not detected in negative-stranded RNA virus infected cells [26]. However during late stages of infection, an NSs deficient virus activated PKR, mostly due to induction of the interferon pathway, whose effects are felt during the second cycle of replication. PKR activation during late stages of infection (10-12h PI) had little impact on the first cycle of infection because by then a majority of the viral protein is made (Fig 2C). Lastly, transient transfection of NSs expression was sufficient to activate PKR upon MLN4924 treatment. Collectively, these data suggest that NSs is both essential and sufficient to activate PKR when NSs mediated PKR degradation is selectively blocked. But how does NSs activate PKR? It is possible that the cellular stress responses triggered by NSs could activate PKR. Many different types of cellular stress, including oxidative stress and DNA damage response induced by NSs could potentially activate PKR [27,28]. Furthermore, NSs suppression of host transcription by RNA pol I and II could activate PKR. Indeed actinomycin D (small molecule inhibitor of transcription) treatment of NSs-deleted RVFV mutant infected cells can activate PKR; although, the exact mechanism has not been defined [13]. Furthermore, direct binding of NSs to PKR may activate PKR; although, the fact that an NSs-R173A mutant that did not bind to PKR still activated PKR, albeit less efficiently, does not favor this hypothesis (Fig 6). Ongoing studies are focused on addressing the mechanism of NSs-mediated PKR activation. Another interesting and important observation made in this study is the dominant role of FBXW11, over its homologous counterpart BTRC, in the regulation of PKR degradation. BTRC and FBXW11 are two homologues of the βTrCP gene that are biochemically indistinguishable and play a redundant role in substrate degradation [29]. At least 35 substrates of βTrCP, with well-described roles in controlling cell cycle and signal transduction pathway, have been described, including CDC25, which is a positive regulator of the cell cycle, β-catenin, and IκB (an inhibitor of NFκB) [23,30]. Despite their important roles in regulating a wide range of cellular functions, the BTRC gene knockout mice (BTRC-/-) did not affect viability or show any gross tissue abnormalities other than impaired spermatogenesis in male mice [31,32]. BTRC-/- MEFs did not accumulate any of the well-known cellular targets of βTrCP. Additionally, siRNA knockdown of FBXW11 by greater than 77% in BTRC-/- MEFs was required to accumulate βTrCP substrates. Thus these data suggest that BTRC and FBXW11have a redundant role in vivo. This property makes FBXW11 an attractive therapeutic target, since knockout of FBXW11 homologue alone may prevent RVFV infection without negatively affecting other immune responses regulated by βTrCP in mice. Recently, the development of floxed FBXW11flox/flox has been reported and future studies with these mice will help to answer these questions [33]. Why is FBXW11 more effective than BTRC in NSs mediated PKR degradation? The dominant role played by FBXW11 in the RVFV infection model is not due to differences in the binding properties between NSs and βTrCP homologues, since both proteins (BTRC and FBXW11) bind with similar affinities to NSs via their WD40 domain (Fig 4F). The lack of redundancy could be explained by differences in the cellular localization of the two homologues and the substrate being targeted. A recent study showed that the more abundant splice variants of BTRC and FBXW11 were expressed in nuclear and cytoplasmic compartments, respectively [34]. PKR is predominantly expressed in the cytoplasm; whereas, NSs expression is observed both in the cytoplasm and nucleus during early stages of infection. NSs in the cellular compartment may recruit FBXW11 targeting PKR proteasomal degradation. NSs in the nuclear compartment may not have access to either PKR or FBXW11. A recent study shows that NSs binds to the p62 subunit of the TFIIH complex via the ΩXaV motif (where Ω = Trp or Phe; X = any amino acid; a = Asp or Glu; and V = Val) present in NSs [35]. The ΩXaV motif (FVEV264) partially overlaps with the degron sequence (DDGFVE263) of NSs. As a result, p62, which is a nuclear protein, could compete with βTrCP to bind to the same site of NSs in the nucleus making it inaccessible to βTrCP. Furthermore, NSs was shown to promote p62 degradation during early stages of infection by assembling the SCFFBXO3 through NSs-FBXO3 interaction [36]. Therefore it is possible that the cytoplasmic fraction of NSs assembled SCFFBXW11; whereas, the nuclear fractions of NSs assembled SCFFBXO3 to target PKR and p62 degradation, respectively. Future work will test these speculations. In summary, we have shown that NSs assembles CRL to target degradation of host factors whose loss of expression is critical to overcome innate immune responses during RVFV infection. Future studies using an animal model of RVF will be required to address the in vivo relevance of these observations. HeLa, 293T, HepG2, mouse embryonic fibroblasts (MEFs) and Vero E6 cells were obtained from the American Type Culture Collection (ATCC) and were maintained under humidified conditions at 37°C, 5% CO2 in Dulbecco' s Modified Eagle Medium supplemented with 10% fetal bovine serum (FBS, Life Technologies). HSAECs (Lonza, USA) were maintained in Ham’s F12 medium supplemented with nonessential amino acids, pyruvate, β-mercaptoethanol and 10% fetal calf serum (FCS). RVFV strains ZH548/MP-12 (MP-12) and ZH501 were obtained from The Salk Institute’s Government Services Division and from Dr. Michael Turell (USAMRIID), respectively. MP-12 was derived from a virulent strain of RVFV (ZH548) and is attenuated in its virulence due to several nucleotide mutations in its genome, but encodes a functional NSs gene [13,14] and can be handled safely under biosafety level (BSL) -2 laboratory conditions. ZH501 is the wild-type strain of RVFV and is fully virulent, requiring use under BSL-3 laboratory conditions. Clone 13 virus was a gift from Dr. Friedemann Weber (Philipps-University Marburg, Germany). Clone 13 was isolated from the 74HB49 strain of RVFV from a human case and encodes a non-functional NSs gene due to a large internal frame deletion that removed 69% of the ORF. VEEV (1CSH3), LASV (Josiah) and MARV (Ci67) were obtained from the USAMRIID collection. All viruses were propagated in Vero E6 cells. Virus-containing supernatants were clarified by centrifugation at 12,000 x g for 30 min prior to storage at -80°C. All virus stock titers were determined by plaque assay on Vero E6 cells as previously described [37]. RVFV ZH501 and VEEV infections were performed in a BSL-3 lab, and MP-12 infections were performed in BSL-2 lab. Work with infectious filoviruses or arenaviruses were performed in BSL-4 labs. All infection assays that required enumeration of infected cells based on immunofluorescence assay (IFA) of viral antigen expressing cells were optimized as described previously [16,17]. The infection conditions were optimized such that at a minimal MOI, greater than 50% of cells expressed viral antigen after multiple rounds of infection. This is to ensure that irrespective of the stage of the viral life cycle targeted by the small molecule or siRNA, the infection rates after multiple rounds of infections are decreased. For all the assays described here, HeLa cells were permissive to infection and were seeded at 20,000 cells per well a day prior to infection in a 96-well plate (black well, clear and flat bottom from Greiner) and infected with virus at the required MOIs in a total volume of 100 μl. Inocula were removed 1h post virus incubation, unless stated otherwise, washed one time with 1x PBS, and replaced with an equal amount of fresh medium. Infection was allowed to proceed for the specified duration of time as indicated in the figure legends of each experiment, followed by fixation in 10% neutral buffered Formalin (Sigma) for one day for BSL-3/4 viruses or 15 min for BSL-2 viruses. Cells were then subjected to immunofluorescence assay (IFA) as described below. IFA of viral antigen expression was carried out as described previously [17]. If permeabilization was required then Formalin-fixed cells were treated with 0. 1% (v/v) Triton-X 100 for 15 min at room temperature prior to blocking. Otherwise Formalin fixed cells were directly blocked and incubated with primary antibody (at 1: 1000 dilution) prepared in blocking buffer containing 3% BSA/PBS for 1h at room temperature (RT). Then cells were washed three times in PBS followed by incubation with the corresponding fluorescent secondary antibody for 1h at RT. Cells were washed in PBS to remove excess antibody. During the last wash, cell nuclei and cytoplasm labelling reagents, which include Hoechst 33342 (Life Technologies) and HCS Deep Red (Life Technologies), respectively, were added to the wash solution at a 1: 10,000 dilution. The following antibodies, 4D4, R3-1D8-1-1a, 1A4A-1,6D8-1, MBG II 9G4-1, L52-161-6 (from USAMRIID hybridoma bank) were used to detect RVFV G, RVFV N, VEE envelope 2 (E2) protein, MARV GP and Lassa virus GP, respectively. For measuring RVFV infections (MOI = 1,24 h) that involved multiple rounds of infection, 4D4 antibody that detects RVFV Gn was used with MP-12 virus. Since the 4D4 antibody does not detect Gn expression of ZH501, N expression was used for measuring infection by ZH501 and clone 13 specifically for most of the study except for Figs 1B and S1A where Gc antibody (5G3) was used. Since the Gc antibody was only available in limited amounts, it was used only for these figures to show that under the conditions of infections used N or G expression did not significantly affect the interpretation of the data. G was used to represent Gn expression throughout the manuscript except in Figs 1B and S1A where Gc was used. As described in the introduction, Gn and Gc form heteromers and are produced by proteolytic processing of the precursor GPC. Therefore Gn or Gc detection for these two figures did not change the interpretation of the data. For NSs-V5 tagged protein and Flag-FBXW11, anti-V5 and anti-Flag (Sigma) antibodies were used. The following secondary antibodies were used: Alexa 488-conjugated goat anti-mouse secondary antibody, Alexa 568-conjugated goat anti-rabbit antibody or Alexa 647-conjugated goat anti-rabbit (Life Technologies) at 1: 1000 dilutions. Confocal images were collected using a Leica TCS-SP5 confocal/multiphoton microscope with a 40x oil objective. High-content quantitative imaging data were acquired and analyzed on an Opera confocal reader (model 3842 [Quadruple Excitation High Sensitivity] or model 5025; PerkinElmer) at two exposures using a 10x air objective. High-content image (HCI) -based analysis (HCA) was accomplished within the Opera or Columbus environment using standard Acapella scripts as described previously [17,38]. Briefly, images of Hoechst stained nuclei were used to draw nuclear boundaries and to count cells, while images of Cell Mask Deep Red stained cells were used for drawing cell boundaries. Images of viral antigen expression were used for enumerating percentage of infected cells. A 10-point dose-response curve (1/3 fold serial dilution from 30 μM) assay was used for dose response curve analyses. Each concentration of the hit compound was tested in triplicate in a 96-well plate format. Cells were treated with the indicated concentrations of MLN4924 for 1h prior to incubating cells with MP-12 or ZH501 at MOI = 1. One hour post virus incubation, cells were washed and replaced with compound containing media. Twenty-four hours post-infection, cells were fixed and subjected to IFA followed by HCA to evaluate the percentage of RVFV infection by enumerating the viral antigen expressing cells. Relative infection inhibition was determined by the ratio of the percentage of infected cells in compound treated cells with mock treated (0. 5% DMSO) and RVFV infected cells. Data were analyzed using the non-linear regression formula: log (inhibitor) vs. response–variable slope (4 parameters) in GraphPad Prism 6. The IC50, defined as the effective concentration resulting in a 50% inhibition of infection, was used to evaluate compound activity. The relative cell number was determined by normalizing the cell number of compound treated + virus-infected cells with mock treated (0. 5% DMSO) + virus-infected cells. Cells infected with rMP-12 (MOI = 10 for 8 h) encoding wildtype NSs-V5 or its mutants served as sources of NSs protein. These lysates were combined with cell lysates expressing the protein of interest. Transient overexpression by plasmid DNA transfection was carried out using Lipofectamine 2000 (Life Technologies). For co-IP assay, cells in a 12 well dish were lysed in IP lysis buffer (Life Technologies, cat # 87787) supplemented with complete protease inhibitor cocktail (Roche) at 4°C for 20 min and clarified at 14,000 x g for 15 min at 4°C. Ten percent of the cleared lysate was kept aside as a lysate control, while the remaining lysate was incubated with 2 μg of the indicated antibody and 25 μl of washed protein A/G plus agarose beads and allowed to rotate mildly for 3h at 4°C. The beads were then washed three times in lysis buffer and one last wash with high salt PBS containing 300 mM NaCl. The bound proteins were eluted by boiling in Laemmli sample buffer, and separated by SDS-PAGE, followed by transfer to a polyvinylidene (PVDF) membrane. The blots were incubated for 1h at RT or overnight at 4°C with the indicated primary antibodies. After three washes, the blots were incubated with the appropriate alkaline phosphatase (AP) -conjugated secondary antibodies (GE Healthcare) according to the manufacturer' s recommendations. The blots were washed and developed by a Western blotting detection system (GE Healthcare). For direct protein analyses, cells in 12-well dishes were lysed in 100 μl of 1X Laemmli sample loading buffer and boiled for 5 min at 95°C. Equal amounts of samples were used for Western blot analysis as described above. The following antibodies were used for Western blot analysis or immunoprecipitation (when required): T446 phosphorylated PKR (Abcam), PKR (BD Transduction), S52 phosphorylated EIF2α (Cell Signaling), EIF2α (Cell Signaling), TFIIH-p62 subunit (Santa Cruz), NEDD8 (Cell Signaling), CUL1 (Santa Cruz), SKP1 (Cell Signaling) and alkaline phosphatase (AP) conjugated V5 antibody (Life Technologies). The antibodies to HA, V5, Flag and Myc were from Sigma. HeLa cells were transduced with lentivirus expressing doxycycline (Sigma) inducible human PKR shRNA (cat # V2THS_170555, GE Healthcare) encoding the antisense sequence 5’-TTTATCTCTGATGTATCTG-3’ and selected in Puromycin (1 μg/ml, from Sigma) containing media. The HeLa stable cell line can be induced with doxycyline (1 μg/ml) to express PKR shRNA. The plasmid DNA for recombinant MP-12 (rMP-12) virus rescue including the S segment encoding plasmid DNA in which the NSs ORF was replaced by a luciferase gene was a kind gift from Dr. Shinji Makino (Univ. of Texas Medical Branch, US). Mutations in the NSs gene, including the addition of a V5 tag to the gene were introduced using Q5 Site-Directed Mutagenesis Kit (New England Biolabs) by following the manufacturer’s recommendations. The virus was rescued as described previously [20]. Briefly, subconfluent BSR-T7/5 cells were co-transfected with an S-genome RNA expression plasmid, such as pProT7-S (+) or its variants encoding NSs mutations and a mixture of pPro-T7-M (+), pPro-T7-L (+), pT7-IRES-vN, pCAGGS-vG, and pT7-IRES-vL using TransIT-LT1 (Mirus Bio Corporation). The culture medium was replaced with fresh medium 24h later. At 5 days post-transfection, the culture supernatants were collected, clarified and inoculated onto VeroE6 cells. The supernatant of infected VeroE6 cells at 2–3 days PI were aliquoted and stored at -80°C. The 70 siRNAs targeting human F-box genes were cherry picked from the Dharmacon ON-TARGETplus SMARTpool siRNA Library—Human Ubiquitin Conjugation Subset 2 or 3 (list in the supplementary data S4 Fig). The smart pool contained 4 different siRNAs targeting individual gene and were screened against RVFV ZH501 in triplicate at a concentration of 15 nM in a 96-well plate format using optimized assay conditions as described previously [9]. Briefly, HeLa cells (12,500 cells per well) were reverse transfected using HiPerfect (Qiagen) transfection reagent (0. 6 μl per well) in a total volume of 100 μl per well in 96-well plates following the manufacturer’s instructions. As a positive control, 3 wells of each plate were transfected with non-targeting siRNAs. Mock infected cells were treated as negative controls. On the following day cells were washed and replaced with fresh media. Forty-eight hours post transfection, cells were subjected to RVFV infection (MOI = 1,24h). The percentage of infected cells was determined using HCA analysis of N expressing cells. Data were represented as the average ±SD. pcDNA3-FLAG-UBC12 (C111S) was a gift from Dr. Tetsu Kamitani [18]. The cullin expression plasmids, pcDNA3-myc3-CUL1, -2, -3, -4A, - 4B, -5, -7 or HA2-PARC/CUL9 were gifts from Dr. Yue Xiong [39–41] (Addgene # 19896,19892,19893,19951,19922,20695,19895 and 20937 respectively). The cullin mutant expression plasmids, pcDNA3-HA-CUL1 ΔC22, L756A/I757A, K720A, Y42A/M43A, Δ610–615 or ΔN53 (19942,19938,19939,19940,19941 and 19950 respectively) were gifts from Dr. Yue Xiong [42,43]. The dominant negative (Dn) cullin expression plasmids pcDNA3-CUL1-3,4A, 4B and CUL5 were a gift from Dr. Wade Harper [21] (Addgene plasmid # 15818,15819,15820,15821 and 15822 respectively). pcDNA3-HA-FBXW11 was a gift from Dr. Yi Sun [44]. pcDNA3- Flag-BTRC/FBXW11-Wt, ΔN, ΔF, ΔD, ΔNF and ΔWD40 constructs were a gift from Dr. Tomoki Chiba [45]. The binding properties of the CUL1 mutant proteins with SKP1 or RBX1 were as follows: the wildtype CUL1 contains 776 amino acid (aa) residues of which the N terminal 249 aa regulate its binding to SKP1, while the remaining 527 aa recruit RBX1 [21,42,43]. The dominant negative Dn-CUL1 (truncated protein expressing the first 452 aa) or CUL1Δ610–615 cannot bind to RBX1, but retains interaction with SKP1. CUL1-ΔN53 or Y42A/M43A mutants, cannot bind to SKP1, but retain their binding to bind to RBX1. Lastly, the CUL1-ΔC22, L756A/I757A, or the NEDDylation defective mutant K720A can bind to both SKP1 and RBX1. NSs-V5 gene expression plasmid was generated by sub-cloning the PCR amplified product from the pProT7-S plasmid expressing NSs-V5 gene into pcDNA3. 1 vector. The siRNAs targeting BTRC (cat# S17109) was from Ambion while FBXW11 (cat # J-003490-07) CUL1 (cat # J-004086-08) and SKP1 (cat # J-003323-15-0002) were from GE Healthcare. pcDNA3-FLAG-UBC12 (C111S) was a gift from Dr. Tetsu Kamitani (Wada, H. , Yeh, E. T. and Kamitani, T. , 2008, J Biol Chem, 275,17008–15). Dominant negative (Dn) Cullin expression plasmids pcDNA3-CUL1-3,4A, 4B and CUL5 were a gift from Dr. Wade Harper (Jin, J. , Ang, X. L. , Shirogane, T. and Wade Harper, J. 2005. Methods Enzymol, 399,287–309.) (Addgene plasmid # 15818,15819,15820,15821 and 15822 respectively). pcDNA3-HA-FBXW11 was a gift from Dr. Yi Sun (Zhao, Y. , Xiong, X. and Sun, Y. 2011. Mol Cell, 44,304–16). The siRNA’s-1, -2 and -3 targeting FBXW11 were from Ambion with cat# S23485, S23486 and S23487 respectively. Relative changes in RVFV RNA levels or mRNA levels of host genes were determined by qRT-PCR on an ABI Prism 7900HT sequence detection system using RNA UltraSense one-step kit and TaqMan probes (Life Technologies). Total RNA was extracted using the RNeasy Plus minikit (Qiagen). Fifty nanograms of RNA was used per qRT-PCR reaction. The target gene mRNA or viral RNA expression levels were normalized to the PPIB housekeeping gene. Relative expression levels were determined using the comparative threshold cycle (CT) method [46]. The taqman probe sequences used for RVFV RNA have been described previously [47]. The student Paired t test was used to determine P values. **** indicated P<0. 0001.

Rift Valley fever (RVF) is a severe disease caused by infection with the Rift Valley fever virus (RVFV) that affects humans and livestock and occurs in large epidemics. Currently there are no FDA-approved drugs or vaccines to treat RVF. Many viruses have evolved unique strategies to overcome host immune responses in order to establish infection. One protein of RVFV called NSs is responsible for over-powering cellular antiviral defenses. NSs is known to degrade double-stranded (ds) RNA-dependent protein kinase (PKR), but neither the mechanism nor the functional significance of this activity has been fully understood. In this study we show that NSs promotes PKR degradation by recruiting PKR to the E3 ligase complex called SCF (SKP1-CUL1-F-box) FBXW11. A short stretch of six amino acids called the degron sequence in NSs regulates the NSs- FBXW11 interaction and is required for the assembly of the SCFFBXW11 complex. We further show that disruption of the SCFFBXW11-NSs complex, with either a small molecule or with NSs degron viral mutants, can block PKR degradation. Surprisingly, when NSs mediated PKR degradation was blocked, NSs was essential and sufficient to activate PKR, causing potent inhibition of RVFV infection by suppressing viral protein synthesis. Therefore early PKR activation induced by inactivation of the SCFFBXW11 is sufficient to induce potent inhibition of RVFV infection. These findings may provide new molecular targets for therapeutic intervention of this important disease.

Abstract Introduction Results Discussion Materials and Methods

medicine and health sciences rift valley fever virus pathology and laboratory medicine hela cells gene regulation pathogens enzymes biological cultures vector-borne diseases microbiology enzymology viruses rna viruses cell cultures microbial genetics ligases bunyaviruses research and analysis methods small interfering rnas infectious diseases proteins medical microbiology gene expression microbial pathogens cell lines viral replication viral genetics biochemistry rna nucleic acids virology viral pathogens genetics viral gene expression biology and life sciences cultured tumor cells non-coding rna organisms

2016

Protein Kinase R Degradation Is Essential for Rift Valley Fever Virus Infection and Is Regulated by SKP1-CUL1-F-box (SCF)FBXW11-NSs E3 Ligase

Transcription factor binding to DNA in vivo causes the recruitment of chromatin modifiers that can cause changes in chromatin structure, including the modification of histone tails. We previously showed that estrogen receptor (ER) target gene activation is facilitated by peptidylarginine deiminase 2 (PAD2) -catalyzed histone H3R26 deimination (H3R26Cit). Here we report that the genomic distributions of ER and H3R26Cit in breast cancer cells are strikingly coincident, linearly correlated, and observed as early as 2 minutes following estradiol treatment. The H3R26Cit profile is unlike that of previously described histone modifications and is characterized by sharp, narrow peaks. Paired-end MNase ChIP-seq indicates that the charge-neutral H3R26Cit modification facilitates ER binding to DNA by altering the fine structure of the nucleosome. Clinically, we find that PAD2 and H3R26Cit levels correlate with ER expression in breast tumors and that high PAD2 expression is associated with increased survival in ER+ breast cancer patients. These findings provide insight into how transcription factors gain access to nucleosomal DNA and implicate PAD2 as a novel therapeutic target for ER+ breast cancer. The nucleosome represents the fundamental unit of chromatin and consists of 147 bp of DNA wrapped ∼1. 7 times around the histone octamer core particle [1]. The N-terminal tails of histones are disordered and reside outside of core nucleosome/DNA structure but they harbor amino acid residues that can be post-translationally modified to regulate many facets of transcription. These modifications can influence transcription factor access to nucleosomal DNA by modulating electrostatic interactions between histones and DNA. However, the means by which different transcription factors (TF) gain access to their DNA elements in the context of chromatin remain to be fully elucidated [2]. Recent studies of nuclear receptors have shown that androgen receptor (AR) [3], glucocorticoid receptor (GR) [4], [5], and progesterone receptor (PR) [6] access their respective DNA elements through extensive reorganization of nucleosomes using chromatin remodeling enzymes that cause concomitant increases in DNase accessibility. While estrogen receptor (ER) also interacts with nucleosome remodelers to maintain the accessible chromatin state [7]–[9], we find that concomitant increases in accessibility at ER binding sites are less prevalent than with other nuclear receptors. We have previously reported that PAD enzymes catalyze the conversion of protein arginine residues to neutrally charged citrulline in a process called deimination or, alternatively, citrullination [10], [11]. More specifically, we found that PAD2-mediated deimination of histone H3 arginine 26 (H3R26Cit) is important for estradiol (E2) -mediated activation of ER-target genes [12]. Here, we examine this E2-induced deamination at high temporal and spatial resolution and test the hypothesis that PAD2 facilitates ER/DNA binding by neutralizing the H3R26 charge of local nucleosomes via deimination, thereby weakening histone tail-DNA interactions and destabilizing DNA interactions with the core nucleosome particle. We previously demonstrated that PAD2 binds to chromatin and regulates gene expression in MCF-7 cells via histone deimination [13]. Additionally, we have shown that the H3R26Cit modification is important for E2 mediated activation ER-target genes [12]. To directly test whether H3R26 deimination occurred at ER binding sites, we examined the genome-wide relationship of ER binding and H3R26 deimination by performing ChIP-seq of ER and H3R26Cit in MCF-7 cells before and after 40 minutes of E2 treatment. We identified 12,301 ER peaks and 28,495 H3R26Cit peaks and found that both sets of peaks largely overlap with previously identified ER peaks in MCF-7 cells [14], [15] (Figure S1). We found only 9 H3R26Cit peaks prior to E2 stimulation, all of which increase after E2 stimulation, suggesting that E2-induced ER binding is directly or indirectly responsible for 99. 97% of the H3R26Cit peaks at any time point. Further inspection reveals that the 9 H3R26Cit peaks prior to E2 treatment are a result of ER binding in the absence of ligand (Figure S2). Ninety-five percent of ER binding sites overlap with H3R26Cit peaks and the intensity of H3R26Cit linearly correlates with ER intensity at ER binding sites (Figure 1A, B). Reciprocally, 47% of H3R26Cit peaks overlap with ER peaks (Figure 1D); and, ER intensity correlates linearly with H3R26Cit intensity at H3R26Cit peaks (Figure 1E). Following E2 treatment, the composite signals for H3R26Cit, H3K4me1, and H3K4me2 are highest at the ER ChIP-seq summit (Figure 1C, F and Figure S3) [3]. In contrast, both AR and GR cause a depletion of modified nucleosomes at the sites of nuclear receptor binding after ligand treatment [3] (Figure S4). These observations support the hypothesis that ER binding occurs at nucleosomal EREs [3]. We next found that the activation of other nuclear receptors, including AR, PR, and GR by their respective ligands does not induce global H3R26 deimination (Figure 2), despite the high expression levels of each receptor in MCF-7 cells (Figure S5). This finding raises the possibility that H3R26 deimination may be unique to ER-activation in MCF-7 cells. In order to further highlight the unique peak structure of the H3R26Cit mark, we next compared the H3R26Cit peak structure with the genomic profiles of other histone marks that are associated with active (H3K4me3, H3K4me2, H3K9ac, H3K36me3, and H3K27ac) and repressed (H3K27me3 and H3K9me3) chromatin [16]. A UCSC genome browser screenshot of a representative locus shows that the H3R26Cit profile is unlike any previously characterized histone modification profile in MCF-7 cells, and looks indistinguishable from that of sequence-specific TFs, such as ER (Figure 3A). We quantified the discrete nature of the H3R26Cit peaks and found that 55% of the ER peaks drop to 25% maximum intensity within 200 bp of the summit, while 67% of H3R26Cit peaks are less than 25% maximum by 200 bp. In contrast, the H3K4me2 and H3K4me3 modifications (Figure 3B) have the next most narrow peak structure with only 33% and 18% of the peaks at 25% max within 200 bp of the peak summit, respectively (Figure 3C). Next, we compared the kinetics of ER binding and H3R26 deimination following E2 treatment to further elucidate the timing of H3R26 deimination relative to ER binding. The ER intensity and H3R26Cit intensities were compared at each H3R26Cit peak for each time point (Figure S6). The Pearson correlation for each E2-induced time point was between 0. 9 and 0. 97, indicating that ER binding intensity is quantitatively correlated with H3R26Cit at H3R26Cit peaks. In fact, H3R26Cit peaks that did not overlap with ER peaks correlate linearly as well, suggesting that ER is binding all H3R26Cit sites with the raw ER signal being below our peak calling threshold. These correlations, combined with other analyses (Text S1) lead us to conclude that an H3R26Cit peak is indicative of a site that is bound by ER after E2 stimulation. Therefore, unless noted, we use this set of peaks for the remainder of the analyses (Dataset S1 and Materials and Methods). While we cannot directly compare ChIP signals between H3R26Cit and ER, we are able to compare each antibody' s ChIP signal as a function of time. We clustered the H3R26Cit peaks by their H3R26Cit signal; using the same order, we plotted the ER signals on an adjacent heat map (Figure 4A). The H3R26Cit derived clusters are clearly delineated using the kinetic ER data, indicating that the classification of ER binding and H3R26Cit cluster intensities are not appreciably different. However, we found that the average H3R26Cit signal across all peaks is highest at five minutes after E2 treatment and gradually decreases through the remainder of the time course (Figure 4B). In contrast, the average ER signal increases gradually until 40 minutes, and then decreases by 160 minutes after treatment (Figure 4B). We also confirmed that there are more H3R26Cit peaks that are highest at 5 minutes than at any other time point (Figure 4C). Over 80% of the ER and 96% of the H3R26Cit signals at the H3R26Cit sites increase at the 2 minute time point, thus making it difficult to definitively determine the temporal sequence of events leading to ER binding. However, the composite profiles and histograms of ER and H3R26Cit signals indicate that H3R26Cit signal reaches maximum intensity at an earlier time point than the ER signal. This finding supports the hypothesis that H3R26Cit plays an important role for the establishment or early maintenance of ER binding. To further test this prediction, we performed ChIP-qPCR at a number of ER binding sites to determine whether stable depletion of PAD2 (the enzyme that deiminates H3R26) resulted in reduced ER binding. Indeed, results show that PAD2 depletion significantly reduces the efficiency of H3R26 deimination and ER binding at these target loci (Figure 5 and Figure S7). Thus, PAD2-mediated H3R26 deimination appears to be required for efficient ER binding to EREs within the context of chromatin. To investigate the mechanism by which the H3R26Cit modification facilitates the binding of ER to DNA, we first tested whether hormone treatment promoted gross remodeling events at ER binding sites by examining changes in DNase hypersensitivity following E2 treatment using ENCODE data [17]. To increase the power to detect differential DNase sensitivity we implemented independent pre-filtering of the data, while using the false discovery rate to control for multiple testing [18], [19] (Figure S8 and Figure S9). In contrast to GR, where 39% of GR binding sites show an increase in DNase sensitivity following dexamethasone treatment (Figure S10B) [20], a much smaller fraction of ER binding sites were significantly increased after E2 treatment (Figure S10A). We also find that over 70% of ER-binding sites are DNase hypersensitive prior to E2 treatment. This corroborates a previous study of ER in ECC-1 and T-47D cells, where 72% and 59% of ER binding sites are hypersensitive prior to E2-treatment [21]; this high degree of pre-existing accessibility may partially account for a lower frequency of ER-induced hypersensitive transitions relative to other nuclear receptors. However, the H3R26Cit modification itself does not appear to facilitate ER-binding by promoting the larger scale changes in nucleosomal organization detected as DNase hypersensitive sites. This conclusion is further supported by our analysis of existing datasets which found that E2 treatment does not result in strong depletion of H3-modified nucleosomes at ER-binding sites [3], [16] (Figure S3). These findings support our principle hypothesis, which is that the H3R26Cit modification enhances the binding of ER to nucleosomal DNA by directly altering nucleosome structure. Previous studies have documented a role for the histone H3 tail in DNA-core particle interaction using MNase digestion [22], [23]. Therefore, we directly tested our hypothesis by performing H3R26Cit ChIP of MNase digested chromatin followed by paired-end sequencing to determine if the H3R26Cit modification altered the MNase-protected region [24]. Results showed that the genomic distribution of protected DNA in the input sample was centered on 149 bp, which is 2 bp larger than would be predicted based on the nucleosome structure [1] (Figure 6A). The MNase protection profile at H3R26Cit peaks prior to E2 exhibited a protection profile with a mode of 155 bp. Importantly, we found that the average protected nucleosome region at H3R26Cit peaks shifts from 155 bp before E2 treatment, to 125 bp after E2-induced deimination (Figure 6B). As a control to test whether other histone modifications are associated with remodeled nucleosomes, we next carried out a control MNase ChIP-seq experiment using an H3K27ac antibody and found that the protected nucleosome region at H3K27ac peaks is 155 bp (Figure 6A) and that this distribution is unchanged upon hyper-acetylation by HDAC inhibition (Figure S11). Note that the protected nucleosome size at ER peaks prior to E2 is also 155 bp because the pre-existing chromatin is largely H3K27-acetylated (Figure S12). These results suggest that PAD2 mediated H3R26 deimination facilitates ER binding by directly altering nucleosomal structure. Taken together, our findings provide novel insight into the mechanisms by which ER is able to efficiently bind to its DNA element in the context of a chromatin. Our current working model posits that, prior to estrogen treatment, histone H3 arginine 26 (H3R26) interferes with ER-nucleosomal ERE binding by electrostatically interacting with DNA at the ERE. Following E2, ER directly or indirectly recruits PAD2 to nucleosomal EREs where PAD2 then deiminates H3R26, thus neutralizing this residue. Following deamination, the H3 tail then no longer occludes this site allowing for more stable ER-nucleosomal ERE binding (Figure 6C). Given that the striking correlation between E2-induced ER binding and H3R26 deimination in MCF-7 cells (a model for ER+ breast cancer), we decided to test whether PAD2 expression and H3R26 deimination may correlate with ER expression in breast tumors. We quantified the coincidence of ER, PAD2, and H3R26Cit staining in serial tumor sections from 21 breast cancer patients. Representative images (Figure 7A) highlight our observation that tumor sections with strong ER signal frequently stain positive for PAD2 and H3R26Cit (e. g. Patients 1 and 2). In contrast, tumors that stain negative for ER rarely stain for PAD2 and H3R26Cit (e. g. Patients 3 and 4). We quantified the staining levels of each section and found that the degree of ER, H3R26Cit, and PAD2 staining between samples is highly correlated (Figure 7B–D). To test whether the observed correlations may have clinical significance, we examined the relationship between PAD2 expression and cancer relapse or overall survival in Luminal A subtype breast cancer patients [25], [26]. We found that PAD2 expression is significantly associated with relapse free survival time [27] (p-value = 0. 0028) and overall survival [28] (p-value = 0. 0053) (Figure 7E, F). This association is also found by an independent, provisional measure of RNA expression from The Cancer Genome Atlas (Figure S13). These data suggest that PAD2, and possibly the H3R26Cit modification, can be useful as predictive biomarkers to further stratify Luminal A breast cancer patients in terms of their risk of recurrence and overall survival. Understanding the mechanisms by which transcription factors (TFs) bind to DNA is central to understanding how transcriptional regulatory networks function. We previously showed that ER target gene activation is facilitated by PAD2-catalyzed histone H3R26 deimination. Herein, we provide insight into how H3R26Cit facilitates the ER-DNA interaction, thus providing a mechanistic explanation for the role of PAD2 in E2-induced ER gene activation. Outcomes from seminal biochemical studies suggested TF access to nucleosomal DNA was hindered by the histone H3 tail, which protrudes from the core histone octamer and overlays nucleosomal DNA [22]. Additionally, these and other studies indicated that neutralization of histone H3 lysine residues via histone acetyltransferase (HAT) -mediated acetylation weakens H3 tail-DNA interactions, thus stabilizing TF-nucleosomal DNA binding. Interestingly, subsequent studies found that deletion of the first 20 amino acids of the H3 tail has no effect on nucleosome stability [29]. However, when more C-terminal regions of the H3 tail were deleted, starting at R26, defects in the wrapping of nucleosomal DNA were observed [29]. Importantly, however, in vitro studies have also found that histone tail hyperacetylation only partially reduces the affinity of histone tails for DNA, likely because the remaining arginine residues are unaffected by acetylation [23], [30]. Outcomes from our study suggest that histone arginine deimination represents a mechanism by which histone H3 arginine residues can be neutralized. Therefore, we suggest that both PAD-mediated histone deimination and HAT-mediated histone H3 tail acetylation are independent and complimentary mechanisms that facilitate TF-nucleosomal DNA binding. Interestingly, a recent study found that global chromatin decondensation occurs as a result of H1 citrullination by PAD4 upon reprogramming and in the pluripotent cell state [31]. This finding is in line with our previous work showing that PAD4-mediated histone deimination affects nucleosome compaction [32]. However, our results indicate that H3R26 deimination does not result in changes in higher order chromatin structure (Figure S10A, C and Text S1). Instead the core nucleosome particle protection size shifts from 149 bp to 125 bp upon H3R26 deimination. Together, these results support the hypothesis that PAD4 mediated histone deimination mediates nucleosome compaction, while PAD2 mediated histone deimination affects nucleosome structure by modulating the wrapping of DNA around the histone octamer. Similar to that found for other histone modifications, our findings highlight the diverse roles that histone deimination can have upon chromatin structure. In addition to our more basic findings, this study also demonstrates that PAD2 and histone H3R26 deimination are positively correlated with ER staining in breast tumor sections. We also found that high PAD2 expression is positively correlated with increased relapse-free survival and overall survival in patients with Luminal A breast cancer. Interestingly, a more detailed analysis of the IHC scoring finds that PAD2 and H3R26Cit staining levels are variable in the ER positive sections, with some of the tumors staining strongly for PAD2 and H3R26Cit (e. g. Figure 7A, Patient 1) compared to other tumors (e. g. Figure 7A, Patient 2) stained more weakly for PAD2 and H3R26Cit. Given our finding linking high PAD2 expression with increased survival in ER+ patients, these results suggest that PAD2 and H3R26Cit could help to further stratify ER+ tumors into clinically relevant subsets with respect to outcome. In sum, our results indicate that the H3R26Cit modification is unlike any previously described histone modification in that the mark is virtually absent from chromatin prior to ligand stimulation and then is observed at ER binding sites across the ER cistrome with over 95% concordance within minutes of estrogen treatment. Additionally, we demonstrate that PAD2-mediated H3R26 deimination appears to facilitate ER-DNA binding by altering the fine structure of the nucleosome. Lastly, our clinical data and analyses indicate that, following further validation, PAD2 and the H3R26Cit modification could be developed into prognostic tools for stratifying ER+ breast cancer patients with respect to survival and treatment. MCF-7 cells were maintained in DMEM supplemented with 10% calf serum. The stable PAD2-depleted MCF-7 cell line was described previously [12] and maintained in the medium containing 1 µg/ml puromycin (Sigma). Before hormones treatment, the cells were cultured for 3 days in DMEM phenol red-free medium supplemented with 10% charcoal-dextran-treated calf serum. ChIP experiments were performed as described previously [12], [33]. Estrogen-starved MCF-7 cells were subjected to E2 treatment at 100 nM for 0,2, 5,10,40 and 160 min, and followed by crosslinking with a final concentration of 2% paraformaldhyde for one minute at 37°C. Crosslinking was quenched in 125 mM glycine on ice for 5 min. Cell lysates were sonicated under conditions yielding fragments ranging from 100 bp to 200 bp. The material was clarified by centrifugation, diluted 10-fold in dilution buffer, and pre-cleared with protein A-agarose beads. The pre-cleared, chromatin-containing supernatant was used in immunoprecipitation reactions with antibodies against H3Cit26 (Abcam, ab19847, lot 135757), ERα (Santa Cruz, sc-542), non-specific rabbit IgG (Upstate 12-370) or without antibody as a control. ChIP-Western analysis confirmed that the H3R26Cit antibody does not recognize ER (Figure S14 and Text S1). The Illumina library preparation was as previously described [34]. Samples were submitted to the Cornell DNA Sequencing and Genotyping Lab and run on the Illumina Genome Analyzer II. Replicates were found to be concordant, with Pearson correlation coefficients all greater than 0. 9, for all E2-stimuated data sets where replicates were performed (Figure S15). Estrogen-starved MCF-7 cells were treated with ethanol or 100 nM E2 for 10 min. Mononucleosomes were prepared as described previously [12]. The crude chromatin was solubilized with a concentration of MNase (NEB M0247) that produced ∼80% mononucleosomes. The mononucleosomes were then immunoprecipitated using anti-H3Cit26 and anti-H3K27ac (Abcam ab4729) antibodies. Advanced Technology Center (Rockville, MD, USA) performed the Illumina library preparation and the paired-end sequencing. The primer sequences used for ChIP-qPCR were summarized in Table S1. Raw sequence reads were aligned to the hg19 genome using bowtie [35]. Replicate concordance was confirmed and replicate files were merged to call peaks using MACS [36] and a mock IP using IgG as a background data set. Broad regions of enrichment were further subdivided using the subpeaks argument in MACS [36]. Subpeaks files from all time points were combined for each antibody. Redundant peaks (those found in more than one time point) were filtered out if a 100 bp window centered on the subpeak summit was within 30 bp of the adjacent subpeak—the subpeak with the most tag counts a the summit position was retained. Raw tag count intensity for each peak coordinate was normalized for each time point and peaks with at least one time point with greater intensity than an independent no antibody ChIP were retained (Dataset S1). We generated composite profile plots by taking the average intensity for a given factor in 20 base pairs steps centered on the peak summits. We used MEME with default parameters and 100 bp of sequence information surrounding the peak summit to identify motifs de novo [37]. Despite the fact that many H3R26Cit peaks do not overlap with any ER peaks, the canonical ERE is the most significant motif found de novo at these H3R26Cit peak summits (Figure S16). DNase data was downloaded from ENCODE (wgEncodeUwDnaseMCF-7Estctrl0hAlnRep*. bam and wgEncodeUwDnaseMCF-7Est100nm1hAlnRep*. bam). We used the program DEseq [19] to statistically determine differentially DNase sensitivity at all H3R26Cit peaks using the raw number of reads at each peak before and after estrogen treatment. We applied an independent gene filter to the peaks in order to increase the power to detect differences and still control for multiple testing (Figure S8). The joint distribution of p-values and read count total remains unchanged after filtering, indicating that unadjusted p-value and total reads counts for each peak are independent variables (Figure S9). Paired-end reads were mapped to the hg19 genome using bowtie2 [35]. Concordant paired-end reads that result in fragment lengths >1 and <500 were considered in the analysis and shown in Figure 6. MCF-7 Cells grown on slides were subjected to hormones treatment at 100 nM for 45 min, including Estradiol (E2), Dihydrotestosterone (DHT), Progesterone (PRO), and Dexamethasone (Dex). Ethanol (EtOH) was used as a control. Confocal microscopy experiments were described previously [11]. In short, after hormones treatment, cells were fixed with a paraformaldehyde fixing solution (1×PBS, 0. 1% Triton X- 100,0. 2% NP-40, and 3. 7% paraformaldehyde) for 10 min at room temperature. After 3 washes (10 min each) with PBST (1×PBS with 0. 2% Triton X-100), cells were blocked with 5% BSA in PBST for at least 1 hr at room temperature. The cells were stained overnight at 4°C with anti-H3Cit26 (Abcam, ab19847) antibody diluted in PBST (1∶100). DNA was stained with DAPI (4,6- diamidino-2-phenylindole) before mounting. Images were collected with LSM 510 laser scanning confocal microscope (Carl Zeiss). PAD2 expression levels were measured by provisional RNA-seq (Figure S13) data from The Cancer Genome Atlas (TCGA) [25], which was curated by UCSC [26], and published TCGA microarray data (Figure 7A, B) [25]. P-values were automatically selected for the best expression threshold cut-off and calculated using the log-rank test statistic. In Figure 7E and 7F, patient samples were split into two groups according to level of PADI2 expression by above and below the 45% quantiles. In Figure S13A, B patient samples were split into two groups according to level of PADI2 expression by above and below the 65% and 60% quantiles, respectively. Four sections from each breast cancer patient were deparaffinized, unmasked, blocked and treated with either rabbit anti-PADI 2 (Proteintech, 12110-1-ap), rabbit anti-ER-α (Santa Cruz, sc-542), rabbit anti-Histone 3 Citrulline 26 (AbCam, ab19847) or non-specific rabbit IgG (Millipore, S-20) using standard methods and developed using the Vectastain Peroxidase Rabbit IgG and Vector DAB substrate kits. Sections stained for ER were examined using a Zeiss AxioObserver microscope and two tumor cell areas on each section which stained positive were chosen and imaged at 5,10 and 20× using a Zeiss Axiophot color camera. These same two areas were imaged on the corresponding sections stained for PAD2, H3R26Cit and the IgG negative control. For patients who were negative for ER in all areas of the section, two areas containing tumor cells were chosen for imaging and corresponding images from the remaining slides were captured as above. In each image the brown DAB staining was scored from 0 to 5, with 0 being no staining and 5 being heavy staining in both nucleus and cytoplasm. The percent of tumor cells in the 10× image at each score level was established and multiplying the stain scores by their corresponding fraction and adding the products together determined a final score for that section area. The following comparisons were made using all of the final scores from both areas of all sections, ER: PADI2, ER: H3R26Cit, PADI2: H3R26Cit. H3R26Cit and H3K27ac ChIP-seq data was deposited into GEO with the accession number GSE58177. ER ChIP-seq data was previously published [38] with accession number GSE54855.

Transcription factors bind to DNA to activate and repress gene transcription. Many transcription factors, particularly nuclear receptors, associate with their cognate DNA element in a highly dynamic manner in vivo. Highly acetylated histone tails and DNase sensitive chromatin are amenable to the initial binding of transcription factors. Upon binding to DNA, transcription factor binding recruits remodelers and coactivators that can cause a concomitant increase in accessibility and acetylation. Herein, we show that estrogen receptor recruitment of a histone deiminase causes the positively charged H3R26 residue to be neutralized. This modification changes the fine structure of the nucleosome particle and facilitates estrogen receptor binding. Lastly, we find that high deiminase expression is associated with increased survival in estrogen receptor-positive breast cancer patients.

Abstract Introduction Results Discussion Materials and Methods

chromatin immunoprecipitation biochemistry genomics proteins genome analysis transcription factors gene expression genetics gene regulation biology and life sciences dna-binding proteins computational biology

2014

Targeted H3R26 Deimination Specifically Facilitates Estrogen Receptor Binding by Modifying Nucleosome Structure

Bacterial microcompartments are large, roughly icosahedral shells that assemble around enzymes and reactants involved in certain metabolic pathways in bacteria. Motivated by microcompartment assembly, we use coarse-grained computational and theoretical modeling to study the factors that control the size and morphology of a protein shell assembling around hundreds to thousands of molecules. We perform dynamical simulations of shell assembly in the presence and absence of cargo over a range of interaction strengths, subunit and cargo stoichiometries, and the shell spontaneous curvature. Depending on these parameters, we find that the presence of a cargo can either increase or decrease the size of a shell relative to its intrinsic spontaneous curvature, as seen in recent experiments. These features are controlled by a balance of kinetic and thermodynamic effects, and the shell size is assembly pathway dependent. We discuss implications of these results for synthetic biology efforts to target new enzymes to microcompartment interiors. While it has long been recognized that membrane-bound organelles organize the cytoplasm of eukaryotes, it is now evident that protein-based compartments play a similar role in many organisms. For example, bacterial microcompartments (BMCs) are icosahedral proteinaceous organelles that assemble around enzymes and reactants to compartmentalize certain metabolic pathways [1–10]. BMCs are found in at least 20% of bacterial species [2,11,12], where they enable functions such as growth, pathogenesis, and carbon fixation [1,10,13–16]. Other protein shells act as compartments in bacteria and archea, such as encapsulins [17] and gas vesicles [17,18], and even in eukaryotes (e. g. vault particles [19]). Understanding the factors that control the assembly of BMCs and other protein-based organelles is a fundamental aspect of cell biology. From a synthetic biology perspective, understanding factors that control packaging of the interior cargo will allow reengineering BMCs as nanocompartments that encapsulate a programmable set of enzymes, to introduce new or improved metabolic pathways into bacteria or other organisms (e. g. [10,20–29]) ]. More broadly, understanding how the properties of a cargo affect the assembly of its encapsulating container is important for drug delivery and nanomaterials applications. Despite atomic resolution structures of BMC shell proteins [1,10,30,31], the factors that control the size and morphology of assembled shells remain incompletely understood. BMCs are large and polydisperse (40-600 nm diameter), with a roughly icosahedral protein shell surrounding up to thousands of copies of enzymes [1,7–9,30,32,33]. For example, the best studied BMC is the carboxysome, which encapsulates RuBisCO and carbonic anhydrase to facilitate carbon fixation in cyanobacteria [1,30,32,34]. BMC shells assemble from multiple paralogous protein species, which respectively form homo-pentameric, homo-hexameric, and pseudo-hexameric (homo-trimeric) oligomers [1,30,31]. Sutter et al. [31] recently obtained an atomic-resolution structure of a complete BMC shell in a recombinant system that assembles small (40 nm) empty shells (containing no cargo). The structure follows the geometric principles of icosahedral virus capsids, exhibiting T = 9 icosahedral symmetry in the Caspar-Klug nomenclature [35,36] (meaning there are 9 proteins in the asymmetric unit). The pentamers, hexamers, and pseudo-hexamers occupy different local symmetry environments. Although the Sutter et al. [31] structure marks a major advance in understanding microcompartment architectures, it is uncertain how this construction principle extends to natural microcompartments, which are large (100-600 nm), polydisperse, and lack perfect icosahedral symmetry. Moreover, the effect of cargo on BMC shell size is hard to interpret from experiments. In some BMC systems, empty shells are smaller and more monodisperse than full shells [23,28,31,37], whereas in other systems empty shells are larger than full ones [38]. Thus, the cargo may increase or decrease shell size. The encapsulated cargo can also affect BMC assembly pathways. Microscopy experiments showed that β-carboxysomes (which encapsulate form 1B RuBisCO) undergo two-step assembly: first the enzymes coalesce into a ‘procarboxysome’, then shells assemble on and bud from the procarboxysome [39,40]. In contrast, electron micrographs suggest that α-carboxysomes (another type of carboxysome that encapsulates form 1A RuBisCO) assemble in one step, with simultaneous shell assembly and cargo coalescence [33,41]. Our recent computational study [42] suggested that the assembly pathway depends on the affinity between cargo molecules. However, that study was restricted to a single shell size, and thus could not investigate correlations between assembly pathway and shell size. Numerous modeling studies have identified factors controlling the thermodynamic stability [43–45] or dynamical formation [46–54] of empty icosahedral shells with different sizes. For example, Wagner and Zandi showed that icosahedral shells can form when subunits sequentially and irreversibly add to a growing shell at positions which globally minimize the elastic energy, with the preferred shell size determined by the interplay of elastic moduli and protein spontaneous curvature. Several studies have also investigated the effect of templating by an encapsulated nanoparticle or RNA molecule on preferred shell size [50,55–57]. However, the many-molecule cargo of a microcompartment is topologically different from a nucleic acid or nanoparticle, and does not template for a specific curvature or shell size. Rotskoff and Geissler recently proposed that microcompartment size is determined by kinetic effects arising from templating by the cargo [58]. Using an elegant Monte Carlo (MC) algorithm they showed that proteins without spontaneous curvature, which form polydisperse aggregates in the absence of cargo, can form kinetically trapped closed shells around a cargo globule. However, there are reasons to question the universality of this mechanism for microcompartment size control. Firstly, several recombinant BMC systems form small, monodisperse empty shells [23,28,31,37], suggesting that the shell proteins have a non-zero spontaneous curvature even without cargo templating. Secondly, when Cameron et al. [39] overexpressed RuBisCO to form ‘supersized’ procarboxysomes, carboxysome shells encapsulated only part of the complex, suggesting that there is a maximum radius of curvature that can be accommodated by the shell proteins. Thirdly, the kinetic mechanism is restricted to systems in which rates of shell association vastly exceed cargo coalescence rates, a condition which may not apply in biological microcompartment systems. Thus, despite this and other recent simulation studies of microcompartments [42,58,59], the factors which control BMC size and amount of encapsulated cargo remain unclear. In this article we use equilibrium calculations and Brownian dynamics (BD) simulations on a minimal model to identify the factors that control the size of a microcompartment shell. Although computationally more expensive than the MC algorithm of Ref. [58], BD better describes cooperative cargo-shell motions and thus allows for any type of assembly pathway. Using this capability, we explore the effect of cargo on shell size and morphology over a range of parameters leading to one-step or two-step assembly pathways. To understand the interplay between shell curvature and cargo templating, we consider two limits of shell protein interaction geometries: zero spontaneous curvature and high spontaneous curvature, which respectively form flat sheets or small icosahedral shells in the absence of cargo. Our calculations find that the presence of cargo can increase or decrease shell size, depending on the stoichiometry of cargo and shell proteins, and the protein spontaneous curvature. For shell proteins with high spontaneous curvature, we observe a strong correlation between assembly pathway and shell size, with two-step assembly leading to larger shells than single-step pathways or empty shell assembly. This result is consistent with the fact that β-carboxysomes tend to be larger than α-carboxysomes. For shell proteins with zero spontaneous curvature, we find that introducing cargo can result in a well-defined shell size through several mechanisms, including the kinetic mechanism of Ref. [58] and the ‘finite-pool’ effect due to a limited number of cargo particles available within the cell. However, spontaneous curvature of the shell proteins allows for robust shell formation over a wider range of parameter space. We simulated assembly dynamics using the Langevin dynamics algorithm in HOOMD (which uses GPUs to efficiently simulate dynamics [81]), and periodic boundary conditions to represent a bulk system. The subunits are modeled as rigid bodies [82]. Each simulation was performed in the NVT ensemble, using a set of fundamental units [83] with σ0 defined as the circumradius of the pentagonal subunit (the cargo diameter is also set to σ0), and energies given in units of the thermal energy, kBT. The simulation time step was 0. 005 in dimensionless time units, and we performed 3 × 106 timesteps in each simulation unless mentioned otherwise. Initial conditions. We considered two types of initial conditions. Except where stated otherwise, simulations started from the ‘homogeneous’ initial condition, in which subunits and (if present) cargo were initialized with random positions and orientations, excluding high-energy overlaps. In the ‘pre-equilibrated globule’ initial condition, we first initialized cargo particles with random positions (excluding high-energy overlaps), and performed 105 simulation timesteps to equilibrate the cargo particles. Shell subunits were then added to the simulation box with random positions and orientations, excluding high-energy overlaps. Systems. We simulated several systems as follows. For shell subunits with spontaneous curvature we set pentamer-hexamer and hexamer-hexamer angles consistent with the T = 3 geometry (see Estimating the shell bending modulus in section S2 Text), and we set εangle = 0. 5. We first performed a set of empty-shell assembly simulations, with 360 hexamers, and varying number of pentamers, in a cubic box with side length 60σ0, with εHH = 2. 6kBT (the smallest interaction strength for which nucleation occurred). These simulations were performed for 107 timesteps to obtain sufficient statistics at low pentamer concentrations despite nucleation being rare. For cargo encapsulation by subunits with spontaneous curvature, we simulated 2060 cargo particles, 180 pentamers, and 360 hexamers in a cubic box with side length 60σ0. Other parameters were the same as for the empty-shell simulations, except that we varied εPH, εSC, and εSC as described in the main text. All simulations with spontaneous curvature used εPH ≥ 1. 3εHH to ensure that the shells with the T = 3 geometry (or asymmetric shells with similar sizes) were favored in the absence of cargo. We note that our results generalize to other ranges of shell interaction parameters, but this choice distinguishes effects due to cargo from those due to changes in the inherent preferred shell geometry. Simulations with strong cargo-cargo and cargo-shell interactions (εCC ≥ 1. 55 and εSC < 8. 75) required a long timescale for pentamers to fill pentameric vacancies in the hexamer shell (discussed in Results). To observe pentamer adsorption, these simulations were run for up to 9 × 106 simulation timesteps. For simulations of ‘flat’ subunits (with no spontaneous curvature), we considered a range of system sizes at fixed steady state cargo chemical potential, with the number of cargo particles varying from 409 to 3275, and the box side length varying from 35σ0 to 70σ0. Since these were NVT simulations, we ensured that the final hexamer chemical potential was the same at each system size by setting the number of hexamers so that the concentration of free hexamers remaining after assembly of a complete shell was constant (10−3 subunits/σ 0 3). The resulting number of hexamers varied from 109 to 581 in boxes with side lengths 35σ0 to 70σ0. The assembly outcomes were unchanged if instead we kept the total hexamer subunit concentration the same across all simulations. For each of these system sizes we performed simulations over a range of εangle to identify the maximum value of κs at which assembly of a complete shell could occur. Simulations were stopped upon completion of a shell or after the maximum simulation time tmax with tmax = 3 × 106 timesteps for boxes with side length ≤ 55σ0 and tmax = 8 × 106 for boxes with side length ≥ 55σ0. The maximum simulation time was increased for large system sizes because the minimum time required for assembly of a complete shell increases linearly with the shell size [84]. To estimate the relationship between the shell bending modulus κs and the parameter εangle we performed additional simulations, in which we measured the total interaction energy of completely assembled shells as a function of εangle (see ‘Estimating the shell bending modulus’ in section S2 Text). Sample sizes. For simulations of shells with spontaneous curvature, we performed a minimum of 10 independent trials at each parameter set. To enable satisfactory statistics on shell size and morphology for parameter sets that result in at most one complete shell in the simulation box 3, we performed additional trials such that at least 10 complete shells were simulated. For flat subunits (Fig 1F and 1G), we identified the maximum εangle for which a complete shell forms at each system size as follows. We first performed independent simulations over a range of εangle values, separated by increments in εangle of 0. 02 for systems with box side length ≤ 55σ0, and increments of 0. 05 for systems with side length ≥ 55σ0. We performed 10 independent trials at each value of εangle. For the largest value of εangle at which at least one of these trials resulted in a complete shell, we then performed 10 additional trials to obtain a more accurate estimate of the shell bending modulus κs at the maximum εangle. We begin by considering shells with T = 3 spontaneous curvature (Fig 1D). To isolate the effects of cargo on shell size, we consider shell-shell interaction parameters which favor pentamer insertion (setting the ratio of pentamer-hexamer and hexamer-hexamer affinities εPH/εHH ≥ 1. 3) so that assembly without cargo results in primarily T = 3 empty shells for our ratio of pentamer to hexamer concentrations, ρp/ρh = 0. 5, and results in shells close in size to the T = 3 geometry at all of the stoichiometries we consider here. A typical assembly trajectory without cargo is shown in Fig 2A. When simulating assembly around cargo, we set the hexamer-hexamer affinity εHH ≤ 2. 2 (while maintaining εPH/εHH ≥ 1. 3) so that assembly occurs only in the presence of cargo, and we vary cargo-cargo εCC and cargo-shell εSC interaction strengths. Throughout this article, all energy values are given in units of the thermal energy, kBT. Except where mentioned otherwise, values of our simulation shell bending modulus κs fall within the range estimated for β−carboxysomes from AFM nanoindention experiments κs ∈ [1,25]kBT (see Ref. [89] and section ‘Determination of parameter values’ in S2 Text; simulations with shell spontaneous curvature use κs = 10 − 16kBT. We now consider the opposite limit: a system of ‘flat’ hexamer subunits, which have zero spontaneous curvature and thus favor formation of flat sheets (Fig 5A). Fig 5B shows a typical assembly trajectory for flat subunits with εCC = 1. 7, in which the cargo rapidly coalesces followed by adsorption and assembly of the hexamers. Interestingly, the shapes of assembly intermediates reflect the lack hexamer spontaneous curvature—hexamers initially assemble into flat sheet wrapped around the globule, deforming the spherical globule into a cigar shape. Eventually the two sides of the sheet meet, creating a seam with an unfavorable line tension due to unsatisfied subunit contacts. As the seam gradually fills in, the elastic energy associated with such an acute deformation forces the complex toward a more spherical shape. As in systems with spontaneous curvature, the hexamer shells exhibit the 12 five-fold vacancy defects required by topology. If pentamers are present they eventually fill these holes (as in Fig 2 above), but for simplicity we consider systems containing only hexamers here. The large shells are roughly but not perfectly icosahedral, presumably reflecting slow defect reorganization on assembly timescales. The size of the assembled shell is limited by the finite system size of our simulations. Importantly, the same limitation occurs within cells when the cargo undergoes phase separation into a single complex whose size is limited by the enzyme copy number (e. g. the procarboxysome precursor to carboxysome assembly [39,40]). We therefore investigated the dependence of assembly morphologies on system size, as a function of the shell bending modulus, κs (controlled by the parameter εangle). Specifically, at each value of κs we performed a series of simulations in which the maximum size of the cargo globule was controlled by changing the system size with fixed total cargo concentration and hexamer chemical potential (section Materials and methods). An example assembly trajectory for a small system is shown in Fig 5C. As shown in Fig 6, we observe a minimum globule size required for complete shell assembly, which linearly increases with κs. We observe complete wrapping for all system sizes above this threshold. Below the threshold size, assembly stalls with one or more open seams remaining; examples of this configuration are shown for a low and high bending modulus in Fig 6. Interestingly, while the pentameric defects are roughly equally spaced within large shells, small shells assembled with extremely low values of κs tend to exhibit adjacent vacancy pairs (Fig 5C, final frame). This defect morphology focuses curvature in a region with no elastic energy (the vacancy) while reducing the number of unsatisfied hexamer edges. To understand these results, in section S2 Text we present a calculation of the equilibrium shell size distribution for subunits with no spontaneous curvature and stoichiometrically limiting cargo. We restrict the ensemble to spherical shells as observed in the simulations. While the aggregates are large and polydisperse without cargo, the calculation shows that cargo leads to a minimum free energy spherical shell size (S8 and S9 Figs). The linear relationship between minimum shell size and bending modulus can be understood from our equilibrium model by comparing the excess free energy difference ΔΩwrap between the complete shell and an unwrapped globule (see section S2 Text). For the simulated conditions, the size and shape of the cargo globule is essentially the same in each of these states, and thus the free energy difference for a globule wrapped by nh hexamers in Eq. (S2. 9) simplifies to Δ Ω wrap = 8 π κ s + Δ G p + Δ μ h n h (1) with Δμh = ghh + ghc − μh, ΔGp as the free energy due to the 12 pentameric vacancies, ghh (εHH) as the hexamer-hexamer interaction free energy, ghc (εSC) as the hexamer-cargo free energy, and μh = kBT log (ρh) the chemical potential of unassembled hexamers at concentration ρh. The term 8πκs describes the bending energy of the complete shell. The minimum globule size n* corresponds to the locus of parameter values at which ΔΩwrap = 0, giving n * = 8 π - Δ μ h κ s + Δ G p - Δ μ h (2) A linear fit to the simulation results for n* results in Δμh = −2. 4 and ΔGp = 80. 5kBT, or 6. 7kBT per pentameric defect. Plugging in ρh = 10−3 subunits/σ 0 3 and ghc = −8. 1kBT for εSC = 7. 0 (using the estimate from Perlmutter et al. [42]) then results in ghh ≈ −0. 45kBT. This value and the fit value of ΔGp are reasonably close to interactions estimated from the relationship between the shell-shell dimerization free energy ghh and potential well-depth εHH for a similar model in Perlmutter et al. [42]. Thus, the simulation results are consistent with the minimum stable shell size predicted by the theory. We have used computational and theoretical modeling to investigate factors that control the assembly of a protein shell around a fluid cargo. We have focused on two limiting regimes of protein interaction geometries—high spontaneous curvature that drives the formation of small shells, and zero spontaneous curvature that favors assembly of flat sheets or polydisperse shells. In both regimes the presence of cargo can significantly alter the size distribution of assembled shells. For high spontaneous curvature, encapsulated cargo tends to increase shell size, whereas for shell proteins with low (or zero) spontaneous curvature cargo templating provides a mechanism to drive shell curvature and thus tends to reduce shell size. These results could provide a qualitative explanation for experimental observations on different systems in which full microcompartment shells were either larger or smaller than empty shells [23,28,31,37,38]. Our simulations identify a combination of kinetic and thermodynamic mechanisms governing microcompartment size control. At equilibrium, the shell size is determined by the stoichiometry between cargo and shell subunits, with an excess of cargo or shell protein respectively favoring larger or smaller shells. Similarly, a high surface energy (high cargo surface tension and weak shell-cargo interactions) favors larger shells whereas a strong shell bending modulus favors shells closer to the preferred size. Although dynamical simulations exhibit similar qualitative trends to these equilibrium results, we observe significant kinetic effects as well. Fast cargo coalescence relative to rates of shell assembly favors larger shells, since closure of an assembling shell prevents further cargo aggregation. Thus, the shell size is strongly correlated to the assembly pathway, with two-step assembly leading to larger shells than single-step pathways. Although many factors likely control shell size in biological systems, this result is consistent with the observations of small empty shell assemblies [23,28,31,37] and the fact that β-carboxysomes (which assemble by two step pathways [39,40]) tend to be larger and more polydisperse than α-carboxysomes (which experiments suggest assemble by one-step pathways [33,41]). Our results for shell proteins without spontaneous curvature build upon Rotskoff and Geissler [58], which identified a kinetic mechanism in which cargo templating drives shell curvature, and shell closure eventually arrests assembly. Their mechanism proceeds by two-step assembly, with initial nucleation of a cargo globule followed by assembly of shell subunits, but requires that rates of subunit arrival are at least 10 times faster than cargo arrival rates [58]. However, it is unclear how many physical microcompartment systems may fit this criteria, and our results suggest other mechanisms may play important roles in microcompartment assembly. Firstly, if cargo is stoichiometrically limiting then the finite-pool mechanism can result in finite shell sizes, with the coalesced cargo still providing a template for shell curvature. Secondly, subunits with spontaneous curvature can form complete shells even under conditions of excess cargo or fast coalescence rates that lead to large cargo aggregates (Fig 3D), as observed for carboxysome assembly in cells [39]. Thus, biological microcompartments with some degree of preferred shell curvature could robustly assemble over a much wider parameter space than systems without spontaneous curvature. Intriguingly, the recent atomic-resolution microcompartment structure from Sutter et al. [31] suggests that different hexamer or pseudo-hexamer species have different preferred subunit-subunit angles, and thus the spontaneous curvature may depend on the shell composition. We will investigate this in a future work. The importance of spontaneous curvature to a particular BMC system could be investigated by comparing our computational predictions to experimental shell size distributions measured for varying cargo/shell protein stoichiometries and interaction strengths. While such tests would be most straightforward to perform in vitro, they could be performed in vivo by varying expression levels of various shell proteins or the enzymatic cargoes. Of particular interest would be a comparison between the shell size distribution in the presence and absence of cargo. However, note that we have focused on extreme limits (high spontaneous curvature or zero spontaneous curvature); systems with moderate shell spontaneous curvature may exhibit less dramatic cargo effects. Also note that the effective shell spontaneous curvature depends on the stoichiometries of different shell protein species; e. g. , overexpressing pentamers would shift the size distribution toward smaller shells (Fig 2D). These results have implications for targeting new core enzymes to BMC interiors. Recent experiments have shown that alternative cargoes can be targeted to BMC interiors by incorporating encapsulation peptides that mediate cargo-shell interactions, but that relatively small amounts of cargo were packaged [21–23,96]. Our previous simulations showed that assembly of full shells requires both cargo-shell and cargo-cargo (direct or mediated) interactions. Here, we see that the strength of cargo-cargo interactions can not only affect the efficiency of cargo loading, but also the size of the containing shell.

Bacterial microcompartments are protein shells that encase enzymes and reactants to enable bacteria to perform vital reactions, such as breaking down chemicals for energy or converting the products of photosynthesis into sugars. Microcompartments are essential for many bacteria, including human pathogens. Thus, there is great interest in understanding how microcompartment shells assemble around their cargo (the interior enzymes and reactants), and what determines the structure and size of a microcompartment. These questions are difficult to answer with experiments alone, because most intermediates in the assembly process are too short-lived to characterize in experiments. Therefore, this article describes theoretical and computational models for microcompartments, which predict assembly pathways and how the sizes of assembled shells depend on factors such as protein interactions and concentrations. The simulations show that the properties of the cargo are an important factor for determining shell size, and suggest an explanation for recent experimental results showing that cargo can either increase or decrease shell size. In addition to helping to understand the natural behavior of microcompartments, the simulations provide guidance to researchers working to reengineer microcompartments to produce drugs or biofuels.

Abstract Introduction Materials and methods Results and discussion

protein interactions condensed matter physics geometry cargo proteins plant science mathematics photosynthesis ribulose-1,5-bisphosphate carboxylase oxygenase thermodynamics proteins chemistry physics biochemistry stoichiometry plant biochemistry curvature biology and life sciences physical sciences nucleation symmetry

2018

The role of the encapsulated cargo in microcompartment assembly

West Nile virus (WNV) —a mosquito-borne arbovirus—entered the USA through New York City in 1999 and spread to the contiguous USA within three years while transitioning from epidemic outbreaks to endemic transmission. The virus is transmitted by vector competent mosquitoes and maintained in the avian populations. WNV spatial distribution is mainly determined by the movement of residential and migratory avian populations. We developed an individual-level heterogeneous network framework across the USA with the goal of understanding the long-range spatial distribution of WNV. To this end, we proposed three distance dispersal kernels model: 1) exponential—short-range dispersal, 2) power-law—long-range dispersal in all directions, and 3) power-law biased by flyway direction —long-range dispersal only along established migratory routes. To select the appropriate dispersal kernel we used the human case data and adopted a model selection framework based on approximate Bayesian computation with sequential Monte Carlo sampling (ABC-SMC). From estimated parameters, we find that the power-law biased by flyway direction kernel is the best kernel to fit WNV human case data, supporting the hypothesis of long-range WNV transmission is mainly along the migratory bird flyways. Through extensive simulation from 2014 to 2016, we proposed and tested hypothetical mitigation strategies and found that mosquito population reduction in the infected states and neighboring states is potentially cost-effective. West Nile disease (WND) is a vector-borne zoonosis which may result from infection by West Nile virus (WNV), a member of the family Flaviviridae, genus Flavivirus. This virus is the most common cause of arboviral disease in the United States [1]. From 1999 to 2017, more than 48 thousands WNV disease cases were reported to the Centers for Disease Control and Prevention (CDC) and more than two thousands of these reported cases resulted in death [2]. WNV is maintained in an enzootic transmission cycle between competent mosquitoes and birds. Birds are the reservoir and amplifying host for this virus. The US Centers for Diseases Control and Prevention (CDC) has identified WNV infection in more than three hundred species of birds. Infected bird movement is likely a key factor that affects the geographic spread of WNV, especially given the different habitats and routes of various species. Although many bird species may be infected with WNV, the American robin is considered an important amplifier of WNV and maybe a driver geographic spread because WNV-infected American robins have low mortality and high viremia [3,4]. Members of the Culex genus of mosquito are the principal vectors of this virus in the United States [5]. Humans, horses, and other mammals can be infected with WNV. However, these infections result in relatively low virus titers (viremia) therefore the infected animals and people are considered dead-end hosts (not capable of infecting feeding mosquitoes). Therefore, they do not have any epidemiological impact on WNV transmission or geographic spread [6]. To understand the transmission dynamics of WNV, several mathematical models have been developed [3,7–10]. These models predict the threshold conditions for WNV spreading in different scenarios. However, most of these models do not consider the spatial dynamics of WNV. Space or geographic spread has a significant role in WNV disease dynamics and modeling of WNV spatial spreading is complex because of the interactions of multiple potential mosquito vectors, avian amplifiers, and mammalian hosts. Liu et al. [9] developed a patchy model to analyze the spatial spreading of WNV, where patches are geographical space. They assumed patches are identical, spatial dispersal of birds and mosquitoes are symmetric within patches, and movement of birds and mosquitoes are only one-dimensional. According to this investigation, long-range dispersal of infected bird populations determines the spatial spread of WNV, not the dispersal of infected mosquito populations. Other investigators proposed a reaction-diffusion model [10], where they have spatially extended the non-spatial model of Wonham et al. [8] to mathematically estimate the spread of WNV. Here, diffusion terms in the reaction-diffusion partial differential equations represent vector mosquito and host bird population movements. They identified traveling wave solutions in their model and calculated the rate of spatial spread of infection. Durand et al. [11] developed a discrete time deterministic meta-population model in order to analyze the circulation of WNV between Southern Europe and West Africa. Another spatial model proposed by Maidana and Yang [12] used a system of partial differential reaction-diffusion equations. They also calculated the speed of disease dissemination by investigating the traveling wave solution of their model. They concluded, mosquito movements do not play an important role in disease dissemination. In addition, they included vertical transmission in their model and determined that vertical transmission is not an important factor for the spatial spread of WNV. Most WNV spread models are mathematical deterministic compartmental models. However WNV spread is highly stochastic because of the demography and movement of hosts and vectors varies between different locations. The major weaknesses of these models are the number and complexity of the compartments required to account for the many host and vector populations. In turn, the number of compartments increases the number of unknown parameters. Approximation of these parameters in any biological system is very challenging and prone to estimation errors which can create inaccuracies in the model outputs. We developed an individual-based heterogeneous network framework to understand WNV geographic spread. To build the network framework, we used the American Robin population density across the contiguous United States. The demographic characteristics of avian host populations and vector populations are not homogenous geographically, so we used a heterogeneous network framework. The transmission intensity of WNV depends on the abundance of WNV-infected vector mosquitoes in a given location. Mosquito population numbers fluctuate with local weather and season throughout the year, therefore we used a temperature dependent transmission rate. Although dead-end hosts cannot spread WNV to mosquitoes, we have quantified WNV case data only for humans, which we used to estimate unknown parameters. To understand the WNV spatial distribution, we proposed distance dispersal kernels, which describes the probability of dispersal with respect to distances. In this framework, we proposed three types of distance dispersal kernels: 1) exponential, 2) power-law, and 3) power-law biased by flyway. Then we compared the three distance kernels using approximate Bayesian computation based on sequential Monte Carlo sampling (ABC-SMC) method [13–18]. After conducting an extensive simulation for 2014-2016, we observed that an adapted fat-tailed or power-law kernel, which has long-distance links in specified directions can best describe the WNV human case data. We tested this network framework for the best kernel with the human case data and found that simulated results for more than 41 states of 49 states are consistent with the reported WNV cases. Our results support previous work on WNV spreading [3], which also modeled WNV spreading with migratory birds. We validate our work computationally from human incidence data. We proposed several theoretical mitigation strategies to control WNV and calculated their estimated costs. From the analysis of mitigation strategies, we suggest that potentially effective mitigation policies would include the application of mitigation control in areas with active transmission and in immediate neighboring states. The study area of this research was the contiguous United States where WNV is considered endemic. We modeled WNV case distributions for 2014-2016. We used three data sets each year to develop our model. The first dataset contained the average monthly temperatures. Mosquito vector abundance correlated with temperature. Temperature data was from the National Centers for Environmental Information [19]. The second dataset contains American Robin population data from eBird [20]. This is a database for bird abundance and distribution, which is formed by the Cornell Lab of Ornithology and National Audubon Society. We used total observation of American Robin in each state of the USA for each month. The robin data set was used to train the network model. The American Robin is abundant throughout the United States and is a preferred food source for many WNV-competent mosquito species [21]. Based on host feeding patterns of the Culex genus of mosquitoes, robins are the most common WNV amplifying host [22–24]. Other important susceptible birds, such as American crow were not used because although they are an indicator species (high crow mortality), they are unlikely to spread virus geographically as they are mostly a residential species. In addition, as an indication of epidemic start point, we used WNV human incidence data. Many species of birds have long-distance migration during the spring and fall. Therefore the network does not focus on one long-distance migrating bird species but aggregates all species along the known flyways. To estimate model parameters we used human case data for WNV from CDC [2], which is the third dataset. To explore WNV long-distance spatial distribution in the USA, we used an individual-based heterogeneous network framework. In this framework, birds are on the individual level, a node represents an individual bird and connection between nodes is the possibility of virus dispersal from one infected bird to another susceptible bird by mosquito vectors. Links or connections are formed by movement of birds or movement of vectors. If there is no link between nodes then infected birds and insects are not moving virus between nodes. All virus transmission occurs by local competent vector mosquitoes. There is some evidence of bird-to-bird transmission, but it likely does not contribute to or maintain outbreaks. We split the bird population into four compartments; susceptible, exposed, infected, and recovered. Although, in the literature most mathematical models do not consider the exposed avian class when modeling WNV [8,12,25,26]. Birds transmit virus to mosquitoes when a susceptible mosquito vector takes an infected blood meal, then the mosquito becomes infectious after the extrinsic incubation period (EIP), or the time needed for the virus to spreads from the mosquito mid gut to the salivary glands; usually this process takes 7 to 14 days [3,27]. In addition, an infected bird can infect many mosquitoes simultaneously and also an infected mosquito can bite many susceptible or infected birds. Therefore, there is some delay in the system, to represent this delay we added the exposed class. We estimated exposed period from data by using the approximate Bayesian computation with sequential Monte Carlo sampling (ABC-SMC) method. After the exposed period, birds entered the infected compartment and an infected bird transitions to recovered after 4-5 days. To simulate this model, we used generalized epidemic mean-field (GEMF) framework developed by the Network Science and Engineering (NetSE) group at Kansas State University [28]. In GEMF, each node stays in a different state and the joint state of all nodes follows a Markov process [28–30]. The node level description of this Markov process is: P r [ x i (t + Δ t) = 1 | x i (t) = 0, X (t) ] = β (T) Y i Δ t (1) P r [ x i (t + Δ t) = 2 | x i (t) = 1, X (t) ] = λ Δ t + o (Δ t) (2) P r [ x i (t + Δ t) = 3 | x i (t) = 2, X (t) ] = δ Δ t + o (Δ t) (3) Here, X (t) is the joint state of all individual nodes at time t. xi (t) is a node state, xi (t) = C means node i is in C compartment at time t, C = 0,1, 2,3 corresponds to susceptible, exposed, infected, and recovered compartment. Yi is the number of infected neighbors of node i, β (T) is the transmission rate from one infected bird to one susceptible bird, which is a function of temperature, λ is the rate for exposed to infectious state, and finally, a node recovers from infectious state at a rate δ. For the spatial dynamic characteristics of WNV transmission, we built a network framework, which has 49 sub-networks one for each adjoining states of the contiguous United States plus the District of Columbia. The number of nodes in each sub-network is proportional to the size of the avian population in that state [20]. We considered the mosquito season June-October for the simulation period. Although the mosquito season is not the same for all states, mosquitoes are active from June to September in all of the states at these times [33]. The network for the avian population is (V, E). Here, V is the set of nodes, which is the union of nodes of all sub-network, V = SN1 ∪ SN2 ∪ SN3 ∪ ………‥ ∪ SN49, here SNi is a set of nodes in the sub-network i and E is the set of links among individual nodes. To build sub-networks, we used the total number of observations of American Robin for states per month in the simulation time period. | S N i | = max m j = m 1: m 2 (O B S m j i) * S c + N 0, here, O B S m j i is the total number of observations of American Robins in state i in month mj, N0 is the error term and N0 ∼ N (5,2) for this model. m1 is the first month after May and m2 is the last month before October when the average monthly temperature is greater than T0. Sc is the scaling constant. In each sub-network, we assumed that nodes are connected through Erdos-Renyi (n, p) random network topology [37]. In this network topology, we created links randomly among nodes with a probability p. Here, n is the number of nodes in a sub-network and p is the probability to form an edge. We set the probability p = R * log (n) /n, here R is a constant (R ≥ 2), as this value is more than the threshold value for the connectedness of an Erdos-Renyi graph [38], so nodes of a sub-network are locally connected. We will refer these networks as a local network in the subsequent sections of this paper. To build connections among sub-networks, we considered long-distance dispersal kernels [7,39], which describe the probability of dispersal with respect to distances. Dispersal kernels provide a simple model of dispersal to model dispersal events. For long-distance events, we used three types of kernel models; 1) Exponential, 2) power-law, and 3) power-law-flyway, which is a power-law kernel biased by flyway. The dispersal phenomenon in this work is not conserved because of long-distance movement of migratory birds and seasonality within bird populations. Some long-distance migratory birds can disperse outside the contiguous United States or outside the network nodes, which are discrete points. The connection probability between two nodes does not represent the probability that a single dispersal event happens rather it represents the probability of contact and subsequent pathogen transmission between them. A simple caricature of the network is shown in Fig 1. There are three sub-networks, A, B, and C. The links, which formed local networks are shown by solid lines. These links are introduced by Erdos-Renyi (n, p) network topology. Dashed lines are inter-links among sub-networks. These links established by using long-distance dispersal kernels. In this framework, we adopted approximate Bayesian computation based on a sequential Monte Carlo sampling (ABC-SMC) method for parameter estimation and model selection [13–18]. The role of mosquito populations in WNV transmission is expressed by disease transmission rate β. This framework used different transmission rates in different parts of the network corresponding to the local mosquito abundance. Using this heterogeneous feature in the framework, we evaluated theoretical mosquito population management measures to reduce the outbreak size or transmission rates in the state level. Some states such as Kansas, do not have statewide mosquito surveillance or management, but in these theoretical scenarios, it is assumed they can develop or benefit from effective statewide mosquito management programs. The framework will simply estimate how much the mosquito abundance is reduced or maintained based on the theoretical outcomes of coordinated control. Furthermore, we realize mosquito control is generally conducted on a county or municipal level, but the human case data is only available on a state level. Therefore the recommendations are for the lowest resolution of the data, which is state level but applies to counties and municipalities as well. If vector management is increased in a sub-network, then transmission rates will be changed by, β r = β R F, here βr is the reduced transmission rate and RF is the reduction factor. Then management costs will be Cost = RF * NSc, here NSc is the number of states where control measures were applied. We considered supplemental management measures with the existing management measures. We used two types of mitigation strategies across the United States, 1) dynamic infected place tracing strategy and 2) static ranked based strategy. In the infected place tracing, we traced the infected states, then plan the mitigation strategies according to them. For this type of mitigation strategies, we considered three cases; 1) case-1: only infected: applied control only in the infected states; 2) case-2: infected & first neighbors: applied control in the infected states with its first neighboring states (whose distance is less than 500km), and 3) case-3: infected & first neighbors & second neighbors: applied control in the infected states with its first neighboring states, and also with its second neighboring states (whose distance is in 500 − 1000km). For infected tracing control measure, we kept track of infected places monthly. If SNi sub-network is infected for month t, then control measures were applied for the month t + 1 based on these three cases. In the static ranked based mitigation strategy, we ranked the states by different variables (for example, temperature, size of the avian population etc.). For this strategy, we considered three cases; 1) temp. : states ranked by temperature, 2) pop. : states ranked by avian population size, and 3) temp. & pop. : states ranked by temperature and avian population size both, then we applied management measures in the top 30% of the states. In this spatial-temporal individual-based heterogeneous network framework, we used three distance kernel models. The fundamental basic WNV epidemic model is the same for all the three network kernels. In the entire network, there are 49 sub-networks representing the 48 adjoining contiguous states plus the District of Columbia. All sub-network nodes are locally connected. The topology of the local network is Erdos-Renyi. The total nodes for the year 2015 was |V| = 7657 and the scaling constant is Sc = 0. 02. Here, E = El ∪ Edd; |El| is the number of total intra-links for all local networks, which is around 167000-170000 and |Edd| is the number of total inter-links among sub-networks. The description of sub-networks is provided in Table B in the S3 Text. We started the epidemic from states with the highest human incidence prior to June. We started the epidemic for the year 2015 by adding two infected nodes, one in sub-network SN4 (California) and another in sub-network SN42 (Texas). Connections among sub-networks are developed by distance dispersal kernels. Parameters for these kernels are estimated from the ABC-SMC method. To test the performance of this framework, we used estimated parameters from Table 1 for power-law kernel influenced by flyway. We set the parameters value; Kpl = 2. 3147, β0 = 0. 0059day-1, λ = 0. 0721day-1, and δ = 0. 2031day-1. The simulation period for the avian population model is from week-23 to week-44. The output of avian population was used as the input of zoonotic spillover compartment. Then we compared the output of zoonotic spillover compartment with human case data for week 24 to week 45. We considered a one-week lag between WNV incidence in birds and WNV incidence in humans. In humans, WNV-infected individuals (approximately 20%) develop a mild febrile illness after 3–6 days [56]. Peak of reporting of dead birds is one week prior than the reporting peak of human incidence [57]. In Fig 5, the mean simulated human case from the 49 sub-networks is compared with the weekly human case data for 2015 for the contiguous USA. The absolute errors between them are shown here. From this whisker plot, we can see that the median of the absolute error for the states is close to zero. In Fig 5, the largest outlier is California (marked by black circles). These outliers result from a mismatch between the simulated peak human incidence time and the observed human incidence peak time possibly because the very long state (north to south) has weather which is very different in southern California (warmer and drier) than northern California (cooler and wetter) causing a difference between peak mosquito seasons in the southern and northern parts. We compared the total yearly incidence of human WNV from this model with the state level reported case data. The results are shown in Fig 6. For 2015, we found that the case data for 42 of 49 locations were within the simulation results. The states where human cases were different from the simulation results were over-reported states (Nevada) and under-reported states (Louisiana, Mississippi, Nebraska, North Dakota, South Dakota, and Washington). The possible reason for this mismatch are reporting error or overwintering of virus in birds or mosquitoes or another bird species (not robins) is the key reservoir species for that state To build a disease prevalence map, we grouped the states in four categories; 1) higher prevalence —incidence is more than 100,2) intermediate prevalence—incidence is in between 50-99,3) moderate prevalence—incidence is in between 25-49 and 4) low prevalence—incidence is less than 25. To group the states, we used the median of the simulation results. The disease prevalence map from the model are presented in Fig 7a and from observed data are presented in Fig 7b. Among 49 locations, 40 locations are in the same prevalence group in both maps. We applied mitigation strategies on the power-law-flyway kernel network model to find the optimal mitigation plan. Fig 8a shows the number of infected states or epidemic size for dynamic infected places tracing. Epidemic size decreased faster with increased reduction factor for case-2 (infected & first neighbors) and case-3 (infected & first neighbors & second neighbors) than case-1 (only infected). The number of states where control measures were applied is displayed in Fig 9, which is proportional to cost. Therefore, the cost was minimal for case-2 than other two cases for RF > 2. From the cost analysis, we concluded that, although the cost for case-1 is less at the beginning of the yearly outbreak, we need to apply management only in the infected places, however by the end of the year the total cost for case-2 will smaller because of the smaller epidemic size. The results of the static ranked based mitigation strategy measure are presented in Fig 8b. We observed that, before RF = 4. 5, number of infected states for temp. & pop. dropped earlier than others. Number of infected states or epidemic size was smaller for temp. than pop. after RF > 3, infected population of a sub-network are more positively correlated with temperature. The NSc is always the same for these three cases. For all mitigation strategies, minimum epidemic size could be 2, as we started the epidemic from two states. We proposed an individual-based heterogeneous network framework and tested three dispersal kernels to understand the spatial spread patterns of WNV human case data across the contiguous United States. This framework requires fewer parameters and has more flexibility to represent the spatial-temporal dynamics of WNV. Adding parameters can make the framework more realistic, for example, more competent bird species, landscape features for habitat preferences of host and vector species, daylight conditions [32], pathogen invasion from outside of USA, variable susceptibility among different hosts and vectors, WNV strain variability, mosquito and virus overwintering, vertical transmission, human movement characteristics etc. However, inclusion of too many factors increases model complexity which makes model optimization difficult given the availability of limited observational data. On the other hand, a simple model may insufficient to represent WNV spatial dynamics. Computational models need to be developed and parameters calculated with sufficient detail to be biologically accurate if they are used to evaluate epidemic management measures. However, for most biological systems, reliable parameter information is unknown. Unknown parameters or inaccurate assumptions add uncertainty to the model. Our framework has only four parameters to estimate (network Parameter K, transmission rate β, transition rate from exposed to infectious state, λ, and human spillover, η). This framework has compartments only for the avian population (susceptible, exposed, infected, and recovered), and is not species specific. We reduced the compartments for vector population by implementing them implicitly through transmission rate between infected nodes and susceptible nodes. The presented framework and dispersal kernel network model has an intermediate complexity that approximate Bayesian computation based on sequential Monte Carlo sampling (ABC-SMC) method successfully calibrated and estimated the parameters with the available data. If more data becomes available, it is possible to add them in this model for improved performance of the model. Furthermore, this framework is flexible and therefore can represent various hosts and vectors including with population seasonality, which plays an important role in WNV dynamics. For host population seasonality, we added a node property Activity, this property allows us to control active host populations in the network in a specific time period. We added vector seasonality in this framework with a temperature dependent transmission rate. This framework proposed one exponential and two fat-tailed distance kernel models for long-distance transmission of WNV with each model having increasing complexity and similarities to natural avian movement. WNV spatial distribution is very complex because WNV can infect more than 300 bird species, some of which are residential birds and short-distance migrators which disperse less than 500 km distances (short connections) whereas some species are long-distance migratory birds creating long connections. The long-distance migratory birds are the long-distance dispersal (LDD) agents for WNV. Previous studies tried to analyze spreading of WNV using a traveling wave with constant velocity, however, WNV spread more rapidly across the North America than would be expected from the assumption of constant velocity traveling wave [58]. Likely this is because traveling wave models unlike distance dispersal kernel models for WNV spreading do not capture the long-distance migrating birds which can have various migratory ranges and distances. Distance dispersal kernels have more flexibility to represent the different bird migration distances and can account for accelerating invasions. However, exponential kernels produce short-connections and therefore like traveling waves are limited to constant expansion, unlike fat-tailed power-law kernels which can generate accelerating invasions by creating the long-distance connections from migratory birds [59]. However, a general fat-tailed power-law kernel makes long-distance links in every direction which does not follow the incidence of WNV. Instead, a power-law-flyway kernel can be used to produce the long connections in the direction of flyways and short links in other directions. Bayesian inference was used to test which of the three kernel models best described WNV distribution on the network for three most recent years (2014- 2016). The power-law-flyway kernel best described the distribution of WNV cases because the long-range WNV transmission was concentrated mainly along the migratory bird flyways. The general power-law kernel overestimated the incidence data in some states because it was creating long-distance links in all directions. The performance for the power-law-flyway dispersal kernel model was evaluated for the three most recent years (2014-2016) when WNV was endemic in the USA. The observed case data for the 49 locations were within the range of the simulated results for 41 states for 2014 (Fig B in S1 Text), 42 states for 2015 (Fig 6), and 45 states for 2016 (Fig D in S1 Text). For all three years, the simulated results were similar to the observed data, except in Colorado, Louisiana, Mississippi, Nevada, Nebraska, North Dakota, and Washington. Nevada was over-reported for 2015 and all others were under-reported. The power law flyway dispersal kernel network model reported more WNV human incidence in Nevada than reported cases, one possible reason for over-reporting cases in Nevada has rural areas, which tend to under report human cases, whereas mosquito control districts and health departments, focused in urban areas, must test birds and mosquitoes, which explains why CDC reported WNV infected mosquitoes in 25% of counties in Nevada. The under-reported states had more human cases than predicted by the model. Under-reporting by the power-law-flyway kernel network model is likely because overwintering of the virus in some states (for example, Louisiana, Mississippi etc.), which was not considered. The overwintering infected Culex mosquitoes can stay in hibernacula such as sewers, houses, caves, and other warm areas in urban, suburban, and rural areas and initiate the outbreak in the spring. Furthermore, there may be under-reporting of cases by the model if robins are not the main reservoir species in a state, which would be predicted between gulf coast states (Louisiana and Mississippi) and northern states such as North and South Dakota and Washington. Stochastic simulations are useful tools to select the optimal future mitigation strategy after outbreaks of invasive species and pathogens. The foot-and-mouth disease (FMD) epidemics in 2001 in the United Kingdom developed by Keeling et al. [60], and mitigation strategies for pandemic influenza in the United States [61] are two well developed models with similarities to the current model. These models explore possible control measures such as culling, vaccination etc. for FMD [60], and vaccination, quarantine etc for influenza [61]. Most of these strategies can be examined with the network framework however, avian culling or vaccination for WNV control is not feasible. Vector control (or mosquito control) is a viable mitigation strategy for WNV, which is not considered by the other two models (FMD and influenza). To be applicable to any pathogens and inclusive of new mitigation methods, the mitigation strategies are non-specific and the predicted effectiveness of the mitigation methods can be adjusted to other methods. In the planning of the mitigation strategies, there is a trade-off between control measures effectiveness and their cost both monetary and loss of life. A stochastic simulation tool can decide the optimal mitigation strategy by dealing with this trade-off. Mitigation strategies for WNV were tested using the power-law-flyway dispersal kernel network model. The mosquito management measures are not specific to larvae or adults, rather simply generally accepted best practices to reduce mosquito abundance for the purpose of reducing pathogen transmission. The mitigation strategy analysis proposes supplemental measures in addition to the existing mosquito management in each state because the states had yearly reported WNV cases despite the existing management methods. To reduce WNV spread, a theoretical policy would be management in neighboring regions and not exclusively in the infected places. Although this approach can cost more at the beginning of the epidemic season however at the end, it can reduce total cost by decreasing the size of the epidemic. If management measures are applied only in the infected states, it is not possible to control the epidemic because of long-distance migratory birds. This is statewide management in a unified effort. We acknowledge that states do not conduct mosquito management in this way, but to test the spillover it was necessary to do the simulation in this way because only state-level data was available. The infected place tracing mitigation technique has been used to control other diseases (for example, FMD, influenza etc.), although their host population and control measure means are different, however, the main concept behind the mitigation techniques are similar. The findings from this research to control WNV epidemic can be useful to select optimal mitigation strategies for other pathogens. This research showed that the inclusion of directional long-distance dispersal of migratory birds improves model representations of the spatial patterns of WNV spread in the United States. The simulation of our framework in the context of long-distance directional dispersal suggested that cooperation and communication can facilitate early treatment and reduced outbreak sizes because of reduced WNV dispersal by American robins.

The underlying pattern of West Nile virus (WNV) geographic spread across the United States is not completely clear, which is a necessary step for continental or state level mitigation strategies to reduce WNV transmission. We report a network model that explains the geographic spread of WNV in the United States. West Nile virus is a mosquito-borne pathogen that infects many avian species with different movement ranges. From our research, we found that migration patterns and routes play an essential role in the WNV spatial distribution. The virus spreads in all directions at short distances because of local birds and short-distance migratory birds. However, the virus also disperses long distances along the avian migratory routes. Our model is designed to be flexible and therefore can be used to explore spreading patterns of other infectious diseases in other geographic locations.

Abstract Introduction Materials and methods Results Discussion

invertebrates medicine and health sciences pathology and laboratory medicine statistics pathogens microbiology vertebrates social sciences animals simulation and modeling viruses animal behavior rna viruses mathematics network analysis infectious disease control insect vectors zoology research and analysis methods infectious diseases computer and information sciences birds medical microbiology behavior mathematical and statistical techniques microbial pathogens monte carlo method disease vectors insects arthropoda mosquitoes psychology animal migration eukaryota west nile virus flaviviruses viral pathogens biology and life sciences species interactions physical sciences amniotes statistical methods organisms

2019

A spatio-temporal individual-based network framework for West Nile virus in the USA: Spreading pattern of West Nile virus

Immunity to the murine cytomegalovirus (MCMV) is critically dependent on the innate response for initial containment of viral replication, resolution of active infection, and proper induction of the adaptive phase of the anti-viral response. In contrast to NK cells, the Vα14 invariant natural killer T cell response to MCMV has not been examined. We found that Vα14i NK T cells become activated and produce significant levels of IFN-γ, but do not proliferate or produce IL-4 following MCMV infection. In vivo treatment with an anti-CD1d mAb and adoptive transfer of Vα14i NK T cells into MCMV-infected CD1d−/− mice demonstrate that CD1d is dispensable for Vα14i NK T cell activation. In contrast, both IFN-α/β and IL-12 are required for optimal activation. Vα14i NK T cell–derived IFN-γ is partially dependent on IFN-α/β but highly dependent on IL-12. Vα14i NK T cells contribute to the immune response to MCMV and amplify NK cell–derived IFN-γ. Importantly, mortality is increased in CD1d−/− mice in response to high dose MCMV infection when compared to heterozygote littermate controls. Collectively, these findings illustrate the plasticity of Vα14i NK T cells that act as effector T cells during bacterial infection, but have NK cell–like behavior during the innate immune response to MCMV infection. The β-herpes murine cytomegalovirus (MCMV) is a well-characterized model of viral infection that results in a non-replicative, chronic infection of immune-competent animals [1]. MCMV is a cytopathic virus that is known to readily infect peritoneal macrophages, dendritic cells (DC) and hepatocytes, inducing significant pathology in both the spleen and the liver [2]–[5]. The acute response to this virus is dependent on natural killer (NK) cell cytotoxicity and IFN-γ production, as animals deficient in perforin or IFN-γ signaling rapidly succumb to infection [4], [6]–[10]. The hepatic immune environment is greatly influenced by the resident cellular subsets and has been shown to be primarily tolerogenic [11], [12]. The major hepatic lymphocyte population in mice is a distinct family of T cells, Vα14 invariant NK T (Vα14i NK T) cells [13], [14]. Vα14i NK T cells are innate lymphocytes that display an effector memory phenotype, expressing CD69 and CD44 constitutively [15]. They are uniquely capable of rapidly producing TH1 and TH2 cytokines in response to antigenic stimulation [16]. The Vα14i NK T cell repertoire is highly restricted, characterized by a Vα14-Jα18 rearrangement with an invariant junction preferentially associated with Vβ8. 2, Vβ7, or Vβ2 [17], [18]. In response to the ligand α-galactosylceramide (α-GalCer), Vα14i NK T cells interact with and activate other immune cells including NK cells, CD8+ T cells, DCs, and macrophages [16]. This immune cell cross-talk is facilitated by direct cell-cell contact and via cytokine release [19]–[22]. Much of the functional significance of Vα14i NK T cell activation in the context of viral infection has been provided by activating the compartment prior to or concomitantly with viral introduction in animal models [23]–[25]. Although this method examines the potential contribution of activated Vα14i NK T cells, it does not examine the physiological function of these T cells in response to viral infection without exogenous stimuli. In the context of other microbial infections, the evidence for direct Vα14i NK T cell involvement is mixed, often being dependent on the type of pathogen [26]–[32]. However, there is indirect evidence that Vα14i NK T cells play a role in anti-viral immune responses. A number of groups have clearly shown that the expression of the antigen-presenting molecule CD1d is often down-regulated by viruses in a myriad of ways, including protein degradation, alterations in transcription, or endosomal sequestration [33]–[35]. Vα14i NK T cells have also been shown to be preferential targets of infection and virus-induced cell death [36], [37]. This indicates that Vα14i NK T cells may have a potential role in the anti-viral response and it is advantageous for the pathogen to prevent their activation. To directly assess the role of Vα14i NK T cells in the innate anti-viral response, their activation status was examined following MCMV infection in vivo. We found that Vα14i NK T cells up-regulate the high affinity IL-2 receptor-α, CD25, produce IFN-γ, but do not undergo proliferation. Importantly, we demonstrate that CD1d is dispensable for Vα14i NK T cell activation and cytokine release in the context of MCMV. However, IFN-α/β and IL-12 are both partially required for optimal activation of the Vα14i NK T cells in response to infection. We also show that in the absence of α-GalCer treatment, Vα14i NK T cells contribute significantly to the overall cytokine response and amplify NK cell-derived IFN-γ production. Collectively, our findings demonstrate a role for the NK T cells in innate sensing of viral pathogens in an unanticipated NK cell-like manner. Inbred C57BL/6 and B6. SJL-Ptprca/BoAiTac mice were purchased from Taconic Laboratory (Hudson, NY). B6. IL-12p40−/− mice were purchased from the Jackson Laboratory (Bar Harbor, ME). B6. CD1d−/− mice (a generous gift from Dr. L. Van Kaer, Vanderbilt University, Nashville, TN) and B6. Jα18−/− mice (kindly provided by Dr. M. Taniguchi, Riken Research Center for Allergy and Immunology, Yokohoma, Japan) were bred, crossed to the B6 (>10 generations) to generate wild-type, heterozygous, and knock-out littermates. Female IFN-α/βR1−/− mice originally generated on the 129. SvEv background and backcrossed on to the C57BL/6 background were kindly provided by Dr. M. Aguet [38] and bred in our facility. All mice, except B6 mice, were bred in pathogen-free breeding facilities at Brown University (Providence, RI). All experiments were conducted in accordance with institutional guidelines for animal care. Stocks of Smith strain MCMV salivary gland extracts were prepared as previously described [39]. Infections were initiated on day 0 with 5×104 plaque-forming units (PFU), administered via i. p. injection. For survival studies, 3×105 PFU were administered via i. p. injection. For antibody-blocking experiments, mice received blocking CD1d mAb (0. 3 mg; clone 1B1; BD Pharmingen) or rat IgG control Abs in PBS at the time of the infection. To obtain splenic lymphocytes, spleens were minced, passed through nylon mesh (Tetko, Kansas City, MO), washed once in 2% PBS-serum and cell suspensions were layered on Lympholyte-M (Cedarlane Laboratories Ltd. , Canada). Hepatic lymphocytes were prepared by mincing and passage through a 70 mm nylon cell strainer (Falcon, Franklin Lakes, NJ). After washing 3 times in 2% PBS-serum, cell suspensions were layered on a two-step discontinuous Percoll gradient (Pharmacia Fine Chemicals, Piscataway, NJ). Splenocytes and hepatic lymphocytes were collected after centrifugation for 20 min at 900×g. CD19-FITC, TCRβ-FITC, CD11b-FITC, CD11c-FITC, NK1. 1-PE, CD1d-PE, B220-PerCP-Cy5, KLRG1-allophycocyanin, CD25-APC, and TCRβ-allophycocyanin were all purchased from eBioScience (San Diego, CA). NK1. 1-PerCp-Cy5. 5, CD11b-PerCp, CD4-PerCp, CD8-PerCp, CD11c-allophycocyanin, B220-allophycocyanin, IFN-γ-allophycocyanin and isotype control were purchased from BD Pharmingen (San Diego, CA). For NK T cell identification, CD1d tetramers were obtained from the National Institute of Allergy and Infectious Disease MHC Tetramer Core Facility at Emory University (Atlanta, GA). Additionally, the following mAbs were purchased from BD Pharmingen and used for ELISA: IFN-γ mAbs (clone R4-6A2, and clone XMG1. 2), IL-4 mAbs (clone 4B11 and BVD6-24G2), IL-2 mAbs (purified JES6-N37-1A12 and biotinylated JES6-5H4) and streptavidin-peroxidase. Hepatic lymphocytes were isolated as described above from congenic C57BL/6. SJL mice. For enrichment of hepatic NK T cells, cells were first depleted of CD8+, CD11c+, CD11b+, and CD19+ cells using the AutoMACS (Miltenyi Biotec) as instructed by the manufacturer. 5–8×106 cells were transferred via tail vein injection into Jα18−/− or CD1d−/− mice. For NK T cell positive selection hepatic lymphocytes were stained with anti-NK1. 1 and anti-CD5 mAbs or with anti-NK1. 1 and CD1d tetramer and sorted using a FACSAria (BD Biosciences). At the time of transfer (1×106 cells per mouse), mice were infected with 5×104 pfu MCMV. At 1. 5 days post-infection, animals were sacrificed and the donor population analyzed for IFN-γ production. For all serum-based measurements, blood was collected via cardiac puncture. Serum was separated from the cellular fraction by centrifugation at 14,000 rpm at 4°C for 30 minutes. Serum levels of cytokines were measured by ELISA or using the cytometric bead array (CBA) mouse inflammation kit (BD Pharmingen). Following lymphocyte isolation, cells were suspended in PBS containing 2% FCS. Cells were then incubated with 2. 4G2 anti-Fc receptor mAb and stained with indicated antibodies. Cells were then fixed in 2% paraformaldehyde in PBS. Intracellular staining for IFN-γ protein was performed using the Cytofix/Cytoperm kit (BD PharMingen). Depending on the experiment and the tissue, 2. 5×105–1×106, events were collected on a FACSCalibur or FACSAria. The data were analyzed using CellQuest software or Diva software (Becton Dickinson, Franklin Lakes, NJ). Statistical significance, designated as a p-value ≤0. 05, was determined by paired, 2-tailed Student' s T-test. It is well documented that NK cells are necessary for the innate anti-viral immune response to MCMV infection [6], [40]. However, it is unclear if naïve Vα14i NK T cells also participate in this innate immune response. To address this issue, wild type B6 mice were infected for 20–40 hrs and the activation status of Vα14i NK T cells was examined in the liver and spleen. In both tissues, Vα14i NK T cells display signs of activation, a decrease in NK T cell numbers (Fig. 1A), and CD25 up-regulation by 20 hours post-infection (Fig. 1B and data not shown). CD1d dependent Ag recognition by Vα14i NK T cells induces their expansion [41]. Additionally, Vα14i NK T cells have been shown to proliferate in response to infection with LPS negative bacteria [31], [42]. However, in the context of MCMV infection, the Vα14i NK T cell compartment does not expand in either number or frequency (Fig. 1D), even at the peak of activation as assessed by CD25 expression (Fig. 1C). We also performed an intra-cellular staining for TCR at different days post MCMV infection. We found that most of the cells were double positive for intracellular and cell surface TCR, ruling out a possible lack of detection of the Vα14i NK T cells due to TCR internalization (data not shown). In contrast to Vα14i NK T cells, NK cells expand during MCMV infection (Fig. 1D). Furthermore, the percentage of CD25+ Vα14i NK T cells rapidly declines in comparison to the protracted decrease in the percent of NK cells positive for the terminal maturation marker, KLRG1 (data not shown). Vα14i NK T cells produce IFN-γ as early as 30 hours post-infection (data not shown), peaking at day 1. 5 post-infection (Fig. 2A). At this time point, the frequency of IFN-γ+ Vα14i NK T cells is comparable to the frequency of IFN-γ+ NK cells in both spleen and liver (Fig. 2B). In the spleen, despite a similar frequency, the number of IFN-γ+ Vα14i NK T cells is lower than the number of IFN-γ+ NK cells. However, the number of IFN-γ+ Vα14i NK T cells is similar to the number of IFN-γ+ NK cell in the liver at day 1. 5 post-infection (Fig. 2C). This indicates that these two subsets of cells contribute equally to the overall amount of IFN-γ in the liver. Notably, Vα14i NK T cells do not produce detectable amounts of IL-4 during MCMV infection in either tissue (data not shown). In order to investigate whether MCMV induced activation of Vα14i NK T cells requires CD1d, B6 mice were treated with a blocking CD1d mAb or control antibody and infected with MCMV. On day 1. 5 post-infection, the percentage of hepatic IFN-γ+ Vα14i NK T cells in mice that received the anti-CD1d mAb or control IgG was comparable (Fig. 3A). Similar results were observed in the spleen (data not shown). To directly assess the contribution of CD1d-mediated Ag presentation to MCMV-induced activation and cytokine production from Vα14i NK T cells, adoptive transfer experiments were performed. Negatively selected (purity >70%) or positively selected hepatic Vα14i NK T cells (purity >95%) from congenic wild-type B6. SJL mice were adoptively transferred into CD1d−/− or Jα18−/− deficient hosts. The recipient mice were simultaneously infected with MCMV for 1. 5 days and the percentage of IFN-γ+ Vα14i NK T cells was determined. Regardless of the host expression of CD1d, donor Vα14i NK T cells produced similar amounts of IFN-γ following MCMV infection in vivo (Fig. 3B & 3C). Taken together, the results indicate that CD1d is dispensable during MCMV induced activation of Vα14i NK T cells. High levels of IFN-α/β and bioactive IL-12 characterize the innate immune response to MCMV infection in vivo [43]. In the absence of either cytokine, the innate anti-viral response is fatally impaired [39], [44]. Here, MCMV infection of IFN-α/βR1−/− and IL-12p40−/− mice further reveals that the activation of Vα14i NK T cells at both 20 and 40 hours post-infection is independent of IL-12 and IFN-α/β, as assessed by the percentage of NK T cells (Fig. 4A & 4B) and CD25 expression (Fig. 4C). However, Vα14i NK T cell-derived IFN-γ is highly dependent on IL-12 in the liver (Fig. 5) and spleen (data not shown), similar to NK cells (Fig. 5). Notably, the Vα14i NK T cell-derived IFN-γ response is reduced by ∼50% in MCMV infected IFN-α/βR1−/− mice, further mimicking NK cell dynamics (Fig. 5). Activated Vα14i NK T cells interact with and activate other immune cells such as NK cells, which subsequently produce cytokines [19], [20], [45]. To investigate the downstream consequences of Vα14i NK T cell absence, we measured both NK cell-derived IFN-γ and serum inflammatory cytokines in infected NK T cell deficient mice. The activation of NK cells in the spleen, as determined by IFN-γ production, was significantly reduced in CD1d−/− mice when compared to heterozygous littermates (Fig. 6A). Similarly, although not significant, a reproducible reduction of NK cell IFN-γ was also observed in MCMV infected Jα18−/− mice compared to Jα18−/+ littermates in five independent experiments (Fig. 6A). Interestingly, an overall reduction of the inflammatory cytokine profile (IL-12, IFN-γ and TNF-α) was seen in the blood of Jα18−/− animals in comparison to heterozygous littermate controls at 1. 5 days post-MCMV infection in vivo (Fig. 6B). Likewise, inflammatory cytokines were diminished in CD1d−/− mice compared to CD1d+/− littermates. However, in the latter case, while IL-12 and TNF-α were reproducibly decreased, only IFN-γ was reduced significantly. Notably, IL-4 and IL-10 were not detectable in the serum of infected animals. To address whether Vα14i NK T cells and/or CD1d participate in the early control of MCMV infection, CD1d−/− and Jα18−/− animals, as well as littermate controls were used in survival studies with high dose MCMV infection. CD1d−/− mice were more susceptible than CD1d+/− mice, as only 50% of the CD1d−/− mice lived beyond day 15. While Jα18−/− animals were not more susceptible than their littermate, heterozygous controls (Fig. 6), they were significantly more susceptible than wild-type B6 mice from outside vendors (not shown). Taken together, these results demonstrate that Vα14i NK T cells influence NK cell activity, the inflammatory cytokine profiles, and that both Vα14i NK T cells and other CD1d restricted T cells are necessary for an optimal immune response to MCMV. The function of NK cells in anti-viral immunity has been documented; however, evidence for a direct Vα14i NK T cell role has not been examined extensively. The results presented in this report show that Vα14i NK T cells sense MCMV infection in vivo, without exogenous stimuli, such as α-GalCer. However, in contrast to bacterial infection, we provide evidence that MCMV induced Vα14i NK T cell activation is TCR independent. Vα14i NK T cells can be activated directly by agonist glycolipids presented by CD1d. For instance, α-GalCer immunization of B6 mice leads to IL-12 independent activation of Vα14i NK T cells [16]. In this case, Vα14i NK T cells release copious amount of IL-4 and IFN-γ and subsequently proliferate. Gram-negative LPS-negative α-proteobacteria, such as Sphingomonas, Ehrlichia, Rickettsia, and Borrelia, express such agonist lipids and can directly activate Vα14i NK T cells [31], [32], [42], [46]. Bacteria that do not express agonist glycolipids have been reported to activate Vα14i NK T cells through up-regulation of self glycolipids and/or IL-12 production through recognition of endogenous lysosomal glycosphingolipids, such as iGb3, presented by LPS-activated dendritic cells [31], [47]. Therefore, there are two major mechanisms for the activation of Vα14i NK T cells against bacteria either via cognate Ag or via self-Ag with APC derived cytokines [48]. Using MCMV infection in vivo, we now demonstrate a novel activation pathway for Vα14i NK T cells, mediated principally by inflammatory cytokines. MCMV induced activation of Vα14i NK T cells clearly differs from the two mechanisms described in response to bacterial infection. First, as opposed to α-GalCer administration [41], [49] and α-proteobacteria infection [31], Vα14i NK T cells do not proliferate nor produce IL-4 following MCMV-induced activation. Second, while IL-12 is not required for optimal stimulation of Vα14i NK T cells in response to α-GalCer or sphingomonas-derived glycolipids [20], [42], here we show that Vα14i NK T cell cytokine production is impaired in IL-12 deficient animals in response to MCMV infection. Finally, in vivo CD1d blocking experiments and adoptive transfer of Vα14i NK T cells into CD1d−/− mice demonstrates that CD1d is dispensable for MCMV induced activation of these lymphocytes. It should be noted that LPS-induced Vα14i NK T cell-derived IFN-γ in vitro does not require CD1d-mediated Ag presentation, instead exposure to IL-12 and IL-18 is sufficient to activate these cells [50]. These data raise the question of why do bacteria such as Salmonella typhimurium and Staphylococcus aureus activate Vα14i NK T cells in both an IL-12 and CD1d dependent manner, while MCMV induced activation of Vα14i NK T cells is CD1d independent? There are several non-mutually exclusive possibilities that could explain this apparent discrepancy. First, viruses unlike bacteria, do not encode enzymatic machinery for lipid synthesis. Second, the peak of the cytokine response to MCMV occurs relatively early when compared to bacteria, allowing for possible Vα14i NK T cell activation to occur prior to cytokine-driven self-Ag activation. Third, while Gram negative bacteria cytokine-driven self-Ag activation of Vα14i NK T cells was demonstrated in vitro using bone marrow derived DCs [31], [47], it has been recently demonstrated that plasmacytoid dendritic cells (pDCs) are the quasi-exclusive source of IFN-α/β, IL-12 and TNF-α early during MCMV infection [51]. It is therefore possible that depending on the pathogen and the source and/or phenotype of recruited CD1d+ DCs may lead to differential activation of Vα14i NK T cells. pDCs and dendritic cells recognize MCMV through TLR9, an essential component of the innate immune defense against MCMV. Tabeta et al have shown that the Vα14i NK T cell response to MCMV is impaired in TLR9−/− mice [52]. Interestingly, the serum level of both IFN-α and IL-12 is reduced in TLR9−/− mice following MCMV infection [52], [53], supporting our findings that the absence of these cytokines impairs the Vα14i NK T cell response to MCMV. However, activation of Vα14i NK T cells by TLR9-stimulated dendritic cells was recently shown to be CD1d dependent [54]. The latter study was performed in vitro using BMDCs grown for 14 days prior to being pulsed with CpG for 16 hours. This procedure clearly differs from MCMV infection in vivo where the peak of the Vα14i NK T cell response is at 1. 5 days post-infection. It is possible that some pathogen-derived products such as CpG may increase endogenous glycolipid presentation during the anti-viral response but that this process is not yet initiated at the peak of the innate response to MCMV. In the context of MCMV infection, we failed to detect an expansion of the Vα14i NK T cells. Instead, there is a gradual loss of these cells in the liver and spleen following infection. Presumably, the lack of TCR engagement by CD1d during MCMV infection promotes the activation of Vα14i NK T cells that release cytokines but do not proliferate and subsequently die. It is also possible that Vα14i NK T cells preferentially undergo virus-induced apoptosis similarly to what has been reported during the anti-viral response against lymphocytic choriomeningitis virus infection in vivo [36]. The early immune response to MCMV is characterized by the production of high levels of inflammatory cytokines [51], [55]. Type I IFNs, which are critical for anti-viral immunity, can be detected very early following MCMV infection and mediate the proliferation and survival of activated lymphocytes [56]. Additionally, the classical TH1-promoting cytokine, IL-12, is also produced early and is necessary for NK cell-derived IFN-γ [57]. Infection of mice deficient in either IFN-α/β signaling or bioactive IL-12 clearly demonstrates that neither cytokine alone is sufficient to mediate an optimal IFN-γ response from Vα14i NK T cells in response to MCMV. IFN-α/β is thought to negatively regulate the production of IFN-γ via inhibition of IL-12 thus ensuring that IFN-γ does not prematurely inhibit proliferation [58]. The timing of IL-12 production may be critical for Vα14i NK T cells as MCMV infected IFN-α/βR1−/− mice produce high levels of IL-12 [56], yet we show that Vα14i NK T cell-derived IFN-γ is impaired in IFN-α/βR1−/− mice. It is also currently unclear if IFN-α/β acts directly on this innate T cell population in the context of MCMV infection. These issues warrant further inquiry. Vα14i NK T cells are widely appreciated for their rapid cytokine production and ability to interact with and activate both innate and adaptive immune cells [15]. The cross-talk between Vα14i NK T cells and NK cells in the context of α-GalCer-mediated stimulation requires IFN-γ and IL-12 production to promote optimal NK cell activation [19], [20]. Vα14i NK T cell-mediated activation of NK cells has been shown to be required for anti-tumor immunity [59], [60] but not for viral infection. We show that in the absence of the Vα14i NK T cell population, the NK cell response to MCMV in vivo is impaired. Notably, Vα14i NK T cells also mediate activation and maturation of DCs and macrophages [61], [62], cells critical for the induction of the anti-viral immune response via production of high levels of key inflammatory cytokines such as IFN-α/β and IL-12 [62]. We are currently investigating the possibility that the function of these subsets in response to MCMV infection may also be altered in the absence of Vα14i NK T cells. Our data suggest that Vα14i NK T cells contribute to the overall immune response to MCMV. However, it has been shown that at a low infecting dose, Jα18−/− mice and B6 wild-type control animals have equivalent viral titers [25]. Using Vα14i NK T cell deficient mice and littermate controls, we examined the pathological impact of Vα14i NK T cell absence during high dose MCMV challenge. While both Jα18−/− and CD1d−/− animals are more susceptible than wild-type B6 mice from outside vendors, only CD1d−/− mice are less resistant than heterozygous littermate control mice to high dose MCMV infection. This suggests that Vα14i NK T cells as well as other CD1d restricted T cells are required for an optimal immune response to MCMV. Collectively, the findings presented in this report indicate that Vα14i NK T cells actively participate in the innate immune response to MCMV in vivo and directly impact the quality of the immune response. However, the mechanism of their activation differs from bacterial induced activation and in this case, NK T cell functions mirror NK cell functions. These data define a previously unappreciated role for CD1d restricted T cells in anti-viral immunity and provide additional insight into the affect of innate immune manipulation on the overall outcome of an immune response.

An efficient immune response to viral infection requires both innate and adaptive immune cells. Natural killer (NK) cells are a critical innate cellular component of the immune response to murine cytomegalovirus (MCMV). Natural killer T (NK T) cells are non-classical T cells that have the potential to bridge the two arms of the immune system. However, the contribution of NK T cells to the anti-viral immune response has not been extensively studied. In the absence of additional stimuli, NK T cells actively participate in the immune response to MCMV infection. Interestingly, in contrast to their response to bacteria, we demonstrate that only the innate NK T cell arm is activated during viral infection while the adaptive branch, TCR engagement by CD1d, is dispensable. NK T cells display signs of activation in response to viral infection, increased expression of CD25, a rapid decrease in cell number, and production of the anti-viral cytokine IFN-γ. The NK T cell response to MCMV also influences the NK cell activity and the inflammatory cytokine profiles. Understanding the physiological function of these unique T cells in the context of infection will aid in the development of novel therapeutic and preventive treatments for viral infections.

Abstract Introduction Materials and Methods Results Discussion

infectious diseases immunology/immune response immunology/innate immunity infectious diseases/viral infections immunology/immunity to infections

2008

NK Cell–Like Behavior of Vα14i NK T Cells during MCMV Infection

The Vibrio cholerae lipopolysaccharide O1 antigen is a major target of bacteriophages and the human immune system and is of critical importance for vaccine design. We used an O1-specific lytic bacteriophage as a tool to probe the capacity of V. cholerae to alter its O1 antigen and identified a novel mechanism by which this organism can modulate O antigen expression and exhibit intra-strain heterogeneity. We identified two phase variable genes required for O1 antigen biosynthesis, manA and wbeL. manA resides outside of the previously recognized O1 antigen biosynthetic locus, and encodes for a phosphomannose isomerase critical for the initial step in O1 antigen biosynthesis. We determined that manA and wbeL phase variants are attenuated for virulence, providing functional evidence to further support the critical role of the O1 antigen for infectivity. We provide the first report of phase variation modulating O1 antigen expression in V. cholerae, and show that the maintenance of these phase variable loci is an important means by which this facultative pathogen can generate the diverse subpopulations of cells needed for infecting the host intestinal tract and for escaping predation by an O1-specific phage. Lipopolysaccharide (LPS) is a prominent constituent of the outer membrane of Gram-negative bacteria. The LPS molecule is divided into three components; lipid A, core oligosaccharide and O-specific polysaccharide (or O antigen). The structure of the O antigen typically defines the serogroup of an organism, and over 200 serogroups of Vibrio cholerae are currently recognized [1]. Interestingly, the O1 serogroup has been and continues to be the dominant cause of both endemic and epidemic cholera throughout the world, though the reasons for this are unknown. The incidence of cholera worldwide is steadily increasing, and the cumulative number of reported cases in 2010 was nearly double what it was in 2009 [2]. When considering the level of gross under-reporting, the actual global disease burden is estimated to be 3–5 million cases and more than 100,000 deaths [3], [4]. The observed increase in reported cases in 2010 is largely due to a recent outbreak of an O1 strain that started in Haiti: Even more concerning is the observation that 53% of the global total of the number of reported deaths from cholera in 2010 occurred in Haiti in a period of only 70 days [2]. These observations highlight the fragile nature of impoverished and tragedy-struck nations to the rapid onset of cholera epidemics. The V. cholerae O1 antigen is composed of 12–18 repeating units of α (1,2) -linked d-perosamine (4-amino-4,6-dideoxy d-mannose) residues, the amino groups of which are acylated with tetronate (3-deoxy-l-glycero-tetronic acid) (Fig. S1) [1], [5]–[8]. The genes currently described as being required for the synthesis of the O1 antigen are located on chromosome 1 of the V. cholerae O1 N16961 genome between open reading frames (ORFs) VC0240 (gmhD) and VC0264 (rjg) (Fig. 1A) [9]. This region (the wbe or rfb region) was originally identified through the heterologous expression of the V. cholerae O1 antigen in Escherichia coli K-12 [10]. Additional genes required for the synthesis of the O1 antigen in V. cholerae were subsequently identified [11], however all O1 antigen biosynthetic genes studied to date have been between the gmhD and rjg flanking genes. The genes responsible for O1 antigen biosynthesis have been placed into the following five groups according to putative function: perosamine biosynthesis (VC0241–VC0244) [12]; O antigen transport (VC0246–VC0247) [13]; tetronate biosynthesis (VC0248–VC0252) [14]; O antigen modification (VC0258) [15], [16]; and additional genes essential for O antigen biosynthesis (VC0259–VC0260, VC0263) [11] (Fig. 1A). A putative pathway for the biosynthesis of perosamine has been proposed by Stroeher et al. [12] (Fig. 1B). In this pathway, which is based solely on homology comparisons, the first step is the conversion of fructose-6-phosphate (F6P) to mannose-6-phosphate (M6P) by ManC (a predicted type II phosphomannose isomerase [PMI]). M6P is then converted to mannose-1-phosphate (M1P) by ManB, and then to GDP-mannose by ManC. GDP-mannose is converted to GDP-4-keto-6 deoxymannose by WbeD and then GDP-perosamine by WbeE. PMIs (E. C. 5. 3. 1. 8) catalyze the reversible isomerization of M6P to F6P and are divided into three families on the basis of amino acid sequence [17]. Type I PMIs are monofunctional enzymes and include proteins from humans to bacteria including E. coli [18] and Salmonella enterica serovar Typhimurium [19]. Type II enzymes are bacterial bifunctional enzymes possessing both PMI and guanosine diphospho-d-mannose pyrophosphorylase (GMP) activity (for the conversion of M1P into GDP-mannose) in distinct catalytic domains [20]. PMIs play critical roles in mannose catabolism and in the supply of GDP-mannose, which is necessary for the mannosylation of various structures including LPS. The distal location of the O antigen extending outward from the bacterial surface positions it at the interface between the bacterium and its environment. As such, the O antigen is important for protection from various environmental stresses including antibiotics and the host immune response [21], [22]. The O antigen is also consequently the target of both the immune system and bacteriophages, which can independently apply powerful selective forces. As such, cell surface structures, including the O antigen, are frequently observed to exhibit high levels of variation [23]. Examples of phase variable surface structures are abundant in bacterial pathogens and include Haemophilus influenzae lipooligosaccharide (LOS) [24]–[26], Neiserria meningitidis LOS [27], Helicobacter pylori LPS [28], [29] and Campylobacter jejuni LOS [30], [31]. The loci responsible for phase variable expression of these structures, often referred to as contingency loci, are thought to offer a preemptive strategy to increase diversity necessary for bacterial adaptation in unpredictable environments [32]. Phase variation can be mediated by DNA polymerase slipped-strand mispairing across simple sequence repeats, and when located in coding sequences, can lead to a frameshift mutation resulting in the production of truncated, often nonfunctional, peptide. Homopolymeric nucleotide tracts are one subset of simple sequence repeats commonly observed to undergo frequent expansion and contraction resulting in reversible heritable phenotypic variation [23], [32]. One variation of the V. cholerae O1 antigen that has been demonstrated and which defines the two serotypes, Ogawa and Inaba, is the presence or absence, respectively, of a terminal methyl group [16]. The two serotypes can undergo serotype conversion during epidemics or in endemic areas [33]–[38]. Spontaneous mutations in the predicted methylase wbeT (VC0258) are linked to this switching phenotype [15] and may be involved in immune evasion as cross-serotype protection is limited. In contrast to the immune pressure being somewhat specific for a given serotype, bacteriophages that target the O1 antigen for use as a receptor may not be serotype-specific. In this regard, it has recently been reported that the absorption of several different O1-specific phages to V. cholerae can be modulated by its cyclic AMP (cAMP) -cAMP receptor protein regulatory system, suggesting that regulatory pathways may exist that alter O1 antigen abundance or surface organization [39]. We recently described ICP1, an O1-specific, but serotype nonspecific V. cholerae phage that is prevalent in cholera patient stool samples in the cholera endemic region of Bangladesh [40]. ICP1 is likely related to the previously described phage JSF4 [41]. We sought to investigate the mechanisms employed by pathogenic V. cholerae O1 to resist ICP1 infection and discovered two phase variation mechanisms by which V. cholerae O1 displays intra-strain O antigen heterogeneity. This heterogeneity is mediated by two contingency loci involved in tetronate biosynthesis (wbeL) and a previously unrecognized PMI (manA) critical for perosamine biosynthesis. Plaques resulting from the infection of a wild type V. cholerae O1 strain with the O1-specific phage ICP1 were routinely observed to have colonies growing in the center indicating the presence of phage resistant isolates. Four independent phage resistant isolates were subjected to whole genome resequencing and the majority of the isolates (three of four) had single nucleotide deletions that mapped to homonucleotide (poly-A) tracts within two genes. Two mutants were found to have a deletion in the poly-A (A8) tract starting at nucleotide position 108 in wbeL, a gene that is predicted to be required for tetronate synthesis [14]. The full-length WbeL protein is 471 amino acids and a single nucleotide deletion within the poly-A tract results in the production of a truncated peptide of 42 amino acids due to a premature stop codon 12 nucleotides downstream of the poly-A tract (Fig. S2). wbeL is unique to V. cholerae O1 strains, and the poly-A tract is 100% conserved in all 37 V. cholerae O1 strains available for analysis through the National Center for Biotechnology Information (NCBI) DNA sequence database. Purified LPS from a wbeL* (A7) phase variant shows a distinct lower molecular weight pattern on a silver stained SDS-PAGE gel (Fig. 2), although the strain exhibits a normal slide agglutination phenotype with anti-Ogawa typing serum (Table 1). The wbeL* strain is completely resistant to infection with ICP1, and accordingly purified LPS from this strain shows no inhibition of ICP1 (Table 1). To test the possibility that the wbeL* A7 frame shift exerts a polar effect on downstream genes, which also contribute to tetronate biosynthesis (Fig. 1B), we performed complementation analysis and found that wild type LPS production and ICP1 sensitivity are restored when the wbeL* mutant is complemented with wbeL in trans (Fig. 2). This indicates that the observed phenotypes are a consequence of the loss of WbeL expression and not due to polar effects of the wbeL* mutation. To further confirm this, an in-frame deletion in wbeL was constructed and complementation analyses were performed. The ΔwbeL strain is devoid of O1 antigen (Fig. 2) and accordingly exhibited no agglutination with anti-Ogawa typing serum (Table 1), and these phenotypes can be complemented with wbeL in trans (data not shown). Since the phenotype of the in-frame deletion mutant is not consistent with the original wbeL* mutation, we hypothesized that the wbeL* allele maintains some O1 antigen biosynthetic function in the cell. Consistent with this, when the ΔwbeL strain is complemented with the wbeL* allele in trans, agglutination in the presence of anti-Ogawa typing serum is restored while the strain maintains complete resistance to ICP1, just like the original wbeL* strain (Table 1). The ΔwbeL mutant expressing wbeL* in trans also produces a small amount of lower molecular weight LPS like the original wbeL* strain (data not shown). To further address the biosynthetic function of the wbeL* allele, we explored two possible explanations for the phenotype of the wbeL* phase variant. First, that the presence of the truncated 42 amino acid peptide produced by the wbeL* allele is necessary and sufficient to allow V. cholerae O1 to elaborate the lower molecular weight LPS observed (Fig. 2). To test this hypothesis we constructed a deletion in the remainder of the coding sequence downstream of the premature stop codon in wbeL*. Purified LPS from a wbeL* strain expressing only the 42 amino acid peptide lacks O1 antigen substituted LPS (Fig. 2), which rules out this hypothesis. The second hypothesis that we tested to account for the biosynthetic role of the wbeL* allele is that it allows some functional WbeL protein to be made because the wbeL* allele is subject to nonstandard decoding (ribosomal frame shifting or transcriptional slippage at the A7 tract). This in turn would allow for a small amount of tetronate modification to occur, resulting in a lower molecular weight but still compositionally (and antigenically) normal O antigen (which is consistent with the observation that the wbeL* phase variant agglutinates in the presence of anti-Ogawa typing serum (Table 1). To test this hypothesis, we made silent point mutations within the A7 tract in wbeL* (A7 to TAAGAAA) designed to prevent non-standard decoding as well as prevent further slipped-strand mispairing during replication and characterized the ability of this strain (referred to as wbeL* PL for phase-locked) to produce O1 antigen substituted LPS. The point mutations in wbeL* PL abolished the ability of this strain to make O antigen substituted LPS, which supports the conclusion that the wbeL* allele maintains biosynthetic function by allowing some functional WbeL protein to be produced through nonstandard decoding, and that this is dependent on the A7 tract. The mechanism responsible is not currently known, nor is it known if there are other sequence motifs within the wbeL* allele that facilitate the +1 frameshifting needed to restore the reading frame. It is important to emphasize that the observed lower molecular weight pattern of purified LPS from wbeL* is not due to reversion of A7 to A8 in the genome of a substantial subset of the population, because if that were the case we would observe a small amount of wild type length O1 antigen and not the unique species observed for the wbeL* strain. These results further suggest that tetronate acylation of the perosamine backbone is necessary for incorporation of the O1 antigen into the LPS molecule, and this may be required for recognition and subsequent transport of the undecaprenyl-linked O antigen polymer to the periplasm by the ABC transporter and/or for efficient ligation of the O antigen polymer to the lipid A-core by WaaL ligase [42]. Importantly, these data demonstrate that the lower molecular weight O1 antigen produced by the wbeL* mutant somehow endows V. cholerae with full resistance to ICP1. The second mutation identified in ICP1-resistant colonies from the center of plaques also mapped to a poly-A tract, and this localized to VC0269, which we designate manA for reasons explained below. manA has two poly-A (A9) tracts and, as illustrated in Fig. 1A, is located approximately 4 kbp downstream of the right junction gene (rjg = VC0264), which was thought to delineate the end of the O1 antigen biosynthetic locus in V. cholerae. ManA shows homology to type I PMIs that catalyze the reversible isomerization of F6P to M6P, which is the first step in perosamine biosynthesis (Fig. 1B). Like wbeL, manA is also specific to V. cholerae O1 strains: All 37 V. cholerae O1 strains available for bioinformatic analysis have manA and it is highly conserved between these strains with the exception of two strains which have the manA* (A8) allele resulting from a single nucleotide deletion in the first poly-A tract (these latter strains are designated 2740–80 and HC-61A1). The full length ManA protein is 399 amino acids, and a single nucleotide deletion within the first poly-A tract is predicted to result in a truncated peptide 81 amino acids long, while a single nucleotide deletion within the downstream poly-A tract (designated manA*) produces a truncated peptide 207 amino acids long (Fig. S3). LPS purified from manA* after overnight growth looks identical to the parental strain (Fig. 2). However, purified LPS from the manA* strain does not bind and inhibit ICP1 as efficiently as a equal amount of LPS purified from the parental strain (Table 1), which is consistent with this mutant being partially phage resistant (Fig. 2 and Table 1). These results leave open the possibility that while manA* produces wild type length O1 antigen, the overall abundance of O1 antigen substituted LPS is less than wild type, and this translates into fewer available receptors for ICP1. Phage infection is a complex process, which can require sequential receptor binding steps. For example T4 infection is initiated by the reversible attachment of at least three of the six long tail fibers to the outer core of the LPS, but does not result in DNA injection unless the six short tail fibers successfully engage their receptors on the inner core of the LPS [43]–[45]. Throughout the course of this study isolates were obtained with a frameshift in either poly-A tract in manA and these isolates were phenotypically indistinguishable from one another with regard to phage sensitivity and agglutination with anti-Ogawa typing serum (data not shown). Similarly, a strain harboring an in-frame deletion of manA was phenotypically indistinguishable from the manA phase variants (Fig. 2 and Table 1), indicating that both frame-shift mutations function as manA nulls. manA does not appear to be required for producing full length O antigen as indicated by SDS-PAGE and silver staining of LPS (Fig. 2), and yet the manA* strain is partially resistant to ICP1, which specifically requires the O1 antigen for infection. These observations led us to speculate that there is likely another gene that can contribute to the conversion of the F6P to M6P in V. cholerae O1. We noted the presence of another annotated type I PMI in the V. cholerae O1 genome, VC1827, hereafter designated manA-2. manA-2 is located immediately downstream of a mannose permease encoded by VC1826 [46], but is not linked to the O1 antigen biosynthetic cluster. manA and manA-2 are 65% identical at the nucleotide level over 70% of their sequence, and at the protein level, they are 59% identical over 97% of their sequence. manA-2 has a higher GC content than manA (46. 1% compared to 42. 1%, respectively) and its GC content is much closer to the overall GC content of the entire V. cholerae N16961 genome (∼47% [47]) suggesting manA may have been recently horizontally acquired. Unlike manA, manA-2 is found in non-O1 V. cholerae. In other Gram-negatives, including E. coli and S. Typhimurium, the manA gene generally maps as an independent gene not associated with the LPS gene cluster due to its role in mannose metabolism [48], [49], although in these organisms the PMI activity of the unlinked manA is required for O antigen synthesis [50], [51]. In organisms that do not metabolize mannose, the manA gene is generally absent [49]. Some bacteria have a bifunctional type II PMI-GMP (typically referred to as ManC) encoded in the LPS biosynthetic cluster. Monofunctional and bifunctional forms of ManC are not clearly distinguishable on the basis of size or even sequence similarity [49]. Indeed ManC (VC0241) in V. cholerae O1 has the bioinformatic designation of a type II PMI, however our results below suggest this enzyme lacks PMI activity. The presence of two putative type I PMIs in V. cholerae O1 led us to investigate the phenotypes of single and double mutants with regards to O antigen biosynthesis. A mutant lacking manA-2 was indistinguishable from wild type with regard to phage sensitivity and LPS pattern on a gel (Fig. 2 and Table 1), indicating that manA-2 is not important for O antigen biosynthesis under the conditions tested. However, in the absence of manA, manA-2 becomes important since the manA* ΔmanA-2 double mutant is completely phage resistant and produces very little fully length LPS (Fig. 2). Another potential source of M6P for perosamine biosynthesis is through the conversion of exogenously acquired mannose to M6P by a hexokinase, which led us to determine if trace mannose present in the growth media (Luria-Bertani [LB] broth) contributed to the small amount of O1 antigen still visible in the manA* ΔmanA-2 double mutant. Indeed, this small amount of O1 antigen substituted LPS is absent when the double mutant is grown in M9-glucose, but is present when the double mutant is grown in M9-glucose plus mannose (Fig. 2), demonstrating that exogenous mannose is responsible for the small amount of O1 antigen substituted LPS in the double mutant. Similar results have been observed in E. coli and S. Typhimurium manA mutants, which are unable to synthesize O antigen without the inclusion of mannose in the growth media [50], [51]. Also consistent with these results, we observed that our double manA* ΔmanA-2 mutant is unable to grow in M9-mannose owing to the critical nature of type I PMIs in the conversion of M6P to F6P as a substrate for glycolysis. These results indicate that manC in V. cholerae O1, which was hypothesized to catalyze the first reaction in biosynthesis of perosamine (Fig. 1B) [12], is not active as a bifunctional PMI-GMP, and is likely only important for the later steps in perosamine biosynthesis for converting M1P to GDP-mannose. As type II PMIs possess two catalytically distinct domains for each PMI and GMP activity [20], attempts were made to construct mutant manC alleles that should be defective for PMI activity alone (data not shown) but all constructs resulted in complete depletion of the O1 antigen, suggesting that these mutations had an inadvertent negative impact on the GMP activity of the protein. The reason behind V. cholerae O1 possessing two type I PMIs is thus not clear, although it suggests that they each have their primary roles: manA in O antigen biosynthesis and manA-2 in mannose metabolism, though it is unclear what factors define those functional roles. A search for other bacteria harboring multiple annotated type I PMIs reveals a limited number of organisms including strains of Vibrio vulnificus (Accession No. NC_005140), Vibrio parahaemolyticus (NC_004605), and Yersinia enterocolitica (NC_015224); however, to our knowledge, the functional roles that these enzymes have in these other organisms are not known. As mentioned previously, the O1 antigen biosynthetic cluster was originally identified through the heterologous expression of the V. cholerae O1 antigen in E. coli [10], and all O1 antigen biosynthetic genes studied thus far have been between the gmhD and rjg flanking genes. manA was likely not identified as part of this biosynthetic pathway because its function was complemented by the E. coli manA gene, permitting expression of the O1 antigen in this host. Additional genes required for O1 antigen biosynthesis were subsequently identified following the initial report by Manning et al. [11], and were similarly likely missed because the phenotype was masked in E. coli. Blokesch and Schoolnik [52] provided some additional support that further extends the O1 wbe region downstream of rjg. They observed that when serogroup conversion of V. cholerae O1 to O139 occurred through uptake of O139 donor DNA during natural transformation, the crossovers were often localized within or downstream of VC0271 at the right junction, and the location of the left junction was within or upstream of gmhD. These results coupled with our observation that manA participates in O1 antigen biosynthesis suggests that the wbe region extends approximately 8 kbp downstream of rjg (Fig. 1B). There are six genes currently annotated in addition to manA in this region (Fig. 1A), however it remains to be seen if these other genes do in fact participate in O1 antigen biosynthesis. The ability of wbeL* and manA* phase variants to colonize the small intestine was assessed in competition assays in the infant mouse model. The wbeL* strain is attenuated over 1000-fold (Fig. 3). We did not anticipate such a high level of attenuation given that this strain still elaborates LPS (although it is a lower molecular weight form, Fig. 2). Interestingly, the altered LPS produced by the wbeL* strain does provide some advantage over not having any O1 antigen, as is apparent by the significantly lower competitive index (CI) for the ΔwbeL strain (p<0. 05, Mann-Whitney U test). To rule out secondary mutations, we chose to use revertant strains in vivo to avoid potential complications concerning plasmid loss and non-wild type gene expression levels during infection. The colonization defect observed for the wbeL* strain is absent when the poly-A (A7) tract is reverted to wild type length (A8) (Fig. 3), demonstrating that the virulence defect is due to the wbeL* allele. The manA* phase variant, which produces an apparently full length LPS but which is partially resistant to ICP1, is over ten-fold attenuated for colonization, and this defect is absent when the A8 tract is reverted to wild type length (A9) (Fig. 3). We anticipated that a manA* ΔmanA-2 double mutant would be significantly more attenuated owing to a major reduction in O1 antigen produced by the strain (Fig. 2). However, although we did observe a further decrease in the CI for manA* ΔmanA-2 compared to manA*, it was not significant (p = 0. 42) (Fig. 3). To try and explain this, we hypothesized that the selective pressure exerted on the manA* ΔmanA-2 double mutant in the small intestine is sufficient to select for spontaneous revertants which would potentiate the observed moderate drop in CI. To test this, we patched manA* ΔmanA-2 isolates recovered from mouse intestines onto M9-mannose agar plates. Consistent with our hypothesis, roughly half of isolates recovered from mice after 24 h of infection had regained the ability to grow on mannose, and subsequent sequencing revealed that these strains had reverted to the wild type, in-frame poly-A tract in manA. Conversely, we were unable to detect spontaneous revertants of the wbeL* or manA* strains among the isolates recovered from mouse intestines infected with those strains, suggesting that the A7 wbeL* allele is less prone to reversion by slipped-strand mispairing, and that the A8 allele in manA* is not under enough selective pressure to revert as long as ManA-2 is functional. Our inability to detect wbeL* revertants is likely due to the experimental limitation that there are so few wbeL* isolates recovered from infected mice that we were only able to test ∼100 CFU. The manA revertants recovered from intestines of mice infected with the manA* ΔmanA-2 strain were effectively equivalent to ΔmanA-2, which we previously observed has no effect on LPS biosynthesis or virulence (Fig. 2 and Fig. 3). Confirming this, we also competed the double mutant with in-frame deletions in both manA and manA-2 and observed the anticipated significant increase in attenuation (>2000-fold) compared to a manA* phase variant alone (p<0. 05 Mann-Whitney U test) (Fig. 3). Furthermore, the double ΔmanA ΔmanA-2 mutant is significantly more attenuated when grown in M9-glucose prior to infection than when it is grown in LB which contains trace mannose (p<0. 05 Mann-Whitney U test). Consistent with this, a comparable CI was observed for the two strains that are devoid of all O1 antigen prior to infection (ΔmanA ΔmanA-2 grown without exogenous mannose, and ΔwbeL) (Fig. 3). In vitro control competitions revealed that the observed defects were specific to the in vivo environment, with the exception of the ΔwbeL mutant, which has a ∼10-fold defect in vitro compared to the wild type (data not shown). The nature of this defect is not known, but it could be due to the build-up of O antigen and/or LPS intermediate products in this strain which, unlike the wbeL* phase variant, is unable to elaborate any O1 antigen on the surface. Consistent with this, mutants harboring transposon insertions in wbeL, which also showed a decreased colonization phenotype in infant mice, also exhibited a decreased CI in in vitro control competitions [53]. Human intestinal epithelial cells produce antimicrobial peptides that are critical components of the host innate defense mechanism [54]. Antimicrobial peptides are inherently structured to target the membrane of bacteria because they are highly basic and have a substantial portion of hydrophobic residues [55], [56]. The net positive charge of these peptides facilitates their electrostatic interaction with negatively charged phospholipid groups or the lipid A anchor of LPS on the Gram-positive or Gram-negative bacterial membranes, respectively, allowing them to induce lysis and bacterial cell death. In order for an antimicrobial peptide to gain access to the Gram-negative outer membrane it must first traverse the barrier presented by the sugar chains of the O antigen layer. Perturbations to the LPS have been shown previously to alter resistance of V. cholerae to antimicrobial peptides; specifically, Matson et al. [57] and Hankins et al. [58] have investigated the structural importance of lipid A with regards to peptide resistance, and Nesper et al. [59] suggested that mutations affecting the LPS core oligosaccharide have a more dramatic affect on antimicrobial peptide resistance than mutations affecting O1 antigen biosynthesis (although only a rough strain was tested in those experiments). We investigated the susceptibility of wbeL* and manA* phase variants and their trans-complemented derivatives to the antimicrobial peptide polymyxin B by determining their survival in killing assays. We observed that the wbeL* mutant exhibited a very low level of survival (Fig. 4A). This phenotype could be complemented and survival levels could be restored to wild type when this mutant was expressing wbeL in trans, again supporting the previous data suggesting there is no polar effect of wbeL* on genes downstream. With regards to the manA* phase variant, we observed polymyxin sensitivity, but in a growth phase-dependent manner. There was intermediate sensitivity of this strain to polymyxin B when the inoculum used for the killing assay was grown up to early exponential phase (OD600 = 0. 15) (Fig. 4A). In contrast, when manA* was grown up to mid-exponential phase prior to treatment with the peptide, wild type levels of survival were observed (Fig. 4A). The phenotype observed at early exponential phase with the manA* phase variant could be fully complemented by expressing manA in trans (Fig. 4A), again indicating the observed phenotypes are due to the loss of ManA. We had also observed that phage sensitivity of the manA* phase variant was growth phase-dependent, and these results parallel the observations in the antimicrobial peptide assay, that is at OD600 = 0. 15, manA* is completely resistant to ICP1 and at OD600≥0. 2, turbid plaques result from ICP1 infection (data not shown). To address this puzzling observation, we purified LPS from manA* and wild type at early and mid-exponential growth phase (OD600 = 0. 15 and OD600 = 0. 5, respectively). In contrast to the seemingly wild type appearance of purified LPS from manA* after overnight growth (Fig. 2) and at mid-exponential phase, the LPS pattern of manA* isolated at early exponential growth phase showed very little O1 antigen substituted LPS (Fig. 4B). Since analysis of the double manA* ΔmanA-2 mutant indicated that manA-2 is important for O1 antigen synthesis only in the absence of manA, we interpret these results to suggest the compensatory activity of ManA-2 is incomplete during early exponential phase growth in LB broth. The reason for this is not known, but may relate to differences in expression, activity or localization of ManA-2. In any event, a complete functional redundancy between ManA and ManA-2 would have been at odds with the observation that all V. cholerae O1 strains have a phase variable manA gene: Specifically, the evolution of a contingency locus would be futile if the encoded protein exhibited complete functional redundancy with a non-phase variable gene. In general, the results of the polymyxin B killing assays (Fig. 4A) parallel the observed in vivo colonization defects (Fig. 3). Strains that produce less O1 antigen substituted LPS (such as wbeL* and ΔmanA ΔmanA-2) are highly susceptible to polymyxin B and are more severely attenuated in vivo than manA*, which accordingly is less susceptible to polymyxin B. These data suggest that the phase variants are defective for in vivo colonization because they are more susceptible the antimicrobial peptides present in the intestinal tract. V. cholerae O1 persists in the environment as a member of the aquatic ecosystem where it is thought to associate with and use the chitinous exoskeletons of zooplankton as a nutrient source [60]. The levels of V. cholerae O1 phages in the environment, including potentially ICP1, have been shown to inversely correlate with disease burden suggesting that phage predation in the natural environment may contribute to the collapse of a given cholera epidemic [41]. We investigated the potential for phage resistance to develop in a simulated natural environment comprised of chitin and pond water. A thousand CFU from independent cultures of wild type V. cholerae O1 were inoculated into pond water with chitin, and ICP1 was added at an MOI of 0. 01. After 24 hours at 30°C, we observed that in all 11 pond microcosms to which ICP1 was added, the phage titer increased at least one million-fold (data not shown). Bacterial levels in the uninfected control had increased 10,000-fold, while the bacterial levels in the infected microcosms varied substantially from below the detection limit to levels nearly comparable to that observed in the uninfected control (Table 2). All isolates recovered from environments to which phage was added were resistant to ICP1 infection. Furthermore, the majority of isolates (63 out of 80 isolates, 79%) from independent microcosms that had become resistant to ICP1 had the wbeL* frameshift allele (Table 2). We only observed heterogeneity in the resistance mechanism for isolates from one microcosm (Table 2, number 10), suggesting that in most cases phage predation resulted in the clonal expansion of a single resistant mutant. We did not observe any mutations in the poly-A tracts of manA, likely owing to the incomplete ICP1 resistance afforded to V. cholerae with those alleles. Control experiments with manA* and wbeL* variants as input strains in the absence of phage confirmed that neither mutant exhibit a decreased ability to grow in the simulated pond microcosm (data not shown). Isolates from all microcosms to which phage were added were also replica plated onto LB agar containing 350 µg/mL polymyxin B (a concentration that permits growth of the wild type parent but not wbeL* or manA*), and all isolates were unable to grow. This indicates that the mutations that occurred less frequently and that did not map to wbeL also likely affected LPS biosynthesis. These results show that in a simulated natural environment phage predation can occur with consequent selection for bacteria with altered O1 antigen, and that the dominant mechanism by which mutational escape is achieved is through mutations in the poly-A tract in wbeL. We also investigated the diversity of phage resistant mutants that appeared in the center of plaques on LB plates by similarly sequencing the poly-A tracts in wbeL and manA in 83 phage resistant isolates. In contrast to the experiment designed to mimic the natural aquatic environment, during selection in LB soft agar overlays in which diffusion of phage and bacteria are relatively limited, the mode of phage resistance is much more varied although a substantial portion can be attributed to phase variation in wbeL and manA (∼20%); we found that six out of 83 isolates were wbeL* phase variants and 11 out of 83 were manA* phase variants (six of the 11 had a deletion that mapped to the first poly-A tract, and the other five to the second poly-A tract). The reason for this difference in frequency of wbeL* and manA* occurrence compared to the pond microcosm is not known. However, as was mentioned previously (and will be addressed below) the manA* mutation was observed in sequenced V. cholerae O1 isolates; therefore it is apparent there are relevant circumstances in which manA* phase variants are selected for. We wanted to determine if O antigen heterogeneity exists in the population of V. cholerae excreted along with ICP1 phage from patients during natural infection. To do this we obtained three ICP1-positive stool samples collected from three patients admitted to the International Centre for Diarrhoeal Disease Research, Bangladesh (ICDDR, B) during a cholera epidemic in 2001. Several representative isolates from all three stool samples were analyzed and were found to be V. cholerae O1 Inaba and were sensitive to ICP1. We tested the ability of these stool isolates to become phage resistant by collecting phage resistant mutants from the centers of plaques resulting from ICP1 infection and readily isolated manA* and wbeL* phase variants (data not shown), showing that these clinical isolates can phase vary at these contingency loci. Next we screened many thousands of colonies from each of these archived stool samples for the presence of manA* and wbeL* phase variants, which we hypothesized might have arisen as a result of ICP1 phage pressure during the patient infection. Since we showed above that manA* and wbeL* phase variants are attenuated in the infant mouse model, we expected that their frequency in the human stool samples would be low or perhaps even undetectable due to decreased fitness during infection of the human small intestine. As we had done previously, we took advantage of the observation that both phase variants exhibit increased sensitivity to polymyxin B compared to the parental strain. We determined that both the wbeL* and manA* phase variants obtained in vitro from these clinical isolate strain backgrounds failed to grow on LB agar plates containing 800 µg/ml polymyxin B after replica plating, while the parental strain maintained its ability to grow. We screened approximately 5,000 colonies from each stool sample by replica plating and did not observe any isolates with increased polymyxin B sensitivity. In the above analysis we were limited by the number of available archived stool samples, since routine practice is to purify a single colony from a stool sample and store that for further examination. While such single colony isolate collections cannot be used to answer the question of whether O antigen heterogeneity exists within the population of V. cholerae excreted from a single patient, it does allow us to determine if O antigen heterogeneity exists between isolates recovered from different patients. We evaluated phage sensitivity of approximately 50 isolates recovered from cholera patients at the ICDDR, B between 2001 and 2005. Three of these isolates displayed resistance to ICP1 and were characterized further. One of the clinical isolates has a mutation in the second poly-A tract in manA. The other two resistant isolates did not have mutations that mapped to the poly-A tracts in either manA or wbeL, and examination of purified LPS from these strains showed that they produce very little full length O-antigen substituted LPS (data not shown), however the nature of the observed defects is not known. From these results it appears that, despite the hypothesized ICP1 phage pressure during infection and despite pressure from the host immune system to reduce or alter the O1 antigen (O1 antigen is the dominant antigen [61], [62]), the intra-patient and inter-patient O1 antigen variability is quite low. In agreement with this, previous studies have detected LPS mutants in mice following coinfection of phage and V. cholerae O1, but at very low frequencies. Zahid et al. , [63] estimated that the frequency of such mutants was on the order of 10−8, and in accordance with our results, failed to detect phage-resistance heterogeneity among V. cholerae O1 directly from human stools. Additionally, despite relatively high levels of O1-specific phage in the stool samples, V. cholerae O1 isolated from the same stool samples remain completely susceptible to phage lysis [63]–[65]. Previous studies of V. cholerae O antigen negative strains or strains with altered LPS structures were also found to be defective in colonization of infant mice [53], [59], [63], [66]–[69]. These results support the assertion that O1 antigen-deficient V. cholerae would be selected against during human infection due to the high fitness cost associated with mutational escape. When put into the context of our findings, it is fitting that mutational escape is frequently conferred through phase variability, as a key feature defining this mode of variability is its reversible nature [23], [32], [70]. Inherently each cell will retain its ability to switch between expression states (the switching rates of phase variable genes are typically between 10−2 to 10−5 [32]), and therefore the phenotype of a clonal population of bacteria capable of phase variation will vary as a function of selection. Predation of V. cholerae O1 in the environment by O1 antigen-dependent lytic phage may rapidly select for the subpopulation with altered O1 antigen mediated by manA* and wbeL* frame-shifted contingency loci, and even though those subpopulations are less suited for life in the intestinal tract, positive selection (for example when this mixed population is ingested by a human) results in enrichment of the subpopulation with full O1 antigen expression. We were able to experimentally confirm the reversibility of frameshift mutations occurring at the poly-A tract in manA, however, we were unable to do so for wbeL likely due to experimental limitations (discussed above). It is possible that the observed mutational escape mediated by the poly-A tract in wbeL is a function of hypermutation and not phase variation (if it is not reversible). However, we have clearly demonstrated the utility of these loci in mediating alterations in the expression of the key V. cholerae antigen and phage receptor. Pathogenic V. cholerae O1 has evolved to live in very diverse environments including fresh water, salt water and the human small intestine. The O1 polysaccharide antigen is the dominant cholera antigen and can induce protective immune responses in humans and animals [61], [71]–[73], and thus is a critical immunogen guiding cholera vaccine development. The extent to which V. cholerae can vary expression of the O1 antigen is not currently appreciated. We demonstrate for the first time that the O1 antigen is subject to phase variation and show that this is mediated by three homonucleotide tracts in two genes (wbeL and manA), which are critical for O1 antigen biosynthesis. The ubiquitous presence of these phase variable homonucleotide tracts in all V. cholerae O1 strains points to the significant role they play in modulating expression of this surface exposed antigen. Moreover, by identifying manA as critical for O1 antigen biosynthesis, we have extended the genome boundaries previously believed to contain all the necessary genes for O1 antigen biosynthesis in V. cholerae. Phase variation mediated by homonucleotide tracts has not been previously well-documented in V. cholerae. To our knowledge, the only prior report of phase variation in V. cholerae was that by Carroll et al. , [74] in which expression of the membrane bound virulence regulator, TcpH, was observed to be subject to phase variation mediated by a poly-G (G9) tract. However, with the growing list of currently available V. cholerae O1 genome sequences, it is clear that this tract is not well-conserved (only three of the available 37 sequenced strains have this tract [data not shown]), and thus this likely does not represent a wide-spread mechanism employed by V. cholerae O1 to alter virulence expression. Examination of the currently available V. cholerae O1 genome sequences may facilitate further exploration of phase variation in this organism; it is interesting to note that there are only twelve homonucleotide tracts of nine or greater nucleotides in length located within coding regions in the V. cholerae O1 N16961 genome, and several of these are located within known virulence factors (data not shown), however the significance this remains to be examined. The biological role of phase variation in mucosal pathogens is frequently anticipated to facilitate immune evasion in the host [75]. However, in the case of the facultative pathogen V. cholerae, our data point to the primary role for O1 antigen phase variation as a strategy for dealing with the strong opposing selective pressures of phage predation in the environment and the strict requirement of O1 antigen for colonization of the intestinal tract. Phase variation of these genes thus allows for a subset of the population of V. cholerae being disseminated from a patient or being ingested in contaminated water, to be resistant to O1-dependent phages or to be virulent, respectively, thus contributing to the overall fitness of this pathogen. We hypothesize that the host immune response represents yet a second strong selective pressure against the O1 antigen, though the effects of this on circulating strains of V. cholerae within immune populations has not been studied. The ubiquitous presence and overall success of ICP1-related phages is likely, at least in part, due to their use of a critical virulence factor as a receptor [40]. Our observation that mutational escape facilitated by wbeL and manA predominates ex vivo strongly suggests that ICP1 is particularly adept at predation of V. cholerae O1 within the human host where the requirement for colonization and virulence necessitates the maintenance of the O1 antigen. This may suggest a mechanism whereby this phage and the human host act synergistically to limit V. cholerae during infection, and perhaps how phage contribute to the overall decline of a given cholera epidemic as has been hypothesized [41], [64]. It remains to be seen if there are additional mechanisms employed by V. cholerae O1 to evade phage predation, specifically within the human intestinal tract, and how this arms race between ICP1 and its bacterial host shapes the evolution of the circulating V. cholerae O1 strains within the endemic region of Bangladesh. Strains were grown on Luria-Bertani (LB) agar or in LB broth at 37°C with 100 µg/ml streptomycin (Sm). When indicated M9 minimal media (supplemented with trace metals, vitamins (Gibco MEM Vitamins, Invitrogen), 0. 1% casamino acids) with 0. 4% glucose and/or 0. 4% mannose was used. Strains containing the pMMB67EH vector were grown in the presence of 100 µg/ml Sm and 50 µg/ml ampicillin (Amp). Expression from the Ptac promoter was induced by the addition of 1 mM isopropyl-β-d-thiogalactopyranoside (IPTG). Phage susceptibility was determined by the soft agar overlay method as described previously [40] and/or by measuring growth of a bacterial isolate in the presence of ICP1 (to an approximate MOI = 1) in LB plus Sm broth culture using a Bio-Tek microplate reader. A wild type V. cholerae O1 strain (E7946) was used in standard plaque assays with phage ICP1 as previously described [40]. Following overnight incubation, colonies were routinely observed in the center of plaques indicating the presence of phage resistant isolates. Four independent colonies were chosen for further analysis including phage resistance assays and whole genome sequencing using an Illumina genome analyzer II (Tufts University Core facility) as previously described [40]. Assembled genomes were aligned to the V. cholerae O1 N16961 [47] and E7946 (unpublished data) reference genomes. Two of the independently isolated phage resistant strains had a single nucleotide deletion in the poly-A tract of wbeL (designated wbeL*), while one phage resistant derivative had a single nucleotide deletion in the second poly-A tract of manA (designated manA*). The other derivative not chosen for further study had a nonsynonymous substitution in manB. PCRs for sequencing and cloning were carried out using EasyA polymerase (Agilent). Primer sequences are available upon request. In-frame unmarked deletions were constructed using splicing by overlap extension (SOE) PCR [76] and introduced using pCVD442-lac [77]. Deletion alleles constructed in this study are missing the entire open reading frame, except for the start and stop codons (with the exception of the wbeL deletion allele which also preserved a single codon immediately upstream of the stop codon). Expression plasmids were constructed by cloning the desired open reading frame (s) (including the predicted ribosome binding site) into the multiple cloning site of pMMB67EH. Expression vectors were transferred into V. cholerae by conjugation with E. coli SM10λpir and selection of SmR AmpR colonies. Strains utilized in this study are shown in Table 3. Slide agglutination tests were performed using V. cholerae O1 Ogawa polyclonal rabbit antiserum (Difco). LPS was extracted from overnight cultures unless otherwise indicated, as described previously [72]. Briefly, cultures were centrifuged and washed twice in TM buffer (50 mM Tris [pH 7. 5], 10 mM MgCl2) supplemented with 1 mM DL-Dithiothreitol before being lysed by bead-beating (BioSpec Products, Inc.) with 0. 1 mm zirconia beads for a total of three minutes with intermittent incubations on ice. Whole cell lysates were treated with proteinase-K (Sigma) at 37°C for 24–48 h as required. Phenol extraction was performed using phase-lock gel light tubes (Eppendorf). Extracts were centrifuged at 75,000× g for 60 min, the pellet was washed with TM buffer and centrifuged as before. Purified LPS was separated on a 4–12% NuPage Bis-Tris gel (Invitrogen) and visualized by silver-staining (SilverQuest, Invitrogen). The concentration of V. cholerae LPS was determined by comparison to a standard curve of E. coli O26: B6 LPS (Sigma) using a Fujifilm FLA-900 scanner as previously described [72]. The ability of purified LPS to neutralize plaque formation was determined as previously described [40]. In vivo competition experiments were done using 4–5 day old CD-1 mice. The dams and their litters were housed with food and water ad libitum and monitored in accordance with the rules of the Department of Laboratory Animal Medicine at Tufts Medical Center. The inoculum was prepared as a 1∶1 mixture of the strain of interest (lacZ+) and the appropriate control strain (ΔlacZ). Mice were infected intragastrically with approximately ∼105 CFU and sacrificed 24 hours post-infection. Small intestines were homogenized in 1 ml LB+16% glycerol, diluted in LB broth, and plated on LB agar plates containing 100 µg/ml Sm and 40 µg/ml 5-bromo-4-chloro-3-indolyl-β-d-galactopyranoside (X-gal). The competitive index was calculated as the ratio of the mutant compared to the control strain normalized to the input ratio. In vitro controls were included in each of these experiments in which the same inoculum was diluted 1∶100 into at least five independent LB cultures and the output ratios of mutant to the control strain were determined on Sm X-gal agar plates as above. Polymyxin B killing assays were done as previously described with minor modifications [57]. Briefly, overnight cultures were subcultured 1∶100 into LB and grown at 37°C to the desired OD (OD600 = 0. 15 and OD600 = 0. 5). 5 µl polymyxin B (Invitrogen) at 500 µg/ml was added to 45 µl of the above culture in a well of a 96-well polypropylene microtiter plate to obtain a final test concentration of 50 µg/ml polymyxin B. After three hours of incubation at 37°C with shaking, serial dilutions of each culture were plated on LB Sm plates. The percent survival was calculated as (CFU (polymyxin treatment) /CFU (untreated) ) ×100. The average percent survival was determined from two biological replicates, each having been done in technical duplicate. Overnight cultures of wild type V. cholerae E7946 were serially diluted in filter sterilized pond water to approximately 105 CFU/ml. 10 µl of diluted culture (103 CFU) was used to inoculate 1 ml chitin solution (1% chitin from crab shells [Sigma] in filter sterilized pond water). To assess the impact of phage on the appearance of phase variants under these conditions, approximately 10 PFU of ICP1 was immediately added following inoculation of bacteria (MOI = 0. 01). The mixture was allowed to incubate for 24 hours at 30°C statically at which time the mixture was vortexed and plated for CFU. ICP1 was enumerated by adding chloroform to a 100 µl aliquot of the above solution, diluted and plated for PFU with E7946 using the soft agar overlay method as described above. Three cholera stool samples collected at the ICDDR, B in 2001 and stored in the presence of glycerol were assayed for the presence of isolates with altered O1 antigen. Single isolates from each sample were found to be O1 Inaba that were sensitive to ICP1. wbeL* and manA* mutants of this clinical O1 Inaba isolate were recovered after plating with ICP1 and used to assess the applicability of replica plating on polymyxin B as a tool to identify heterogeneity within a stool sample. Both the wbeL* and manA* isolates in this background failed to grow on LB agar plates containing 800 µg/ml polymyxin B, while the parental strain maintained its ability to grow. Each stool sample was plated on LB agar containing 100 µg/ml Sm and incubated overnight at 37°C. Plates were then replica plated onto polymyxin B plates and incubated overnight at 37°C to identify polymyxin B sensitive isolates in the stool sample. Approximately 5000 colonies were analyzed per stool sample. All animal experiments were done in accordance with NIH guidelines, the Animal Welfare Act and US federal law. The experimental protocol using animals was approved by Tuft University School of Medicine' s Institutional Animal Care and Use Committee. All animals were housed in a centralized and AAALAC-accredited research animal facility that is fully staffed with trained husbandry, technical, and veterinary personnel.

The O1 serogroup of Vibrio cholerae is the most common cause of the potentially fatal diarrheal disease cholera, which remains a significant global health burden worldwide. The O1 antigen is a constituent of the lipopolysaccharide portion of the outer membrane, and its location on the bacterial surface makes it a major target of both the immune system and bacteriophages. We used an O1-specific bacteriophage as a tool to understand if, and how, V. cholerae can alter O1 antigen expression. We discovered that two genes, which are critical for O1 antigen biosynthesis, are subject to phase variation. Additionally, one of the phase variable genes we identified was not previously known to play a role in O1 antigen biosynthesis in V. cholerae. Phase variation is a well-recognized mechanism many other bacterial pathogens use to generate variable expression of surface components, and this is generally thought to help these organisms evade the immune system. Phase variation has not previously been described as a widespread mechanism used by V. cholerae, furthermore, this is the first report that V. cholerae O1 is capable of generating diverse populations with variable and unique O1 antigen expression.

Abstract Introduction Results/Discussion Materials and Methods

biology microbiology

2012

Phase Variable O Antigen Biosynthetic Genes Control Expression of the Major Protective Antigen and Bacteriophage Receptor in Vibrio cholerae O1

Microsaccades aid vision by helping to strategically sample visual scenes. Despite the importance of these small eye movements, no cortical area has ever been implicated in their generation. Here, we used unilateral and bilateral reversible inactivation of the frontal eye fields (FEF) to identify a cortical drive for microsaccades. Unexpectedly, FEF inactivation altered microsaccade metrics and kinematics. Such inactivation also impaired microsaccade deployment following peripheral cue onset, regardless of cue side or inactivation configuration. Our results demonstrate that the FEF provides critical top-down drive for microsaccade generation, particularly during the recovery of microsaccades after disruption by sensory transients. Our results constitute the first direct evidence, to our knowledge, for the contribution of any cortical area to microsaccade generation, and they provide a possible substrate for how cognitive processes can influence the strategic deployment of microsaccades. Microsaccades, which frequently occur during gaze fixation, translate retinal images by only a few photoreceptors. Despite their modest size, microsaccades strongly impact visual perception [1–5] and visually guided behavior [6–8]. Indeed, visual responses in a number of brain structures are dynamically influenced by either the production or consequence of microsaccades, with responses being enhanced immediately before microsaccades [9], suppressed during or just after microsaccades [7,10,11], and then subsequently enhanced [10–15]. While some mechanisms underlying microsaccade generation have been elucidated in the superior colliculus (SC) [16–20], cerebellum [21], and brainstem saccadic burst generator [22–24], no study has addressed the involvement of any cortical area in microsaccade generation. This gap in knowledge is all the more surprising given the strategic deployment of microsaccades in tasks requiring high visual acuity [25,26], the impacts of microsaccades on visuomotor processing noted above, and the interest in microsaccades as a potential biomarker for visuospatial attention [27–31]. Here, we directly examined the causal role of the frontal eye fields (FEF), a key cortical oculomotor structure that projects strongly to the SC [32,33], in microsaccade generation. To address this, we reversibly inactivated large volumes of either the unilateral or bilateral FEF using cryoloops implanted in the arcuate sulcus and examined the changes in microsaccade behavior, focusing primarily on how FEF inactivation alters the well-known evolution of microsaccades that occurs following peripheral stimulus onset [28,29,34,35]. Our results show that the role for the FEF in microsaccades is distinct from that for the SC, and that the FEF provides a plausible substrate for how microsaccades can be strategically deployed. Across our sample, we found consistent alterations in microsaccade amplitude and peak velocity, regardless of whether the microsaccade was generated before or after peripheral cue onset in our tasks. Fig 1A shows the effects of unilateral FEF inactivation on ipsilesional and contralesional microsaccade amplitude for our representative dataset. When the FEF was not inactivated, microsaccades had relatively small amplitudes (median: 0. 51°) compared to the fixation window, which likely relates to the small size of our fixation cue [38]. During FEF inactivation, the amplitude distributions for both ipsi- and contralesional microsaccades were shifted toward larger amplitudes (p < 0. 0001, Wilcoxon rank sum test), with greater shifts for contralesional (increased by 21%) versus ipsilesional (increased by 10%) microsaccades. Regardless of increases in microsaccade amplitude during FEF inactivation, the vast majority of microsaccades remained <1. 5°. Across our sample, contralesional microsaccade amplitudes increased significantly in four of five configurations; ipsilesional microsaccade amplitude increased significantly in two of five configurations (Fig 1B). During bilateral FEF inactivation, the amplitudes of both leftward and rightward microsaccades also increased significantly (Fig 1B). Importantly, microsaccade amplitude increased regardless of whether microsaccades were generated before or after cue presentation (Fig 1C, and see Materials and Methods for definitions of pre-cue and rebound periods). Could these increases in microsaccade amplitude be a simple consequence of a biased fixation position? We analyzed fixation position with and without FEF inactivation and found that unilateral FEF inactivation biased fixation position by less than 1° toward the intact visual hemifield (S1A Fig). However, this bias persisted before, during, and after cue presentation (S1B and S1C Fig), meaning that any changes in microsaccade behavior due to FEF inactivation were not sufficient to correct for a biased fixation position. This observation is consistent with the idea that FEF inactivation introduces a new balance point for eye position, as observed during SC inactivations [19,39], rather than a mechanism that acts to correct for the biased fixation position, since unilateral FEF inactivation also increased ipsilesional microsaccade amplitudes (Fig 1). Similarly, we observed increased microsaccade amplitudes in both directions during bilateral FEF inactivation (Fig 1B and 1C), despite a fixation position bias only toward one side (S1C Fig). More compelling evidence against a simple compensatory mechanism based on a bias in fixation position is provided by microsaccade peak velocity, which decreased independent of increased microsaccade amplitude or fixation offset. Such decreases in peak velocity are shown in the velocity-amplitude main sequence relationships in Fig 2A for both ipsilesional and contralesional microsaccades; note how both main sequence relationships are shifted downward during unilateral FEF inactivation. To determine the significance of such changes, we fitted a linear regression to 5,000 bootstrapped samples of microsaccades for both the FEF warm and FEF cool conditions and then extracted peak velocities from each relationship at amplitudes of 0. 4° to 1. 0° with 0. 1° increments. We found significantly decreased peak velocities across this entire range of amplitudes for both ipsilesional and contralesional microsaccades (insets of Fig 2A, each p < 0. 01, Welch' s t tests). Fig 2B shows how FEF inactivation alters the kinematic profiles of microsaccades matched for radial amplitudes (e. g. , between 0. 40° and 0. 45°; see shaded region of Fig 2A) by lowering peak velocity and significantly increasing microsaccade duration (ipsilesional, p < 0. 05; contralesional, p < 0. 0001, Wilcoxon rank sum test). Such changes in kinematics and duration are consistent with FEF inactivation altering the drive to brainstem circuits generating microsaccades. To analyze any changes across our sample, we extracted peak velocities at 2° and found significant decreases of 9% and 23% for ipsilesional and contralesional microsaccades during FEF inactivation, respectively (Fig 2C, each p < 0. 0001, Welch' s t tests). Unilateral FEF inactivation significantly decreased contralesional peak velocity in all five configurations and significantly decreased ipsilesional peak velocity in three of five configurations (Fig 2C). Bilateral FEF inactivation significantly decreased peak velocity for both leftward and rightward microsaccades in both monkeys (Fig 2C). Once again, such bilateral decreases in microsaccade peak velocity occurred regardless of whether microsaccades were generated in the pre-cue or rebound period (Fig 2D) and despite a unilateral bias in fixation position during bilateral inactivation (S1C Fig). Therefore, even prior to any task-related stimulus, FEF inactivation had a measurable impact on microsaccade metrics and kinematics, with such an impact often influencing even ipsilesional microsaccades. We next describe an even larger impact of FEF inactivation on the rate of cue-induced microsaccades. Microsaccade rate shows robust and highly repeatable modulations after peripheral cue onsets, decreasing ~50 ms after cue onset and then rebounding before returning to baseline [4,29,35]. To analyze the effects of FEF inactivation on microsaccade rate, we divided our data into three periods: pre-cue, microsaccadic inhibition, and rebound (Fig 3; see Materials and Methods for how these periods were defined; microsaccade rate was calculated within a sliding ±50 ms window with step size of 5 ms). As shown in our representative data, the influence of unilateral FEF inactivation was largely specific to the rebound period (Fig 3A and 3B): note how the rate of such rebound microsaccades decreased from 1. 08 to 0. 70 microsaccades/s with FEF inactivation and then recovered to 0. 87 microsaccades/s with FEF rewarming (both changes significant, p < 0. 0001, Welch' s t test). In contrast, microsaccade rate in the pre-cue period decreased from 1. 11 to 0. 91 microsaccades/s (p < 0. 0001) with FEF inactivation and decreased further to 0. 76 microsaccades/s (p < 0. 01) when the FEF was rewarmed, suggesting that this effect may have been due to satiation with increasing trial count. Microsaccade rate during the inhibition period was unchanged. We also analyzed both the start and end of microsaccadic inhibition and the timing of the first rebound microsaccade after cue presentation (see Materials and Methods). FEF inactivation had no influence on the start of microsaccadic inhibition following cue onset (59 ms for both pre- and peri-cool) or its end (152 versus 158 ms for FEF pre- versus peri-cool, p = 0. 74, Wilcoxon rank sum test). In contrast, the timing of the first rebound microsaccades increased from 264 to 291 ms with FEF inactivation and then recovered to 266 ms when the FEF was rewarmed (S2A Fig, both p < 0. 0001; Wilcoxon rank sum test). For each of the three monkeys, unilateral FEF inactivation systematically decreased microsaccade rate during the rebound period (Fig 3D, significant in four of five configurations) but not the pre-cue period (Fig 3C). Bilateral FEF inactivation further reduced microsaccade rate during the rebound period (Fig 3D) but, unlike unilateral inactivation, also significantly decreased microsaccade rate during the pre-cue period (Fig 3C). Thus, with bilateral inactivation, there was a generalized decrease in microsaccades. Across our sample, the decrease in microsaccade rate in the rebound period averaged 24% with unilateral inactivation and 54% with bilateral inactivation. Consistent with a generally exacerbated effect of bilateral versus unilateral FEF inactivation, we also found a relatively greater increase in the timing of the first microsaccade in the rebound period during bilateral (44 ms) versus unilateral (7 ms; S2B Fig). These results indicate that FEF integrity is critical for cue-induced microsaccades and that larger bilateral inactivation volumes can further impact microsaccades generated before cue presentation. These effects on cue-induced microsaccade rate are also categorically different from those reported during pharmacological inactivation of the SC, where cue-induced microsaccade rates remained unchanged [18]. Because FEF inactivation also introduced a bias in fixation position, we wondered whether this could explain the changes in microsaccade rate during the rebound period. To investigate this, we repeated our analysis of microsaccade rate after performing a median split of FEF warm and FEF cool trials based on their radial fixation error in the pre-cue period (Fig 4A). This analysis exploits the substantial overlap in fixation positions with or without FEF inactivation (S1A Fig), and, in fact, fixation error was significantly larger for the higher-than-median FEF warm trials than for the lower-than-median FEF cool trials (Fig 4A, fixation error for FEF warmhigh = 0. 82°; fixation error for FEF coollow = 0. 53°). As shown in Fig 4B, the robust decrease in rebound microsaccades during FEF inactivation persisted regardless of this split in fixation error. To quantify this across our sample, we calculated the change in rebound microsaccades from FEF warmhigh trials to FEF coollow trials. If the changes in rebound microsaccades during inactivation arose because of a greater fixation error, then we should observe no decrease in rebound microsaccades across these subsets of data, because fixation offset was greater in FEF warmhigh trials. However, as shown in Fig 4C, we still observed a profound decrease in microsaccade rate during the rebound period with FEF inactivation. Therefore, the effects of Fig 3 cannot be due to fixation error. Finally, these analyses led us to investigate whether FEF inactivation impacted eye position drift, and not just overall bias. Even though our eye tracker was not well suited to study drift at a higher resolution, to the extent that we could measure it, eye position drift before cue onset was not influenced by FEF inactivation (S3 Fig). FEF inactivation also did not influence dependencies between drift and microsaccades. Specifically, we analyzed the relationship between eye position drift and microsaccades, as previous work has shown that drift speed is lower before as compared to after a microsaccade [40]. However, we observed no systematic influence of FEF inactivation on this relationship (S4 Fig). Thus, we conclude that the effects shown in Fig 3 on FEF inactivation on microsaccade rate could not be attributed to biases in fixation position or drift in the pre-cue period. Next, we investigated whether the effects of FEF inactivation on microsaccade rate only occurred when cues were presented contralateral to the side of inactivation. To our surprise, we found that unilateral FEF inactivation decreased microsaccade rate during the rebound period regardless of the side of the cue (Fig 5A, shown for our representative data during left FEF inactivation). Despite an idiosyncratically higher rate of rebound microsaccades for cues presented in the intact (left) hemifield even before inactivation, unilateral FEF inactivation significantly reduced microsaccade rate during the rebound period for cues presented in both the intact (left; 41% decrease, p < 0. 001, Welch’s t test) and affected (right; 12% decrease, p < 0. 05) hemifield (see arrows). We observed such robust decreases in microsaccade rate during the rebound period during unilateral FEF inactivation across our sample, regardless of the side of the cue relative to the side of inactivation (Fig 5C). Such decreases were even greater during bilateral FEF inactivation (Fig 5B; both hemifields are presumably affected in this configuration; Fig 5C). We also compared the influence of unilateral or bilateral FEF inactivation on microsaccade rate during the pre-cue and rebound periods (Fig 5D). This analysis again revealed that each cooling configuration robustly decreased the rate of microsaccades in the rebound period regardless of the side of the cue, but that only bilateral FEF inactivation decreased microsaccade rate before cue onset. Together with Fig 3, these results demonstrate the importance of FEF integrity when microsaccades are deployed, particularly after cue onset. Cue presentation is known to briefly bias microsaccade direction toward and then away from the cue [28,29,34,35]; will FEF inactivation alter such directional modulations? Our monkeys exhibited strong idiosyncratic tendencies in microsaccade direction even before FEF inactivation, which complicates our interpretation. However, we still observed consistencies across our sample, especially when examining how FEF inactivation impacted the fraction of microsaccades toward the cue during the rebound period. In general, unilateral FEF inactivation biased microsaccades toward the affected side, with this bias becoming more pronounced following contralesional cues. For example, before FEF inactivation, monkey DZ had a strong idiosyncratic tendency to make leftward microsaccades, which was perturbed for ~400 ms after cue presentation (Fig 6A). Left FEF inactivation increased the tendency for rightward microsaccades even before cue onset (i. e. , the blue line lies above the red line for rightward cues, but below the red line for leftward cues), perhaps to correct for an altered fixation position. During the microsaccade rebound period, FEF inactivation exacerbated this effect only when cues were presented in the affected hemifield (arrow in Fig 6A). To quantify this effect, we measured how FEF inactivation altered the fraction of microsaccades toward cues in the rebound period, segregated by the side of the cue. This fraction significantly increased from 40% to 62% during unilateral FEF inactivation for cues in the affected hemifield (p < 0. 0001, Welch' s t test) but only increased from 6% to 11% for cues in the intact hemifield (p = 0. 08). Across our sample, such an increase was seen in three of five unilateral configurations for cues in the affected hemifield but never for cues presented in the intact hemifield (Fig 6C). Thus, FEF inactivation influenced microsaccade directionality after contralesional but not ipsilesional cues. Interestingly, this directional profile differs from how FEF inactivation influenced microsaccade rate for both contralesional and ipsilesional cues (Figs 3 and 5). For bilateral FEF inactivation, pre-existing biases in microsaccade direction following cue onset were delayed and magnified. For example, in monkey DZ (Fig 6B), bilateral FEF inactivation did not alter the general idiosyncratic tendency of the monkey (e. g. , before or long after cue onset), but it instead prolonged and exaggerated the transient modulation of microsaccades after the cue. Quantitatively, the fraction of microsaccades toward the cue during the rebound period changed from 7% to 4% for leftward cues (p = 0. 08, Welch' s t test) and from 45% to 7% for rightward cues (p < 0. 0001, see arrow). Across our sample, unilateral FEF inactivation biased microsaccade directions toward the affected side even before cue onset (Fig 6D). This is again different from SC inactivation [18], but it could be related to the altered fixation position. However, the bilateral inactivation data shows that altered fixation position does not produce directional biases in the pre-cue period. Interestingly, changes in microsaccade directionality after cue onset only occurred in the affected hemifield during unilateral inactivation (Fig 6C), but even these changes were not consistently seen in our sample. Taken together, all of our results emphasize that the main effect of FEF inactivation on microsaccade deployment is through modulations of rate rather than microsaccade directionality. This profile differs completely from that seen following inactivation of the SC, which robustly and consistently altered microsaccade direction without influencing microsaccade rate [18]. Microsaccades can be strategically deployed in tasks requiring high visual acuity [25,26], but the substrates responsible for such top-down influences on microsaccades are poorly understood. Previous studies have demonstrated a critical role for the SC in microsaccade generation and deployment [16,18,19]. Frontal inputs into the SC may enable cognitive processes to influence microsaccade behavior. The results of FEF inactivation, which preferentially impacted the rate of microsaccades following peripheral cue onset, are largely consistent with this idea. Furthermore, the FEF has been implicated in covert visuospatial attention [43,44] and sends visual, cognitive, and saccade-related signals directly to the SC [45]. We are not suggesting that the FEF is the sole source of frontal input into the SC for microsaccade control, but it may well serve as an important interface between the SC and other prefrontal areas. We were also intrigued by the differences in how FEF or SC inactivation impacted microsaccades generated after peripheral cue onset. Specifically, FEF inactivation decreased microsaccade rate without having an equally large impact on direction; in contrast, inactivation of the caudal SC (which represents peripheral cue locations) impacted microsaccade direction without influencing rate [18]. There are a number of potential, and not mutually exclusive, explanations for these comparative results. They may speak to a particularly important role for the FEF in intervals requiring top-down control of microsaccades, in this case providing signals related to when a microsaccade should be generated. Alternatively, our cryogenic inactivation of the FEF inactivated a much larger volume of tissue compared to the more focal pharmacological techniques used to inactivate the SC (see below). Such differences in inactivation volume are likely even more important considering that the strength of topographic organization in the FEF tends to be less than that observed in the SC. FEF neurons tuned for small retinal errors, which are akin to those found in the rostral SC, tend to be diffusely distributed throughout the FEF and not just confined to the most lateral portion [45,46]. Thus, methodological differences between inactivation techniques hinder the functional conclusions about the comparative role of each area in microsaccade behavior. Nevertheless, the impact of FEF inactivation on the microsaccade rate signature can help us better understand the underlying neural mechanisms. The microsaccade rate signature describes the well-known and highly replicable inhibition and then rebound of microsaccade rate following presentation of any stimulus [28,29,34,35,47]. Despite the large volume of inactivated FEF tissue, unilateral inactivation delayed and blunted microsaccade production only during the rebound period (Fig 3), without affecting the baseline rate of microsaccades before cue onset or the start of the microsaccadic inhibition period. Perhaps most surprisingly, such effects were observed regardless of the side of cue presentation. Hence, they cannot simply be explained by impaired processing or neglect of a contralesional stimulus. The temporal specificity of FEF inactivation, impairing the rebound but not baseline or inhibition periods, demonstrates that recovery of microsaccades after disruption by sensory transients requires frontal inputs and, hence, is not simply a passive process. Consistent with this, the direction of microsaccades in the rebound period tends to be opposite to those preceding reflexive microsaccades directed toward the cue [31]. Following the same logic, FEF inputs do not seem to be involved in the onset of microsaccade inhibition, as inhibition onset was not influenced by FEF inactivation. Based on our results, it appears that different portions of the microsaccade rate signature are attributable to different neural substrates (e. g. , non-frontal inputs to microsaccade inhibition and frontal inputs to the rebound). Interestingly, a role for frontal inputs in the first microsaccade after inhibition is consistent with recent models regarding microsaccade generation. In a model by Hafed and Ignaschenkova [35] that utilizes the framework of a recurring rise-to-threshold process, the process initiating the first microsaccade after inhibition has a faster rate of rise to threshold. Similarly, in a model by Engbert [48,49] that considers spatiotemporal dynamics of SC activity, the rebound from inhibition is associated with a change in threshold that integrates sensory and attentional inputs. While both of these models are agnostic as to the source of signals that change microsaccade behavior in the rebound period, attributing either a faster rate of rise (in the Hafed and Ignaschenkova model) or attentional signals (in the Engbert model) to frontal sources is broadly consistent with the impact of FEF inactivation on the microsaccade rate signature. Note, however, that our reasoning regarding the impact of unilateral FEF inactivation on the Engbert model hinges on the assumption that a large unilateral inactivation can produce bilateral effects (see below). In support of our contention that frontal sources are involved in rebound microsaccades, simulations of the Hafed and Ignaschenkova model in which we reduced the rate of rise related to the first microsaccade after inhibition produced results very similar to those produced by FEF inactivation (S1 Text and S5 Fig). While we observed strong influences of FEF inactivation on microsaccade rate, we observed little systematic influence of FEF inactivation on microsaccade direction. This finding may be attributable to the idiosyncrasies of our subjects, but perhaps more fundamentally, we only studied microsaccades during the performance of delayed-saccade tasks. The strongest evidence linking microsaccade direction to the allocation of visuospatial attention has come from tasks in which covert attention needs to be allocated precisely to perform the task [4,28–30,50]. Recent psychophysical results have demonstrated a dissociation between microsaccade rate and direction effects in attention paradigms [30], and experiments inactivating the SC have shown that rate and direction may not necessarily be affected by the same neural mechanism [18]. In light of these findings, it will be of considerable future interest to see the comparative effect of FEF inactivation on microsaccade rate and direction in other paradigms. The FEF has an important role in the generation of saccades and deployment of visuospatial attention into the contralateral visual hemifield [43,44], and the effects of FEF inactivation on contralesional microsaccades are consistent with the extension of this role for the FEF into the range of the smallest amplitude saccades. How then do we explain the impact of FEF inactivation on the peak velocity of ipsilesional microsaccades, and on microsaccades deployed after the onset of ipsilesional cues? The response fields for neurons in the rostral SC can cover portions of both contralateral and ipsilateral fields [16]. If homologous FEF neurons tuned for small amplitudes also cover both hemifields, then inactivation of such neurons may contribute to the decreases in ipsilesional microsaccade peak velocity with unilateral FEF inactivation. Consistent with this, the lateral portion of one FEF, which preferentially represents small amplitude saccades, also projects to the contralateral SC [51]. A preferential projection from the lateral but not medial FEF to the contralateral SC may explain why unilateral FEF inactivation did not decrease the peak velocity of larger ipsilesional saccades [36]. Furthermore, recent findings suggest a more nuanced role in how the FEF contributes to spatially guided behavior. For example, while focal FEF inactivation increases or decreases the reaction times of contralesional or ipsilesional saccades respectively [43,52,53], larger volume temporary or permanent lesions of the FEF raise saccade reaction times bilaterally [36,54]. In our previous work [36], we estimated that the volume inactivated via cooling is conservatively at least four times larger than that typically achieved using pharmacological modulations or optogenetics [18,19,55]. In light of this large inactivation volume, we speculated [36] that bilateral reaction time increases may arise from differences in how the FEF commits to a saccadic decision via widespread disinhibition of the intact FEF or to the presence of diffusely distributed FEF neurons with ipsilateral response fields [56] whose contribution is only revealed by large-volume inactivation. Similarly, inactivation of diffusely distributed FEF neurons tuned to small retinal errors [45,46] may delay the generation of microsaccades during the rebound period, regardless of the side of the cue (e. g. , S4 Fig), thereby delaying and blunting the rebound period. The FEF has been implicated in the deployment of covert visuospatial attention via top-down signals to extrastriate visual cortex [57–59] and was recently shown to contribute to pupil dilation [60]. Our discovery of a role for the FEF in microsaccade deployment raises the interesting possibility that the FEF can also influence visual processing in still more ways, for example, by strategically deploying microsaccades or via pre-microsaccadic modulations that shape visual processing before the arrival of re-afferent visual input [9]. Our findings set the stage for future experiments that distinguish how cognitive processes optimize visual processing via the preparation and generation of microsaccades or by coordinating such microsaccades with other components of the orienting response [27]. Three male monkeys (Macaca mulatta, monkeys GB, DZ, and OZ, weighing 11. 1,9. 8, and 8. 6 kg, respectively) were used in these experiments. Only monkey GB contributed data to our previous manuscript [36]. All training, surgical, and experimental procedures conformed to the policies of the Canadian Council on Animal Care and National Institutes of Health on the care and use of laboratory animals, and were approved by the Animal Use Subcommittee of the University of Western Ontario Council on Animal Care (2007-099-10). We monitored the monkeys' weights daily, and their health was under the close supervision of the university veterinarians. Each monkey underwent one surgery to enable reversible cryogenic inactivation of one or both FEFs. Monkeys DZ and OZ were implanted with bilateral FEF cryoloops, whereas monkey GB was only implanted with unilateral FEF cryoloops in the right hemisphere. Our surgical procedures of implanting cryoloops in the arcuate sulcus have been previously described [36,61]. Briefly, we performed a small, 2. 25 cm2 craniotomy at the stereotaxic coordinates of the arcuate sulcus spur and implanted two customized, stainless steel cryoloops (each 5 to 8 mm in length and extending 3 mm into the sulcus) into each arcuate sulcus, which allowed for the cooling of tissue adjacent to the superior and inferior arms of the arcuate sulcus. Cryoloop temperatures of 3°C silence post-synaptic activity in tissue up to 1. 5 mm away without influencing the propagation of action potentials in nearby axons [61]. For this manuscript, we only collected data using the cryoloop in the inferior arm of the arcuate sulcus, which provided an estimated volume of inactivation of 90 mm3 in the anterior bank of the arcuate sulcus. Cooling only the cryoloop in the inferior arm of the arcuate sulcus produced the expected triad of contralateral saccadic deficits (i. e. , decreases in peak velocity, accuracy, and increases in reaction time), which was approximately 70% of the total saccadic deficits observed from cooling both cryoloops [36]. Head-restrained monkeys were placed in front of a rectilinear grid of 500+ red LEDs covering ±35° of the horizontal and vertical visual field. We conducted experiments in a dark, sound-attenuated room and recorded each monkey' s eye position using a single, chair-mounted eye tracker (EyeLink II). The behavioral tasks were controlled by customized real-time LabView programs on a PXI controller (National Instruments) at a rate of 1 kHz. A single experimental dataset consisted of a pre-, peri-, and post-cooling session, with each session containing the same number of correct trials. The number of trials for a given dataset ranged from 180 to 480 correct trials, depending upon the number of cue locations. Our experimental procedure for cryogenic inactivation of the FEF has been previously described [36]. Briefly, following the completion of the pre-cooling session, chilled methanol was pumped through the lumen of the cryoloops, decreasing the cryoloop temperature. Once the cryoloop temperature was stable at 3°C for at least 3 min, we began the peri-cooling session. Upon finishing the peri-cooling session, we turned off the cooling pumps, which allowed the cryoloop temperature to rapidly return to normal. When the cryoloop temperature had reached 35°C for at least 3 min, we started the post-cooling session. Because we simultaneously recorded neurons in the intermediate layers of the superior colliculus (iSC) with FEF inactivation, it was necessary to minimize the amount of time for transitions (i. e. , shorter than 3 min between pre- and peri-cooling and peri- and post-cooling sessions) to ensure continued isolation of an iSC neuron throughout the full dataset. However, cryoloop temperatures rapidly decreased or increased when the cooling pumps were turned on and off, respectively, and we still found similar effects on saccadic behavior with slightly reduced transition durations. The effects of FEF inactivation on neuronal activity within the iSC will be described in a future manuscript. Monkeys performed memory and visually-guided saccades toward peripheral cues after a delayed response period. Following a variable fixation period of 750 to 1000 ms during which monkeys maintained fixation within a +/- 3° window of a central cue, a peripheral cue appeared in either visual hemifield. Monkeys were required to maintain fixation of the central cue and delay their saccadic response until the central cue was extinguished. Note that despite the large fixation window in our experiments, our central cue was 0. 63° in diameter, explaining why most microsaccades were significantly smaller than 1° (Fig 1A). Peripheral cues were either extinguished 150 or 250 ms after onset or remained on for the ensuing memory or visually guided saccade, respectively. After a delayed response period of at least 750 ms, monkeys were rewarded with a liquid reward if they generated a saccade toward the location of the remembered or persistent peripheral cue within 1,000 ms of the offset of the central cue. This response window allowed us to differentiate trials with increased saccade reaction times from neglect of the peripheral cue during FEF inactivation, although monkeys had very few saccade reaction times >500 ms. When we were also recording iSC activity, the location of one peripheral cue coincided with the peak of the response field of an isolated iSC neuron; the other peripheral cue was placed in the diametrically opposite position. In this report, peripheral cues were always located within 45° radial angle relative to the horizontal meridian and more than 5° in radial eccentricity from the central cue. Analysis of microsaccade rate and directionality in the 500 ms window surrounding cue onset revealed no differences depending on the location of the peripheral cue or depending on whether the peripheral cue remained illuminated or not. Accordingly, we pooled all trials together, subdividing data based only on the side of the cue relative to the side of FEF inactivation. Offline, we screened all trials for microsaccades in a customized graphics user interface made in MatLab (Mathworks) that automatically detected microsaccade onset and offset using velocity (10°/s) and acceleration (600°/s2) criteria. We only accepted trials in which the monkey maintained fixation of the central cue for the full delayed period and removed any trial in which we identified any blinks or other aberrant changes in eye position or velocity (e. g. , due to fatigue or inattention). We verified the onset and offset marks for each microsaccade and removed any microsaccades with amplitudes greater than 3° or severe curvatures in their trajectories (i. e. , ratio of maximal to final displacement greater than 2). To differentiate microsaccades from drift, we also removed any microsaccades with onset accelerations lower than 1,000°/s2. We considered all microsaccades generated for each monkey actively fixating the central cue (i. e. , fixation and delayed response periods) regardless of whether they correctly looked to the location of the peripheral cue. Similar results were observed if we constrained our analysis only to successfully performed trials. While our amplitude limit of 3° is very liberal, we wanted to ensure that any reduction of microsaccade occurrence during FEF inactivation (see Results) was not due to a coinciding increase in microsaccade amplitude above an arbitrary limit. Despite this liberal definition of microsaccade amplitude, and despite the specifics of our task and fixation window size, for each monkey, we found that the distribution of microsaccade amplitudes (e. g. , median microsaccade amplitude of 0. 51° for our example monkey in the FEF warm condition, see Fig 1A) was in good agreement with previous studies in monkeys and humans (reviewed in [62]). Perhaps most importantly, all of the results of FEF inactivation still held if we reduced our amplitude limit to 2°. We investigated the contribution of the FEF to multiple aspects of microsaccade behavior in this manuscript. Microsaccade rate was defined as the number of microsaccades within a sliding ±50 ms rectangular window (in steps of 5 ms) divided by the number of all acceptable trials. Based on observations across monkeys, we used fixed time windows to quantify the microsaccade rate for the pre-cue period (i. e. , 200 ms preceding cue onset), microsaccadic inhibition period (i. e. , 60–140 ms after cue onset), and rate rebound period (i. e. , 140–400 ms after cue onset; see Fig 3B for depiction of these periods). In order to investigate the timing of cue-induced microsaccades, we defined the microsaccade response time as the mean latency of the first microsaccade generated following cue onset during the rate rebound period. We defined the microsaccade amplitude as the angular vectorial displacement from microsaccade onset to offset. The microsaccade peak velocity was defined as the maximal vectorial velocity during its movement. To characterize changes in peak velocity, we constructed velocity-amplitude main sequence relationships and then extracted the peak velocity for 2° microsaccades from a fitted linear regression. We also investigated microsaccade directionality as the fraction of microsaccades toward the cue (i. e. , sum of microsaccades toward the cue divided by the sum of microsaccades directed either toward or away from the cue); therefore, microsaccade directionality was independent of rate. Microsaccades directed within ±45° of the cue or diametrically opposite location of the cue were classified as toward or away from the cue, respectively. Finally, we also determined the specific timing of the microsaccade rate signature for each monkey. For this analysis, we first counted microsaccades across the full trial duration in ±50 ms bins and then calculated a threshold number of microsaccades that corresponded to 20% of the mean number in the pre-cue period. We determined the start of microsaccadic inhibition and rebound periods by incrementing bins backward and forward from 100 ms after cue onset in 1 ms steps, respectively, to find the next bin that exceeded the threshold number. To determine the time course and statistics of microsaccade rate and directionality, we performed sliding window analyses in which we calculated a given measure within a ±50 ms window, and incrementally shifted this window every 5 ms for the full trial duration. The 95% confidence intervals of the mean microsaccade rate and peak velocity at 2° were calculated using 5,000 bootstrapped samples of randomly selected trials with replacement, while for directionality we used a binomial probability function. For statistical comparisons of specific time periods and/or conditions between bootstrapped distributions, we performed Welch' s t tests (p < 0. 05). For all other microsaccade measures, we determined statistical significance using Wilcoxon rank sum tests (p < 0. 05).

Microsaccades are small, fixational eye movements that precisely relocate the visual axis. Despite evidence that microsaccades can be strategically controlled in high-acuity visual tasks, impacting visual processing, and considerable knowledge about how microsaccades are generated by the oculomotor brainstem, little is known about the cortical substrates that control microsaccades. To address this gap, we examined microsaccades generated by non-human primates before, during, and after large-volume reversible unilateral or bilateral inactivation of the frontal eye fields, a key oculomotor area in the frontal cortex. In support of a role for the frontal eye fields in microsaccades, microsaccade metrics and kinematics were altered during frontal eye fields inactivation. More surprisingly, frontal eye fields inactivation also impaired the generation of microsaccades following presentation of peripheral cues, regardless of the side of the cue or inactivation configuration. To our knowledge, our results constitute the first direct evidence for the contribution of any cortical area to microsaccade generation and suggest that the frontal eye fields can provide the top-down signals to the oculomotor brainstem needed to strategically guide microsaccades.

Abstract Introduction Results Discussion Materials and Methods

velocity cognitive science medicine and health sciences classical mechanics reaction time vertebrates social sciences neuroscience animals mammals primates cognitive neuroscience vision eyes neuronal tuning sensory physiology monkeys animal cells head physics visual system psychology cellular neuroscience eye movements anatomy cell biology physiology neurons biology and life sciences ocular system sensory systems sensory perception physical sciences cellular types amniotes organisms motion

2016

A Causal Role for the Cortical Frontal Eye Fields in Microsaccade Deployment

Schistosomiasis remains a global public health challenge, with 93% of the ∼237 million infections occurring in sub-Saharan Africa. Though rarely fatal, its recurring nature makes it a lifetime disorder with significant chronic health burdens. Much of its negative health impact is due to non-specific conditions such as anemia, undernutrition, pain, exercise intolerance, poor school performance, and decreased work capacity. This makes it difficult to estimate the disease burden specific to schistosomiasis using the standard DALY metric. In our study, we used Pediatric Quality of Life Inventory (PedsQL), a modular instrument available for ages 2–18 years, to assess health-related quality of life (HrQoL) among children living in a Schistosoma haematobium-endemic area in coastal Kenya. The PedsQL questionnaires were administered by interview to children aged 5–18 years (and their parents) in five villages spread across three districts. HrQoL (total score) was significantly lower in villages with high prevalence of S. haematobium (−4. 0%, p<0. 001) and among the lower socioeconomic quartiles (−2. 0%, p<0. 05). A greater effect was seen in the psychosocial scales as compared to the physical function scale. In moderate prevalence villages, detection of any parasite eggs in the urine was associated with a significant 2. 1% (p<0. 05) reduction in total score. The PedsQL reliabilities were generally high (Cronbach alphas ≥0. 70), floor effects were acceptable, and identification of children from low socioeconomic standing was valid. We conclude that exposure to urogenital schistosomiasis is associated with a 2–4% reduction in HrQoL. Further research is warranted to determine the reproducibility and responsiveness properties of QoL testing in relation to schistosomiasis. We anticipate that a case definition based on more sensitive parasitological diagnosis among younger children will better define the immediate and long-term HrQoL impact of Schistosoma infection. Schistosomiasis remains a public challenge globally with 93% of the estimated 237 million infections occurring in Africa [1]. Its transmission is mainly influenced by exposures to environmental factors (contact with infested water, distance to infected water bodies), individual characteristics (treatment history, sex, and age) and socioeconomic factors (occupation and education) [2]–[5]. Schistosomiasis is rarely fatal but due to its recurring nature is manifested as a persistent chronic disorder in endemic areas, resulting in significant health burden [6]–[8]. It is estimated that people living in Schistosoma -endemic areas carry the infection one-third to one-half of their lives [9], yet they may only rarely exhibit the advanced morbidities that are classically associated with schistosomiasis, such as advanced hepatic fibrosis with portal hypertension (for S. mansoni), or bladder and kidney deformity, bladder cancer, or infertility (for S. haematobium) [7]. In reality, much of the negative health impact is due to less obvious or specific conditions such as anemia, undernutrition, abdominal pain, exercise intolerance, poor school performance, and lowered work capacity [7], [8], [10]. The non-specificity of chronic infection symptoms, manifested as these subtle morbidities, makes it difficult to accurately estimate the specific disease burden due to schistosomiasis. As pointed out by King and Bertino [11], the present disability-adjusted life year (DALY) system of the World Bank and the World Health Organization (WHO) [12] not only ignores these morbidities, but also disregards the pervasiveness of co-morbidities, including polyparasitism, in Schistosoma-endemic areas. While the debate on DALY estimates of disease burden persists, attention is turning to the use of patient-reported outcomes, such as health-related quality of life (HrQoL) [6], [7], [11], [13]. Use of quality of life (QoL) assessment tools in evaluating schistosomiasis and other neglected tropical disease (NTD) burden is gaining greater standing [13]–[16]. Quality of Life is defined as an individual' s perception of one' s position in life in the context of culture and value systems, as well as in relation to one' s goals, expectations, standards and concerns [17]. The World Health Organization defines health as being “not only the absence of disease and infirmity but also the presence of physical, mental, and social well-being” [18]. HrQoL therefore refers to the physical, psychological, and social scales of health seen in functional areas influenced by a person' s experiences, beliefs, expectations, and perceptions. Measuring HrQoL is an important outcomes indicator in evaluating health-care interventions and treatments, in understanding the burden of disease, in identifying health inequalities, in allocating health resources, and in epidemiological studies and health surveys [19]. Unlike in other chronic conditions such cancer and sickle cell disease where HrQoL tools have been widely used, for NTDs these tools have only been used to assess the burden of advanced schistosomiasis (using EQ-5D plus or WHO QoL-bref questionnaires) [13], [15], [20], soil-transmitted helminthes (using both EQ-5D and SF-12) [16], and echinococcosis (using SF-12) [21], and to assess the burden of polyparasitism in Cote d' Ivoire (using SF-36v2 questionnaire but only the physical scale) [14]. SF-12 has also been used to describe the impact of acute schistosomiasis on quality of life in a group of travelers returning from a luxury safari trip in Tanzania [22]. Results of most of these studies suggest that the burden of schistosomiasis has been consistently underestimated [13], [15], [21]. Moreover, in two of these studies by Jia and colleagues, by only targeting people with chronic schistosomiasis [13] or advanced schistosomiasis [15] it was not possible to contrast HrQoL findings among schistosomiasis patients to people living in same localities without schistosomiasis. Additionally all these studies used QoL tools that do not capture the changes in QoL that occur in different developmental stages of children, the most important epidemiological demographic for active Schistosoma infection [23]. The choice of the HrQoL tool depends mainly on the purpose (the health conditions being investigated) and the target population (the general population, adults only or children only) [24], [25]. Measurement of HrQoL in children is particularly difficult because of the need for different instruments in different age groups, and the need for instruments that accommodate the different cultures. Two types of HrQoL measures have been developed, generic and condition-specific instruments. Generic or non-categorical instruments typically include global or summary ratings of multiple scales or health profile approaches. In contrast, condition-specific measures of HrQoL address the challenges associated with a particular illness, such as cancer. Since our target population in this study was children aged 5–18 years, we decided to use the Pediatric Quality of Life Inventory (PedsQL). PedsQL is available in two generic instruments; one that comprises 23 items (PedsQL) with forms for adult (over 26 years), young adult (18–25 years), adolescent (13–18 years), child (8–12 years), young child (5–7 years) and toddler (2–4 years) and a shorter instrument with 15 items (PedsQL 4. 0 SF15) which has forms for adolescent, child, young child, and toddler. Furthermore, there are PedsQL disease-specific modules including modules for arthritis, asthma, cerebral palsy, cardiac, diabetes, family impact, family information, oral health and transplant among others (http: //www. pedsql. org/ [26]). The validity and reliability of the instrument have been confirmed as a population health measurement tool and in different child populations with chronic illnesses in descriptive and evaluative studies [27]–[30]. The PedsQL has been used in children with different debilitating conditions such as those associated with asthma, transplant recipients, attention-deficit hyperactivity disorder (ADHD) and neuromuscular disorders [31]–[34]. The present study evaluated health related quality of life (HrQoL) in children living in a S. haematobium endemic area in coastal Kenya, and determined the utility of the pediatric quality of life inventory short form (PedsQL 4. 0 SF15) in assessing HrQoL. Additionally, we determined the impact of local transmission features and socioeconomic standing, which are considered potentially important modifiers of S. haematobium–related disease burden. Ethical clearance was obtained by the Institutional Review Board at the University Hospital Case Medical Center of Cleveland and the Ethical Review Committee of the Kenya Medical Research Institute (KEMRI). Children were eligible if they were residents of the area for at least two years, were between 5–18 years old, and had provided child assent and written parental consent. The study population comprised children aged 5–18 years old. who participated in both parasitological and nutritional studies in the five selected endemic rural villages (Milalani, Magadzoni, Gwadu, Dzitenge and Kinango A) in the three districts (Msambweni, Kinango, and Kwale) of Kwale County in Coast Province, Kenya [10], [35]–[37] (Figure 1). This sub-study was embedded in a larger study of the ecology of transmission of vector-borne parasitic infections (the ‘Polyparasitism Project’). This project enrolled participants through house-to-house demographic surveys in May–June 2009 for Milalani and in June, July, August and August–September 2010 for Gwadu, Dzitenge, Kinango A, and Magadzoni, respectively. Parasitological testing and anthropometric measurements were conducted simultaneously in July–August 2009 for Milalani, in October–November 2010 for Gwadu, Dzitenge, and Kinango A, and in April–May 2011 for Magadzoni. Treatment for any parasite infections detected during survey testing was provided immediately after the completion of each village survey. Due to logistical constraints, the PedsQL 4. 0 SF15 questionnaires (see below) were administered at varying (3–16 month) intervals afterwards: in December 2010 for Milalani, in April–May 2011 for Gwadu, Dzitenge, and Kinango A, and in July 2011 for Magadzoni. The numbers of participants per village at enrolment, at parasitological testing/anthropometric assessment, and at HrQoL assessment are detailed in Figure 2. All children with full parasitological and anthropometric results were eligible for inclusion in the study. For Milalani and Magadzoni villages, we randomly selected 92 children who had been S. haematobium egg-positive and 91 who had been S. haematobium egg-negative in each community for PedsQL 4. 0 SF15 tool administration. For the other villages, we targeted all eligible children, irrespective of their initial egg-output status, for PedsQL 4. 0 SF15 administration. Subjects submitted one midday sample for examination for S. haematobium infection. Ten-milliliter aliquots from well-mixed urine samples were subjected to standard Nucleopore filtration [38]. A single stool examination by Kato-Katz method [39] was used to identify infection by hookworm and other soil-transmitted helminths. All children included in this analysis provided finger prick blood for hemoglobin (Hb) measurement (Hemocue, Ängelholm, Sweden) [10]. Anemia was categorized according to WHO criteria by age and sex [40]– for ages <12 years, Hb<11. 5 g/dL; for ages ≥12 years, Hb<12 g/dL; but for males ≥15 years, Hb<13 g/dL. Our standardization and measurement procedures used for anthropometric assessments have been detailed elsewhere (Bustinduy, et al. , [10]). The nutritional indicators; height-for-age (HAZ) and body-mass index (BMI) -for-age (BAZ) were computed using the World Health Organization' s Anthro-Plus software for ages 5–19 years (WHO, Geneva, Switzerland) based on reference growth standards from the year 2006 [41], [42]. Stunting and wasting were defined for values≤−2 for HAZ and BAZ, respectively, according to WHO standards [43]. The PedsQL 4. 0 SF 15 used in this study includes parallel child self-reports (age range 5–18 years) and parent proxy-reports. It differs from PedsQL 4. 0 generic core scales instrument by the number of items in the total scale and in the various subscales. The PedsQL 4. 0 SF15 consists of 5,4, 3, and 3 for physical, emotional, social and school functioning respectively, making a total scale of 15 items whereas the PedsQL 4. 0 generic core scale comprises of 8,5, 5, and 5 for physical, emotional, social and school functioning, respectively, for a total of 23 items. Details are available at http: //www. pedsql. org/about_pedsql. html. The survey' s preliminary instructions indicate to the subject that survey questions ask how much of a problem each item has been during the past one month. A five-point Likert-like response scale is used. The response scale for each item was “never” (0), “almost never” (1), “sometimes” (2), “often” (3), and “almost always” (4). Responses were transformed to 100,75,50,25, and 0, respectively, resulting in a scale range of 0–100, with the higher number scores indicating better HrQoL. Overall scores and sub-scale scores were computed as the sum of the items divided by the number of items answered (this accounts for missing data). If more than 50% of the items in the scale were missing, the scale score was not computed [44]. Two summaries and one overall score were computed. Of note, the physical health summary score (5 items) is the same as the physical functioning subscale. The psychosocial health summary score (10 items) is computed as the sum of items divided by the number of items answered in the emotional, social, and school functioning subscales. Total Scale Scores for child self-report and parent proxy-report are also presented. In preparation for the study, two forward translations (English-Swahili) and one backward translation (Swahili-English) translations were done for the PedsQL 4. 0 SF15. The approved final Swahili translations of the PedsQL 4. 0 SF15 were first pre-tested on 23 young children (5–7 years old), 39 children (8–12 years old) and 36 teens and their parents from a neighboring village. These were all non-participants in the present study but similar to those included in this study. The illiteracy levels in the study area were estimated at about 60% and thus the PedsQL 4. 0 SF15 instrument was interview administered [21], [45] for both children and their parents. All the PedsQL 4. 0 SF15 questionnaires in this study were administered by two well-trained research assistants who were native speakers of Swahili, under the supervision of the first author. Feasibility of the PedsQL 4. 0 SF15 generic version was determined from the average percentage of missing responses. The percentage of all possible item responses left unanswered was calculated for each subject on each single and summary scale and averaged over subjects. Utility of the instruments in terms of distributional coverage overall and by subscale was evaluated by calculating the percentage of subscale-level average responses reaching the minimum (floor) or the maximum (ceiling) of the scoring scale. In QoL studies, floor and ceilings effects are used to evaluate the depth of a health problem being measured. If floor effects exist, it means the QoL tool is showing a lower than actual HrQoL, and if ceiling effects exist, QoL tools may be underestimating QoL or the magnitude of the problem being measured. Studies with small floor or ceiling effects (1–15%) are considered to meet acceptable measurement standards, whereas studies with moderate floor or ceiling effects (>15%) are considered less precise in measuring latent constructs at the extremes of the scale [25], [47]. Internal consistency is a measure of the extent to which items in a questionnaire (subscale or scale) are correlated (homogeneous), thus measuring the same concept. It is an important measurement property for questionnaires that intend to measure a single underlying concept (construct) by using multiple items such as PedsQL questionnaire. Cronbach' s alpha coefficient was utilized to determine scale internal consistency reliability [48]. Scales with reliabilities of 0. 70 or greater are recommended for comparing patient groups, while a reliability criterion of 0. 90 is recommended for analyzing individual patient scores [49]. A low Cronbach' s alpha indicates a lack of correlation between the items in a scale, which makes summarizing the items unjustified. A very high Cronbach' s alpha (>0. 95) indicates high correlations among the items in the scale, i. e. , redundancy of one or more items [25]. Construct validity was determined utilizing the known-groups method, which compares scale scores across groups known to differ in the health construct being investigated. In our study the known groups are the children who tested positive and negative for S. haematobium egg output, who will henceforth be referred to as ‘Sh egg-positive’ and ‘Sh egg-negative’, respectively. Known groups validity was examined through a comparison of these egg-positive and egg-negative groups, of children from families of lower versus higher SES, of stunted versus non-stunted children, and those from high prevalence (high risk) villages versus moderate prevalence (lower risk) villages, using independent t-tests. To complement statistical testing, effect sizes are presented to assist in the interpretation of the relative degree of between-group score differences by indexing these differences to within-group score variation [50], with lesser import if the between-group score difference is small relative to the within-group variation in scores. Effect size utilized in these analyses was calculated by taking the difference between the score means for either cases vs. controls, stunted vs. non-stunted children, children from low SES vs. high SES, or children from high-risk vs. moderate risk villages, divided by the pooled standard deviation of the egg-negative/high SES/not stunted/lower-risk village categories, as appropriate [51]. Effect sizes for differences in means are designated as small (0. 20–0. 49), medium (0. 50–0. 79), and large (≥0. 80) in magnitude [52]. Agreement between child self-report and parent proxy-report was determined through 2-way mixed-effect model (absolute agreement, single measure) intraclass correlations [53]. Intraclass correlation results are generally interpreted as follows: ≤0. 40, poor to fair agreement; 0. 41–0. 60, moderate agreement; 0. 61–0. 80, good agreement; and 0. 81–1. 00, excellent agreement [54]. Following our observation of significant differences in group-wise mean HrQoL scores in the analyses described above, our next objective was to determine the independent contributions of age, sex, village risk for schistosomiasis (high vs. moderate), S. haematobium infection, hookworm infection, anemia, SES, stunting, and wasting in nested models of HrQoL outcomes, for both total and psychosocial sub-scale scores. To do this, we used generalized multivariable linear modeling, adjusted for covariance at the village level using generalized estimating equation (GEE) technique (SPSS). Stepwise backward removal of non-significant variables was used to create ‘best fit’ parsimonious models based on Akaike information criteria (AIC) retaining explanatory variables with P-values<0. 1. Significant multiply-adjusted parameter estimates are reported (with 95% CI and corresponding p-values) for covariates remaining in the final models. In total, 1580 children, aged 5–18 years old, from five villages participated in parasitological and anthropometric testing. Their mean age was 10. 6±3. 5 and 51% were female. The majority of the children in Milalani, Magadzoni, and Gwadu (the more rural villages) came from families of low SES while most children in Kinango A and Dzitenge (the more urban villages) came from families of higher SES (Table 1, Figure 1). The overall S. haematobium infection prevalence was 42. 2% (766/1580), similar for males and females, but significantly different by village (Table 1) and by age group (Table 2). School-age infection prevalence was significantly greater in Milalani, and Gwadu villages (here referred to as high-prevalence villages, according to WHO guidelines [55]) compared to Magadzoni, Dzitenge, and Kinango A villages (here referred to as moderate prevalence villages). S. haematobium infection intensity was highest in 8–12 year olds and lowest in 5–7 year olds (P<0. 01, Table 2), and varied significantly by village (Table 1) with an overall geometric mean intensity of infection of 37. 8 eggs/10 mL of urine. Males had significantly heavier infection than females (P<0. 02). There were also significant inter-village and across-age group differences in the proportion of anemic children and in mean Hb levels (Tables 1 and 2). Many study children had either acute undernutrition, as measured by wasting (BAZ score ≤−2), chronic undernutrition, as measured by stunting prevalence (HAZ score ≤−2), or both. The highest malnutrition levels were recorded in villages closer to the coastline (Milalani, Magadzoni, and Gwadu) compared to the more inland villages (Dzitenge and Kinango A) (Table 1). Wasting and stunting were lowest in the 5–7 year age group compared to older age groups (Table 2). Significantly more males were stunted (56% vs. 44%, Χ2 = 9. 2, P<0. 01) or wasted (60% vs. 40%%, Χ2 = 14. 8, P<0. 001) as compared to females. This is a first attempt to measure self-rated multidimensional HrQoL related to urogenital schistosomiasis in children. In our initial analysis, we observed a clear trend toward lower HrQoL in all measurement scales (except physical scale) among children with S. haematobium egg-positive status, but the differences were not significant. This trend was confirmed in parent-proxy reports, which indicated the same trend in all performance scales, and for the social scale, parents of egg-positive children did report significantly lower HrQoL scores than parents of egg-negative children. Significant differences were observed in other group-wise analysis, indicating lower HrQoL for children resident in high-S. haematobium prevalence villages as compared to those living in moderate-prevalence villages, and significantly lower HrQoL for children with growth stunting. Previous investigations described strong associations between growth stunting and low intensity Schistosoma infections [6], [56]–[58]; thus the joint association of lower HrQoL with local prevalence and stunting was an expected finding. Our results are in contrast to recent studies of other NTDs that suggest no clear differences in HrQoL between infected and non-infected health states [14], [16]. Increasing severity of intestinal schistosomiasis, whether caused by S. japonicum in P. R. China [13], [15] or by S. mansoni in Egypt [20], has been associated with significantly greater reductions in self-reported quality of life scores. It is possible that inclusion of a larger range of age-groups, and patients with more advanced complications of S. haematobium would have resulted in more pronounced differences in our observed outcomes. We found significant associations between either i) lower socio-economic status, or ii) residence in high risk villages, with lower PedsQL 4. 0 SF15 scores across all scales, except for physical and school functional scales. Socio-economic status is known to affect the risk of Schistosoma infection by either limiting water-use options and/or access to health care, or through other poverty-related factors [11], [57], [59], [60]. In our study, the initial lack of significant differences in HrQoL between children with and without egg output could be attributed to confounding by the distribution of low SES and undernutrition across the village levels. Our multivariable analysis indicated a strong village-type effect (with high-prevalence status having significantly lower HrQoL scores), but with significant interaction between village type and all other covariates tested. A further stratified analysis indicated that S. haematobium egg positivity was significantly associated with reduced HrQoL in moderate prevalence villages, whereas in high-prevalence villages, it was associated with higher HrQoL. This difference is likely to reflect issues with the imperfect sensitivity of our standard parasitological diagnosis of urogenital schistosomiasis (urine egg output [61]), and covariation between age- and SES-related infection risk. In addition, for schistosomiasis and other neglected tropical diseases, the time from the onset of active infection to the onset of clinical disease may be protracted. Also, a single urine exam may miss 20–30% of active infections that would be found by repeated urine exams [61]. Serological testing indicates that among children in high-prevalence villages, active or recent Schistosoma infection prevalence is close to 100% [23]. Light intensity infections are more easily missed in parasitological testing. However, these can still result in significant inflammation-mediated morbidity. We believe that in high-prevalence villages, infection status is already saturated, so that egg-positivity reflects mainly those with heavier infections. In the high-prevalence setting, this sub-group includes older children with greater mobility [62], who may therefore have better scores on the PedsQL scales. In moderate prevalence villages, where egg testing is more likely to reliably distinguish infected and uninfected children, S. haematobium egg-positivity was significantly associated with lower HrQoL scores. Overall, uncertainty (negative study bias) about the status of egg-negative controls may explain why we observed significant ‘infection-related’ differences in moderate-risk villages but not in high-risk villages. Finally, because PedsQL 4. 0 SF15 questionnaire administration was not administered at the same time as the parasitological assessment, some of the ‘controls’ identified during parasitological assessment may have turned into ‘cases’ (i. e. , those who were now or formerly S. haematobium infected) by the time HrQoL assessment was done. Because treatment was given before administration of the Peds QL survey, acute impacts of active infection may have gone missed, and these may have had an impact on our study findings. We recommend that similar studies in future carry both assessments concurrently, and include supplemental serologies (circulating antigen testing, anti-Schistosoma IgG4), in order to minimize the misidentification of the true infection status of ‘cases’ and ‘controls’. The mode of administration of the PedsQL 4. 0 SF15 questionnaire was interviewer-based; thus, the high response rate was not surprising. However, it may have contributed to the lack clear differences between the cases and controls, because interviewer-based modes of questionnaire administration suffer from a well- recognized problem of limited willingness to acknowledge a problem [24]. Since all interviews across all villages were conducted by two trained research assistants using one-to-one administration, we believe the interviewer effect was not significant. For the study children, inclusion of a parent-proxy survey, in addition to the direct patient survey, was considered to add external evidence of the validity of the answers we obtained from the children. Floor effects were largely absent in the study population, except in the school scale where negligible (<1. 3%) floor effects were observed, especially among controls. On the other hand, substantial ceiling effects were evident in almost all scales, and were more prominent in physical and school scales particularly among controls (Sh egg - negative). This means that the PedsQL 4. 0 SF15 questionnaire likely underestimated HrQoL, especially for children who tested negative for S. haematobium egg output. While ceiling effects are a common phenomenon, they restrict the ability of the HrQoL tool to detect change or describe health above the average in more healthy populations [16], [63]–[65] We consider the PedsQL to be a suitable tool for assessing quality of life in children with schistosomiasis. There is evidence that its reliabilities were high (alphas generally ≥0. 70), floor effects were acceptable and identification of children from both low SES and ‘high risk’ villages was valid. PedsQL 4. 0 SF15 was an effective tool for measuring quality of life in children living in schistosomiasis- (and likely other neglected tropical disease-) endemic areas. The clear capability of PedsQL 4. 0 SF15 tool to identify geographical areas with different transmission intensities and SES groups is particularly important in schistosomiasis control. In large scale schistosomiasis control programs, its practicability (administered within 5 minutes) means it can appropriately be used to rapidly tease out high transmission localities for further adjustment. However, further research is needed, especially on its reproducibility and responsiveness (ability to detect clinically important changes over time) in relation to schistosomiasis. The difficulties in measuring physical and school health illustrated here also point to the need for further research and the development of a schistosomiasis specific PedsQL tool to enhance assessment of Schistosoma infection-related health impact.

Because urogenital schistosomiasis is a multi-decadal chronic disease that begins in early childhood, and because it is a disease that may affect nearly everyone in endemic communities, its impact on personal health-related quality of life (HrQoL) has been difficult to gauge accurately. In order to provide a more precise estimate of schistosomiasis' impact on overall health status, we used a standardized questionnaire, the PedsQL SF15 Inventory, to interview children, aged 5–18 years, and their parents, to quantify their reported physical, social, emotional, and scholastic performance status. Scores were significantly lower in villages having high Schistosoma prevalence, as compared those having moderate prevalence. In adjusting for age, sex, socioeconomic standing, undernutrition, anemia, and hookworm parasites, we found that relative poverty, stunting, wasting, and S. haematobium infection were significant correlates of HrQoL scores, with differential effects in high- and moderate-prevalence communities. The greatest differences were noted in the psychosocial domains of performance. We conclude that exposure to urogenital schistosomiasis has an overall detrimental effect on HrQoL at a level of 2–4% impairment. New implementation of better diagnostics for children is expected to refine our estimates of this association, as will follow-up studies of HrQoL following effective individual and community deworming.

Abstract Introduction Methods Results Discussion

medicine public health and epidemiology growth retardation social and behavioral sciences infectious disease epidemiology sociology pediatrics social epidemiology nutrition clinical epidemiology global health neglected tropical diseases epidemiological methods malnutrition disease mapping infectious diseases disease ecology environmental epidemiology epidemiology pediatrics and child health social welfare child health public health health screening schistosomiasis survey methods environmental health pediatric epidemiology socioeconomic aspects of health child development

2013

Evaluation of the Health-related Quality of Life of Children in Schistosoma haematobium-endemic Communities in Kenya: A Cross-sectional Study

Rabies, resulting from infection by Rabies virus (RABV) and related lyssaviruses, is one of the most deadly zoonotic diseases and is responsible for up to 70,000 estimated human deaths worldwide each year. Rapid and accurate laboratory diagnosis of rabies is essential for timely administration of post-exposure prophylaxis in humans and control of the disease in animals. Currently, only the direct fluorescent antibody (DFA) test is recommended for routine rabies diagnosis. Reverse-transcription polymerase chain reaction (RT-PCR) based diagnostic methods have been widely adapted for the diagnosis of other viral pathogens, but there is currently no widely accepted rapid real-time RT-PCR assay for the detection of all lyssaviruses. In this study, we demonstrate the validation of a newly developed multiplex real-time RT-PCR assay named LN34, which uses a combination of degenerate primers and probes along with probe modifications to achieve superior coverage of the Lyssavirus genus while maintaining sensitivity and specificity. The primers and probes of the LN34 assay target the highly conserved non-coding leader region and part of the nucleoprotein (N) coding sequence of the Lyssavirus genome to maintain assay robustness. The probes were further modified by locked nucleotides to increase their melting temperature to meet the requirements for an optimal real-time RT-PCR assay. The LN34 assay was able to detect all RABV variants and other lyssaviruses in a validation panel that included representative RABV isolates from most regions of the world as well as representatives of 13 additional Lyssavirus species. The LN34 assay was successfully used for both ante-mortem and post-mortem diagnosis of over 200 clinical samples as well as field derived surveillance samples. This assay represents a major improvement over previously published rabies specific RT-PCR and real-time RT-PCR assays because of its ability to universally detect RABV and other lyssaviruses, its high throughput capability and its simplicity of use, which can be quickly adapted in a laboratory to enhance the capacity of rabies molecular diagnostics. The LN34 assay provides an alternative approach for rabies diagnostics, especially in rural areas and rabies endemic regions that lack the conditions and broad experience required to run the standard DFA assay. Rabies is an acute progressive viral encephalitis characterized by central nervous system disorder that ultimately leads to death [1]. Rabies is the only disease considered to have nearly 100% mortality, and remains a major public health problem in Asia and Africa with 70,000 human deaths annually, of which approximately 21,000 occur in India alone [2,3]. Rabies results from an infection by different species of the genus Lyssavirus, family Rhabdoviridae, with Rabies virus (RABV), the type-species of the genus responsible for the majority of deaths [4]. Lyssaviruses are grouped into at least three phylogenetic groups: Phylogroup I (RABV, Aravan virus [ARAV], Khujand virus [KHUV], Bokeloh bat lyssavirus [BBLV], Duvenhage virus [DUVV], European bat lyssavirus 1 [EBLV-1], European bat lyssavirus 2 [EBLV-2], Australian bat lyssavirus [ABLV], and Irkut virus [IRKV]), Phylogroup II (Mokola virus [MOKV], Shimoni bat virus [SHIBV], and Lagos bat virus [LBV]) and Phylogroup III (Ikoma lyssavirus [IKOV], West Caucasian bat virus [WCBV] and recently proposed Lleida bat virus [LLBV]); and the novel Gannoruwa bat lyssavirus species [5–8]. The significant diversity in genome sequences between RABV, especially among lyssaviruses, make it difficult to develop a single, robust, easy to use diagnostic assay. Currently, a definitive rabies diagnosis can only be confirmed by post-mortem testing. The World Health Organization (WHO) and the World Organization for Animal Health (OIE) have defined the direct fluorescent antibody (DFA) test as the gold standard for rabies diagnosis [9]. The DFA test is a rapid and sensitive method, but its accuracy depends on the quality of brain tissue, availability of high-quality anti-rabies diagnostic conjugates, accessibility to a fluorescence microscope and, most important, an experienced diagnostician [10]. Virus isolation is recommended by OIE in cases of inconclusive DFA results and for detailed molecular characterization of the virus [4]. The laboratory criteria for ante-mortem diagnosis of human rabies include a clinically compatible case that must be confirmed by the detection of lyssavirus antigens, specific antibody, or viral RNA in a clinical specimen at a state or federal public health laboratory [11]. Molecular methods have several advantages in rabies diagnostics compared to other methods, including improved sensitivity, higher throughput, and the potential for variant typing. RT-PCR assays are essential for the detection of Lyssavirus RNA in tissues with low viral load such as saliva, nuchal skin biopsy, and eyewashes; moreover, RT-PCR assays can be used to assess the efficacy of experimental therapeutics as well as for risk assessment to prevent nosocomial transmission [12,13]. In the past, several gel-based conventional RT-PCR assays using simple, nested, or hemi-nested approaches for Lyssavirus RNA detection in clinical samples have been published [14,15]. The amplicons generated by conventional RT-PCR assays require sequencing for confirmation and for phylogenetic analyses. These assays are reported to be highly sensitive when using a hemi-nested approach [16]; however, conventional RT-PCR, especially hemi-nested RT-PCR, is prone to non-specific amplification, leading to false-positive results. An amplicon sequencing step, which is both time consuming and labor intensive, is necessary to rule out potential false positive cases. Real-time RT-PCR based assays have improved sensitivity and specificity compared to other methods, highlighting the potential of a real-time RT-PCR assay to become a leading assay for rabies laboratory diagnosis. Over the last decade many studies have been published assessing real-time RT-PCR assays for the detection of RABV and other lyssaviruses via SYBR Green or TaqMan probe [17–21] methods. TaqMan probe based real-time RT-PCR assays often have lower sensitivity relative to SYBR Green based assays due to the sensitivity of TaqMan probes to the diversity of target sequences in the genomes of different RABV and other lyssaviruses [18,22–24]. SYBR Green technology based real-time RT-PCR assays may provide greater sensitivity but can carry significant risk of false positive results when performed without a confirmatory sequencing step [20]. Typically, a combination of two or more assays is used to improve assay sensitivity and account for the breadth of lyssavirus diversity [25]. In this study, we investigated the potential of a new real-time RT-PCR assay, which utilizes degenerate, multiplex primers and probes and minor-groove-binding protein (MGB) or locked nucleotide (LNA) modified probes to overcome the limitations of prior real-time RT-PCR rabies diagnostic assays. This new assay is able to detect representatives from all formally accepted RABV variants and all other Lyssavirus species with superior sensitivity and specificity, compared to traditional hemi-nested RT-PCR methods. Brain tissue samples from rabid animals were obtained via routine surveillance activity of the Centers for Disease Control and Prevention (CDC; Atlanta, GA, USA). Institutional Animal Care and Use Committee (IACUC) or ethics committee approval was not necessary because moribund animals, dead animals, or animals involved in human rabies exposure were collected by the US State Health Departments, US Department of Agriculture and veterinary laboratories during routine surveillance and diagnostic service. The authors did not perform any animal sampling during this study. Primer and Probe Design: The LN34 assay amplifies a 165 nucleotide region including the leader sequence, the transcription initiation signal, and part of the coding sequence of the nucleoprotein (N) gene. The forward primers are a multiplex of two oligos that target the beginning of the RABV genome, whereas the reverse primer targets the N gene and contains 6 degenerate nucleotide modifications. The LN34 probe targets the transcription initiation signal sequence and contains a single degenerate pyrimidine (C or T) at position 65. Probes were modified by either LNA or by MGB. The sensitivities and specificities of the LNA and MGB modified probes were compared in the validation process. Table 1 contains primer and probe sequences and modifications. Degenerate nucleotides are indicated using the IUPAC nucleotide ambiguity code. Two LNA modified probes, LN34 and LN34a, contain a 5′FAM florescent label indicated in parentheses, LNA modified nucleotides indicated by a plus preceding the location of the LNA base in the sequence (e. g. +A, +G, +C, +T), and a 3′ BHQ1 quencher (Table 1, in parentheses). The LN34lago probe is specific for Lagos bat virus lineage B and C subspecies and was mixed with the LN34 probe in a multiplex format; LN34m is an MGB modified probe with a 5′ 6FAM florescent label and 3′ MGB and NFQ quencher. An artificial positive control RABV RNA was developed for the assay based on a previous publication [26] and was used to identify potential contamination. The sequence contains 127 bases: oLPC-rabies3-4: GCA CAG GGT ACT TGT ACT CAT ACT GAT CTG AAT CCA TTG TAG AGG TGT TAG AGC ACG ACA GGT TTC CCG ACT GGA TCT TTC TTT GAT CTG GTT GTT AAG CGT TCG CCC TAT AGT GAG TCG TAT TAC A A previously described β-actin real-time RT-PCR assay was used in this study, with slight modifications, as an internal or negative control [23]. βactin probe: (HEX) -TCC ACC TTC CAG CAG ATG TGG ATC A- (BHQ1) β-actin forward primer: CGATGAAGATCAAGATCATTGC β-actin reverse primer: AAGCATTTGCGGTGGAC The reaction conditions were optimized for multiple factors, including annealing temperature, length of reverse transcription and PCR reaction steps, as well as the ratio of primers, probes, and master mix components. The optimized reaction conditions are as follows: Ag-Path ID One-Step RT-PCR Kit (Life Technologies) was used with either the LN34 assay primer and probe sets (LN34 and LN34lago) or the β-actin assay primer and probe set. One femtogram (about 10,000 copies) of the artificial RABV RNA template was used as positive control for the LN34 assay. One μl of forward and reverse primer stocks (10 μM) and 1. 0 μl of probe (5 μM) were used in the 25 μl reaction set up following the directions of the commercial kit. The cycling conditions are as follows: reverse transcription at 50°C for 30 minutes, followed by RT inactivation / initial denaturation at 95°C for 10 minutes, and amplification for 45 cycles at 95°C for 1 second and 56°C for 20 seconds on a ViiA7 real-time PCR system (Applied Biosystems). We used the One-step RT-ddPCR advanced kit (Bio-Rad, CA), which delivers improved efficiency and specificity for precise RNA target quantification by Droplet Digital PCR. Details of the reaction set up and cycling conditions are provided in S1 Text. All clinical samples were tested for the presence of RABV antigens using the DFA test using fluorescein-isothiocynate (FITC) -conjugated anti-RABV monoclonal antibodies (Fujirebio Diagnostics Inc. , Malvern, PA, USA Cat 800-092 and EMD LIGHT DIAGNOSTICS Rabies DFA Reagent Cat 5100) [9]. Total RNA was extracted from brain, skin, cornea, saliva, or tissue culture supernatant using Trizol reagent (Invitrogen) according to the manufacturer’s instructions. Reverse transcription, amplification of the viral RNA and sequencing of the N gene were performed as described by Hughes et al. 2006 [27] BioEdit software (http: //www. mbio. ncsu. edu/BioEdit/bioedit. html) was used for the alignment of LN34 target sequences. N gene sequencing was performed as described by Hughes et al. 2006 [27]. Eighty-eight whole N gene nucleotide sequences included in the validation panel were subjected to multiple sequence alignments using CLUSTALW (http: //www. ebi. ac. uk/clustalw/index. html). All sequences were edited to 1350 bp fragments using BioEdit (S2 Text) [28]. A similarity matrix was calculated using the 88 aligned sequences with MEGA 6 and is summarized in S1 Table [29]. A phylogenetic tree was constructed using the full N gene sequences of the 88 samples in the validation panel and 104 sequences retrieved from GenBank (S1 Fig) using the neighbor-joining method and Kimura-2 parameter substitution model with 1000 bootstrap replicates. We sought to design and develop an assay capable of amplifying and detecting a region of the genome conserved across divergent Lyssavirus species. The LN34 assay amplifies a 165 nucleotide region including the 58 nucleotide leader sequence, the 12 nucleotide transcription initiation signal, and the first 95 nucleotides of the coding sequence of the N gene (Fig 1). The leader region and transcription initiation signal are strictly conserved in length in all known lyssaviruses [30]. The forward primers target the first 25 nucleotides of the leader sequence with a low level of degeneracy, while the reverse primers use 6 degenerate nucleotides to cover the significant diversity of its target sequence within the N gene. The LN34 probe targets the transcription initiation signal sequence from position 59 to 75 and utilizes a single degenerate pyrimidine (C or T) at position 65. This duplex probe design matches perfectly to all published RABV and other Lyssavirus sequences except a few striped skunk RABV variants, MOKV, IKOV, IRKV and LBV subspecies lineages B and C. Fig 1 summarizes the sequence variation in the genomic region corresponding to the probe target sequence from over 200 lyssaviruses. To compensate the low melting temperature of the 17 nucleotide probes, the probes were modified either by MGB protein (LN34m) or LNA (LN34). The melting temperatures of the modified probes were calculated using either the software Primer Express 3. 01 for the MGB probe (Life Technology, California) or the website at http: //www. idtdna. com/calc/analyzer for the LNA modified probes (Table 1). Several LNA modified probes containing 5 to 7 LNA nucleotides at different positions of the probe sequence were evaluated using a number of RABV and other Lyssavirus samples. Of these, two LNA modified probes (LN34 and LN34a) yielded the best results overall, and the LN34 probe was selected for comparison with an MGB modified probe (LN34m). The CDC rabies repositories have limited RABV samples from Asia (India and Sri Lanka) and Africa (lineage 1) for the validation process. We used in silico analysis to overcome this shortcoming; the sequences of representatives from all major clades or lineages of South-East Asia RABV strains, African RABV lineage 2 and 4 (Fig 1, sequences with blue labels) were analyzed by sequence alignment. As the LN34 assay target sequences are highly conserved or identical among those RABV representatives labeled in blue in Fig 1, LN34 is likely to be able to detect other RABV variants within the same phylogenetic cluster. The global phylogenetic tree in S1 Fig was generated using sequences from 192 RABV and other lyssaviruses. These 192 sequences were chosen based on their comprehensive coverage of Lyssavirus species, variants, and geographic locations. The sequences labelled in red correspond to samples that were directly tested in the LN34 assay validation experiments. A multiple sequence alignment revealed that all the RABV sequences in the global tree have conserved LN34 targeting sequences. The sequences used in this alignment were selected from over 280 published Lyssavirus genome sequences in Genbank; only sequences from representatives of RABV and other lyssaviruses that exhibited variation at the probe targeting positions or were representative of unique clades not available in the validation panel are displayed in the sequence alignment. The available BBLV genome sequence lacks the first 29 bp. The available LLBV genomic sequence has limited N gene coverage and is not shown in the sequence alignment. The positions of the forward primer (For), reverse primer (Rev), and probe are listed in Table 1. For the 31 sequences included in Fig 1: samples used in the validation panel are labeled in red with the corresponding Genbank accession number, Lyssavirus species, three letter country code and host, if known. Samples that were not available for validation but are representative of unique clades of the global phylogenetic tree (S1 Fig) are labeled in blue. After initial validation of the LN34 and LN34m probes and parameter optimization using a small set of selected RABV RNA, the LN34 and LN34m probes were evaluated using a panel of RABV samples from the CDC repository. This validation panel (n = 88) was selected to include representative variant samples from around the world (Table 2) including 59 variants of highly divergent RABV isolates from Asia, Africa, Europe, South America and North America. The validation panel also included representatives of 13 Lyssavirus species, but did not include LLBV or the recently identified Gannoruwa bat lyssavirus species [8]. Both LN34m and LN34 probes were able to detect all variants from the validation panel except the subspecies lineage B and C of LBV (see below) [32]. For the majority of the isolates in the validation panel, the LN34m and LN34 probes demonstrated similar sensitivity, as shown by their similar cycle threshold (Ct) values (Table 2). However, for those isolates with known sequence variations from the probes, the performances of the LN34m and LN34 probes differed considerably. The stripped skunk RABV and IRKV each has a single nucleotide polymorphism (SNP) at position 69 of the probe sequence (Fig 1). Surprisingly, the Ct values produced using the LN34m probe were reduced by at least 10, compared to that of the LN34 probe for one of the Texas skunk isolate (Table 2, S. No 31) and an IRKV isolate (Table 2, S. No 75). This difference in Ct values equates to at least 100 fold difference in sensitivity. Thus, the MGB modified probe is more sensitive to single SNP changes than the LNA probe. Similarly, MOKV and IKOV (Fig 1) each has a single SNP at position 63 that also led to Ct value differences of 8 to14, which also translates to a significant sensitivity loss when using the LN34m probe (Table 2, S. No. 77–80,88). As the goal of this assay is to attain high sensitivity to detect a broad range of Lyssavirus species, the LNA-modified LN34 probe was used for all future validation. The sequences from lineages B and C subspecies of LBV have two SNPs relative to the probe target sequence, and the validation panel revealed that those subspecies were not detected by either the LN34m or the LN34 probe. Therefore, a specific LNA modified probe (LN34lago, Table 1) was synthesized for the LBV lineage B and C subspecies and combined with the degenerate LN34 probe to make up the LN34 assay. This updated LN34 assay was able to detect all 4 lineages of LBV in our validation panel (Table 2), and the addition of the LBV probe did not affect the sensitivity or the specificity of the LN34 assay. The multiplexed LN34 assay was able to detect all known positive samples present in the validation panel (Table 2) and exhibited better overall sensitivity and specificity compared to the RT-PCR and real-time RT-PCR assays used previously in RABV diagnostics and validation. Next, the LN34 assay was further evaluated using a thorough comparison with a conventional RT-PCR assay and DFA test using clinical samples. The LN34 assay was able to identify all the rabies positive clinical samples identified by the DFA test (Table 3, n = 13), yielding a sensitivity of 100%. The LN34 assay also correctly ruled out negative clinical samples confirmed by the DFA test and conventional RT-PCR (n = 193), leading to 100% specificity for this assay. Many among those negative samples were initially reported as inconclusive or unable to be diagnosed by DFA due to issues related to reagent quality. The conventional RT-PCR assay, run in parallel with the LN34 assay, produced amplification of products from several of the negative samples that exhibited molecular weights similar to the positive control, but these products were determined to be the results of non-specific amplification by sequencing. The LN34 assay performed well on multiple sample types. The negative clinical samples described above included skin biopsies (n = 14) and saliva (n = 15) from patients with neurological symptoms consistent with rabies as well as DFA-negative or inconclusive animal brain samples of varying quality (n = 164). The positive clinical samples included human ante-mortem skin biopsies (n = 2), saliva (n = 4), cornea (n = 1), human post-mortem brain biopsy (n = 1) and animal brain samples (n = 5) (Table 3). Compared to the brain samples, the RABV viral load and/or RNA levels were much lower in the skin biopsy, saliva, and cornea samples, as indicated by their much higher Ct values in both the LN34 and βactin assays. In addition to the source tissue, the quality of rabies diagnostic samples often varies greatly due to storage conditions. Three ante-mortem samples (Lab ID A15-0728, A15-0729, and A15-0731) were collected from the same patient and had been stored at room temperature for an extended period of time due to a transportation delay. Although these samples did not meet the required standards for rabies testing due to sample condition, the results from the LN34 assay revealed high viral RNA load in the brain biopsy sample (Table 3). Interestingly, all three positive human ante-mortem saliva samples had similar LN34 Ct values, indicating similar viral load or viral RNA quantity, although they were collected in different clinical settings and storage conditions. The LN34 assay was also used to test 58 field samples collected during US annual surveillance testing and passive surveillance by wildlife services that were determined to be unfit for the DFA test due to limited tissue or the unavailability of appropriate conjugates. The LN34 assay revealed that 18 (31. 03%) samples were positive (Table 4). A major concern with diagnostic assay development is the prevention of false diagnoses due to sample degradation or sample preparation error. To minimize these risks, we included two different controls within the multiplex assay: detection of the housekeeping gene β-actin by real-time RT-PCR and a positive control for LN34 assay amplification. The β-actin assay was adapted from a previous publication with minor modifications [23] and was implemented to assess the presence of RNA in the sample. A LN34 negative result requires the β-actin assay to be positive for a negative rabies diagnosis. The β-actin assay utilizes labelling by fluorescent HEX and has same running conditions as the LN34 assay. Thus, a diagnostic sample can be run on the same plate or even in the same well in a multiplex format. An artificial positive RABV control RNA was also adapted from a previous publication [26] with minor modifications: a short LacZ gene sequence was added to allow for differentiating between the artificial RNA and a positive sample, which allows for the easy identification of potential carry-over contamination. The artificial positive control RNA generates a 127 bp amplicon in the LN34 assay that can be differentiated from the 165 bp amplicon produced by RABV and other Lyssavirus positive samples on an agarose gel. The artificially generated positive control RNA can be stored for an extended period in a stabilization buffer and can be used as a standard for the comparison of LN34 assay performance between different laboratories and running conditions. We next examined the sensitivity of the LN34 assay, as one of the key benefits of real-time RT-PCR over other antigen-based diagnostics is increased sensitivity. The efficiency of the LN34 assay was estimated to be greater than 92% by testing serial dilutions of RNA extracted from samples positive for RABV strain CVS11 (93%), DUVV (92%), and LBV (98%). The detection limit of the LN34 assay was measured by droplet digital PCR assay using both artificial positive control RNA and RABV strain ERA; the LN34 assay was capable of detecting a single digit number of RNA copies for both RNA samples (S1 Text). Lastly, we evaluated the feasibility of using the LN34 assay to detect all known lyssaviruses by in silico analysis. The N gene of all samples used in the validation panel were sequenced, and the newly generated sequences were deposited in Genbank (S2 Text). A similarity matrix of N gene sequences among validation panel lyssaviruses was calculated using Mega6 software. The N gene sequences of RABV representatives exhibited approximately 80% or higher nucleotide identity with other RABV sequences. The lowest identity values were found among RABV from southeastern North America, including those isolated from the vampire bat of Mexico and the raccoon and South Central skunk of the US. This finding is in agreement with previous RABV variant classifications that suggest American indigenous RABV are one of the most divergent strains [30]. The N gene sequences of the examined representatives from non-RABV lyssaviruses had lower nucleotide identity values ranging from 68% to 79%. Also, according to this analysis, the most divergent Lyssavirus was IKOV, which is also in agreement with previous studies [6]. The combined sensitivity and robustness of the LN34 assay to detect lyssaviruses from all available species highlights its utility as a diagnostic tool applicable for laboratory rabies diagnosis. Real-time RT-PCR based diagnostics are not currently recommended by international agencies for routine diagnosis of RABV, but considerable developments in both PCR platforms and PCR chemistry have made this technology increasingly attractive for decision making in rabies case management. The LN34 assay represents a major improvement over previously published real-time RT-PCR assays for the detection of RABV and other lyssaviruses, as it can detect a broad range of phylogenetic diversity with superior sensitivity and specificity. The LN34 assay has shown equal (or possibly better) specificity and sensitivity to the DFA test, and the one step reaction format reduces the chances of false positive results that are a risk inherent to two step and SYBR Green based RT-PCR assays. Also, through implementing strict quality control procedures (i. e. , the β-actin internal control and artificial positive control), one can identify and reduce both false negative results in degraded samples and carry-over contamination from positive controls [26]. Initial assay development for the LN34 assay was based on more than 280 published full genome sequences from lyssaviruses of all known species (except LLBV), covering all major geographic based clades: American indigenous, Indian, South-East Asian, African lineage 2, African lineage 3, Arctic-related and cosmopolitan strains. The LN34 forward primer consists of degenerate nucleotides near the 3′ end as the first 9 nucleotides of the leader sequence are identical among all examined lyssaviruses; the reverse primer also targets a relatively conserved region of the N gene that has been used for the design of conventional rabies RT-PCR primers previously [19]. The reverse primer used in the LN34 assay is longer and uses more degenerate nucleotides compared to those used in previous studies in an attempt to maximize the universal amplification from divergent lyssaviruses while maintaining a high level of specificity [19]. The sequence of the LN34 probe targets one of the most conserved regions across individuals from all species of lyssaviruses. A well-known rabies RT-PCR primer (JW12) was developed from the same region as the LN34 probe [15], but the LN34 probe is shorter than JW12 and has identity with most known RABV sequences when a single degenerate site is accounted for (excluding some stripped skunk RABV variants). Additionally, genomic sequences from representatives of at least 8 lyssaviruses, which are more diverse than RABV isolates, exhibited 100% nucleotide identity with the LN34 probe (Fig 1) and could be amplified efficiently using the LN34 primers (Table 2). The low level of sequence degeneracy designed in the LN34 primers and probes allows the LN34 assay to maintain an optimal degree of sensitivity and specificity for rabies diagnostics. Previous studies have shown that LNA and MGB probes have similar sensitivities [33]. Our evaluation of the MGB and LNA modified LN34 probes revealed similar sensitivities for most of the RABV and other lyssaviruses tested, which agrees with sequence alignments showing that probe target sequence is identical in most of the validation panel samples. However, we observed differences in the sensitivities of the two probes for a few variants with genomic sequence variations that did not correspond to the probe sequence (Fig 1). The LN34 MGB probe was very sensitive to single nucleotide changes in the probe’s target sequence, whereas the LN34 LNA probe exhibited a higher tolerance to single nucleotide mismatches. To allow for detection of a broader range of lyssaviruses, the LNA-modified probe was implemented in the LN34 assay. It remains unclear if the differences in probe tolerance to mismatches is specific to the LN34 probe or it if may affect other probe designs. The LN34 assay was validated using a large and highly diverse panel containing representative members of major RABV variants and 13 other lyssaviruses. The LN34 Ct values corresponding to all RABV variants tested were highly positive even when samples in the validation panel were diluted 10 fold from the original samples. Since LBV lineage B and C subspecies could not be detected by the initial probe design, a multiplex format containing a specific probe for this subspecies was developed to cover this detection flaw. The additional probe did not affect assay performance, suggesting that the LN34 assay can accommodate changes as novel lyssaviruses emerge. In fact, the newest Lyssavirus LLBV, phylogenetically grouped with IKOV (Phylogroup III), was not validated in the LN34 assay [6] because the available sequences do not cover the LN34 probe region and, thus, could not be evaluated in an in silico analysis. However, as LLBV sequences and samples become available, a LLBV specific probe can be added to the LN34 assay if the region contains polymorphisms. The average Ct values of samples in the validation panel were 24. 490 and 22. 886 using the LN34 and LN34m probes, respectively, after a 10 fold dilution of most RABV samples, testing both brain and cell culture samples. The samples from the annual inter-lab testing 2015 panel, had the lowest Ct values. The range of Ct values observed was most likely due to differences in sample quality, as some of the samples in the validation panel had been frequently used and have degraded over time. Some viral strains, i. e. DUVV and EBLV-1, were prepared in different batches or at different times and varied considerably in their LN34 Ct value (Table 2). For those samples where Ct values differed by more than 3, the samples were re-run on the same plate, and only RABV isolates from Gabon Africa (Table 2 S. No. 15) and a bat associated RABV from the US (Table 2, S. No. 18,27) still exhibited Ct value differences greater than 2, which may be due to variations in the sequences targeted by the primers. We also noticed slight variations in Ct values produced by the LN34m and LN34 probes, which were likely caused by the assays being completed independently in separate batches. A key benefit of RT-PCR based assays is the ability to detect limiting quantities of substrate, which has large implications when clinical sample quality and quantity are sub-optimal. Our analysis revealed that the sensitivity of the LN34 assay was consistently superior or equivalent to that of the conventional hemi-nested RT-PCR assay for multiple validation panel samples, although not all samples were tested. Based on the absolute quantification using ddPCR and taking into account assay efficiency, the LN34 assay was able to detect a single digit copy of the RNA template from rabies clinical samples. As the majority of available RABV and other Lyssavirus genomes have sequences identical to those of the LN34 probes and primers, the amplification efficiency and the limit of detection of the LN34 assay should remain relatively consistent for most lyssaviruses. The high sensitivity of the LN34 assay is important for the diagnosis of samples that have been stored or transported under sub-optimal conditions and cannot to be diagnosed by other methods. The standard protocol for DFA testing requires that samples be maintained under cold chain conditions, which is not always feasible. Diagnosis based on ante-mortem samples is particularly challenging, as viral shedding is not only low but is also intermittent; frequent monitoring is necessary, which can be rapidly facilitated using the LN34 assay. In this study, the LN34 assay proved capable of diagnosing several samples that were considered unfit for DFA testing or that were inappropriate for analysis by the DFA test, including surveillance samples as well as multiple ante-mortem samples, despite relatively low level of rabies RNA. All LN34 positive ante-mortem samples were confirmed by direct sequencing of the amplicons produced in the LN34 assay, which could be used to generate genotyping results for the diagnostics. Each year, rabies surveillance in the United States involves testing over 100,000 suspected rabid animal samples, of which over 6,000 rabid animals were diagnosed in 2014 [34]. The inherent high throughput nature of the LN34 assay has the potential to improve the speed of testing and reduce the workload for diagnostic laboratories. Furthermore, for those thousands of samples unfit for DFA testing due to limited sample quantity or sample degradation, the LN34 assay can be used as the diagnostic test of choice, since it requires very little material. We tested a total of 58 samples considered unfit for DFA testing. Of these, 18 samples (31%) were positive by the LN34 assay (Table 4). By applying the real-time RT-PCR LN34 assay, the number of samples unsuitable for testing can be further reduced, which will improve the surveillance and control of rabid animals in the US. Current rabies sample collection and storage methods can be tedious, using real-time RT-PCR based diagnostic methods such as the LN34 assay can reduce the burden of sample collection, transportation, and storage for rabies diagnostics and surveillance. This is of particular importance in rural areas or developing countries which often have limited resources and lack experienced laboratorians. Further, nucleic acid based detection techniques are comparatively easier to perform and the technology is deployable to be used in the field. Since this technique has the potential to be used as a high throughput assay, it can thus can be used as a tool for real-time rabies surveillance to check the spread of the virus in wildlife species as well to monitor the usefulness of control programs like oral rabies vaccination. The LN34 assay allows for testing samples that have been stored in stabilization buffers, which means that clinical samples can be stored for extended periods of time under less stringent conditions and significantly reduces the need for cold chain and its associated costs. Furthermore, the collection method can be modified or simplified to reduce the potential for contamination as well as mitigate the risk to the person obtaining the sample; for example, the implementation of a real-time RT-PCR assay would allow for use of the OIE approved rapid sample collection of brain samples which avoids the process of opening the skull and collecting the sample from the occipital foramen (which is particularly challenging in large animals) [35]. Additionally, the artificial RNA control can be produced in large quantities using commercial kits and stored for an extended time using an RNA storage buffer, and the RNA control can be distributed to different laboratories to monitor the performance of the LN34 assay and improve quality control for rabies diagnostics. WHO and OIE set a goal for canine rabies eradication by 2030; a key component for the success of this goal is to improve rabies diagnostics and surveillance, especially in developing countries. LN34 assay is able to overcome many hurdles for rabies diagnostics in resource limited areas: it reduces the logistical burdens associated with the rabies sampling process, storage, and transportation and the quality of test standards and reagents can be quickly established across laboratories around the world, as the LN34 assay can leverage the current real-time RT-PCR platforms and skills in most laboratories with limited additional training required. Our validation results and the preliminary results in a pilot program implementing the LN34 assay in multiple laboratories show that the LN34 assay increases the accuracy of rabies diagnostics. We expect that the LN34 assay will improve rabies diagnostic capacities in many laboratories around the world.

Rabies is a preventable disease–but is still responsible for approximately 70,000 human deaths worldwide each year. The majority of human deaths occur in Asia and Africa where there is a lack of diagnostic resources and expertise, making it difficult to develop effective prevention and control strategies. In recent years, several real-time RT-PCR based diagnostic assays have been introduced to many developing countries in an effort to control the H1N1 pandemic flu, Ebola outbreak, and other tropical viral infections. In an effort to further improve rabies diagnostics, we developed a pan-lyssavirus Taqman real-time RT-PCR assay called LN34 for the detection of all known RABV variants and other lyssavirus species. The LN34 assay uses a combination of degenerate nucleotides, multiplex primers and probes, and unique probe modifications to achieve superior sensitivity and specificity compared to previously published RT-PCR based rabies diagnostics. Equally important, the LN34 assay is simple to set up, high throughput, combines multiple standard controls and can be used directly in widely available real-time RT-PCR systems. The LN34 assay was validated using a broad and comprehensive panel of highly diverse RABV variants and other lyssaviruses. A validated universal rabies diagnostic assay will be important in regions where RABV and other lyssaviruses co-circulate and for establishing a widely accepted diagnostic protocol. Over 200 clinical samples (including ante-mortem, post-mortem, and field derived samples) were tested with the LN34 assay, and the assay achieved 100% diagnostic sensitivity and specificity in our laboratory. Over 300 published genome sequences from representatives of RABV and other lyssaviruses were found to contain the conserved LN34 primer and probe targeting sites in an in silico analysis. We are expanding the validation of the LN34 assay to multiple domestic and international laboratories and expect the LN34 assay will drastically improve rabies diagnostic capacities globally.

Abstract Introduction Materials and Methods Results Discussion

reverse transcriptase-polymerase chain reaction sequencing techniques medicine and health sciences pathology and laboratory medicine pathogens split-decomposition method tropical diseases microbiology vertebrates animals mammals dogs viruses multiple alignment calculation rabies rna viruses neglected tropical diseases molecular biology techniques research and analysis methods rabies virus sequence analysis infectious diseases zoonoses artificial gene amplification and extension bioinformatics sequence alignment medical microbiology microbial pathogens molecular biology nucleotide sequencing lyssavirus computational techniques polymerase chain reaction viral pathogens database and informatics methods biology and life sciences viral diseases amniotes organisms

2017

A Pan-Lyssavirus Taqman Real-Time RT-PCR Assay for the Detection of Highly Variable Rabies virus and Other Lyssaviruses

The origin of syphilis is still controversial. Different research avenues explore its fascinating history. Here we employed a new integrative approach, where paleopathology and molecular analyses are combined. As an exercise to test the validity of this approach we examined different hypotheses on the origin of syphilis and other human diseases caused by treponemes (treponematoses). Initially, we constructed a worldwide map containing all accessible reports on palaeopathological evidences of treponematoses before Columbus' s return to Europe. Then, we selected the oldest ones to calibrate the time of the most recent common ancestor of Treponema pallidum subsp. pallidum, T. pallidum subsp. endemicum and T. pallidum subsp. pertenue in phylogenetic analyses with 21 genetic regions of different T. pallidum strains previously reported. Finally, we estimated the treponemes' evolutionary rate to test three scenarios: A) if treponematoses accompanied human evolution since Homo erectus; B) if venereal syphilis arose very recently from less virulent strains caught in the New World about 500 years ago, and C) if it emerged in the Americas between 16,500 and 5,000 years ago. Two of the resulting evolutionary rates were unlikely and do not explain the existent osseous evidence. Thus, treponematoses, as we know them today, did not emerge with H. erectus, nor did venereal syphilis appear only five centuries ago. However, considering 16,500 years before present (yBP) as the time of the first colonization of the Americas, and approximately 5,000 yBP as the oldest probable evidence of venereal syphilis in the world, we could not entirely reject hypothesis C. We confirm that syphilis seems to have emerged in this time span, since the resulting evolutionary rate is compatible with those observed in other bacteria. In contrast, if the claims of precolumbian venereal syphilis outside the Americas are taken into account, the place of origin remains unsolved. Finally, the endeavor of joining paleopathology and phylogenetics proved to be a fruitful and promising approach for the study of infectious diseases. The widespread availability of effective antimicrobial therapy, in association with population screening, resulted in a considerable decline in the frequency of syphilis in the middle of the last century [1]. The mass campaigns around 1960 treated 46 out of 152 million people screened and, as a result, endemic treponematoses (bejel, pinta and yaws) were eradicated in many regions of the world. However, reservoirs persisted and expanded into poor communities with deficient hygiene and health care [2]. Recently, outbreaks of syphilis occurred in different subpopulations, also in association with AIDS. One decade old worldwide estimates report that there are 12 million new cases of syphilis per year [3], whereas the total number of people affected with the non-venereal treponemal diseases reached 2,5 million [4]. Due to venereal syphilis, 157,000 deaths, 460,000 abortions or stillbirths, 270,000 low-birth-weight babies and 270,000 cases of congenital syphilis were registered solely in Africa in 2002 [5]. The complete genome sequence of Treponema pallidum, the advancement of PCR technology, and the use of penicillin, were thought to make possible the eradication of venereal syphilis [6]. Nevertheless, syphilis and the endemic treponematoses are still a heavy burden today, but WHO' s new goal to eradicate yaws by 2012 will hopefully change this situation [7]. This also means that the spread, virulence as well as the origin of syphilis needs to be further investigated. Eros and Thanatos are linked not only in mythology, but also in the study of sexually transmissible diseases. The achievements of the sexual revolution (Eros) have been undermined by the fear of AIDS (Thanatos). After initially blaming homosexuals and Africans, research on the origin of AIDS is more objective today, facilitating treatment and quality of life. Similarly, during the last centuries, this phenomenon transformed the life threatening and shameful syphilis in an easily treatable disease. Artistic representations of syphilis clearly reflect this transition [8]. Shortly after the XVIth century explosion of a sexually transmitted and disfiguring disease in Europe, Venus is shown as a source of contamination. Thereafter, women continue to be blamed for syphilis, in a satirical as well as stigmatizing way. More recently, not only artists, but also physicians and other researcher were inspired by the origin of syphilis. And for a long time not only women but also “savage” people discovered in the New World were blamed for syphilis. In contrast to the huge efforts to treat AIDS, syphilis and the other treponematoses are neglected diseases today, with a still mysterious origin. Treponemal diseases are caused by bacteria from the genus Treponema. Currently it is believed that different Treponema cause different diseases: T. carateum is responsible for pinta, T. pallidum subsp. endemicum leads to bejel (non-venereal syphilis or endemic syphilis), T. pallidum subsp. pertenue causes yaws, whereas T. pallidum subsp. pallidum is responsible for venereal and congenital syphilis (as reviewed in [9], [10]). These diseases show overlapping clinical manifestations and tend to occur in distinct geographical settings, but today they all affect mainly people from poor communities with deficient health services and poor hygiene conditions. Syphilis is the only venereal treponemal disease, whereas pinta, bejel, and yaws are transmitted through skin and/or oral contact. A brief description of each of the four treponematoses follows (as reviewed in [9]–[11]). Pinta is the only treponemal disease which does not affect skeletal tissues. Therefore it will not be further mentioned, since our purpose here is to use osteological data to estimate, along with a phylogenetic analysis, the origin of treponematoses. Yaws occurs in hot regions with high humidity. In the late stages it is very destructive on bone (Figure 1), skin and mucous tissues. The most distinctive alteration in bone is the “saber shin”, where the tibia develops an abnormal buildup of bone particularly on the anterior and medial surface of the tibia. Bejel occurs in dry, hot and temperate climates. Inflammations and destructive changes of tissues, including bone, occur only in the advanced stages of the disease. The clinical manifestations are intermediary between yaws and venereal syphilis. Venereal syphilis is the most destructive of the treponemal diseases and today shows a worldwide distribution. Syphilis can result in painful osseous alterations, leading to “caries sicca” of the skull and also to “saber shin” tibiae. This disease is the only treponematosis that is able to affect inner organs, causing nervous system damage and cardiovascular complications. Venereal syphilis, much more than other treponematoses, can get through the placenta, causing congenital syphilis. Short and narrow central incisors (Hutchinson' s incisors), small and dome shaped first molars (Moon molars), periostitis on long bones, “saber shin” tibiae and saddle shaped noses are the most important signs of congenital syphilis in the juvenile skeleton. One of the most popular hypotheses on the origin of syphilis states that it came from the Americas and was spread to Europe by Columbus' seamen [12], [13]. This is the Columbian hypothesis, which gained support through old ethnographic reports on treatment of syphilis with native plants in the New World [9]. Other hypotheses have also been proposed to explain the evolutionary history of the four treponemal diseases. The Pre-Columbian hypothesis argues that syphilis and the other treponemal diseases were present throughout the New as well as the Old World in pre-Columbian times, but was misdiagnosed as leprosy in Europe [14]. Hackett, a convinced pre-Columbianist, proposed a scheme of mutational development linking the four treponemal diseases in 1963 [14]. Accordingly, a) pinta arose somewhere in Africa or Asia from an animal infection about 17,000 years before present (yBP) and spread to the rest of the world; b) mutations on the pinta causing microbe lead to yaws about 12,000 yBP, and spread to the world except to the Americas; c) bejel arose from yaws about 9,000 yBP in arid climates; and finally d) mutations on the bejel causing treponeme originated venereal syphilis about 5,000 yBP in south-west Asia and then spread to Europe and the rest of the world. Finally, as discussed by some paleopathologists, the Unitarian hypothesis, states that treponematoses always had a worldwide distribution, where every social group had the kind of treponematosis appropriate to its geographic and climatic conditions and its stage of cultural development. Thus, according to this hypothesis, yaws, bejel, pinta and venereal treponematosis (or syphilis) are seen as adaptive responses of Treponema pallidum to peculiarities of environment, culture, and contact with other populations [14]. New data, instead of clarifying matters, contribute to an even more intricate scenario. For instance, osseous evidences of syphilis on pre-Columbian individuals from Europe [15]–[17], point against a New World origin of syphilis, while molecular data contradict the Unitarian hypothesis [18]. Consequently, although the issue has been discussed for five centuries, the origin of syphilis is not yet clear. Why? Do we have enough data to answer this question? Are there novel approaches to gain a better understanding of the when, how and where syphilis began troubling humans? Recently, the study of the origin of syphilis has gained notoriety, when Harper et al. (2008) employed molecular genetics to analyze 21 genetic regions of Treponema stemming from geographically distinct areas [19]. According to the authors, “T. pallidum arose in the Old World, in the form of non-venereal infection, before spreading with humans to the Middle East/Eastern Europe, in the form of endemic syphilis, and then to the Americas, in the form of New World yaws”. Then, “a T. pallidum strain from the Americas was introduced back into the Old World, probably as a result of the European exploration of the Americas, becoming the progenitor of modern syphilis-causing strains” [19]. However, this work resulted in controversy, particularly due to the restricted number of DNA polymorphisms (SNPs) analyzed in two samples collected in a single location [20]. The approach used herein differs from that employed by Harper et al. (2008) [19]. Albeit it is based on the same genetic sequences, here we combine a Bayesian inference approach with time calibration points obtained from treponemal diseases in ancient human bones, resulting in confidence intervals for the likeliness of the origin of this human-pathogen interaction. Paleopathology studies origin and distribution of diseases based on historical accounts, iconography, ethnography, and ancient human remains. The most reliable source of information on health and disease of past populations are the lesions certain diseases, such as treponematoses, leave on ancient human skeletons [21]. Until recently, the great majority of studies on treponematoses were carried out on North American osteological collections [11]. However, the last years have seen a growing contribution of research on material stemming from other regions of the world [11], [16], [22]–[26]. Thus, exploring paleopathology to aid unraveling the origin of syphilis is mandatory. A totally different approach refers to molecular biology and phylogenetics. The advances and joint study of molecular biology and phylogenetics allow the development of different analytical approaches which permit potentially more secure inferences based on genetic data. Recently this approach was used to a) unravel the degree of admixture of Neanderthal and modern man [27], b) to molecularly confirm cases of tuberculosis in ancient human remains [28], and c) to pinpoint treponemal diseases in mummies [29]. As more and more genomes of disease causing microorganisms are sequenced and published, more strength the molecular approach gains in expanding our knowledge on the history of the interaction between us and other species. Therefore, studying the origin of infectious disease using phylogenetics is indeed promising. The integration of paleopathology and molecular biology harbors not only the advantages of each of these approaches alone, but enhances analytical quality, since it allows studying phylogenetics with the knowledge of the when and where certain bony lesions appeared first in the history of our ancestors. Syphilis and the other treponematoses represent natural candidates to test this approach. Relying on the construction of a worldwide map containing all accessible reports on palaeopathological evidences of treponematoses before Columbus' return to Europe, we first explored the osteological data alone. Then we used some of the oldest dates of treponemal infection found in different regions of the globe to calibrate the time of the most recent common ancestor (TMRCA) of Treponema pallidum subsp. pallidum, T. pallidum subsp. endemicum and T. pallidum subsp. pertenue, using previously published sequences [19]. More specifically, we explored three different scenarios where this integrative approach is investigated (Figure 2). First a) we tested if T. pallidum emerged before modern humans evolved, using treponemal evidence in Homo erectus [30]. Therefore we set a minimum bound of 1. 6 million yBP on the root of the T. pallidum group. Furthermore, b) we explored the possibility that T. pallidum subsp. pallidum emerged from less virulent strains in Europe after Columbus' conquest, as recently suggested [19], setting the TMRCA of T. pallidum subsp. pallidum to 450–550 yBP. Finally, c) we analyzed if T. pallidum subsp. pallidum emerged in the time range between 16,500 yBP and 5,000 yBP. These dates refer to the entrance of the first humans into the New World [31], and to the most ancient probable osseous evidence of syphilis [32], respectively. The sources used to create the worldwide treponematoses map include primary publications (and the references within) on osteological evidences of treponematoses prior to the European conquest of the New World. These sources included: a) online published reports, b) traditional journals of free access through the University of São Paulo, c) books from the library of the Laboratório de Antropologia Biológica IBUSP, and d) reprints and pdf files requested from authors. The reported diagnoses of treponemal diseases were included in the table and map as published in the peer reviewed articles (Figure 3, Table S1). Wherever cited in the original reports, a specific diagnosis is presented in our map, such as (venereal) syphilis, congenital syphilis, bejel and yaws. However, in many cases, the authors did not distinguish between the different treponematoses and thus this broader term was used. All sources were divided into three main periods according to the intervals reported in the book The Myth of Syphilis [11]. In case a diagnosis or date of a certain treponematosis case was inconclusive, this was remarked (see legend of Figure 3, and Table S1). We used the nearest location possible to plot the sites on Google Maps. Special attention was given not to duplicate points reported more than once. Statistical analyses were carried out to test temporal and geographic distribution of pre-Columbian cases of venereal syphilis, bejel and yaws. Cases diagnosed solely as treponematoses, and those reported in the original article with either uncertain age and/or diagnosis were excluded from the statistical analyses. The tests employed were Chi square and Fisher exact test, considering p<0,05 as statistically significant. Only significant p values are cited in the text. Sequences from 20 Treponema pallidum isolates (9 T. pallidum subsp. pallidum, 9 T. pallidum subsp. pertenue and 2 T. pallidum subsp. endemicum) were obtained from previously published data [19], consisting of 21 genes and/or intergenic regions. GenBank accession numbers for all sequences used in this study are listed in Table S2. Some regions were not included in the analysis, since they could be involved in recombination events and/or present an elevated number of non-synonymous substitutions [19]. Sequence alignments were manually constructed and the final data set consisted of 5,412 base pairs (bp) of concatenated sequences. Since interspecific variation could lead to considerable overestimation of the divergence times for intraspecific data [33], no outgroup was included in the analyses. Thus, the approach used in this part of the study aims at explaining the existing genetic variation (and comparing the resulting rates of evolution with other bacteria) rather than inferring the order of evolutionary events. Since T. pallidum is essentially a clonal population and there is no evidence of recombination in the database used in the present study, it is possible to estimate divergence times using a coalescent approach with appropriated calibration points. To avoid using a molecular clock rate estimated for bacteria not related to T. pallidum, such as E. coli [34], we used the palaeopathological record to calibrate the molecular clock. From the 128 osteological cases employed for the paleopathological analysis, we used six for the phylogenetic analyses. They refer to the oldest treponematosis case registered in the genus Homo and to the reports on the oldest cases of probable bejel, venereal syphilis and congenital syphilis in different parts of the World (see below). These records provided minimum and/or maximum age estimates to constrain nodes in the trees, as suggested by Achtman (2008) [35] and shown for our database in Figure 2. Therefore, the different hypotheses on the origin of T. pallidum subspecies were tested using the Bayesian phylogenetic software BEAST V1. 4. 7. Given that evolutionary rates can vary between different organisms, we employed a coalescent model with constant population size under a relaxed molecular clock, which accounts for rate variation among tree branches [36]. Posterior distributions of parameters were obtained by two independent Markov Chain Monte Carlo (MCMC) analyses (chain length of 10 million, sampling every 1000 chains). The chains were compared to ensure convergence and merged to obtain final results. For each hypothesis tested, we estimated the divergence times and substitution rates of the whole group of treponemes and of T. pallidum subsp. pallidum. To warrant a more conservative approach, we assumed either the lower (most recent/slower) or the upper (most ancient/faster) high posterior density (HPD) value, depending on the context. Finally, we also tested hypotheses on the origins of T. pallidum subsp. pallidum using published data from whole genome comparison of the two T. pallidum subsp. pallidum strains available: T. pallidum subsp. pallidum S. Nichols [37] and T. pallidum subsp. pallidum SS14 [38]. Comparative genome analysis of these two strains performed by Matejkova et al. (2008) revealed a total of 327 SNPs within the 1. 1 mega base genome [38]. Since genetic divergence (K) between two sequences is calculated as the number of substitutions per base pair, the K between these two strains is 2. 96×10−4 substitutions per base pair (327/1. 1×106 base pairs). Assuming that the evolutionary rate does not vary over time, the genetic divergence can be used to calculate the time since their most recent common ancestor (TMRCA) using the following equation: When the divergence time is known, the equation above can also be used to estimate the substitution rate per site per year. Therefore, we used two different divergence times to estimate rates of evolution in T. pallidum subsp. pallidum: (i) 500 yBP, as proposed by Harper et al. (2008) and others as described above [19], and (ii) the most ancient osseous evidences of syphilis, as for example the case from Colombia dated to circa 5,000 yBP [39] (Figure 3 and Table S1). At least three transitions of the human-microbe interaction occurred after about 10,000 yBP [40]. The first, called the Neolithic revolution, led to domestication of plants and animals, and to the passage from nomadic to sedentary life in the Middle East. Many microbes originating from husbanded animals would have had contact with humans and failed to prosper. Some of them, however, would have survived causing infections such as influenza, smallpox, cholera, and tuberculosis. The second transition refers to empowered Eurasian civilizations that came in economic and military contact 1,500 to 3,000 yBP, swapping their germ pools with often disastrous results [42]. These included the Justian Plague of AD 542 that devastated Constantinople and massive epidemics in China at about the same time. The third transition was consequence of the European imperialism from the XVth century onwards. It led to the trans-oceanic spread of a big repertoire of infections brought by the conquerors [42]. This did not only happen in the Americas but also as a consequence of the European exploration of Asia and Australia and of the slave-trade from Africa to various parts of the world. This favored the spread of numerous parasitic diseases with specific strains according to geographic area. The main question asked herein is where, in this intricate sequence of events, should the origin of syphilis be placed. As a first attempt to answer this question, we plotted all available pre-Columbian evidences of bony treponematoses on a world map. This meant we accepted the risk of different methods and criteria for the establishment of differential diagnosis reported in the previously published studies. This also meant that we acknowledge the possibility that the following interpretations on the osteological analysis alone might be flawed. Our literature review produced a total of 128 pre-Columbian treponematoses cases worldwide (Figure 3, Table 1, and Table S1). Excluding cases with uncertain diagnoses or age and those with “probable” diagnoses (and thus considering only original reports with clear diagnoses and dating), there are almost twice as many treponematoses cases in the New (N = 56), than in the Old World (N = 33). This is not surprising if considering that systematic studies regarding these pathologies have been more frequently undertaken in North America than elsewhere. If excluding from these cases all those with the broad diagnosis of treponematosis, there are significantly more cases of bejel in the Old (Old World: 8/16, New World: 7/42; p = 0. 0172) and significantly more cases of yaws in the New World (Old World: 1/16; New World: 18/42; p = 0. 0108), while venereal syphilis is evenly distributed (Old World: 7/16; New World: 17/42; p = 0. 8210), as seen in Figure 4A. This probably reflects climatic conditions that favor the manifestation of bejel in the more arid Old World and yaws in the often hot and humid areas of the New World. However, this could also be due to a bias caused by a higher number of publications by scientists (e. g. Rothschild) whose strong belief is that yaws was the only treponematosis present in the New World at the time. On the other hand, this also means that the reports of syphilis in Europe and Russia long before the XVIth century contradict the assumption that syphilis was the “New World' s revenge upon Columbus and his crew” [43]. Thus, considering only osseous evidences of treponematosis, it appears that venereal syphilis afflicted humans thousands of years ago in the New as well as in the Old World (Figures 3 and 4a, Table 1). The number of established differential diagnoses (or clear cases of venereal, syphilis, bejel, yaws and treponematosis in the peer reviewed articles) increases substantially from the more ancient (6000–1000 BC, N = 15) to the two more recent periods (1000 BC–AD 1000, N = 29; AD 1000-contact, N = 45), as seen in Table 1. This is due to the higher number and the better state of conservation of more recent skeletal collections compared to those exhumed from more ancient sites. The progressive decrease of yaws (6/13; 8/21,5/24) and bejel cases (5/13; 7/21; 3/24), although suggested in Figure 4B, is not statistically significant. However, there is a huge and significant rise especially in venereal syphilis cases from 1000 BC to the time of Contact (6/21 to 16/24; p = 0. 0169), as already noted for North America in the book “The Myth of Syphilis” [11]. While this may be attributed to an increase of the venereal form of treponematosis, especially in the New World (5/16 to 11/18; p = 0. 1005), the Old World syphilis cases cannot be ignored. So, considering solely these palaeopathological reports, the Columbian hypothesis seems not to explain the origin of syphilis. On the other hand, it also seems clear that treponemal diseases in general were well established in pre-Columbian times all over the world. The oldest dates we are aware of refer to bejel in Sudan 15,000 yBP, yaws in Florida some 7,900 yBP, and treponematosis in Peru more than 8,000 yBP (Figure 3, Table 1, Table S1). Nevertheless, the microevolutionary sequence proposed by Hackett in 1963, although a good working hypothesis, does also not account for all osseous evidences of the treponematoses evaluated [14]. According to Hackett (1963) [14], bejel is seen as typical of arid environments, but there are evidences of bejel in the area of former Yugoslavia [44] and one in today' s Canada [45]. Additionally, Hackett proposed that bejel arose from yaws about 9,000 yBP, but the 15,000 years old bejel case from Sudan [46] pushes this transition back some six thousand years. These data suggest that the scheme proposed by Hackett does not match the evidences anymore. Finally, the Unitarian hypothesis predicts that each social group had the treponemal disease compatible with the prevailing environmental and cultural conditions. However, no clear climatic, ecological or cultural pattern emerges when analyzing Figure 3, Table 1 and Table S1. It is important to remark that there is controversy regarding the criteria for the establishment of differential diagnosis of treponematoses in osteological collections [11], [47], [48]. So far, there are two main schools of paleopathological interpretation of skeletal lesions representative of treponemal diseases: one very specific [49] and another extremely broad [50]. Thus, the conclusions based solely on paleopathology reports have to be taken very cautiously. This also means that there is an urge to standardize diagnostic criteria, if the history of treponematoses is to be unraveled. Until this “gold standard” is established, tested and accepted (vis-à-vis the fragmentary nature of the archaeological record and the osteological paradox [51]), a reassessment of all claims of treponemal diseases registered until the present would be unfruitful. Furthermore, despite the enormous efforts, there still remains controversy on the differential diagnosis of treponematoses also in living humans, due to partially overlapping symptoms and modes of transmission - that seem to be defined by opportunity rather than biology [20]. Another important issue refers to the changing virulence of syphilis over time. First mentioned as evil pocks in the Edict of the Holy Roman Emperor Maximilian, the devastating epidemics in the late XVth and early XVIth century were assumed by some scientists and historians to have been venereal syphilis. Although it seems clear that it was caused by a sexually transmitted disease, the evidence that it was syphilis is less certain [11]. In any case, it rapidly turned to be a very aggressive disease, killing within a few years [14]. Today, however, untreated patients with syphilis can survive for decades. This strongly suggests that this disease, also called the great imitator or the great impostor, changes with time. This would mean either the bacteria lost virulence during time or that syphilis killed off the most susceptible population right in the beginning of its devastating existence. As discussed above analyzing pre-Columbian cases of treponematoses based on osseous remains renders no clear picture on the origin and spread of these infectious diseases. This suggests the necessity of employing other methodologies. Here we test if using osseous evidences to calibrate phylogenetic analyses of the treponemes yields plausible results. In order to explore the efficacy of this new integrative method we investigate the three scenarios discussed below. Using 1. 6 million years as the date of the first evidence of treponematosis in the genus Homo, as that reported for H. erectus [30], we found a mean rate of 6. 35×10−10 substitutions/site/year for the T. pallidum group (Table 2). This is about ten times slower in comparison to other bacteria (E. coli: 4. 5–5. 0×10−9, Buchnera 8. 2×10−9 subs/site/year [34]), and even slower than the rate observed in humans (Eukaryotes and humans: 2. 5 to 0. 4×10−9 [52]). Since in bacteria generation time is much smaller and effective population size is many times bigger than in humans, and as genetic diversity increases proportionally with generation time and effective population size [34], this slow evolutionary rate seems very unlikely for the T. pallidum group. Furthermore, this analysis shows a temporal overlap of the most recent common ancestor of T. pallidum subsp. pertenue and T. pallidum subsp. pallidum from 175,000 to 87,600 yBP (Table 2). This suggests that both these treponemes existed since at least about 88,000 years ago, meaning that yaws as well as venereal syphilis should have accompanied successful human migrations since that time. In case the lesions found in the H. erectus fossil really can be attributed to treponematosis and if our phylogenetic analysis is correct, we expect to encounter evidences of treponemal diseases, including a true syphilitic Eve, in bones as ancient as 88,000 years ago, not only in our own species, but also in our cousin species, such as in the Neanderthals. A fact that turns this possible is that related treponemes have been found to affect other than the human species, including African monkeys [53] and rabbits [54]. Thus, it is feasible that the treponeme species affecting humans was originally a zoonosis [55], as already suggested by Hackett in 1963. Despite the observation of treponematosis in H. erectus 1. 6 million years ago [30], no other sign of treponemal diseases was found until now, older than about 15,000 years ago (Table S1). On the other hand, there are some evidences of pre-Columbian venereal syphilis outside the New World, i. e. in Europe, Russia, and Japan (Figures 3,4A, Table 1 and Table S1). In France, this evidence is strong since there is an infant with signs of congenital syphilis buried during the Late Roman Empire, amongst other cases such as the fetus from Costebelle [10], [15], [56] (Table S1). Other pre-Columbian Old-World cases of congenital syphilis were found in the ancient Greek colony Metaponto (dated to 580–250 BC) in today' s Italy [57] and in XIIIth century Turkey [23]. That is, all continents, except Africa and Oceania show pre-Columbian evidences of either venereal or congenital syphilis. That Africa shows no cases as yet of ancient syphilis is easily explained by the scarce paleopathology studies undertaken there until now. However, in Oceania many such studies have been carried out, so this argument is not valid. Although there is only one case of treponematosis plotted for Australia on Figure 3, we are aware of more such cases on this continent prior to the arrival of the Europeans in the XVIIIth century. This includes not only endemic forms found in the arid parts of this continent, but also some cases that could match a diagnosis of venereal syphilis [58]. Since the exact number and location of skeletons with treponematosis buried prior to the arrival of the Europeans was unavailable, Australia is underrepresented in our analyses. Considering the unlikely evolutionary rate found using the evidence of treponematosis in H. erectus to calibrate phylogenetic analysis and the absence of osseous treponemal disease older than 15,000 BP anywhere in the world, it is far from probable that the treponematoses, as we know them today arose more than a million years ago. The treponematosis diagnosed in H. erectus [30] can thus have been caused not by the treponeme strains analyzed herein, but by ancestral forms still unknown. As a next step, we tested the hypothesis that T. pallidum subsp. pallidum emerged in Europe from less virulent strains caught in the New World by Columbus' crew about 450–550 yBP [14], [19], [59]. Assuming this hypothesis, the most recent common ancestor of the whole group of T. pallidum, was found to be dated to 3,900 yBP (Table 3). This would mean either that the treponematoses originated in the New World about 4,000 years ago, or that they emerged at this time in the Old World and were then brought to the New World. In the first case, treponematoses would have spread from the New World to the rest of the globe mainly during the Great Voyages. However, this interpretation would not explain the existence of treponematoses dated to before 4,000 yBP in the Old World, such as the cases from Sudan, Egypt, India and Russia (Figure 3, Table S1). In the case treponematoses emerged in the Old World by 4,000 yBP and were brought by more recent migrational waves to the New World, this would explain the pre-Columbian treponematoses finds in Europe and Asia (Figures 3 and 4, Table S1), but does not allow treponematoses in New World bones older than 4,000, as for example the case of yaws registered in bones as old as 7,900 yBP excavated in Florida [39] or the treponematosis from a circa 5,000 yBP riverine shell mound in Brazil [22]. Additionally, the results of this analysis imply that T. pallidum evolved exceptionally fast, with a substitution rate of 1. 41×10−6 substitutions/site/year (95% HPD of 2. 96×10−7 to 2. 98×10−6). If assuming the lower HPD as true, T. pallidum evolved at a rate at least 100 times faster than E. coli (4. 5–5. 0×10−9 subs/site/year [34]), but similar to that found in Helicobacter pylori (6. 2×10−7 to 9. 2×10−7 [60]). Since this last organism lacks a DNA repair system [61] one would expect that T. pallidum, which has a repair system [37], would show a rate slower than that of H. pylori. Thus, a treponeme substitution rate in the order of 10−6 to 10−7 would be unlikely. Moreover, despite limited, comparative studies with T. pallidum sequences have shown a small amount of variability in the treponeme genome, suggesting a high conservation degree between the various subspecies [62], [63]. A high conservation degree has also been observed among sequences of Tpr genes of T. pallidum subsp. pallidum strains collected at different times (in the early 1900s and in the final years of the 20th century) [18]. Since the Tpr are the most variable genes of the T. pallidum genome, this result corroborates the idea that the substitution rate in T. pallidum is possibly not high. Consequently, this points to a slow substitution rate in the treponemes as a whole. Therefore, the scenario that venereal syphilis emerged as recently as 500 years ago seems utterly unlikely, be it in the New or the Old World. Using the root constrains of 16,500 yBP, when the first humans are believed to have entered the Americas [31], and 5,000 yBP (the oldest osseous evidence of venereal syphilis in the New World [32]), we found that the upper 95% HPD of the most recent common ancestor for the treponeme group is 77,400 yBP (Table 4). This result does not allow yaws in H. erectus, but could explain the claims of pre-Columbian treponematoses in H. sapiens across the world. Therefore, according to the phylogenetic approach used herein, our data do not contradict that syphilis might have emerged in the New World [19]. However, based on our paleopathology survey, we cannot exclude that venereal syphilis emerged in the Old World. Additionally, the estimated evolutionary rate was found to be 8. 82×10−8 substitutions/site/year (Table 4). This rate is plausible if compared to other bacterial rates. But this result has to be taken carefully, since the evolutionary rate varies between organisms, as consequence of differences in life traits [34]. On the other hand, there is evidence that variation in generation time does not affect evolutionary rates in some bacterial lineages [64]. Consequently, our estimations may be reliable. Therefore, to independently test this evolutionary rate, we used the oldest and most strongly supported osseous evidences (bejel in Sudan 15,000 yBP [46] and congenital syphilis in France 1,600 yBP [10], [15], [56], [65]) to constrain the TMRCA of the whole group of T. pallidum and the TMRCA of T. pallidum subsp. pallidum. The resulting rate was 4. 09×10−8 (varying from 6. 82×10−8 to 2. 62×10−9 substitutions/site/year – data not shown). Despite the difference, these rates partly overlap, suggesting that the range estimated testing hypothesis c) is likely. In addition to the hypotheses tested and discussed above, we also compared the rate of substitutions between the two whole genome strains of T. pallidum subsp. pallidum reported by Matejkova et al. (2008) [38]. If considering 5,300 yBP at Indian Knoll in the United States [59] and 5,000 yBP in Colombia [32] as the oldest and “safest” osseous evidences of venereal syphilis in the World (Figure 3 and Table S1), the resulting rate is 2. 1×10−8 substitutions/site/year. When, however, considering the oldest osseous evidence of congenital syphilis as being that reported for Costebelle in France circa 1,600 yBP [65] and for Metaponto in Italy about 2,400 yBP [57], these values correspond to 9. 25×10−8 and 6. 16×10−8 substitutions/site/year respectively (data not shown). Finally, if considering that venereal syphilis emerged only some 500 years ago from the less virulent strains brought to Europe by Columbus, we found 2. 92×10−7 substitutions/site/year. Whereas the first three rates vary in the same range as that reported for E. coli and Buchnera [34], the rate of 2. 92×10−7 is far too fast. This, again, argues against a very recent origin of venereal syphilis, in contrast to Harper et al. (2008) and others [19]. Our study harbors the back draws that limit every paleopathological as well as phylogenetic study. More specifically, the limits of this work are as follows: First, although the worldwide map on pre-Columbian treponemal bones contains as much data as we could collect, totalizing 128 cases, many reports on ancient treponematoses could not be included in our study. Second, since the diagnostic criteria differ between the reports on treponematoses in ancient bones, one cannot exclude that reassessing all possible cases with standardized methods would lead to different results. Third, since there is little sequence variation within and between the Treponema strains available, our highest probability density (HPD) values are large, reflecting the uncertainty inherent in almost all molecular date estimates. Forth, the absence of a suitable molecular clock rate for the T. pallidum group also limits the possibilities of estimating more accurate dates for the emergence of specific clades of treponemes. Therefore, we can not exclude that the evolutionary rate of T. pallidum lies outside the range observed for other bacteria. Last but not least, we could not advance knowledge of the place of origin of syphilis. Thus, our results have to be taken cautiously. Despite these limitations, we believe that the major strength of this work is the cross-fertilization of two so distinct areas as paleopathology and molecular phylogenetics. This novel approach enabled us to study phylogenetics of the treponemes with the advantage of knowing where the characteristic bony lesions appeared first in the history of our ancestors, thus significantly enhancing analytical quality. Despite the decades of enormous research efforts to describe and classify the different treponemes and treponematoses, both in active infections and in archaeological collections, some important hallmarks have to be reached to unravel the origin and spread of venereal syphilis. These include the establishment of diagnostic criteria to clearly distinguish between the three treponemal diseases that affect the skeleton, as well as selecting candidate cases in understudied osteological collections. After the establishment of this “gold standard” a reassessment of the well contextualized osteological collections worldwide has to be undertaken using a population approach. On the other hand, more phylogenetic studies of the whole treponeme group, including the saprophytic, zoonotic, and commensal species, as also those affecting primates and other mammals have the potential to aid in this effort. The more frequent use of new and complementary techniques, such as immunohistochemical studies of mummies [29] can also significantly enhance our comprehension on the origin and spread of this challenging disease. The use of evidences from both, paleopathology and molecular phylogenetics, provided a broad and integrative approach to the study on the origin of treponematoses. If our analysis using treponematosis in H. erectus is correct, we expect to encounter very ancient evidences of treponemal diseases, including a true syphilitic Eve. Since the oldest evidence reaches only 15,000 yBP and given that the evolutionary rate found is far to low compared to other bacteria and also to humans, treponematosis, as we know it today, seems very unlikely in a 1. 6 million years old H. erectus. On the other hand, when constraining the TMRCA of T. pallidum subsp. pallidum to Columbus' return to Europe some 500 years ago, the estimated evolutionary rate was considerably higher than expected for bacteria, and does not explain the majority of the osseous record. Thus, both these hypotheses seem improbable. Alternatively, the estimated evolutionary rate reaches values comparable to other bacteria, when we fix the TMRCA to between 16,500 and 5,000 yBP. Consequently, venereal syphilis seems to have emerged around this time. This would allow the claim of venereal syphilis in pre-Columbian times throughout the world. Most importantly, rather than trying to pinpoint the emergence time of venereal syphilis, we used treponematoses to test our transdisciplinary approach aimed at reconstructing the evolutionary history of different human infectious diseases. As described above, we independently explored three scenarios. The retrieved results corroborate the currently most accepted views on the origin of syphilis. This opens the possibility to use this approach to shed light on other diseases that afflict humankind since immemorial times.

Syphilis is a reemerging disease burden. Although it has been studied for five centuries, its origin and spread is still controversial. Did it accompany the evolution of the genus Homo and does it date back to more than a million years or did it emerge only after Columbus' s return to Europe? Initially, to test the validity of a new interdisciplinary approach we constructed a worldwide map showing precolumbian human skeletons with lesions of syphilis and other related diseases (also caused by different treponemes). Then, we selected the oldest cases to estimate the timing of the treponemes' history, using their DNA sequences and computer simulations. This resulted in treponeme evolutionary rates, and in temporal intervals during which these microorganisms could have emerged. Based on comparisons with other bacteria, we concluded that treponematoses did not emerge before our own species originated and that syphilis did not start affecting mankind only from 1492 onwards. Instead, it seems to have emerged in the time span between 16,500 and 5,000 years ago. Where syphilis emerged, however, remains unsolved. Finally, the endeavor of joining as distinct fields as paleopathology and molecular biology proved to be fruitful and promising to advance our understanding of the rise and fall of the infectious diseases that have afflicted humans across time and space.

Abstract Introduction Materials and Methods Results/Discussion

evolutionary biology/human evolution infectious diseases/neglected tropical diseases microbiology/microbial evolution and genomics non-clinical medicine/history of medicine infectious diseases/sexually transmitted diseases molecular biology

2010

Syphilis at the Crossroad of Phylogenetics and Paleopathology

Babesiosis is considered an emerging disease because its incidence has significantly increased in the last 30 years, providing evidence of the expanding range of this rare but potentially life-threatening zoonotic disease. Babesia divergens is a causative agent of babesiosis in humans and cattle in Europe. The recently sequenced genome of B. divergens revealed over 3,741 protein coding-genes and the 10. 7-Mb high-quality draft become the first reference tool to study the genome structure of B. divergens. Now, by exploiting this sequence data and using new computational tools and assembly strategies, we have significantly improved the quality of the B. divergens genome. The new assembly shows better continuity and has a higher correspondence to B. bovis chromosomes. Moreover, we present a differential expression analysis using RNA sequencing of the two different stages of the asexual lifecycle of B. divergens: the free merozoite capable of invading erythrocytes and the intraerythrocytic parasite stage that remains within the erythrocyte until egress. Comparison of mRNA levels of both stages identified 1,441 differentially expressed genes. From these, around half were upregulated and the other half downregulated in the intraerythrocytic stage. Orthogonal validation by real-time quantitative reverse transcription PCR confirmed the differential expression. A moderately increased expression level of genes, putatively involved in the invasion and egress processes, were revealed in the intraerythrocytic stage compared with the free merozoite. On the basis of these results and in the absence of molecular models of invasion and egress for B. divergens, we have proposed the identified genes as putative molecular players in the invasion and egress processes. Our results contribute to an understanding of key parasitic strategies and pathogenesis and could be a valuable genomic resource to exploit for the design of diagnostic methods, drugs and vaccines to improve the control of babesiosis. Babesiosis is a worldwide emerging infectious disease [1] caused by a protozoan parasite of the genus Babesia which is naturally transmitted by ixodid ticks and infects vertebrate erythrocytes. Parasite infection of natural hosts such as in cattle causes bovine babesiosis, resulting in high economic losses. Recently, the severity of human infection has also become apparent [2]. The four identified Babesia species confirmed to infect humans to date are B. microti, B. divergens, B. duncani, and B. venatorum [2,3]. B. microti has a worldwide distribution while B. divergens is a causative agent of babesiosis in humans and cattle in Europe [3]. Symptoms range from asymptomatic to flu-like symptoms in healthy individuals. However, in asplenic, immune-compromised patients, children, elderly people and patient who are experiencing B. divergens infection, the disease can be severe. Thus, human babesiosis results in hemolytic anemia, jaundice and hemoglobinuria due to the lytic effects of parasite multiplication in the red blood cells (RBC). B. divergens infections are therefore considered medical emergencies and patients require immediate treatment [3]. As evident from its status as an emerging disease, the number of human babesiosis cases has experienced a global increase [1]. This number may still not fully reflect the actual numbers of infections as many cases go unreported or are misdiagnosed because the non-specific symptoms of babesiosis can hinder accurate diagnosis. European cases in particular may reflect this situation because babesiosis is not a notifiable disease in the continent and several cases likely go unreported. This is supported by seroprevalence studies that indicate that the percentage of the European population infected with Babesia is much higher than the low clinical incidence and case reports to date [4]. Notably, these anti-babesial antibody studies also indicate that the detected antibodies were acquired by people who were exposed to the Babesia parasites [4]. In fact, antibody testing is important in identifying asymptomatic and mild cases when parasites may not be detectable in a blood smear or by PCR, suggesting the need for active surveillance programs to accurately detect babesiosis in Europe. Originally endemic in just a few regions on the world, including Europe [3,4] and EEUU [2] the steady increase of cases in other areas of the world (Asia, Africa, Australia, Canada, Egypt and South America) indicates its zoonotic potential. Notably, the presence of B. divergens outside Europe and North Africa was recently confirmed in China and Japan [5,6]. Onset of symptoms in the host results from the continuous invasion and destruction of RBCs by the parasite. The invasion of RBCs by the B. divergens free merozoite is an active process [7] that exploits an original locomotion mechanism called gliding motility, an intimate tight connection between the parasite and the erythrocyte known as a moving junction and a coordinated secretion of proteins from apical organelles (rhoptries, micronemes, dense granules or spherical bodies according to the species) [8–10]. Once hidden within the host cell, the intraerythrocytic lifestyle contributes to protect parasites from the interactions of anti-babesial antibodies capable of recognized and block the antigens of the free merozoite surface. Parasites gain nutrients and multiply by binary fission resulting in a considerably complex pleomorphic process that yields different intraerythrocytic parasite stages leading to the development of merozoites that exit to invade new RBCs and thus propagate [7,11]. In vitro, this proliferative cycle [12] has been studied to investigate biological processes such as invasion, multiplication and egress [7,12–14]. However, the dynamics and regulation of these processes are still largely unexplored at the molecular level in B. divergens. To gain new insight into this emerging but neglected pathogen, the genome of B. divergens human isolate (Bd 87 strain) was recently sequenced and assembled by our group using reads produced by three next-generation sequencing (NGS) platforms. The draft genome (publicly available under accession number CCSG01000001 to CCSG01000514) was approximately 10. 7-Mb in 514 scaffolds and revealed over 3,741 protein coding-genes [15]. In this study, we significantly improve the quality of assembly of the B. divergens genome by exploiting our previous sequence data using new computational tools and assembly strategies. We also report transcriptional data from two different stages of the B. divergens in vitro asexual lifecycle: i) free extra-cellular merozoite and ii) intraerythrocytic stage. Using genomic and transcriptomic data we provide an annotation-directed improved genome sequence assembly [16] which enable us to explore the synteny between B. divergens and other Babesia species. Moreover, this functional and comparative genomics approach alongside the gene expression profile of the two stages, highlight molecular features associated with key processes to the survival of the parasite, including invasion, gliding motility, moving junction formation and egress. Previously, we reported a draft genome [15] for B. divergens, assembled using a hybrid approach with three different platforms and sequencing depths: Illumina (~310x); 454 (~20x); Pacific Biosciences (PacBio) (~20x). Reassembly of the genome delivered larger contigs and scaffolds with chromosome lengths. As mentioned, the assembly continuity has been improved here, as shown in Table 1, where the number of contigs and scaffolds has been reduced and their average length has increased. The latest version of the genome can be found under the accession number GCA001077455 (CCSG02000001-CCSG02000141). The new assembly showed a better continuity that can be inferred from the reduced number of scaffolds and contigs and their higher average length. The total bases in the assembly was reduced since several redundant contigs that were generated in the previous assembly, were accurately resolved, although the contig N50 value decreased which is attributable to a more uniform length distribution and probably because some contigs were broken in misassembled regions. In the previous assembly, we generated 5,000 gene models, but in this work, the gene prediction was improved by using RNA sequencing (RNAseq) data from both free merozoites and intraerythrocytic parasite stages (details are explained below) and the corroboration of predicted genes by comparing their products against proteins from 17 different Babesia species. From a total of 40 million paired-end reads, 90% of these reads were successfully mapped to the genome, which allowed the validation of the 4,546 gene models. Additionally, genes were corroborated by ensuring that they had a protein ortholog with at least other Babesia species (S1 Table). Interestingly, from all the 17 Babesia species analyzed in the protein orthologous comparison, B. microti and B. bovis had the most orthologous in common with B. divergens (3,251 and 3,738, respectively). Taking in account this orthology information, we could support the annotation of 3,386 proteins in B. divergens which we considered as high confidence annotation. We also analyzed the clustering of orthologous proteins between the three species and found 2,660 groups where 1,427 were the core genes in the three species (S1 Table). According to our analysis, B. bovis presented 18 clusters with gene duplications while B. microti and B. divergens presented only 7 clusters with duplicated genes. As a result of the ortholog comparison from the gene model prediction improvement, we decided to compare our latest assembly against the genomes of B. bovis and B. microti to explore the synteny between organisms. In Fig 1, the one-to-one orthologous connecting the chromosomes from B. divergens to B. bovis and B. microti are depicted, allowing us to observe that some of our scaffolds have similar chromosome resolution reported for the other species [17,18]. Interestingly, the arrangements in B. divergens are more similar to the organization of B. bovis than B. microti, especially between chromosomes 2 and 3 of the former species. Contigs 1 and 2 (~2. 2 and 1. 4 Mb) of our improved assembly have a high correspondence to B. bovis chromosomes 3 and 2, respectively. Contig 3 (1. 12 Mb) has correspondence to the largest contig of chromosome 4, where we can observe a possible inversion and some regions rearranged in Contigs 4 and 6 (1. 07 and 0. 87 Mb). Notably, we observed some correspondence to chromosomes 1 and 3 from B. microti, with Contigs 3,4, 5 (1. 06 Mb), 6 and 7 (0. 63 Mb), suggesting significant rearrangements between the two species. Chromosome 2 from B. microti presented some correspondence to Contig 1, but not as much as observed with B. bovis chromosome 3. It can be inferred that B. microti has more genomic rearrangements in comparison to B. bovis and B. divergens, which presented a higher number of syntenic regions. In fact, previous phylogenetic investigations confirmed that B. microti constitutes a piroplasmid lineage distinct from B. bovis and B. divegens [19]. It is noteworthy that the three genomes were sequenced and assembled using different sequencing technologies and assembly algorithms, which could hinder the synteny comparison. Since none of them have a finished genome quality, there will be limitations due to missing parts in each genome. Additionally, we performed a pulse field gel electrophoresis (PFGE) for the genomic DNA and the results showed 4 probable chromosomes with the following approximate sizes of 1. 05 (chromosome 1), ~1. 7–1. 8 (chromosome 2), ~2. 9–3. 0 (chromosome 3) and 4 Mb (chromosome 4) (S1 Fig). The size and number of chromosomes are also similar to those reported for the B. bovis genome [20] where chromosomes 1,2, 3 and 4 are estimated to be 1. 4,2. 0,2. 8 and 3. 2 million base pairs, respectively. These results suggest that the probable ordering for our assembly, where we could assign contigs/scaffolds into chromosomes, is very similar to that observed in B. bovis. It is important to mention that although the B. bovis genome is the best Babesia reference genome available, only chromosomes 2 and 3 are finished and the rest are reported in several contigs [17]. In these cases, the use of genetic markers or FISH techniques [21] are necessary to corroborate the correspondence of the assembly into chromosomes and have achieve better synteny resolution. Free merozoites and intraerythrocytic parasites (Fig 2) represent the result of the cyclic processes of invasion and egress that B. divergens undergoes during its asynchronous asexual life cycle. Hence, we have sequenced the transcriptomes of both free extracellular parasites and intraerythrocytic stages using Illumina sequencing (S2 and S3 Tables). To profile both transcriptomes, we isolated mRNA from two biological replicates to generate two independent RNAseq libraries for each stage. We compared the RNAseq expression profile of both free merozoite and intraerythrocytic parasite stages using three different algorithms: DEseq, EdgeR and NOIseq. The NOIseq analysis yielded 1,441 differentially expressed genes, 734 of them were up-regulated and 707 genes down-regulated in the intraerythrocytic parasite stage (S4–S6 Tables). Based on these data and considering that molecular studies in term of host cell invasion and egress are limited in Babesia [22], we restricted our results to 45 genes of interest putatively related to both processes. These genes were manually inspected and curated using Artemis. Of note, 33 of the genes were confirmed by at least one of the three different algorithms and their expression levels were validated for free merozoite and intraerythrocytic stages using real-time quantitative reverse transcription PCR (qRT-PCR) (Table 2). The real-time qRT-PCR analysis revealed that 26 genes putatively related to the components of the molecular invasion machinery and 3 genes related to the egress process (Table 2) [22–25] were upregulated between 0. 2 and 2-fold in the intraerythrocytic stage. This result showed a moderately increased expression level of these genes in the intraerythrocytic stage compared with the free merozoite stage (Table 2 and Fig 3). However, 4 genes encoding proteases putatively involved in invasion or egress: i) BDIVROU_0278701. t1. 2 (rhomboid-like protease 4. 1, ROM4. 1), ii) BDIVROU_0230400. t1. 2 (rhomboid-like protease 4. 4, ROM4. 4), iii) BDIVROU_0280000. t1. 2 (subtilisin-like serine protease, BdSUB1) and iv) BDIVROU_0161200. t1. 1 (mac/perforin protein 2, MAC2) [26–29] were modestly induced between 0. 2 and 0. 5-fold in the free merozoite (Table 2 and Fig 3). RNASeq and real-time qRT-PCR findings indicated that genes of specialized B. divergens functions could be present in both the free merozoite and intraerythrocytic stages, suggesting that the majority of the genome is utilized throughout the complete asexual lifecycle of this parasite. Notably, the upregulation of some genes in the intraerythrocytic stage could imply a moderately increased need for invasion, gliding motility, and egress protein expression, before these processes occur (Table 2). Alternatively, a moderate or less marked induction of the same genes in the free merozoite would be required to maintain baseline expression level during the processes. Moreover, this expression profile suggests that the intraerythrocytic parasite could harbor proteins in its apical organelles that are required during the rapid processes of invasion and egress, thus ensuring that the complex and specialized mechanisms of invasion, gliding, and egress are primed and function successfully at the appropriate time. Surprisingly, the free merozoites also exhibit qualitatively a similar expression profile, although the transcripts are at lower levels. This may be linked to the short lifecycle of the parasite and may reflect that these transcripts need to be present to accommodate the requirement for the various proteins as soon as the parasite invades. It will be of interest to check if the protein levels are also equivalent between these two parasite stages. On the other hand, since the asexual life cycle of B. divergens is asynchronous, there are simultaneously intraerythrocytic parasites at different stages of their development, ranging from parasites which have just invaded new erythrocytes to parasites which are dividing or egressing from the host cell. As a result, there are not apparent abrupt changes in the expression profile between intracellular and extracellular stages. This is reflected in the data dispersion observed for free merozoites replicates in the principal component analysis (PCA) and multidimensional scaling (MDS) plots (S2 Fig.) generated by DESeq2 and EdgeR, respectively. However, for those proteins transcribed in two different phases of the lifecycle, i. e. when the parasite is outside the RBC and when it is inside the RBC, regardless their expression level, the research could lead towards the development of vaccines and new treatments capable of targeting several different stages at the same time, as previously suggested for P. falciparum [30]. Finally, on the basis of this genomic approach and in the absence of models of invasion and egress available for B. divergens, we have proposed the following potential molecular models according to the B. divergens genes identified and validated in this work (S7 and S8 Tables) and following the general prototypes suggested for the apicomplexan parasites Plasmodium and Toxoplasma [22–25]. a) B. divergens genes associated with calcium (Ca2+) in invasion Since Ca2+-dependent proteins and Ca2+ signaling are important for microneme exocytosis and invasion in apicomplexan parasites [22], we identified the following B. divergens candidates genes that encode proteins related to Ca2+: i) BDIVROU_0339410. t1. 2 (protein kinase G, PKG), ii) BDIVROU_0124800. t1. 2 (phosphoinositide phospholipase C, PI-PLC), iii) BDIVROU_0071500. t1. 1 (diacylglycerol kinase 1, DGK1), iv) BDIVROU_0291600. t1. 1 (acylated pleckstrin homology domain containing protein, APH), v) BDIVROU_0224310. t1. 2 (Ser/Thr phosphatase, PP1 or calcineurin), vi) BDIVROU_0130101. t1. 2 (calcium-dependent protein kinase 4, CDPK-4) and vii) BDIVROU_0416400. t1. 1 (double C2 domain protein, DOC2) (Table 2 and Fig 4). The identification of these genes suggests that Ca2+ and calcium–dependent proteins are involved in B. divergens invasion in a similar manner to other apicomplexans. Although PKG activates microneme secretion and parasite egress in a calcium-independent manner [25], this protein is associated with the release of Ca2+ from the endoplasmic reticulum (ER) to the cytoplasm and is mediated by cyclic guanosine monophosphate (cGMP). PI-PLC is then activated by cGMP-dependent PKG, thus leading to hydrolysis of phosphatidylinositol 4,5-bisphosphate (PIP2) to diacylglycerol (DAG) and inositol 1,4, 5-trisphosphate (IP3), each of which plays a different role [31,32]. DAG is converted to phosphatidic acid by DGK1. Phosphatidic acid is a mediator of microneme secretion and is associated with APH, a surface microneme protein that acts as a phosphatidic acid–detecting protein. Depletion of APH leads to blockade of micronemes and, consequently, to inactivation of the motile machinery of the parasite or glideosome during invasion [33]. Moreover, IP3 stimulates an influx of Ca2+ from the ER surface to the cytoplasm [31,34]. Simultaneously, this influx of Ca2+ activates calcium-dependent proteins as calcineurin, DOC2 and CDPK4 which interacts with PKG, thus favoring attachment of the merozoite to the RBC, the secretion of several micronemal proteins [35] and the control of the actomyosin motor to ensure an efficient gliding motility for the parasite [36]. b) B. divergens genes related to actin dynamics, glideosome and moving junction The actomyosin-based molecular motor or glideosome and the moving junction facilitate the entry of apicomplexan parasites into the host cell. These molecular systems are well characterized in Plasmodium and Toxoplasma [23,25] and we found the following six implicated genes from both within B. divergens genome: i) BDIVROU_0025700. t1. 2 (actin); ii) BDIVROU_0303400. t1. 1 (formin-a, FMRa); iii) BDIVROU_0036900. t1. 1 (formin-b, FMRb); iv) BDIVROU_0096510. t1. 2 (actin-depolymerizing factor, ADF or cofilin); v) BDIVROU_0279300. t1. 2 (profilin); and vi) BDIVROU_0073701. t1. 2 (F-actin–binding protein or coronin) (Table 2 and Fig 4). The identification of these genes suggests that the movement of B. divergens during invasion is mediated by actin in a similar manner to other apicomplexan parasites [37]. Actin is found as a globular monomer or G-actin and polymerizes into filaments of F-actin via a nucleation phase mediated by formins that localize at the apical pole of invading merozoites [38]. In contrast, ADF and profilin help to maintain a high actin content in the form of a monomer (G-actin) and a low concentration of polymerized actin (F-actin) by sequestering free G-actin [39–42] while coronin participates in the organization of F-actin filaments, the regulation of the motility [43] or mediating multiple parasites functions [44]. Actin dynamics and glideosome power the parasite locomotion [25], thus we identified six B. divergens genes related to the glideosome machinery: i) BDIVROU_0242901. t1. 2 (myosin heavy chain A, MYOA); ii) BDIVROU_0412101. t1. 2 (myosin light chain 1, MTIP for Plasmodium spp and MLC1 for T. gondii); iii) BDIVROU_0339200. t1. 2 (gliding associated protein 50, GAP50); iv) BDIVROU_0272901. t1. 2 (gliding associated protein 45, GAP45); v) BDIVROU_0052401. t1. 2 (gliding associated protein 40, GAP40); and vi) BDIVROU_0152200. t1. 2 (glideosome-associated connector, GAC) (Table 2 and Fig 4). Similar to other apicomplexan parasites, the glideosome of B. divergens may be located between the plasma membrane and the inner membrane complex (IMC) of the parasite and comprises a similar molecular structure formed by MYOA, which in association with MLC1 [45,46], GAP45, GAP50 and GAP40 generates the traction force for motility and entry [38,47]. As an alternative to the MYOA motor, we predicted a second gene BDIVROU_0133800. t3. 2 that encodes the myosin heavy chain B (MYOB), and it may play a similar role in parasite motility, thus suggesting that the movement in B. divergens could be governed by alternative myosin motors [48]. GAP proteins anchor the B. divergens glideosome to the IMC and the cytoskeleton [47,49,50], while the GAC connector attaches it to the plasma membrane [51]. This last protein specifically connects the glideosome to a transmembrane adhesin (TgMIC2 for T. gondii) that may interact with a host cell receptor to propel the parasite forwards [52]. In the B. divergens genome and transcriptomes, we found at least five adhesins as potential candidates to connect the parasite to the RBC: i) BDIVROU_0156200. t1. 2 that encodes the orthologous thrombospondin-related apical membrane protein (TRAP for Plasmodium and TgMIC2 for T. gondii, mentioned above) [24]; ii) BDIVROU_0126000. t1. 2 that codes for the apical membrane antigen BdAMA1, that binds to a trypsin- and chymotrypsin-sensitive receptor on the RBC [53]; iii) BDIVROU_0182900. t1. 2 that encoded the glycosyl phosphatidyl inositol (GPI) anchored protein Bd37 involved in the penetration of B. divergens using a potential endocytosis mechanism [54,55]; iv) BDIVROU_0183000. t1. 2 that codes for a putative 50-kDa surface protein (BdP50) and v) two identical copies (BDIVROU_0113100. t1. 2 and BDIVROU_0112800. t1. 1) of the genes that code the rhoptry-associated protein 1, RAP-1, a protein that binds to a RBC receptor [56]. In addition, we confirmed two more genes from the RAP family: i) BDIVROU_0112900. t1. 2 that encoded a protein of 492 aa (RAP-1b) with an estimated molecular mass of 56. 38 kDa, the amino acid sequence exhibited significant percentages of similarity to other known RAP proteins as B. bovis RAP-1 (83%) [57], B. divergens RAP-1 (72%) [56], B. ovata RAP-1 (79%) [58] and B. gibsoni RAP-1b (63%) [59] and ii) BDIVROU_0117600. t1. 2 encoding a ropthry protein [60] that showed 34. 4% similarity to the RAP-1 related antigen (RRA) of B. bovis, a protein that may contribute to erythrocyte invasion [61]. The amino acid sequence of RAP-1, RAP1-b and RRA from B. divergens exhibited the common features of the RAP family [61] such as a signal peptide, the presence of the strict conserved 4 cysteine residues and the 14 amino-acid motif. The genome sequence showed the B. divergens rap-1 locus extending over about 7. 3 Kb and contains two identical rap-1 genes interspaced with the rap-1b gene. Rap-1 and rra genes were located on the same contig but the rra gene was found around 96. 8 kb downstream from the RAP-1 locus. None of the genes were flanked by other rap-1 like genes or any gene related to the invasion or egress processes. Moreover, for successful cell invasion, apicomplexan parasites establish a moving junction that moves from the apical to the posterior end of the parasite [23]. We identified at least three B. divergens genes implicated in the formation of this system: i) BDIVROU_0259100. t1. (rhoptry neck protein 2, RON2 [56]); ii) BDIVROU_0191501. t1. 2 (rhoptry neck protein 4, RON 4); and iii) BDIVROU_0142300. t1. 2 (rhoptry neck protein 5, RON5) (Fig 4). Similar to Plasmodium and Toxoplasma, B. divergens RON proteins could constitute a ring-shaped structure in the host cell membrane. This structure together with AMA1 enables a tight connection between P. falciparum or T. gondii and their respective host cells leading to the formation of the moving junction [23,25,62–65]. However, as we mentioned above B. divergens AMA1 binds to a specific receptor on the RBC providing an attachment between the merozoite and the host cell [53], thus suggesting that the constitution of the moving junction core could be different between B. divergens and other species and could involve a host cell receptor. Alternative or multiple roles for B. divergens AMA1 and specific adaptation of this protein to a particular B. divergens merozoite-host cell interaction could be also considered [52]. Interestingly, we identified the gene BDIVROU_0165301. t1. 2 (Table 2) that belongs to the spherical body protein family (SBP) of the genus Babesia which seem to be involved in stabilizing the environment after invasion and in aiding parasite growth [66]. We also identified the gene BDIVROU_0161200. t1. 1 that encodes a subtilisin-like serine protease (BdSUB1) that play a relevant role during invasion [27] and five novel genes: i) BDIVROU_0278701. t1. 2 (ROM4. 1), ii) BDIVROU_0280300. t1. 1 (ROM4. 2), iii) BDIVROU_0439300. t1. 1 (ROM4. 3), iv) BDIVROU_0230400. t1. 2 (ROM4. 4) and v) BDIVROU_0449600. t1. 1 (ROM4. 5) that showed homology with rhomboid-like protease ROM4, suggesting the presence of a novel B. divergens ROM family. The homologous T. gondii ROM4 and the P. falciparum subtilisin-like serine protease PfSUB2 cleaves the transmembrane domain of adhesins (AMA1, and TRAP) to disengage their interactions with the host cell for parasite entry [26,67,68]. Thus, the ROM protease family detected in the B. divergens asexual life cycle could aid in the shedding of surface adhesins as well. In addition to shedding, P. falciparum AMA1 is phosphorylated by the cAMP-dependent protein kinase (cAMP-PKA) an enzyme that was recently shown to be involved in merozoite invasion [69]. This gene has also been identified for B. divergens in the present study (BDIVROU_0165500. t1. 1). c) B. divergens genes related to egress of the parasite Egress is an ordered, regulated process assisted by a cascade of proteolytic activities that occur seconds before the parasite exits the host cell [24,70]. For example, perforins, that exhibit calcium-dependent lytic activity, form pores in the parasitophorous vacuole membrane of Plasmodium and in the plasma membrane of the host cell facilitating the egress process [28,29,71]. Interestingly, four B. divergens genes with homology to mac/perforin proteins (MAC) were also found in this work, these are: i) BDIVROU_0064900. t1. 1 (MAC1), ii) BDIVROU_0161200. t1. 1 (MAC2), iii) BDIVROU_0065210. t1. 2 (MAC3), and iv) BDIVROU_0173601. t1. 2 (MAC4) (Table 2). As reported in the literature, cysteine proteases are also involved in the asexual life cycle. Specifically, papain proteases play a dual role in haemoglobin digestion and parasite egress [72–74]. We identified eight B. divergens genes encoding cysteine proteases. Three of them belong to the papain family: i) BDIVROU_0308600. t1. 1 (papain- 2), ii) BDIVROU_0202300. t2. 1 (vignain like protease) and iii) BDIVROU_0197200. t1. 1 (papain-like cysteine protease). Interestingly, we also have curated the remaining 5 genes: i) DIVROU_0190900. t1. 1 (calpain 7), ii), BDIVROU_0315500. t1. 1 (CPC1-like protease), iii) BDIVROU_0028600. t1. 1 (CPC2-like protease), iv) BDIVROU_0012300. t1. 1 (B. divergens OTU‐like cysteine protease, BdOTU) and v) BDIVROU_0273700. t1. 1 (B. divergens autophagy-related peptidase, BdATG4). These genes, have been previously described in P. falciparum, as essential for cell cycle progression [75] apicoplast homeostasis [76] and growth [77]. Therefore, the presence of these homologs suggests a similar role in B. divergens. Our study provides novel and comprehensive information on the identification of a large panel of proteins involved in the critical processes of invasion and egress of B. divergens. The strategy developed shows how genome sequencing and RNAseq can further our understanding of transcriptional patterns of B. divergens blood stages. Although apicomplexan genomes differ in terms of sequencing data and assembly strategy, orthologous proteins related to biological processes that have been reported for other parasites can be found in B. divergens. Thus, we have proposed multiple invasion and egress molecular candidates for B. divergens guided by the cross-comparison with the related Apicomplexan parasites, Plasmodium and Toxoplasma. Based on our results, we confirmed significant similarities with both parasites but also some key differences in the composition of the B. divergens invasion machinery (Fig 4). Therefore, future studies are needed for defining conserved versus evolutionary adaptations of the principal molecules involved in these biological processes between Apicomplexan species. Moreover, our transcriptomic analysis revealed the expression profile of the B. divergens genome and showed which genes are active at a particular point in time. However, the genetic expression landscape observed in B. divergens free merozoite and intracellular life stages are pieces of a complex jigsaw puzzle that needs to be solved by using other techniques such as proteomics and metabolomics. Furthermore, we will extend our analysis to the transcriptome of other clinical samples and/or isolates, especially those from hosts infected by B. divergens. Thus, we will be able to compare differences between the functional expression profile of the parasite in vivo during active infection and in vitro. Finally, we consider that the results generated in this work, can guide other efforts in the improvement of diagnosis and detection, drug target and vaccine development for the control of this emerging and neglected disease. Human A+ blood from healthy volunteer donors was used to maintain B. divergens blood stage cultures. Protocols for the use of this blood were approved by the Blood Transfusion Center, Madrid, Spain. Donors provided informed written consent for use of their blood for research purposes. B. divergens (Bd Rouen 1987 strain) was propagated in vitro in human A+ RBCs. High parasitemia asynchronous parasite cultures were prepared at 40% parasitemia according to previously published protocols [9,78]. PFGE was used to determine chromosome number and size of the B. divergens genome. Asynchronous cultures were centrifuged at 1224 x g for 5 min and washed with RPMI 1640 (Life Technologies Corporation, Carlsbad, CA). The resultant pellets, containing intact cells, were embedded in 0. 8% agarose plugs using the clamped homogeneous electric field (CHEF) mammalian genomic DNA plug Kit and treated with proteinase K according to the manufacturer’s instructions (Bio-Rad Labs Inc. , Hercules, CA). The CHEF-DR III system (Bio-Rad labs Inc.) was used to separate the intact chromosomes of B. divergens. The PFGE conditions used for a 0. 1–2. 0 Mb DNA size range were: 6 V cm-1, pulses of 60–120 s for 24 h and an angle of 120°. The conditions used for a 1. 8–4. 6 Mb DNA size range were: 4. 5 V cm-1, pulses of 200 s for 48 h and an angle of 106°. The gels were stained and photographed under ultraviolet transillumination. The de novo genome assembly was performed using a hybrid approach with three different sequencing technologies. The first de novo assembly was obtained using the Newbler software [79] with 454 sequencing data and default parameters. Then, a second de novo assembly with Illumina paired-end and mate paired sequencing data, was performed using AllPaths-LG assembler [80] using default parameters. Then, 454 and Illumina scaffolds were merged using the genome assembler, reconciliation and merging (GARM) meta-assembler [81] with default parameters to generate a consensus assembly. Subsequently, SSPACE [82] and iterative mapping and assembly for gap elimination (IMAGE) [83] were used to improve scaffolding and gap filling, respectively. Finally, we used PacBio sequencing data that was previously corrected with Iterative Correction of Reference Nucleotides (iCORN) [84] using five iterations with Illumina reads, to increase the continuity of the assembly using the software Patch [80] with default parameters. The latest version of the genome can be found with the accession number GCA001077455 (CCSG02000001-CCSG02000141) and associated to the BioProject PRJEB6536. For gene prediction, we used Braker v2. 0. 4 and AUGUSTUS v3. 0. 3 using a set of proteins that were obtained from the previous version of the assembly [15] and with the support of RNAseq data. We corroborated the predicted gene models by comparing them against proteins from other Babesia species deposited in the NCBI database and filtering them to 90% identity to remove redundancy. Using Proteinrtho v5. 15 [85], we retained those models with at least one ortholog in other species. The following list of Babesia species were used: B. bennetti, B. major, B. odocoilei, B. ovata, B. orientalis, B. capreoli, B. ovis, B. canis, B. rossi, B. rodhaini, B. equi, B. caballi, B. divergens, B. bigemina, B. gibsoni, B. microti and B. bovis. Finally, protein annotation was achieved using the Trinotate v3. 0. 1 pipeline [86]. For the comparative genomic analysis, the genomes available at Genbank for B. bovis (GCA_000165395. 1) and B. microti (GCA_000691945. 2) were used. The Proteinortho v5. 15 [85] was used to obtain protein clusters of Babesia genomes analyzed here with the following parameters: E-value for blast: 1e-05, min. percent identity of best blast alignments: 25, min. coverage of best blast alignments in percent: 50, min. similarity for additional hits: 0. 95. The results from the protein orthologous was also used to verify the annotation. Additionally, we used the Circos [87] package for circular data visualization and we were able to observe chromosome rearrangements among B. divergens, B. bovis and B. microti. Free merozoites and intraerythrocytic parasites were collected from two highly parasitized independent asynchronous B. divergens cultures, 75 ml each at parasitemias of ≈ 40%. Cultures were centrifuged at 600 x g and 4°C for 5 min in fixed-angle rotor heads to collect the supernatant containing the free merozoites and pellets containing the intraerythrocytic forms. Free merozoites were isolated [9] and resuspended in Trizol LS Reagent (ThermoFisher). Pellets containing the intraerythrocytic forms were resuspended in RPMI 1640 and washed 3 times by centrifugation at 600 x g for 10 min in order to remove possible traces of free merozoites. Pellets were diluted to 1: 3 with RPMI 1460. This suspension was then layered on top of a 1. 122 g/ml isosmotic Percoll solution prepared according to the manufacturer´s instructions (GE Healthcare, Freiburg, Germany) and centrifuge (2000 x g, 10 min). The infected erythrocytes collected at the interface were washed 3 times in PBS at 4°C and 600 x g for 10 min and resuspended in Trizol LS Reagent (ThermoFisher). Total RNA from B. divergens free merozoites and intraerythrocytic parasites was prepared using Trizol LS Reagent (ThermoFisher) and chloroform extraction. DNA was removed using the QIAGEN RNase-Free DNase Set (Qiagen Inc. Valencia CA) and cytoplasmic ribosomal RNA was removed using the Ribo-Zero removal kit (Illumina, San Diego, CA, USA). The RNA was measured on an Agilent 2100 Bioanalyzer using an RNA 6000 Pico Chip kit (Agilent Technologies, Inc, Santa Clara, California) and used for libraries preparation and real-time PCR assays (S2 Fig.). Libraries were prepared using the Illumina TruSeq RNA Sample Prep Kit v2 (Illumina) following the manufacturer’s protocol. High quality RNA samples from free merozoites (FM1 and FM2) and from intraerythrocytic stages (IE1 and IE2), with RIN numbers ranged from 8. 8 to 10, were used for NGS library construction (S2 Fig). FM1, FM2, IE1 and IE2 biological replicates were used to prepare two independent libraries for each stage. A third technical replicate was also prepared for each stage mixing FM1 and FM2 and IE1 and IE2 respectively. The libraries were sequenced using the Illumina HiSeq2000 platform with a paired-end configuration using 202 cycles (2x101 reads). RNAseq data was deposited at NCBI under the Biosample accession number SAMN12187113 and included in the BioProject PRJNA552284. Differential expression analysis was performed using two biological replicates through the integrated differential expression analysis multi-experiment (IDEAmex) website (http: //zazil. ibt. unam. mx/ideamex/) using three different differential expression packages: edgeR [88], DESeq2 [89] and NOISeq [90]. EdgeR and NOISeq were performed by applying trimmed mean of M values (TMM) [91] as the normalization method. To identify differentially expressed genes, genes whose p value was less or equal to 0. 01 and log fold change (FC) greater or equal than 1. 5, were selected for each method. The genes reported as differentially expressed in Table 2, were inspected and curated manually using Artemis [92] and selected as candidates for further analysis. Transcriptomic analyses were validated by real-time qRT-PCR using the same RNA extractions that had been prepared previously for RNAseq sequencing. Two μg each of free merozoite and intraerythrocytic parasite RNA was used for cDNA synthesis using the High Capacity RNA-to-cDNA kit (Applied Biosystems by ThermoFisher) following the manufacturer’s instructions. Controls without reverse transcriptase were used to investigate potential gDNA contamination by PCR using actin-specific primers. A set of 33 genes was tested by real-time qRT-PCR. Primers (S9 Table) were designed using Primer3 software [93] and were validated by testing amplification efficiency using 10-fold dilutions of cDNA by real-time qRT-PCR and melt curve analysis. Real-time qRT-PCR measurements were performed using the Qiagen Rotor-Gene-Q Real-Time PCR Detection System. Reactions were prepared in triplicated in a total volume of 20 μl using the SYBR Premix Ex Taq (Takara-Clontech, Otso, Japan), 20 ng of cDNA and primer concentrations of 0. 2 mM. Transcript expression levels were calculated with the 2−ΔCt method using the endogenous control gene BDIVROU_0242600. t1. 1 encoding for B. divergens endonuclease/exonuclease/phosphatase family domain containing protein, EnExPh as references.

Babesiosis has long been recognized as an economically important disease of cattle, but only in the last 40 years has Babesia been recognized as an important pathogen in humans. Babesiosis in humans is caused by one of several species (B. microti, B. divergens, B. duncani and B. venatorum). The complete Babesia lifecycle requires two hosts, the ixodid ticks and a vertebrate host. It is the parasite' s ability to first recognize and then invade host erythrocytes that is central to the pathogenesis of babesiosis. Once inside the cell, the parasite begins a cycle of maturation and growth, resulting in merozoites that egress from the red blood cells (RBCs) and seek new, uninfected RBCs to invade, perpetuating the infection. To better understand this asexual lifecycle, the authors focused on the parasite genome and transcriptome of the asexual erythrocytic forms of B. divergens. Through this functional and comparative genomic approach, the authors have identified genes putatively involved in invasion, gliding motility, moving junction formation and egress, providing new insights into the molecular mechanisms of these processes necessary for B. divergens to survive and propagate during its life cycle.

Abstract Introduction Results and discussion Methods

sequencing techniques parasite groups parasitic protozoans parasitology membrane proteins developmental biology apicomplexa molecular motors protozoans molecular biology techniques rna sequencing cellular structures and organelles motor proteins research and analysis methods proteins life cycles merozoites comparative genomics molecular biology cell membranes biochemistry eukaryota cell biology babesia genetics biology and life sciences genomics computational biology organisms parasitic life cycles

2019

Comparative and functional genomics of the protozoan parasite Babesia divergens highlighting the invasion and egress processes

Accumulation of filamentous actin (F-actin) at the immunological synapse (IS) is a prerequisite for the cytotoxic function of natural killer (NK) cells. Subsequent to reorganization of the actin network, lytic granules polarize to the IS where their contents are secreted directly toward a target cell, providing critical access to host defense. There has been limited investigation into the relationship between the actin network and degranulation. Thus, we have evaluated the actin network and secretion using microscopy techniques that provide unprecedented resolution and/or functional insight. We show that the actin network extends throughout the IS and that degranulation occurs in areas where there is actin, albeit in sub-micron relatively hypodense regions. Therefore we propose that granules reach the plasma membrane in clearances in the network that are appropriately sized to minimally accommodate a granule and allow it to interact with the filaments. Our data support a model whereby lytic granules and the actin network are intimately associated during the secretion process and broadly suggest a mechanism for the secretion of large organelles in the context of a cortical actin barrier. Natural killer (NK) cells are lymphocytes of the innate immune system that function in clearance of tumor and virally infected cells [1]. Elimination of susceptible target cells is tightly regulated and follows ligation of germline-encoded activation receptors [2]. As NK cells do not require receptor gene rearrangement, they are constitutively enabled for cytotoxicity. Thus, NK cell activation must be tightly regulated to ensure that healthy cells remain unharmed. Efficient lysis requires the tight adherent formation between the NK cell and the target cell termed the immunologic synapse (IS). The formation of a mature, cytolytic synapse between an NK cell and a target cell occurs in stages that can be thought of as checkpoints in the activation process [3]–[5]. Major cytoskeletal steps that are required in this process include the rearrangement of filamentous actin (F-actin) and the polarization of the microtubule organizing center (MTOC) [6]–. These events culminate in the directed secretion of lytic granule contents at the IS, which is prerequisite for NK cell cytotoxicity. F-actin accumulation at the synapse is the first major cytoskeletal reorganization event and is critical to subsequent steps and function of the IS [5]. Inhibiting proper F-actin dynamics in NK cells with the actin targeting drugs cytochalasin [6], [9], latrunculin [10], or jasplakinolide [3] inhibits their cytotoxicity. Furthermore, NK cells from patients with Wiskott-Aldrich Syndrome (WAS) who have mutations in the actin regulatory protein, WAS protein (WASp), are poorly cytotoxic [9]. This defect is attributable to improper reorganization of F-actin at the IS. Additionally, the actin nucleator Arp2/3 complex, which is enabled by WASp, is also required for cytotoxicity [10]. Cytochalasin treatment, Arp2/3 complex depletion, or WASp deficiency prevent the normal accumulation of F-actin at the synapse [5], [9], [10]. One question that arises from the creation of a dense polarized network at the IS is how secretion of lytic granules occurs through a potential barrier. The traditional view of granule delivery through the actin network holds that granules reach the synaptic membrane through a void of actin in the center of the network. This model is based on the observation from 3-D confocal microscopy that actin forms a dense peripheral ring around the IS [5], [11]. There is a caveat to the seemingly unobstructed access to the membrane that this “ring” provides: the actin motor protein, myosin IIA, is required for secretion and, more specifically, for granule delivery to the synaptic plasma membrane [12], [13]. These data are at apparent odds with one another as a requirement for myosin IIA for secretion necessitates a requirement for actin. One explanation is that granules are secreted at the periphery of the synapse where the traditional model depicts the location of F-actin. Another explanation is that the center of the synapse actually contains F-actin but does so at a level that has been undetectable by conventional 3-D confocal microscopy. Here we use microscopy techniques that provide enhanced sensitivity and resolution over those used previously to investigate the NK cell IS. We show that F-actin is present throughout the synapse and that lytic granules likely navigate and are secreted through the filamentous network by accessing minimally sufficiently sized clearances. These data demonstrate a previously unappreciated distribution of F-actin at the NK cell IS and redefine granule access to the synaptic membrane and functional secretion. Visualization of the synaptic actin network has relied on 3-D reconstructions of confocal slices [5], [11]. Here, we took advantage of the superior resolution of imaging in the XY plane to investigate the polarization and distribution of actin. First, we evaluated GFP-actin expressing NK-92 cells conjugating with the susceptible and adherent cell line, mel1190. These cells conjugated in a manner that afforded us the ability to image the synapse in the XY plane. GFP-actin was polarized toward the contact site (Figure 1A) and surprisingly displayed a diffuse distribution across the synapse (Figure 1B). This distribution was quantitatively analyzed and confirmed using a radial intensity profiling algorithm, which demonstrated that the intensity throughout the contact site was substantially above the background. To more directly image the cortical region of the NK cell immunologic synapse, we used total internal reflection fluorescence microscopy (TIRFm), which has the benefit of an improved signal to noise ratio over confocal microscopy and is limited to visualization within the first membrane proximal 100 nm [14]. Cells were activated via crosslinking of NKp30, a natural cytotoxicity receptor whose ligand is expressed on tumor cells [15], and CD18, a member of the heterodimeric integrin lymphocyte function-associated antigen-1 (LFA-1). Both integrin receptor and activation receptor activation are critical for polarized secretion of granule contents [16]. This combination of signals resulted in robust activation, which was demonstrated by degranulation as measured by enzymatic activity of granzyme A in the supernatant (Figure S1). TIRFm imaging of activated NK-92 cells demonstrated a distribution of F-actin throughout the synapse (Figure 1C). Quantitative analysis using radial profile plotting confirmed the presence of F-actin throughout the cell contacts (Figure 1E). To ensure that these findings were not particular to the NK-92 cell line, we activated and imaged freshly isolated ex vivo NK cells. Similar to NK-92 cells, synaptic F-actin in ex vivo NK cells was identified throughout the contact (Figure 1D, F). These results demonstrate that the NK cell synapse is defined by an abundant, diffuse F-actin network. To evaluate the kinetics of actin accumulation at the activated synapse, NK-92 cells expressing GFP-actin were imaged using TIRFm after contacting an activating surface. Actin accumulated quickly, within 5 min, and was sustained over the period of observation (50 min) (Figure S2A, Video S1). There was an initial paucity of actin at the synapse followed by a rapid filling in, as demonstrated by the separation of peak contact area and mean fluorescence intensity (MFI) of GFP-actin in that region (Figure S2B). The decrease in MFI over time was due to photobleaching as separate imaging of fields at 10 and 40 min did not show MFI differences (unpublished data). Importantly, actin was diffusely accumulated prior to timepoints at which granule contents were detected in the supernatant (Figure S1). Thus, actin was present as a potential barrier to lytic granule access to the plasma membrane. Because there was abundant actin present at the synapse, we wanted to determine if lytic granules might utilize relative clearances in the actin network to access the synaptic membrane. To address this, GFP-actin expressing cells were loaded with LysoTracker Red dye, which enables tracking of lytic granules and definition of their position relative to actin, and followed in real time after activation. Numerous granules were identified in the synaptic actin network using two-color TIRFm. Although some relative hypodensities were apparent in the synaptic actin network (Figure S3A–C), the LysoTracker labeled granules did not necessarily appear in these relative voids of actin (Figure 2A, Video S2). To quantitatively analyze this observation across all synaptic granules in an NK cell, the actin intensity in the region of the synaptic granule was compared to that of the entire synapse by dividing the MFI of the respective intensity values to produce a ratio measurement. This ratio, when compared to minimum and maximum potential ratios, demonstrated that on average granules approached the membrane in areas of actin (Figure S4A, B). Combining measurements of all granules in the synapse over 1 h from 14 cells defined the mean granule ratio value as 1. 0 (Figure S4C). Although there was a range of actin intensities present throughout the synapse as measured by the ratios of minimum and maximum intensity values to the MFI, few granules were present in areas of particularly low or high actin content. Thus, the colocalization of lytic granules with mean actin signal suggested that granules access the synapse in close proximity to the actin network. The MTOC is known to deliver lytic granules to the immunological synapse in NK cells [17]. To investigate the relationship among granules, the MTOC, and the synaptic actin network, we imaged the synapse using both confocal microscopy and TIRFm. The MTOC was present in the plane of the synapse and granules that were also in the plane of the synapse were present at a mean distance of 2. 05 µm from the MTOC (Figure 2B, C, E), a distance consistent with granules that are converged to the MTOC [17]. To note, the MTOC was not present in any distinct clearance of F-actin. TIRFm demonstrated more clearly than confocal microscopy the varying density of the synaptic actin network (Figure 2D). These data suggest that the MTOC delivers granules to the synaptic actin network. Because there was variability in colocalization between synaptic actin and granules (Figure S4), we considered the possibility that an approximated granule might not necessarily be capable of degranulation. Specifically, we reasoned that granules that ultimately degranulate represent a subpopulation of approximation events. Furthermore, we hypothesized that granules capable of degranulation might be those present within focal actin hypodensities. In order to study this directly, we developed a novel degranulation indicator for use in live cells. Lysosomal-associated membrane protein 1 (LAMP1, CD107a), which is sorted to lytic granules [18], is routinely used to detect cells that have degranulated by its appearance on the cell surface [19], [20]. Although previous investigations used antibody to LAMP1 to visualize degranulation [21], we adopted a cell-intrinsic approach by targeting a reporter fluorophore to the lytic granules. We fused pHluorin, a pH sensitive mutant of GFP that does not fluoresce at acidic pH [22], to the cytoplasmic tail of LAMP1 (Figure S5A, B) and obtained stable expression in NK-92 cells. As expected with localization of the pHluorin-LAMP1 construct to lytic granules, treatment with concanamycin A (which effectively neutralizes lysosomal pH by inhibiting the vacuolar-type H+ ATPase [23]) resulted in a robust increase in green fluorescence as measured by flow cytometry (Figure 3A). Since degranulation is an activation-induced process, we also treated pHluorin-LAMP1 expressing cells with the phorbol ester, PMA, and calcium ionophore, ionomycin, and found a rapid increase in pHluorin fluorescence, consistent with LAMP1 surface upregulation (Figure 3A). To better define pHluorin-LAMP1 localization to acidic granules, LysoTracker Red loaded pHluorin-LAMP1 expressing cells were studied using TIRFm after activation. Individual LysoTracker Red labeled granules could be identified at the synapse and were observed to undergo a shift from red to green fluorescence (Figure 3B, Video S3). This event is consistent with the granule fusing with the synaptic membrane, releasing its contents, and encountering a pH neutral environment. These data are consistent with lytic granule targeting of pHluorin. We next used pHluorin-LAMP1 expressing NK-92 cells to address whether granule approximation results in degranulation. LysoTracker Red loaded, pHluorin-LAMP1 expressing cells were imaged over time using TIRFm (Figure S6A, Video S4). There were significantly more approximation events than degranulation events (mean = 31 and 8 per cell, respectively) over 1 h (Figure S6B, C). Thus only a subset of granules that approximate the synaptic membrane result in a degranulation event. To directly investigate where degranulation occurs relative to the synaptic actin network, we stably coexpressed pHluorin-LAMP1 and mCherry-actin in cells and imaged them following activation using two-color TIRFm. Timelapse imaging demonstrated that degranulation events occurred in areas of at least some actin fluorescence, similar to that which was seen with granule approximations (Figure 3C, Video S5). We quantitatively evaluated the actin intensity at degranulation points by dividing the intensity values of the actin signal at the point of degranulation events by that of the entire cell contact (ratio of MFIs), thus generating a normalized, comparable value. Degranulations were identified in regions of actin that had slightly lower signal than the mean actin signal of the cell footprint (ratio = 0. 965) (Figures 3D, S7). To provide an additional measure, the ratio of MFIs for points adjacent to the degranulation and the entire cell contact were calculated. When compared to ratio of MFIs for the degranulation point itself, the ratio of MFIs for adjacent points were significantly higher. This indicated that degranulation events were occurring in areas of locally hypodense actin. We additionally calculated the minimum and maximum potential values of the ratio measurement. The degranulation values were between the minimum and maximum values (Figure S7). This indicated that degranulation events did not occur in areas of minimal nor maximal actin but rather that they occurred in close proximity to the actin network. To further characterize the local actin network at the point of degranulation in consideration of focal hypodense regions, we quantitatively evaluated actin fluorescence in the entire immediate vicinity of degranulation events. Measurements of actin fluorescence were made along sequential pixel radii emanating from the centroid of individual degranulations extending approximately 1 µm outwards (Figure 3E). The fluorescent intensities of actin as well as that of pHluorin were quantified in concentric circles along these radii (Figure 3F). In general, as the pHluorin signal diminished from a degranulation centroid, the mCherry-actin signal increased suggesting that degranulation occurs in a focal actin hypodensity (Figure 3F). To measure multiple events, the change in intensity between consecutive radiating circles was determined and plotted for all observed degranulation events (Figure 3G). The mean value of the intensity change demonstrated reduced actin intensity at each of the innermost four radii compared to the neighboring outer radius, thus reflecting the example in Figure 3F. These values were significantly different from the baseline value of zero, which would have indicated no change in the actin network. This indicates that in moving from the periphery of the region of the degranulation event to its center the actin intensity decreased to a detectable degree. Thus, degranulation tends to occur in locally hypodense areas of the actin network (i. e. , in regions with some but relatively less actin). In the presence of an actin barrier at the plasma membrane, therefore, hypodense regions of actin provide a potentially more accessible route to the synaptic membrane. Since granules are in contact with at least some actin during degranulation, we next investigated the role of actin dynamics in degranulation. We inhibited actin polymerization and dynamics with drugs that prevent F-actin assembly (latrunculin A and cytochalasin D) or disassembly (jasplakinolide). Inhibitor addition at the time of activation almost completely inhibited degranulation (Figure 4A), a result that is consistent with the previously reported requirement for initial actin reorganization in synapse formation and maturation [5]. In order to avoid inhibiting the initial, requisite actin reorganization (which we observed by 5 min—Figure S2), we treated cells with the inhibitors following 10 min of activation (a point at which minimal degranulation had occurred—Figure S1). Addition of the actin inhibitors after 10 min of activation resulted in an approximately 50% decrease in bulk degranulation (Figure 4A). Interestingly, addition of inhibitors at 20 min had only a marginal effect on degranulation. Thapsigargin, which has the net effect of elevating intracellular calcium levels, was used as a positive control and increased degranulation. To further evaluate the effects of the inhibitors on NK cells, we imaged F-actin at the synapse following inhibitor treatment. Jasplakinolide had no effect on F-actin presence or distribution; latrunculin A completely depleted the actin network; cytochalasin D had variable effects on cells, with some cells appearing unaffected while others showing a relative depletion of some filaments (Figure 4B, and unpublished data). To quantitatively evaluate these effects, cells were analyzed using radial intensity profile measurements. This analysis demonstrated no major depletion in F-actin following jasplakinolide or cytochalasin D treatment and robust depletion of the actin network following latrunculin A treatment (Figure S8A, B). Collectively, these results indicate that F-actin presence and reorganization immediately prior to the start of degranulation but after large-scale actin accumulation has occurred is critical for subsequent granule release. To determine if the synaptic actin network was dynamic at the time corresponding to degranulation, GFP-actin expressing cells were imaged by TIRFm and evaluated at both early and late timepoints of activation. Subtle but consistent changes in actin intensity were visualized over the course of imaging (Figure S9A, Video S6). Although most peaks and troughs of intensity did not appear to change over time, some shifts were detected as highlighted by surface plot rendering of intensity values (Figure S9B, Video S6). To quantify this, images from 5-min series at early and late timepoints of activation were evaluated for changing fluorescent intensities by plotting line profiles across the synapse. Temporally sequential line profiles were overlaid together on one graph, which demonstrated a constant trend in the fluorescence over time with imperfect alignment (Figure S9C, D). The variability was observed during both timeframes of activation and signified a dynamic state of actin where small changes in intensity were occurring. Variation was consistent at both 10 and 30 min after activation as defined by the standard deviation of mean intensity for each pixel across a linear profile (Figure S9E). We were able to inhibit actin dynamics to a measurable degree by using jasplakinolide to stabilize filaments and latrunculin A to prevent new filament formation (Figure S9F). Thus at the level of TIRFm, the synaptic actin network was dynamic at both early and late timepoints. This is consistent with the requirement for actin function at times after which synaptic actin accumulation has occurred. The resolution of fluorescence microscopy is diffraction limited to around 200 nm [24]. While we did detect focal hypodensities in the actin network corresponding with regions of degranulation, it was unclear if these represented true openings among the actin filaments. Thus we pursued super-resolution of the synaptic actin network to surpass the limit of diffraction. Stimulated emission depletion (STED) microscopes that utilize continuous wave (CW) fiber lasers can image with spatial resolution below 60 nm and thus provide the opportunity to investigate synaptic structures with superior resolution [25]. Using CW-STED microscopy, we first imaged Citrine-actin expressing NK-92 cells conjugated to mel1190 cells (Figure 5A). Actin was present throughout the contact and had a distribution that was similar to that seen with our confocal microscopy images (with cells activated on glass and with mel1190 cells) and TIRF images (activated on glass). This leads us to conclude that our method of activating and imaging NK cells on glass does not induce a distribution or architecture of the actin network that is distinct from that seen with actual target cells. We next asked whether there was an activation-dependent change in the synaptic actin architecture. To this end, we imaged F-actin in cells that had been stimulated with antibody to CD18, which does not induce degranulation, or in combination with antibody to NKp30, which robustly induces degranulation (Figure S1). Cells stimulated through CD18 alone had a dense synaptic network while the addition of antibody to NKp30 resulted in a more diffuse architecture with many observable clearances in the network (Figure 5B, C). To quantify this observation, clearances were identified based on measured granule diameters. Granules were separately identified using STED microscopy and had a range of diameters with a mean value of 333±103 nm (Figure S10A). The majority of granules were in the 250–499 nm range of diameters, with a smaller number in the 500–749 nm range. These two ranges, which represent areas that are minimally sufficient in size to allow granules access to the plasma membrane, as well as the larger 750+ nm range, were used to categorize clearance area. The number of these clearances in cells significantly increased when a degranulation signal (anti-NKp30) was incorporated (Figure 5D–F). To note, there were many clearances in the 250–499 nm range, fewer in the 500–749 nm range, and still fewer to none in the large 750+ nm range upon activation (Figure 5D–F). This trend in the frequency of clearances reflects the frequency of granule diameters. Furthermore, the mean clearance size divided by the mean granule size resulted in a value of 1. 48±0. 48, demonstrating that on average clearances were only slightly larger than granules (Figure S10B). Thus, full activation for degranulation results in an actin network with many access points that are minimally sufficient in size to accommodate granules. Having defined the presence of clearances in the synaptic actin network upon activation, we next sought to identify granule localization relative to the network. Thus we simultaneously imaged the actin network by STED microscopy and granules by laser scanning confocal microscopy. Granules within activated cells displayed a range of interaction with the actin network and all that were present at the synapse had at least some (Figure 5G). Colocalization of granules with F-actin was first measured as a percent of the granule area that contained an actin signal. There was 58±31% colocalization at the synapse, compared to 11±18% colocalization at a distance of 0. 5–1 µm from the synapse, indicating that the granules were extensively contacting the actin network after they were delivered to the synapse. To further quantitatively analyze these interactions, we measured line profiles of intensity values for F-actin and perforin across the granule (Figure 5H). Granules were localized within minimally sized clearances, in slightly larger clearances, or directly atop filaments. In all cases, there was an association of the granule with actin as defined by granule and actin line profiles intersecting and/or overlapping. Thus granules use both minimally sufficiently sized clearances and only slightly larger clearances to gain access to the plasma membrane, and they do so in direct interaction with the actin network. To obtain unprecedented, nanometer resolution of actin filaments at the synapse, we used platinum replica electron microscopy. In order to expose the inner surface of cells, corresponding to the synapse, for metal coating, cells were “unroofed” by mechanical removal of the bulk of the cell body with the nucleus. Images of platinum replicas of the activated synapse confirmed our earlier light microscopy data that the F-actin network exists throughout the synapse and contains small granule-sized clearances (Figure 6A, B). Filaments within the network, however, were present in varying densities. Consistent with the requirement of WASp for synaptic maturation and previous studies defining the localization of the Arp2/3 complex to the synapse [9], [10], branched arrays of filaments were detected (Figure S11), although the branching frequency appeared lower, as compared to a typical branched network in lamellipodia, and many long filaments were also present. In order to determine whether the abundance or distribution of clearances changes over time, the filamentous network at the synapse was evaluated after 10 and 30 min of activation. Quantitative assessment of the actin network in multiple cells defined a similar total contact area and total area occupied by filaments (filament density) between the two timepoints (Figure S12A, B). We also measured the individual groups of clearances in the actin network that were appropriately sized to allow lytic granule passage. We were able to detect clearances that would accommodate a granule at both timepoints (Figures 6C–E, S13). The number of clearances at each of the size thresholds was similar at the two activation timepoints tested. The distance of these clearances from the cell centroid, however, was different between the timepoints. After 30 min of activation the clearances were closer to the cell center than at 10 min (Figure S12C), suggesting that a fine-tuning of the actin network occurs and may be a necessary process in the maturation leading to degranulation. Thus ultra-resolution imaging of the synaptic actin network further defined the presence of minimally sufficiently sized clearances that would allow granule passage and demonstrated their changing position over time. Our TIRF data suggested, and our STED data demonstrated, that granules approximate the synapse in close association with the actin network. To confirm this observation on the nanometer level we used platinum replica electron microscopy to image “unroofed, ” membrane-intact NK cells. Granule-sized organelles could be identified on the intracellular face of the actin network in approximation with filaments as well as intercalated within filament clearances (Figure 7A, B). These data are consistent with a model of granule approximation and degranulation whereby granules transit to the membrane through an interaction with the F-actin network and utilize multiple minimally sufficiently sized clearances instead of a single large opening (Figure 7C). Actin accumulation defines an early stage in the maturation of the NK cell IS and is required for subsequent cytolytic function. Without proper reorganization of the actin cytoskeleton, lytic granules fail to polarize to the synapse and NK cells display inadequate cytotoxicity. Upon polarization, lytic granules require myosin IIA function to approximate the plasma membrane and have their contents directly secreted. The dependence on an actin motor protein for the secretory process suggests a requirement for the actin network itself. We have defined the distribution of actin at the IS using techniques that provide unprecedented sensitivity and resolution. We have done so both by live cells conjugated to target cells as well as a more flexible model system. We have also shown that lytic granules reach the plasma membrane and are secreted in areas of actin. Furthermore, our data suggest that secretion events likely occur in minimally yet sufficiently sized clearances in the actin network. While our studies relied upon activating an NK cell line using immobilized antibody we were able to obtain physiologically relevant supporting data. Firstly, NK-92 cells activated in this manner released contents of lytic granules (Figure S1). Secondly, the distribution of actin at the IS in NK-92 cells was similar on immobilized antibody and on flat, living target cells even when using enhanced resolution (Figures 1B, C, S2A, 5A, C, 5G, 6A, B). Thirdly, to show direct ex vivo supporting evidence, the F-actin distribution at the IS of freshly isolated human NK cells was consistent with that seen in NK-92 cells (Figure 1). Thus the actin network we imaged using our model activation method was unlikely to be an artifact. A benefit of our model approach, however, was the ability to utilize platinum replica electron microscopy to image the synapse with nanometer resolution, and then correlate these findings with our fluorescence microscopy data. The F-actin network in secretory cells was first believed to be a barrier to exocytosis [26]. This hypothesis was supported by data from numerous cell types including neuroendocrine cells (rev. in [27], [28]), neurons (rev. in [29]), platelets [30], and goblet cells [31] among others where loss of cortical F-actin correlated with an increase in secretion. In many of these cell types it has become increasingly appreciated that F-actin also serves as a required facilitator of secretion [27]–[29], [32]. In cells of the immune system, there is evidence for both models of secretion. Mast cells require some degree of F-actin present for agonist-induced secretion as robust depletion of F-actin with latrunculin interfered with normal release [33]. Conversely, F-actin is reported to be a barrier to secretion for neutrophil granule secretion [34]. Investigations with F-actin and granule secretion at the IS in cytotoxic lymphocytes have thus far been limited to T cells. In cytotoxic T lymphocytes, it has been suggested that actin is not a barrier to secretion [35], [36]. This includes the observation that actin is cleared from portions of the IS and granules are delivered to the synaptic membrane by the microtubule organizing center [35]. Because of 3-D confocal microscopy depicting F-actin rings at the NK cell IS [5], [11], the delivery of lytic granules is believed to occur in actin-devoid regions as in T cells. However, the reliance for granule secretion on myosin IIA in NK cells has called this model into question [12]. Specifically, that myosin IIA associates with lytic granules and its function is required for approximation to the synaptic membrane [13] strongly suggests that the mechanism for granule secretion is different in NK cells. Our data support a role for F-actin as a facilitator of secretion rather than a barrier. Inhibiting actin polymerization with cytochalasin D or latrunculin A after activation resulted in diminished secretion of lytic granule contents rather than an increase. This suggested that actin was not a barrier to secretion, but that its dynamics were required. The latter hypothesis is supported by the results showing a similar inhibitory effect when cells were treated with jasplakinolide. Interestingly, the effect of the actin inhibitors was most pronounced at 10 min following activation and less so at 20 min following activation, and thus defined a critical window of actin reorganization that facilitates degranulation (Figure 4). One explanation for this is that reorganization is required for the bulk of granules to approach the membrane and dock, which is accomplished by 20 min. Regulation of actin reorganization as a requirement for secretion has been proposed in other cell types. In muscle cells the Arp2/3 complex and cofilin function in the required actin reorganization for GLUT4 vesicle exocytosis [37]. In chromaffin cells, Cdc42, neural-WASp, and the Arp2/3 complex are proposed to function together at the plasma membrane to facilitate secretion through generation of new filaments [38]. Our work defines an actin network that is more pervasive at the NK cell IS than previously thought. Although this could serve as a potential barrier, we have identified abundant granule-sized clearances that could function as sufficient access points to the plasma membrane. These could provide functionality by allowing granules to pass between filaments and to simultaneously interact with them, whereby myosin IIA could exert force in squeezing granules between filaments or in post-fusion expulsion of granule contents. This latter possibility has been suggested in chromaffin cells where myosin II function was required for appropriate release of catecholamines [39]. Another consideration is that granules may use clearances smaller than their equatorial area by squeezing through adjacent filaments. Although we cannot rule out this possibility, it is nevertheless a potential mechanism that is consistent with our hypothesis. Thus we propose that degranulation events at the NK cell IS represent a coupled interplay between actin filaments and clearances that presents additional opportunities for regulatory steps important to NK cell cytotoxicity. NK-92 and GFP-actin expressing NK-92 cell lines were a kind gift from K. Campbell and were maintained in Myelocult (StemCell) media supplemented with 100 U/mL penicillin and streptomycin (Gibco) and 100 U/mL IL-2 (Hoffman-La Roche). mCherry-actin, Citrine-actin, and pHluorin-LAMP1 expressing cells were generated by retroviral transduction of NK-92 cells as described [40]. Briefly, 2–4 µg of plasmid DNA were transfected into the Phoenix packaging line using Fugene (Roche) lipofection reagent. Supernatant was harvested on day 2 post-transfection. NK-92 cells, Polybrene (Sigma), and supernatant were mixed and spun in a well of a 6 well plate at 1,000×g for 90 min at 32°C. Following overnight incubation at 32°C, cells were spun down and resuspended in supplemented Myeolocult. Cells were grown for 3 d prior to the introduction of puromycin (2 µg/mL) (InvivoGen) or hygromycin B (150 µg/mL) (Cellgro). mCherry-actin and Citrine-actin expressing cells were sorted for high expression by the University of Pennsylvania Cell Sorting Facility. Ex vivo NK cells were prepared from concentrated whole blood as described [41]. The pHluorin-LAMP1 retroviral plasmid was generated by BioMeans, Inc. by inserting the sequence for pHluorin (a kind gift from G. Miesenböck) between the signal sequence and the transmembrane domain of IL-2Rα linked to the cytoplasmic tail of LAMP1 (a kind gift from M. Marks). A flexible GS linker was added between pHluorin and the transmembrane domain sequences. The entire construct was subsequently cloned into the MIGR1-puromycin vector. The mCherry-actin retroviral plasmid was generated by PCR amplifying mCherry-actin from a pmCherry plasmid with 5′ BglII and 3′ EcoRI restriction site overhangs. The PCR product was digested and ligated into the pMSCV-Hygromycin plasmid (a kind gift from W. Pear), which had an EcoRI site in the Hygromycin resistance gene sequence eliminated by site-directed mutagenesis. The Citrine-actin retroviral plasmid was generated by amplifying the Citrine sequence from the pRSET-b Citrine plasmid (a kind gift from R. Tsien) with 5′ and 3′ BglII overhangs. The product was digested and ligated into a similarly digested mCherry-actin retroviral plasmid, effectively removing the mCherry sequence and inserting the Citrine sequence. Proper orientation of insert was verified by DNA sequencing by the Children' s Hospital of Philadelphia Research Institute sequencing core facility. Flow cytometry was performed to verify pHluorin-LAMP1 expression. Cells were untreated or treated with phorbol myristate acetate (PMA, 100 ng/mL, Sigma) and Ionomycin (1 µg/mL, Sigma) for 30 min or Concanamycin A (CMA, 100 nM, Sigma) for 90 min and samples were run on a BD FACSCalibur. Cells were washed and resuspended in supplemented Myelocult prior to use. For imaging of lytic granules, cells were incubated with 100 nM LysoTracker Red DND-99 (Molecular Probes) for 30 min at 37°C, washed once, and resuspended in supplemented Myelocult. ΔT dishes (Bioptechs) were coated with 5 µg/mL anti-NKp30 (Beckman-Coulter) and 5 µg/mL anti-CD18 (Clone IB4) for 1 h at 37°C, washed with PBS, and prewarmed prior to imaging with 1 mL dye free R10 (dye free RPMI 1640 (Gibco), 10% fetal bovine serum (Atlanta Biologicals), 10 mM HEPES (Gibco), 100 U/mL penicillin and streptomycin, 100 µM MEM nonessential amino acids (Gibco), 1 mM sodium pyruvate (CellGro), and 2 mM L-glutamine (Gibco). 4×105 cells were added to the dishes, which were maintained at 37°C with a heated stage and lid (Bioptechs). For live cell imaging of actin dynamics following inhibitor treatment, cells were activated as above for 10 min before addition of media containing DMSO or jasplakinolide (1 µM, Calbiochem). Following 5 min of incubation, media containing DMSO or latrunculin A (10 µM, Sigma) was added to the dish. After 5 min of further incubation, cells were imaged. For fixed cell experiments, 1×105 cells were adhered to No. 1 glass coverslips coated with antibody as described above. Samples were fixed and stained with Alexa Fluor 488 phalloidin or 568 phalloidin (Molecular Probes) as described [41]. For experiments with inhibitor treatments, cells were activated on coverslips for 10 min or 20 min, treated with inhibitors for 5 min, and then fixed and stained. Cytochalasin D (Sigma) was used at 10 µM. Samples were imaged through a 1. 49 NA, oil immersion, 60×, APO N TIRFm objective or a 1. 45 NA, oil immersion, 100×, PlanApo TIRFm objective (Olympus) when noted. 488 nm (Spectra-Physics) and 561 nm (Cobalt) diode lasers were launched through a two-line combiner of an LMM5 (Spectral Applied Research) into a rear mounted TIRF illuminator (Olympus) on an Olympus IX-81. Lasers were aligned for total internal reflection prior to each experiment. Images were captured using Volocity (PerkinElmer) to control a C9100 EM-CCD camera (Hamamatsu). Mel1190 cells were plated into ΔT dishes 1 d prior to use. Cells were stained with CellMask Deep Red (Invitrogen) according to manufacturer' s instructions just prior to imaging. GFP-actin expressing NK-92 cells were added to the dishes, which were maintained at 37°C, and imaged for up to 1 h. Cells were imaged through a 63×1. 4 NA Plan-APOCHROMAT objective (Zeiss) on a Zeiss Observer. Z1 using a C10600 ORCA-R2 camera (Hamamatsu). The microscope was equipped with a CSU10 spinning disk system (Yokogawa). 491 nm (Cobalt) and 655 nm (CrystaLaser) diode lasers were launched through an LMM5 (Spectral Applied Research). NK-92 cells were activated on antibody coated glass coverslips as described above for 30 min and then fixed and stained with rabbit anti-human Pericentrin (Abcam), Alexa Fluor 488-phalloidin, and anti-Perforin-Alexa Fluor 647 (Biolegend). The secondary antibody to anti-Pericentrin was a goat anti-mouse Pacific Blue (Molecular Probes). Cells were imaged in three dimensions on a spinning disk confocal Olympus DSU IX-81 microscope. Mel1190 cells were grown in a monolayer overnight on No. 1. 5 glass coverslips. Citrine-actin expressing NK-92 cells (106) were resuspended in media and incubated on Mel1190 targets for 30′ at 37°C. Cells were fixed with 2% paraformaldehyde and mounted with ProLong antifade reagent (Invitrogen). Cells were imaged at the plane of the interface between NK and target cells. Separately, for imaging of granules, Citrine-actin expressing NK-92 cells were immobilized to bound antibody as described above, then fixed, permeabilized, and stained with anti-perforin Alexa Fluor 488 (Biolegend). For visualization of actin and perforin in NK-92 cells, cells were immobilized to bound antibody as described above, then fixed, permeabilized, stained with Alexa Fluor 488 phalloidin and anti-perforin Alexa Fluor 647, and imaged at the plane of the glass or 0. 5–1 µm above it. All samples were mounted with Prolong anti-fade reagent (Invitrogen). Cells were imaged through a 100×1. 4 NA HCX APO objective on a Leica TCS STED CW system controlled by Leica AS AF software. Alexa Fluor 488 and Citrine were excited using a 488 nm Argon laser and STED depletion was achieved using a 592 nm continuous wave fiber laser. Alexa Flour 647 was excited using a HeNe 633 laser and imaged using the laser scanning confocal modality of the system. Fluorescence was detected with HyD detectors (Leica). Cells were washed, resuspended in supplemented Myelocult, and allowed to adhere to 0. 15 glass coverslips coated with antibody as described above. For imaging of the actin network alone, samples were prepared following a modified protocol described in [42]. After a period of incubation at 37°C, samples were washed once in PBS, dipped into a 1∶3 dilution of PEM buffer (0. 1 M PIPES (Sigma), pH 6. 9,1 mM EGTA (Sigma), 1 mM MgCl2 (Sigma) ) in dH20 for 10 s, sonicated in PEM for 1–2 s at a 45° angle to the probe, and incubated with 1% Triton X-100 (Sigma) in PEM for 1–2 min before fixation in 2% glutaraldehyde (Fluka) in PEM. For imaging of granules, cells were “unroofed” by applying a nitrocellulose membrane that had been wet in PEM to the coverslip for 45–60 s, removing it, and then fixing the sample in 2% glutaraldehyde in PEM. All samples were processed for EM as described [43]. Platinum replicas were imaged on a JEM 1011 transmission electron microscopy (JEOL USA) at 100 kV. Images were captured using an ORIUS 835. 10W CCD camera (Gatan) and are presented in inverted contrast. Immulon 4HBX 96 well flat bottom plates (Thermo) were coated with murine IgG (BD), anti-human NKp30, anti-human CD18, or both anti-NKp30 and anti-CD18 at 5 µg/mL in PBS overnight at 4°C. Plates were washed twice with PBS and blocked for 1 h at room temperature with R10 media. 1×105 cells were added to wells and plates were spun at 1,000 rpm for 2 min before incubation at 37°C. For the timecourse degranulation assay, supernatants from spontaneous and activated wells were harvested at indicated times. For inhibitor treatments, media containing inhibitor or vehicle (DMSO) was added at indicated times. Thapsigargin (Calbiochem) was used at 1 µM. Supernatants were harvested after 60 min. For total release, cells were lysed in 0. 5% Nonidet P40 (Accurate Chemical and Scientific). Supernatants were assayed by mixing 20 µL supernatant with 200 µL of a solution containing PBS, 9. 8 mM HEPES (Gibco), 196 µM Z-L-Lys-SBzl hydrochloride (BLT, Sigma), and 218 µM 5,5′-dithiobis (2-nitrobenzoic acid) (DTNB, Sigma). Samples were incubated for 30 min at 37°C and absorbance was measured immediately at 405 nm. Percent total release was measured by subtracting the spontaneous release value from activated release values (A–S) and the total release value (T–S), and then dividing (A–S) by (T–S). Images and timelapse series were analyzed using either Volocity or the FIJI package of ImageJ (http: //pacific. mpi-cbg. de). Using Volocity, fluorescently tagged actin footprints were identified using a classifier that identifies objects above a selected standard deviation above the mean intensity (usually 0–1). LysoTracker and pHluorin positive events were similarly identified (4–10 standard deviations above mean intensity), with the additional exclusion of events smaller than 0. 05 µm2. For actin fluorescence ratio measurements, the MFI at the location of granule approximation or degranulation was divided by the MFI of the actin footprint. Maximum and minimum potential actin intensities were also measured by dividing the maximum and minimum pixel intensities of the actin footprint by the MFI of the footprint. For the “adjacent points” measurement, the region of interest (ROI) that was identified for the degranulation (pHluorin) signal was moved to four neighboring locations. The actin intensities were measured, averaged, and then divided by the MFI of the actin footprint to generate a single value. Distance of granules from the MTOC was determined as described [17] by inputting the coordinates of the MTOC centroid and a granule centroid into the Pythagorean equation (a2+b2 = c2: (Xcentroid−Xclearance) 2+ (Ycentroid−Yclearance) 2 = c2, where c is the distance from the MTOC). FIJI was used to generate radial intensity profiles using the radial profile plugin (http: //rsbweb. nih. gov/ij/plugins/radial-profile. html). For a profile of the entire cell footprint an ROI circle was drawn around the cell and data from running the plugin were exported to Excel (Microsoft). For profiles of degranulation events images were first split into red and green images and then a circle with a radius of 8 pixels (1. 08 µm) was drawn around each event. Data were generated for 6 radii from the center of each event. To determine the radial intensity change, an outer radial intensity value was subtracted from an inner radial intensity value. Thus a negative value indicates that the outer circle has higher mean intensity than the inner circle. To measure changes in the actin network over time, sequential images were imported into FIJI, a line was drawn across the center of the cell, and line intensity profile data were generated for each time point at the same location within the cell. Data were exported to Excel and standard deviations calculated. Surface plots were generated using the surface plot function in FIJI. Images of granules taken using STED microscopy were analyzed in Volocity and diameters measured by drawing lines across the center of the granules. STED microscopy images of the actin network were imported into FIJI and processed before analysis. Background was subtracted using the Rolling Ball Subtraction algorithm with the radius set to 150 pixels and then pixel intensities were squared twice. An ROI was drawn around the interior of the cell and clearances were identified using the default autothreshold with “dark background” unchecked. Clearance areas were sorted, grouped, and counted in Excel based on size. Dividing the number of clearances per cell by the area measured normalized the values. Area cutoffs were implemented by using granule diameter as a reference. The smallest two sizes calculated assume that granules are uniformly spherical and require a clearance that has an area that would accommodate the equatorial area of the granule. The larger size categorizes all clearances that are larger than most granules. Colocalization analysis was performed in Volocity using the “Find objects” tool to identify granules and a fixed intensity threshold, which was adjusted to generate interfilament or intercellular black space on a per field basis, to identify the actin network. The colocalized area of granule and actin network staining was divided by the area of the granule to yield a value denoting the percent of the granule area colocalized with actin. Five granules from 10 cells over 2 experiments were measured (N = 50 granules). Line profiles of granules and the actin network were performed in FIJI after channels were separated. All STED images were measured in their raw form, but shown after processing in Leica AS AF with the median noise reduction feature. Images from platinum replica electron microscopy were inverted and linearly contrast enhanced using Photoshop (Adobe) and imported into FIJI for processing and analysis. Background was subtracted from each image using the Rolling Ball Subtraction algorithm with the radius set to 25 pixels. Pixel intensities were subsequently squared. Cells were identified using the default autothreshold and cell contact area and cell centroid were measured with “include holes” checked. ROIs were drawn around the interior of the cell to more accurately identify filaments and to avoid debris. Filaments were identified using the default automatic threshold with “dark background” checked. Clearances in the filamentous network were identified as mentioned above. Distance from the cell centroid was determined by inputting the coordinates of the cell centroid and a clearance centroid into the Pythagorean equation as described above for granule to MTOC distance. All data were plotted using Prism (Graphpad). Statistical significance was determined using Prism to perform one- or two-sample, unpaired or paired, two-tailed Student' s t tests. Unless otherwise indicated all tests were two-sample, two-tailed, and unpaired. Where noted, n. s. denotes differences that have p values >0. 05 and therefore considered not significant.

The immune system' s natural killer cells eliminate diseased cells in the body. They do so by secreting toxic molecules directly towards the diseased cells, so causing their death. This process is essential for the host organism to defend itself against infectious diseases. The interface between the natural killer cell and its target—the lytic immunological synapse—forms by close apposition of the surface membranes of the two cells. It is characterized by coordinated rearrangement of proteins to allow lytic granules, which contain the toxic molecules, to fuse with the cell surface at the synapse. Given the large size of the granules, one challenge the natural killer cell faces is how to contend with network of actin filaments just under the cell surface, which potentially could pose a barrier to secretion. The current model proposes large-scale clearing of actin filaments from the center of the immunological synapse to provide granules access to the synaptic membrane. By using very high-resolution imaging techniques, we now demonstrate that actin filaments are present throughout the synapse and that natural killer cells overcome the actin barrier not by wholesale clearing but by making minimally sufficient conduits in the actin network. This suggests a model in which granules access the surface membrane by means of specific and facilitated contact with the actin cytoskeleton.

Abstract Introduction Results Discussion Materials and Methods

cellular structures subcellular organelles molecular cell biology immune cells cell biology nk cells immunology biology cytoskeleton

2011

Natural Killer Cell Lytic Granule Secretion Occurs through a Pervasive Actin Network at the Immune Synapse

Current drugs to treat African sleeping sickness are inadequate and new therapies are urgently required. As part of a medicinal chemistry programme based upon the simplification of acetogenin-type ether scaffolds, we previously reported the promising trypanocidal activity of compound 1, a bis-tetrahydropyran 1,4-triazole (B-THP-T) inhibitor. This study aims to identify the protein target (s) of this class of compound in Trypanosoma brucei to understand its mode of action and aid further structural optimisation. We used compound 3, a diazirine- and alkyne-containing bi-functional photo-affinity probe analogue of our lead B-THP-T, compound 1, to identify potential targets of our lead compound in the procyclic form T. brucei. Bi-functional compound 3 was UV cross-linked to its target (s) in vivo and biotin affinity or Cy5. 5 reporter tags were subsequently appended by Cu (II) -catalysed azide-alkyne cycloaddition. The biotinylated protein adducts were isolated with streptavidin affinity beads and subsequent LC-MSMS identified the FoF1-ATP synthase (mitochondrial complex V) as a potential target. This target identification was confirmed using various different approaches. We show that (i) compound 1 decreases cellular ATP levels (ii) by inhibiting oxidative phosphorylation (iii) at the FoF1-ATP synthase. Furthermore, the use of GFP-PTP-tagged subunits of the FoF1-ATP synthase, shows that our compounds bind specifically to both the α- and β-subunits of the ATP synthase. The FoF1-ATP synthase is a target of our simplified acetogenin-type analogues. This mitochondrial complex is essential in both procyclic and bloodstream forms of T. brucei and its identification as our target will enable further inhibitor optimisation towards future drug discovery. Furthermore, the photo-affinity labeling technique described here can be readily applied to other drugs of unknown targets to identify their modes of action and facilitate more broadly therapeutic drug design in any pathogen or disease model. Protozoan parasites of the Trypanosoma genus cause widespread disease and death across large regions of the developing world. In sub-Saharan Africa Trypanosoma brucei gambiense and T. b. rhodesiense are the causative agents of human African trypanosomiasis (HAT, or African sleeping sickness) in humans while several species cause disease in livestock and wild animals, creating a major socio-economic burden to the African continent. The parasites are spread through the bites of infected tsetse flies and, if left untreated, infection is usually fatal. Over 65 million people who live in the tsetse fly habitat are at risk of infection and each year there are an estimated 15–20,000 new cases [1]. In the early 1900’s African trypanosomes became one of the first subjects of “modern drug discovery” when Paul Ehrlich, following his observations on differential cell stains, hypothesised that some molecules could be developed to target pathogens but not their hosts (a term he coined “chemotherapy”), and screened a library of synthetic dyes in trypanosome-infected animals to find a “magic bullet” [2,3]. Through a combination of rational synthetic chemistry and phenotypic screening his pioneering work led to the discoveries by others of suramin in 1917 and melarsoprol in 1949 [4], both of which are still front-line drugs for the treatment of early stage (suramin) and late stage (melarsprol) infection by T. b. rhodesiense [5]. Pentamindine, which is currently the first-line treatment for early stage infection by T. b. gambiense [5], was likewise developed from the anti-diabetic synthalin in 1937 [6,7]. However, HAT has been neglected over the past half century and all of these antiquated non-oral drugs are difficult to administer, are sometimes ineffective and are themselves toxic, often causing undesirable side effects with melarsoprol causing the death of up to 5% of those treated [5,8]. Furthermore, melarsoprol resistance is a growing issue [9–14] and new drugs are therefore urgently needed, particularly for late stage infection. Despite their antiquity and widespread use, the targets and modes of action of these currently used drugs are poorly understood, making it difficult to design to safer analogues. Investment from the pharmaceutical industry has been slow in forthcoming for this and related neglected diseases, which affect many of the poorest and most underdeveloped countries in the world, and efforts so far have been driven instead by charities and non-profit organisations. Advances in automated liquid handling, cell culture and detection technology has allowed researchers and the pharmaceutical industry to return to phenotypic screening-based practices, as those pioneered by Ehrlich, for the latest drug discovery efforts. We recently reported the total synthesis and trypanocidal activity of the acetogenin, chamuvarinin [15,16] and non-natural bis-tetrahydropyran 1,4-triazole (B-THP-T) analogues thereof including compound 1 ([17]; Fig 1A) using a phenotypic screening approach. Acetogenins are a family of over 400 structurally related fatty acid-derived natural products isolated from tropical plants of the Annonaceae family (for review, see [18]), and characteristically bear one to three tetrahydropyran (THP) and/or tetrahydrofuran (THF) rings flanked by a terminal γ-lactone head and a hydrophobic tail. Many members have been reported to display high inhibition of mitochondrial complex I [19–21], making them cytotoxic to a wide range of organisms [22,23], and their particularly high potency against ATP-hungry tumour cells (reviewed in [24]) has led to their investigation as potential anti-cancer chemotherapeutics; despite mammalian cells requiring complex I activity, pre-clinical trials with select acetogenins are encouraging, with some proving as effective and selective as Taxol, a first-line treatment for some cancers, at reducing solid tumours in mice [25]. Cytotoxic activities vary among acetogenins and between cell lines/organisms but several studies have demonstrated that both γ-lactone and THP/THF moieties are essential for complex I inhibition [26–28]. Intriguingly, chamuvarinin and B-THP-Ts are toxic to procyclic form (PF) and bloodstream form (BSF) T. brucei [15–17] with EC50 values in the low micromolar range (Fig 1), however, complex I is not essential in either form of the parasite [29,30], and our B-THP-Ts lack the terminal γ-lactone indicating that our compounds must have a different mode of action in kinetoplastids. In order to further optimise the potency and selectivity of our trypanocidal B-THP-T compounds this study set out to determine their precise target (s) in T. brucei. The identification of targets of lead compounds is often necessary to allow subsequent lead optimisation, but target identification is a major bottleneck in the drug discovery pipeline following phenotypic screens. Photo-affinity labelling provides an elegant solution by covalently attaching a bifunctional photo-affinity probe analogue of the lead compound to the target for its ultimate detection (for review see [31] and Fig 1A & 1B) and can readily be achieved by incorporating photoreactive and reporter moieties to the lead compound. Azides, benzophenones and diazirines make excellent photoreactive substituents as they decompose upon UV-irradiation to form short-lived reactive nitrenes, diradicals or carbenes respectively, which rapidly react with neighbouring molecules to form covalent bonds. Hence, by photo-activating the probes once they have been trafficked to their targets, these protein targets can be preferentially labelled. Several radiolabled leads have been functionalised with photoreactive azides [32,33], benzophenones [34] or diazirines [21,35,36] and their targets detected by autoradiograph following protein separation, i. e. SDS-PAGE. A similar, but radiolabel free, methodology has also been employed by functionalising an inhibitor with both a photoreactive diazirine and a fluorescent reporter moieties, allowing the target to be visualised both in gel and within intact cells, thus adding the benefit of target localisation determination. The major drawback of using a radiolabel or fluorescent tags is that, while it allows one to visualise the target, it rarely provides a means to isolate or identify the target. Thus, many [37–44] have used bi-functional photo-affinity probes containing a photoreactive moiety coupled with biotin affinity tag which allows the photo-affinity labelled target to be purified from other proteins using streptavidin beads and identified by tandem mass spectrometry. More versatile yet is the functionalization of the lead compound with photo-affinity group and alkyne handle which allows the subsequent attachment of any desired reporter, such as fluorophore [45–47] or biotin [48,49] using the Cu (II) -catalysed azide-alkyne cycloaddition, also known as the “click reaction” [50] once the photo-affinity probe had been UV-cross-linked to its target. To identify the target (s) of B-THP-T compounds we synthesised two photo-affinity probe analogues of our lead compound (Fig 1A, synthesis to be reported elsewhere): mono-functional compound 2 contains a photoreactive diazirine, but has no reporter capability; while bi-functional compound 3 contains the same photoreactive diazirine and an alkyne handle to allow subsequent identification through appendage with biotin affinity or Cy5. 5 fluorescent reporter tags. Importantly, the diazirine and alkyne encompass only minor changes to our lead B-THP-T, compound 1, and all compounds show similar levels of toxicity (Fig 1 and S1 Fig) indicating that the photo-affinity probes work through the same mode of action and are thus suitable for target identification purposes. Here we employed an in vivo pulse-chase photo-affinity labelling methodology to covalently attach bi-functional compound 3 to its target within its biological context and, coupled with liquid chromatography-tandem mass spectrometry identified proteins that are potential targets of B-THP-T compounds. Subsequently, using several metabolic and biochemical assays with lead B-THP-T, compound 1, we validated two of those protein hits as binding partners and now allows us to begin work to optimise the potency and selectivity of our inhibitors towards these targets. Furthermore, it demonstrates the strength of this photo-affinity labelling technique, which can be readily applied to any drug discovery program to accelerate the drug discovery process. To identify the target (s) of the trypanocidal B-THP-T, compound 1, we developed compound 3, a bi-functional photo-affinity probe analogue of compound 1 (Fig 1A and 1B, synthesis to be published elsewhere). In addition to the core structure of compound 1, compound 3 contains a diazirine to covalently bind to its target upon UV-activation and an alkyne handle to which an azide reporter, such as a fluorophore or affinity tag can subsequently be added via copper catalysed azide-alkyne cycloaddition, often referred to as the ‘click reaction’ [31,50]. As a negative control, we also developed the mono-functional compound 2 (Fig 1A), which contains the diazirine, but lacks the alkyne and therefore cannot conjugate the azide reporter. Crucially, the addition of the diazirine and alkyne moieties encompass only minor changes to the parent molecule and have minimal effect on overall compound potency (Fig 1A and S1 Fig), suggesting that compounds 1–3 have the same mode of action and that photo-affinity probe 3 is therefore suitable for the identification of target (s) of our B-THP-T compounds. Compounds 1–3 were pulse-chased to their targets in live PF T. brucei cells and UV-irradiated to covalently attach the diazirine-containing compounds to their targets (Fig 1B). The pulse-chase methodology was employed rather than labelling in cell lysates so that inhibitors would be trafficked to their target (s) and labelling would therefore occur in its biological context, minimising the likelihood of identifying high affinity targets that do not form in vivo. Once the extracted proteins had been subjected to azide-alkyne cycloaddition with Cy5. 5-azide, proteins that had been appended with Cy5. 5 were detected by SDS-PAGE coupled with fluorescence imaging (Fig 1C). As expected, when either the diazirine or alkyne were absent (as in compounds 1 and 3) no proteins fluoresced as there was no mechanism for them to covalently bind the Cy5. 5 reporter molecule, and this absence of fluorescence indicated that there was minimal non-specific binding of Cy5. 5 reporter to cellular proteins. Proteins only incorporated the Cy5. 5 reporter tag when diazirine and alkyne were both present on the B-THP-T compound (i. e. , with the bi-functional compound 3, lane 3 of Fig 1C) confirming their essentiality for photo-affinity labelling of protein targets. 18 protein bands were detected to varying degrees, with most prominent bands around 70 and 55 KDa, suggesting major targets of these sizes. To identify the cellular compartments housing B-THP-T protein targets, Cy5. 5-azide was “clicked on” to B-THP-T-conjugated proteins in fixed whole cells following pulse-chase photo-affinity labelling, and detected by fluorescence microscopy (Fig 2). Bi-functional compound 3 without UV-conjugation and lead compound 1, were used as negative controls. As with Fig 1C, Cy5. 5 labelling only occurred when compound 3 was cross-linked by UV-activation, again indicating specific labelling of target proteins (s). Cy5. 5 labelling was absent from the nucleus, but showed good co-localisation with the MitoTracker staining, suggesting that B-THP-T compounds predominantly target proteins associated with the mitochondria. Next a series of pull-down experiments were performed to identify the proteins tagged with bi-functional photo-affinity probe, compound 3. For pull-down of the target (s), biotin azide was used as reporter and “clicked” on to the B-THP-T-labelled proteins following the same pulse-chase in vivo photo-affinity labelling methodology as used above. Tagged proteins were then enriched with streptavidin-agarose, digested on-bead with trypsin, analysed by LC-MSMS, and identified through Mascot searches of the T. brucei proteome. Compound 1 and mono-functional compound 2, which both lacked the alkyne handle required for biotin-azide cycloaddition, were used as negative controls and any proteins pulled down with either of these were eliminated from the bi-functional compound 3 list of potential protein targets as they represented proteins bound non-specifically to the beads rather than through the biotin-streptavidin interaction. Following elimination of low scoring proteins and unknown proteins (which would be difficult to validate in the timeframe of this study due to their unknown nature), twenty-five T. brucei proteins were identified as potential targets in three replicate pull-downs from whole cell protein extracts (S1 Table). Of those, seventeen (68%) utilised nucleotides as a co-factor, indicating that B-THP-Ts may mimic nucleotides. As localisation studies indicated the target was primarily localised in the mitochondria, attention was focused on the six mitochondrial proteins identified in the pull-down data (Table 1). Pleasingly, bands corresponding to proteins of the masses in Table 1 appear to be Cy5. 5-labelled via compound 3 cross-linking in lane 3 of Fig 1C, with the most intense bands (~70 KDa and ~55 KDa) matching closely with the masses of two of the top three hits (72 KDa mitochondrial heat shock protein 70, mHSP70; and 56 KDa ATP synthase F1 β-subunit). While this supports the likelihood that the proteins identified in Table 1 were captured via the streptavidin-biotin/compound 3 interaction rather than non-specific interaction with the beads, the list only serves as a guide and each of the hits needs to be validated before being identified as a bona-fide target. The most intense Cy5. 5-labelled band in Fig 1C matched the mass the top-scoring hit in Table 1 (72 KDa mHSP70), which is under investigation as a B-THP-T target and will be reported elsewhere. Of the remaining five mitochondrial pull-down hits, four (hits 2–5) have roles in the production of ATP through proline metabolism (Fig 3) and are the subject of further investigation herein as they can be validated or eliminated through a series of related follow-up assays. Proline is one of the principle amino acids used as a carbon source for PF T. brucei in the insect midgut [51,52]. It is taken into the PF mitochondrion and converted to pyruvate through eight enzymatic steps, and then further catabolised to alanine or acetate end products (Fig 3, [53]). ATP is produced at two points during this process: first through substrate-level phosphorylation at succinyl-CoA synthetase (Fig 3, step 5) as intermediate succinyl-CoA is converted to succinate by succinyl-CoA synthetase; and second through oxidative phosphorylation (the coupling of the electron transport chain (ETC) with ATP production) at the FoF1-ATP synthase (mitochondrial complex V) as the reduction of succinate to fumarate by mitochondrial complex II feeds electrons into the electron transport chain (ETC). Two of the pull-down hits, δ-pyrroline-5-carboxylate dehydrogenase (δPCDH, hit 4 in Table 1, Fig 3) and dihydrolipoamide succinyltransferase (DLST, hit 2 in Table 1, Fig 3) have roles prior to substrate-level ATP production, while two other hits, the F1 α- and β-subunits (hits 5 and 3 respectively in Table 1, Fig 3), form the ADP-binding regulatory and ADP-binding catalytic domains respectively of the FoF1-ATP synthase and have direct roles in ATP production from oxidative phosphorylation downstream of substrate-level ATP production. Targeting of any of these four hits by B-THP-T compounds would consequently impact cellular ATP production to varying extents. Therefore, a series of biochemical phenotypic studies were undertaken to validate or exclude the pull-down of these four hits. We first set out to determine if B-THP-T compounds affected cellular ATP levels. PF T. brucei were incubated in buffered PBS with/without proline as the sole carbon source and various inhibitors for 2 h. Cells were harvested and their ATP content was determined by bioluminescence assay. Two ETC inhibitors were used as positive control: malonate is a competitive inhibitor of mitochondrial complex II and prevents the reduction of succinate to fumarate and entry of electrons to the ETC; and antimycin A (AA) is a mitochondrial complex III inhibitor, which prevents the transfer of electrons to cytochrome c and ubiquinone recycling. Both ETC inhibitors significantly inhibited cellular ATP production by 62 ± 15% and 71 ± 9% respectively (Fig 4), consistent with them shutting down oxidative phosphorylation but leaving upstream substrate-level phosphorylation at succinyl-CoA synthetase unaffected. Compound 1 inhibited ATP production in a dose-response manner to similar levels (by 55 ± 6%), supporting the hypothesis from pull-down data that its target is involved in the ATP production pathway from proline. We next sought to identify the point at which ATP production was blocked using digitonin-permeabilised cells. Digitonin permeabilises the plasma membrane of trypanosomatids, but crucially, leaves the glycosomes and mitochondria intact and fully functional, allowing them to be probed with compartment-specific ATP-yielding substrates [54,55]. Digitonin-permeabilised PF T. brucei were probed with succinate, which is readily transported across the PF T. brucei mitochondrial membrane by the dicarboxylic acid carrier [56] and can yield only oxidative phosphorylation-derived ATP through the FoF1-ATP synthase (complex V). As expected, the mitochondrial complex II inhibitor, malonate, completely eliminated all ATP production from succinate, while downstream inhibitors antimycin A (AA), complex V inhibitor oligomycin A (OA) and protonophore 2,4-dinitrophenol (DNP) almost abolished ATP production (Fig 5A). Lead B-THP-T, compound 1, decreased ATP production from succinate to similar levels, while rotenone and salicylhydroxamic acid (SHAM), which inhibit the ETC at different entry points and are therefore irrelevant inhibitors, had no significant effect, indicating that compound 1 inhibits oxidative phosphorylation, and is consistent with the F1 α- and/or β-subunits being targets. However, this result on its own does not confirm that the F1 subunits are targets, as inhibition of substrate (ADP or succinate) uptake or any of the mitochondrial complexes could have the same effect. Next, mitochondria were incubated with α-ketoglutarate, which is also taken into the mitochondria by the dicarboxylic acid carrier, to examine the effects of B-THP-T compounds on substrate-level phosphorylation. The multiprotein α-ketoglutarate dehydrogenase complex, of which dihydrolipoamide succinyltransferase (DLST, hit 2 in Table 1) is a component, converts α-ketoglutarate into succinyl-CoA (Fig 3), allowing succinyl-CoA synthetase to generate substrate-level ATP. In addition, electrons from resulting succinate can enter the ETC to generate oxidative phosphorylation-derived ATP. As complexes III-IV export six protons for every α-ketoglutarate metabolised and the FoF1-ATP synthase imports only four protons with every ATP synthesised, each mole of α-ketoglutarate can yield 1. 5 moles ATP through oxidative phosphorylation. Thus, 40% of the ATP produced from α-ketoglutarate is by substrate-level phosphorylation and 60% is through oxidative phosphorylation. With α-ketoglutarate as substrate, the four oxidative phosphorylation inhibitor controls only blocked up to 48 ± 12% of ATP production (Fig 5B) as they were unable to block upstream substrate-level ATP production by succinyl-CoA synthetase. Compound 1 inhibited ATP production by 67 ± 13%, achieving a significantly lower level of inhibition than that achieved when succinate was used as substrate (Fig 5A), suggesting that the blockade is not due to inhibition of substrate uptake, but is due to inhibition of one of the mitochondrial complexes. The greater level of inhibition achieved by compound 1 over the other oxidative phosphorylation inhibitors suggests that compound 1 may also be inhibiting substrate-level phosphorylation to some extent, and is, perhaps, an indication that compound 1 also targets DLST (pull-down hit 2 in Table 1). However, the more significant effect of compound 1 is, by far, on oxidative phosphorylation and the F1 subunits are therefore likely to be more significant targets. Use of the cytosolic oxidative phosphorylation substrate glycerol-3-phosphate (Gly-3P) allowed us to establish whether substrate uptake was a target of our compounds. Mitochondrial Gly-3P dehydrogenase (mGPDH) is located in the intermembrane space of the mitochondrion and converts cytosolic Gly-3P into dihydroxyacetone phosphate (DHAP) while transferring electrons to the ETC via ubiquinone [54]. Gly-3P can therefore be used to probe oxidative phosphorylation at a different entry point from succinate and is not reliant on substrate uptake. When Gly-3P was used as substrate, malonate had no effect on ATP production (Fig 5C) as complex II is irrelevant at this ETC entry point. Likewise, other irrelevant inhibitors rotenone and SHAM had no effect on ATP production. However, AA, OA and DNP, which all inhibit oxidative phosphorylation downstream of mGPDH and mitochondrial complex II, reduced ATP production by 75 ± 14%, 65 ± 21% and 67 ± 13% respectively. The inability of oxidative phosphorylation inhibitors to abolish ATP production from Gly-3P has been noted previously [54] and may be the result of glycolytic ATP production following glycosomal uptake of Gly-3P or its product, DHAP. Lead B-THP-T, compound 1, acted similarly, inhibiting ATP production by 82 ± 16%. This confirmed that inhibition of oxidative phosphorylation by compound 1 is not due to inhibition of substrate uptake as Gly-3P is not taken into the mitochondrion and, furthermore, it indicates that the inhibition is downstream of electron entry to the ETC, or more specifically, that compound 1 inhibits one of complexes III-V. The potency of lead B-THP-T, compound 1, was determined against oxidative phosphorylation in dose-response assays in digitonin-permeabilised PF T. brucei. The IC50 of compound 1 was 57 ± 15μM and 39 ± 5 μM when succinate and Gly-3P were used respectively (Fig 5D). Although these values are higher than in whole cell EC50 assays (8. 4 μM), it is noteworthy that they cannot be compared directly as the concentrations of proteins and substrates differ greatly between experiments, as do the time-scales of the experiments: digiton-permeabilised cells were plated at over 1000-fold greater density than live cells and used an acute incubation (30 min) rather than the chronic 72 h incubation of the cell viability assay. To determine which mitochondrial complex of oxidative phosphorylation was targeted by B-THP-T compounds, their effects on the mitochondrial membrane potential (Δψm) were monitored in live PF T. brucei (Fig 6A). MitoTracker Red CMXRos, like tetramethylrhodamine and safranine, accumulates in active mitochondria and the degree of accumulation is dependent on the Δψm [57,58], however, accumulation of MitoTracker Green FM is independent of the Δψm and the ratio of red/green uptake can therefore be used to determine the Δψm normalised to the number and size of mitochondria present [59]. In PF T. brucei mitochondrial complexes III-IV generate a Δψm by pumping protons out of the mitochondrion during the ETC (Fig 3), and thus ETC inhibitors (malonate or AA) prevent proton export and decrease the Δψm (Fig 6B). Conversely, mitochondrial complex V (FoF1-ATP synthase) uses this potential to generate ATP as it pumps protons back in (Fig 3), so inhibition of complex V by OA prevents proton import and elevates the Δψm (Fig 6B and [58]). Similarly, compounds 1 and 3 elevated the mitochondrial membrane potential (by 2-3-fold), confirming that they, like OA, reduce oxidative phosphorylation by the inhibition of the FoF1-ATP synthase. The F1 α- and β-subunits of the FoF1-ATP synthase were both identified as potential B-THP-T targets through biotin pull-down (Table 1), and the FoF1-ATP synthase was shown above to be inhibited by lead B-THP-T compound 1, but to confirm whether both, or just one of the F1 subunits is a target of our compounds, we generated cell lines endogenously expressing α- or β-subunits with C-terminal GFP-PTP tags. The PTP affinity tag comprises epitopes for proteins C and A separated by a TEV protease cleavage site for efficient purification of the target protein and the cloning methodology facilitates recombination within the endogenous targeted gene rather than exogenous integration, allowing for endogenous levels of expression [60]. The endogenous expression produces tagged protein at physiological levels, allowing it to be trafficked and form complexes in the same way as untagged endogenously expressed protein, as previously demonstrated through the purification of intact FoF1-ATP synthase with a C-terminally TAP-tagged F1 β-subunit [61]. In addition, we cloned green fluorescing protein (GFP) into the tag to act as an additional reporter and for antibody-free localisation studies. Fluorescence imaging (Fig 7A) confirms that both tagged subunits correctly localise to the mitochondria. Pulse-chase in vivo photo-affinity labelling with bi-functional compound 3 (and lead compound 1 as negative control) was performed on cells from both cell lines expressing GFP-PTP-tagged F1 α- or β-subunits. Tagged subunits were immobilised on IgG-sepharose beads, purified, and Cy5. 5 fluorescent reporter clicked on, and proteins were detected by western blotting (Fig 7B). Mouse anti-GFP coupled with anti-mouse-DyLight-800 identified tagged F1 α-GFP-PTP and F1 β-GFP-PTP subunits migrating as expected at approximately 89 KDa and 101 KDa respectively. In trypanosomatids (but not in other organisms) the native α-subunit is expressed as a 63 KDa protein which is then cleaved by an unknown protease to 20 KDa N-terminal and 43 KDa C-terminal domains, both of which remain associated with the FoF1-ATP synthase complex [61,62]. Pleasingly the 109 KDa GFP-PTP C-terminally tagged α-subunit also appears to be cleaved at the same position yielding the 89 KDa GFP-PTP-tagged C-terminal domain of the α-subunit, indicating that it is localised and processed within the cell as the native untagged protein. Fig 7B shows that approximately equal amounts of GFP-PTP-tagged protein had been captured between compound 1/3 pairs and between α- and β-subunits (Fig 7B), indicating that α- and β-subunits were expressed (and captured) to similar levels. However, for both α- and β-subunits, Cy5. 5 fluorescence at 700 nm was only detected appreciably in samples photo-affinity labelled with bi-functional compound 3, and not with lead compound 1. This indicates that compound 3 can UV-conjugate to both F1 α- and β-subunits and corroborates their pull-down in Table 1. F1 of mitochondrial complex V is a multi-protein complex composed of α-, β-, γ-, δ- and ε-subunits with stoichiometry 3: 3: 1: 1: 1 [63,64]. The α- and β-subunits form a catalytic heterohexamer of alternating subunits with ADP/ATP binding sites at their interfaces (S2 Fig and Fig 8A and 8B) while γ- and ε-subunits form an asymmetrical central stalk connected to the Fo moiety which together, driven by the proton-motive force, rotate within the α/β hexamer to drive ATP production by introducing conformational changes at the catalytic sites and changing their affinities for ADP and ATP [63,65–67]. X-ray crystal structures of F1 have been determined for several species, but no high-resolution structure of trypanosomatid F1 have been solved. Consequently, to determine likely B-THP-T binding sites over the entire F1 moiety, B-THP-T compound 1, which displays only modest selectivity towards T. brucei over mammalian HeLa cells (Fig 1), was docked into the entire F1 crystal structures from Sacharomyces cerevisiae (PDB entry 2WPD [68]) and Bos taurus (PDB entry 1BMF [63]) using AutoDock Vina [69]. For both crystal structures, catalytic and regulatory ATP binding sites were identified as sites of lowest binding energy and therefore most likely binding positions of B-THP-T compounds. The F1 α- and β-subunits are paralogues of one another with sequence identity of 20–25% (value depending on species), and while the β-subunit has catalytic activity, the α-subunit has lost this ability and instead plays a regulatory role in ATP synthesis and hydrolysis. They share a common RecA-like protein fold along with other classes of ATP-binding proteins including kinases, phosphatases, ATPases, heat shock proteins, transfer/transport ATPases, and permeases [70,71], in which the adenine of the nucleotide binds within a sequence-non-specific hydrophobic pocket and the B-phosphate within the amine nest of the conserved Walker A motif (sequence GxxxxGKT/S) (Fig 8A and 8B). Although the mode of ATP-binding is similar for α- and β-subunits, their three-dimensional shapes differ due to differences in amino acid sequence. The three ATP-binding regulatory sites of the α-subunits (top panel of S3A Fig) are all relatively similar, however, the three ATP-binding catalytic sites of the β-subunits (lower panel of S3A Fig) are different as their shapes are dependent on the position of the central stalk (and thus state of bound nucleotide). The most significant differences are in the distances between Tyr345 and Phe424, and in the position of the neighbouring α-subunit which completes the catalytic site. The empty site of the β-subunit (βE) is open, while the ATP-bound (βTP) and ADP + Pi-bound (βDP) conformations are closed and somewhat more similar to one another. Compound 1 docked similarly within all three regulatory ATP-binding pockets of the yeast α-subunits (αE, αDP and αTP, top panel of S4 Fig and Fig 8C). The triazole H-bonded with the amine-rich phosphate nest, THP2 occupied a similar position to the ATP ribose and hydrophobic tail bound within the hydrophobic adenine-binding pocket. THP1 could potentially form an H-bond via a water bridge, while the terminal hydroxyl could H-bond with Thr214, Lys211 of the α-subunit or Ala327 of the neighbouring β-subunit. Interactions between compound 1 and the bovine regulatory subunits were similar. Compound 1 docked similarly within the nucleotide-bound β-subunits (βDP and βTP, lower panel of S4 Fig and Fig 8D) of yeast F1 with THP2 sandwiched between Tyr345 and Phe424 and the hydrophobic tail buried within the hydrophobic adenine-binding pocket lined by V165, F166, P346 and F418. The triazole adopted a similar position to the nucleotide’s ribose, allowing THP1 to sit close to the phosphate-binding Walker A nest and the terminal hydroxyl to form extensive H-bonds with Pi-coordinating Arg190, Glu189, Arg260 and Arg’375 (or Adp256 and Thr164). However, in the open BE conformation, Tyr345 and Phe424 lie too far apart to sandwich THP2 and the triazole instead forms interactions with the Walker A nest and THPs potentially forming H-bonds via water bridges. Docking of compound 1 to the ATP-binding sites of the α- and β-subunits reveals that compound 1 can form similar interactions as ATP and suggests that B-THP-T compounds mimic ADP, a finding which is mirrored by the pull-down of predominantly nucleotide utilising proteins during target identification (Table 1). Intriguingly, only one other protein (phosphoenolpyruvate carboxykinase–a glycosomal protein shown in S1 Table) identified as a potential target during pull-down experiments has the Walker A motif, suggesting that the B-THP-T compounds probably do not target all members of this protein family, but instead bind to specific nucleotide-binding sites. Until recently drug discovery efforts focussed on the identification of compounds targeting proteins specific to the target organism with the view that such compounds would have low toxicity to the human hosts. Such studies typically began with screening and lead optimisation against recombinant target protein, but projects often derailed during subsequent cell-based testing when compounds were shown to be equally toxic to mammalian cells or compounds failed to reach their targets within the cells. Phenotypic screening removes all previous bias towards targets and brings into play targets that were previously though undruggable by testing compounds in their context within the cell. However, the major bottleneck in phenotypic screening lies in identifying the target after initial screens, often making it difficult to further improve compound potency and selectivity. We previously identified compound 1 as a trypanocidal compound during phenotypic screening and this study has successfully identified its protein target and mode of action using a photo-affinity labelling approach. Using a compound 1-like bi-functional photo-affinity probe (compound 3), we showed that the target is localised to the mitochondria and biotin pull-down experiments identified the α- and β-subunits of the FoF1-ATP synthase (mitochondrial complex V) as potential targets. We next used compound 1 to validate our photo-affinity labelling experiments and showed that compound 1 inhibits cellular ATP production by blocking oxidative phosphorylation at the FoF1-ATP synthase. Photo-affinity labelling with bi-functional compound 3 confirmed that both α- and β-subunits are targeted and modelling suggests that compound 1 binds at their ATP-binding regulatory and catalytic sites respectively. While it is possible that both sites could be targeted, it must be noted that regulatory and catalytic sites sit at the interfaces between α- and β-subunits and it may therefore be possible for photo-affinity probe compound 3 to covalently bind both of the subunits from either catalytic or regulatory site. Regardless, future investigations can target both sites to tailor new compounds that bind specifically to each, and this study highlights the usefulness of photo-affinity labelling in target identification for structure-based drug discovery. The overall architecture of the FoF1-ATP synthase is well conserved among different species. Although the α- and β-subunits of T. brucei F1 share 47% and 67% sequence identity respectively with their mammalian homologues, structural differences exist between them. In particular, the trypanosomatid α-subunit is cleaved into 20 KDa N-terminal and 44 KDa C-terminal chains and the recent low-resolution structure of T. brucei FoF1 obtained by cryo-electron tomography [62] suggests that this causes a major shift in the position of the α-subunit within the complex with respect to the β-subunit. Crucially, as the catalytic and regulatory binding sites are positioned at the interfaces between α- and β-subunits this is likely to have an effect on the overall shapes and sizes of those sites and a high-resolution structure is warranted to facilitate future structure-based drug design. The essentiality of the mitochondrial respiration complexes in human hosts may make the FoF1-ATP synthase (complex V) appear an unattractive target despite the possibility of significant differences between human and trypanosomatid homologues. However, the mitochondrial complexes are being investigated as targets in other diseases. For example, complex I-targeting acetogenins have shown promise as anti-cancer agents in preclinical studies [25], while complex V-targeting Bedaquiline has also recently been approved for the treatment of tuberculosis [72]. The FoF1 ATP synthase has been shown to be essential in both PF [61] and BSF [73,74] T. brucei, although its role in each parasite form differs. In the procyclic form (and in mammalian hosts), it is chiefly involved in oxidative phosphorylation whereby mitochondrial complexes III and IV generate a mitochondrial membrane potential (Δψm) by exporting protons from the mitochondrion during the electron transport chain (ETC), and mitochondrial complex V (the FoF1 ATP synthase) generates ATP while pumping protons back in. Furthermore, the Δψm is used for other functions, such as protein trafficking [75] and tRNA import [76]. BSF T. brucei lacks much of the oxidative phosphorylation machinery (namely, complexes III and IV, [77]) as it generates sufficient ATP through glycolysis [78] and generates the essential Δψm by running FoF1-ATP synthase in reverse, hydrolysing ATP instead of producing it [73,79,80]. The difference in function between mammalian and BSF T. brucei FoF1 ATP synthase (ATP production versus mitochondrial membrane polarisation respectively) could be exploited for the creation of synergistic drug combinations that do not affect the mammalian hosts. PF double marker strain 29–13 were maintained as reported previously [81] in full growth medium (SDM-79 medium supplemented with 10% foetal bovine serum (FBS) (Gibco), 2 g/L sodium bicarbonate, 7. 5 mg/L haemin, 15 μg/mL neomycin (G418) and 50 μg/mL hygromycin) at 27°C with 5% CO2. For experiments using low-glucose growth medium, PFs were adapted to glucose-free SDM-79 supplemented with 10% FBS, 2 g/L sodium bicarbonate, 7. 5 mg/L haemin and maintained as above. ATP production assays were carried out in duplicate, and three independent assays were conducted. Cells were digitonin-permeabilised as previously described [55]. Briefly, cells were washed in PBS and permeabilised for 5 min at a density of 1x108 cells/mL in SoTE buffer (20 mM Tris. HCl (pH 7. 5), 2 mM EDTA pH 7. 5,600 mM sorbitol) supplemented with 0. 015% (w/v) digitonin. Permeabilised cells were pelleted at 5000 g for 3 min at 4°C and suspended to 8. 4x107 cells/mL in ATP assay buffer (20 mM Tris. HCl (pH 7. 4), 15 mM KH2PO4,0. 6 M sorbitol, 10 mM MgSO4,2. 5 mg/mL fatty acid-free BSA. ATP production assays were carried out in 75 μl volumes in 96-well plates with a final PF T. brucei cell density of 6. 7 x 107 cells/mL as previously reported [55]. Unless otherwise stated, inhibitors were used at high concentration (exceeding their EC90 values in live cells) because plating density was >1000-fold higher than with live cell EC50 determination and to ensure that the target was fully inhibited. B-THP-Ts, antimycin A (AA) and oligomycin A (OA) were used at 200 μM, Dinitrophenol (DNP) at 1 mM, malonate at 5 mM, rotenone at 100 μM and salicylhydroxamic acid (SHAM) at 100 μM. 59. 25 μL digitonin-permeabilised cells (prepared above) were added to 0. 75 μl inhibitor dissolved in DMSO (or 0. 75 μL DMSO without inhibitor for no inhibitor control) and incubated for 10 min to allow inhibitors to take effect. Substrate was added to 2. 5 mM and incubated for 10 min. Reactions were started with the addition of ADP to 60 μM and incubated at RT for 30 min. Reactions were quenched by the addition of 1. 75 μL of 60% perchloric acid and the plate was incubated on ice for 30 min. The acid was then neutralised with 11. 5 μL of 1 M KOH. ATP production was quantified using ATP bioluminescence assay kit CLS II (Roche). 10–20 μL from each well was transferred in duplicate to a black 96-well plate. Volumes were made to 50 μL with 0. 5 M Tris-acetate pH 7. 75. 50 μL luciferase was added to each well, mixed, and precise luminescence was recorded with a Spectramax Gemini XPD (Molecular Devices) set to high sensitivity. Background luminescence was subtracted from each well and ATP levels calculated relative to the uninhibited controls. Effects of compounds on proline metabolism were determined using a methodology adapted from Podolec et al [85]. Briefly, glucose-free-conditioned PF T. brucei cells were washed into buffered PBS and incubated in 200 μl volumes with inhibitors for 10 min. 3 mM proline was added as the sole carbon source and cells were incubated for 2 h, after which cultures were spiked with 1 mM citric acid (as internal standard) and cells were pelleted. Supernatants containing metabolic end-products and citrate standard were analysed to evaluate effects on proline metabolism (see below), while the ATP content of cells was determined as follows. Cells were resuspended in 75 μL PBS and lysed with the addition of 1. 75 μL 60% perchloric acid. Reactions were incubated on ice for 30 min, then neutralised with 11. 5 μL 1 M KOH. ATP content was determined as above. F1 α- and β-subunits were purified using a similar procedure to that previously described for TAP-tagged subunits [61]. Briefly, PF T. brucei cells from 50 mL culture endogenously expressing GFP-PTP-tagged F1 subunits were photo-affinity probed with 50 μM compound 3 or compound 1 as outlined above, except cells were lysed in IP buffer (50 mM Tris. HCl (pH 7. 6), 150 mM NaCl, 1% TX-100) supplemented with ‘Complete’ EDTA-free protease inhibitors (Roche). GFP-PTP-tagged proteins were solubilsed for 1 h with end-over-end agitation at 4°C and insoluble material pelleted at 16000 g for 10 min at 4°C. The supernatant was applied to 20 μL IgG-sepharose (GE Healthcare Life Sciences) and incubated with end-over-end agitation for 1 h at 4°C. Beads were pelleted at 16000 g for 1 min and non-bound material was discarded. Beads were washed 4 times in 1 mL IP buffer to remove contaminating proteins. Compound 1 was docked into F1 from Bos taurus (PBD entry 1BMF [63]) and Saccharomyces cerevisiae (PDB entry 2WPD [68]) using AutoDock Tools [90] and AutoDock Vina [69] as per the user manuals. Briefly, the energy-minimised 3-dimensional coordinates of compound 1 were generated with ChemBio 3D (Perkin Elmer) and converted to PDBQT format with AutoDockTools. All non-protein atoms (e. g. , waters, nucleotides and metals) were removed from the F1 PDB coordinates and hydrogens were added with AutoDockTools. Initially compound 1 was docked into the entire F1 complex using a grid-box of dimensions 126 x 126 x 126 Å containing the entire complex and an exhaustiveness of 48 to account for the large search area. Compound 1 was next docked into the ATP-binding sites of each α- and β-subunit using a grid-box of dimensions 17 x 15 x 23 Å covering the binding site. Positions of docked compound 1 were evaluated using Pymol (Schrodinger).

Millions of people are at risk of developing African sleeping sickness through infection with the parasite Trypanosoma brucei, which is fatal if untreated. Current therapies are antiquated and inadequate, requiring hospitalization to deliver them through complex regimens, thus new effective, cheap and easy to administer therapies are urgently required. We recently reported the generation of a new compound that is selectively active against the parasite responsible for the disease, but its mechanism of action was unknown. This body of work identifies the specific target within the parasite as the essential FoF1-ATP synthase (mitochondrial complex V), first by analysing the proteins bound to a tagged analogue of our active compound (in a process known as photo-affinity labelling), and secondly by determining the effects of our active compound on cellular metabolism. With the target now identified work can begin to improve inhibitor potency and selectivity with the aid of rational drug design and structural optimisation towards future drug discovery.

Abstract Introduction Results and discussion Materials and methods

b vitamins medicine and health sciences chemical compounds biotin metabolic processes parasitic protozoans organic compounds oxidative phosphorylation protozoans alkynes amino acids mitochondria energy-producing organelles bioenergetics pharmacology cellular structures and organelles hydrocarbons proteins chemistry vitamins cyclic amino acids proline drug discovery azides biochemistry trypanosoma cell biology organic chemistry drug research and development biology and life sciences physical sciences trypanosoma brucei gambiense metabolism organisms

2017

Photo-affinity labelling and biochemical analyses identify the target of trypanocidal simplified natural product analogues

APOBEC3s (A3s) are potent restriction factors of human immunodeficiency virus type 1/simian immunodeficiency viruses (HIV-1/SIV), and can repress cross-species transmissions of lentiviruses. HIV-1 originated from a zoonotic infection of SIV of chimpanzee (SIVcpz) to humans. However, the impact of human A3s on the replication of SIVcpz remains unclear. By using novel SIVcpz reporter viruses, we identified that human APOBEC3B (A3B) and APOBEC3H (A3H) haplotype II strongly reduced the infectivity of SIVcpz, because both of them are resistant to SIVcpz Vifs. We further demonstrated that human A3H inhibited SIVcpz by deaminase dependent as well independent mechanisms. In addition, other stably expressed human A3H haplotypes and splice variants showed strong antiviral activity against SIVcpz. Moreover, most SIV and HIV lineage Vif proteins could degrade chimpanzee A3H, but no Vifs from SIVcpz and SIV of gorilla (SIVgor) lineages antagonized human A3H haplotype II. Expression of human A3H hapII in human T cells efficiently blocked the spreading replication of SIVcpz. The spreading replication of SIVcpz was also restricted by stable A3H in human PBMCs. Thus, we speculate that stably expressed human A3H protects humans against the cross-species transmission of SIVcpz and that SIVcpz spillover to humans may have started in individuals that harbor haplotypes of unstable A3H proteins. Simian immunodeficiency virus (SIV) naturally infects many species of African Old-World monkeys, such as African green monkeys, mandrills and red-capped mangabey [1,2]. However, these viruses appear to be nonpathogenic in their natural hosts [2,3]. Chimpanzees (cpz), which are the evolutionarily closest extant primate to Homo sapiens, are infected by SIVcpz [4]. The common chimpanzee includes four subspecies, only two of which, Pan troglodytes troglodytes (Ptt) and Pan troglodytes schweinfurthii (Pts), are infected by SIVcpz (SIVcpzPtt and SIVcpzPts, respectively) [4]. Genome analysis of SIVcpz indicates that SIVcpz originates from the cross-species transmission and recombination of three different SIV strains: SIVrcm from the red-capped mangabey (rcm), SIVgsn/mus/mon from the greater-spot-nosed (gsn), mustached (mus), and mona monkeys (mon), respectively, and a currently unidentified SIV [5–7]. SIVcpzPts is thought to be the origin of SIVcpzPtt after intra-chimpanzee transmission [5]. SIVcpz is of particular interest because it is the ancestor of human immunodeficiency virus (HIV) -1. HIV-1 M and N groups originated from zoonotic transmission of SIVcpzPtt from west-central Africa [8,9]. Additionally, recent studies indicate that SIVgor from gorillas (gor) is the origin of HIV-1 groups O and P [10,11]. The HIV-1 M group is the pandemic virus, whereas viruses of groups N and P are only found in a few infected individuals [12,13]. The HIV-1 O group is mainly distributed in west-central Africa and has a low prevalence rate (less than 1% of global HIV-1 infections) [14,15]. The other HIV lentivirus, HIV-2, resulted from cross-species transmission of SIV from the sooty mangabey monkey (SIVsmm) [14]. Human intrinsic cellular antiviral factors may have direct relevance for the zoonotic infection of humans and the human-to-human spread of SIVs. Several restriction factors have been identified that repress lentiviral replication [16–18]. The family of human APOBEC3 (A3) restriction factors is formed by seven different proteins, A3A–D and A3F–H. Virion encapsidated APOBEC3D (A3D), A3F, A3G, and A3H inhibit HIV-1 that lacks the gene vif (HIV-1Δvif) by deaminating cytidines in the viral single-stranded DNA that is generated during reverse transcription, thereby introducing G-to-A hypermutations in the coding strand [19]. To achieve productive infections, lentiviral Vif proteins directly interact with A3s and recruit them to an E3 ubiquitin ligase complex to induce A3 degradation by the proteasome [20–22]. Several studies have investigated how A3G serves as a barrier for cross-species transmission of lentiviruses [23–25]. Human A3H represents the most evolutionarily divergent A3 gene; it includes seven haplotypes and several splice variants [26–28]. The protein stability of human A3H is one determinant of its antiviral activity [29–31]. The human A3H haplotype (hap) II, in contrast to A3G, is only sensitive to specific HIV-1 Vifs and adaptation of HIV-1 Vif to A3H hapII has been described [32–35]. Thus, the polymorphism of human A3H has relevance for HIV-1 infection and AIDS progression [36,37]. To investigate how human A3s may affect the replication of SIVcpz, we generated novel luciferase reporter viruses based on two SIVcpz strains (SIVcpzPtsTAN1 and SIVcpzPttMB897). This system revealed that SIVcpz transmission to humans may have been significantly affected by the presence of stable A3H. To test SIVcpz, we first generated SIVcpz nanoluciferase (NLuc) reporter viruses using two SIV strains (SIVcpzPttMB897 and SIVcpzPtsTAN1; S1 Fig). SIVcpzPttMB897 was isolated from wild chimpanzee (Pan troglodytes troglodytes) in southern Cameroon in 2007 [9,38], and this strain is regarded as the ancestor of the pandemic HIV-1 M group [14]. SIVcpzPtsTAN1 was derived from chimpanzee subspecies Pan troglodytes schweinfurthii, and this strain does not cause sustained infections of humans [39]. The SIVcpz reporter constructs were generated by replacing most of the nef gene with NLuc. Additionally, the vif gene of SIVcpz was inactivated (S1 Fig). SIVcpz-NLuc reporter viruses pseudotyped with the glycoprotein of the vesicular stomatitis virus (VSV-G) were produced by plasmid transfection of 293T cells. When infected with the VSV-G pseudotyped viruses, SIVcpzPttMB897-NLuc and SIVcpzPtsTAN1-NLuc, 293T cells showed high luciferase counts while very low nanoluciferase activity was detected when the viruses were not VSV-G pseudotyped (Fig 1). The luciferase activity of SIVcpzPtsTAN1-NLuc was around 10-fold less than SIVcpzPttMB897-NLuc, even when equal amounts of virions normalized for reverse transcription activity were used for infection (Fig 1). Thus, these two novel SIVcpz reporter viruses transmitted the luciferase enzyme activity via glycoprotein-dependent infection. Four SIV-luciferase reporter viruses based on SIV of macaques (SIVmac), African green monkeys (SIVagm) and chimpanzees (SIVcpzPts, and SIVcpzPtt) were used to investigate the antiviral activity of chimpanzee A3s. We found that cpzA3C, D, F, G, and H reduced the infectivity of SIVmacΔvif (Fig 2A). SIVmac Vif fully antagonized restrictions of cpzA3C, G, and H, and to a large extent overcame cpzA3F, but it did not inhibit the restriction of cpzA3D (Fig 2A). Chimpanzee A3s showed a similar restriction pattern against SIVagmΔvif, but SIVagm Vif only abolished the restriction of cpzA3C and partly inhibited the restriction of cpzA3H. Even in the presence of SIVagm Vif, cpzA3D, F, and G significantly reduced the infectivity of SIVagm (Fig 2A). The expression of the cpzA3C, F, G, and H was detectable by immunoblotting using HA-tagged specific antibodies, while cpzA3D was not detectable using our immunoblotting system (S2A and S2C Fig). In the absence of Vifs, cpzA3C reduced the infectivity of SIVmac and SIVagm by 5–10 fold and weakly inhibited SIVcpzPttMB897 and SIVcpzPtsTAN1 by 1–2 fold (Fig 2C and 2D). cpzA3D, F, and H inhibited SIVcpzΔvif by 10–15-fold, while cpzA3G reduced the infectivity of both SIVcpzPttMB897Δvif and SIVcpzPtsTAN1Δvif to an even greater extent (Fig 2C and 2D). SIVcpzPttMB897 and SIVcpzPtsTAN1 Vifs are able to counteract all cpzA3s, but not all in cases to the same level, e. g. cpzA3F, the full viral infectivity (vector control) was restored, consistent with previous study [40] (Fig 2C and 2D). Taken together, these data indicate that chimpanzee A3s, such as cpzA3D and cpzA3G, can protect chimpanzees from infection with SIVs of rhesus macaques and African green monkeys. Our data and a previous study indicate that chimpanzee A3s, especially cpzA3D, play an important role as a barrier to cross-species transmission of SIVs from monkeys to chimpanzees (Fig 2A and 2B) [40]. Next, we asked whether human A3s (hA3s) form a barrier to SIVcpz infection of humans. Thus, we analyzed the anti-SIV activity of human A3s by using the four SIV reporter systems. Similar to chimpanzee A3s, hA3C, D, F, G, and H (including hapI and hapII) inhibited SIVmacΔvif and SIVagmΔvif infections, and SIVmac Vif abolished most of these restrictions but was only weakly active against hA3D (Fig 3A). SIVagm Vif only significantly overcame the restriction of hA3C, hA3H hapI, and hA3H hapII (Fig 3B). hA3D, hA3F and hA3G displayed resistance to SIVagm Vif counteraction, indicating that these three factors may protect humans against infection by SIVagm (Fig 3B). Consistent with a previous study, hA3B strongly reduced the infectivity of SIVmac and SIVagm regardless of Vif [41]. However, hA3A showed only a low-level inhibition of SIVmac and SIVagm and this restriction was resistant to both SIVmac and SIVagm Vifs (Fig 3A and 3B). The expression of hA3s in transfected 293T cells was detected by immunoblotting (S2B and S2C Fig). In the absence of Vif, hA3D, F, and G reduced the infectivity of SIVcpz, while Vif proteins from both SIVcpzPttMB897 and SIVcpzPtsTAN1 antagonized these hA3s (Fig 3C and 3D). In contrast to the experiments with SIVmacΔvif and SIVagmΔvif, no antiviral activity of hA3C was seen against SIVcpzΔvif (Fig 3C and 3D). Interestingly, two human A3s (hA3B and hA3H hapII) showed strong inhibition of SIVcpz regardless of Vif expression (Fig 3C and 3D). While hA3H is expressed in primary CD4+ lymphocytes and has the ability to inhibit HIV-1 [26,35], hA3B is not found in HIV target cells [35]. Together our data indicate that hA3H hapII may block SIVcpz cross-species transmission to humans. To characterize the interaction between SIVcpz and A3H in more detail, the incorporation of cpzA3H and hA3H hapII into SIVcpz viral particles was analyzed by immunoblotting. In the absence of Vif, both cpzA3H and hA3H hapII were encapsidated into SIVcpzPttMB897 and SIVcpzPtsTAN1 (S2D Fig). Vif from both SIVcpz strains reduced the cpzA3H protein level in the cell lysate by depletion and decreased the cpzA3H incorporation into viral particles (S2D Fig). In agreement with the infectivity data of SIVcpz with cpzA3H (Fig 2C and 2D), SIVcpzPtt Vif was more active against cpzA3H than SIVcpzPts Vif [40]. However, the steady-state expression and particle encapsidation of hA3H hapII did not change in the presence of SIVcpz Vif, which corresponds with hA3H’s hapII antiviral activity against wild-type SIVcpz (Fig 3C and 3D and S2D Fig). Furthermore, we investigated whether the cytidine deaminase activity is required for hA3H hapII inhibiting SIVcpz. We introduced the E56A mutation in the cytidine deaminase domain of hA3H hapII, which was previously reported to completely abolish the protein’s deaminase activity [42]. The E56A mutant of A3H lost significantly anti-viral activity compared with wild-type hA3H hapII, but remained a 10-fold inhibitory activity against SIVcpzΔvif (Fig 4A). Next we analyzed the presence of G-to-A mutations indicative of A3 deamination in the viral genome by amplifying a 700-bp fragment of the viral genome 12 h post-infection. Viruses prepared without co-expression of A3 showed no detectable G-to-A mutations. However, in the presence of hA3G, we found a hypermutation rate of around 2. 8% in the SIVcpz genome (Fig 4B). Viruses made in the presence of hA3H hapII contained a mutation rate of around 0. 9%, while hA3H hapII E56A did not edit the SIVcpz genome (Fig 4B). The sequence plots, confirmed that hA3G preferred GG motif (mutated G is underlined, which is CC in the deaminated minus strand), while hA3H hapII mutated a GA motif (TC in the minus strand) predominantly (Fig 4B), which is consistent to previous studies [42–44]. Taken together, these data indicate that hA3H hapII inhibits SIVcpz by both deaminase dependent and independent mechanisms. To further characterize the level of anti-SIVcpz activity mediated by hA3H, different amounts (5–200 ng) of hA3H hapI or hA3H hapII expression plasmids were co-transfected with SIVcpzPttMB897 wild-type or Δvif reporter constructs and the viral infectivities were determined. The results indicate that the anti-SIVcpz activity of hA3H hapII increased with the dose of transfected hA3H hapII plasmid regardless of Vif expression (Fig 5A and 5B). Even a low level of hA3H hapII (5 ng) displayed around 10-fold inhibition of SIVcpzPttMB897Δvif and Vif was not able to overcome this restriction. We also found that hA3H hapI showed around 20-fold inhibition of SIVcpzPttMB897, when 200 ng hA3H hapI expression plasmid was transfected (Fig 5A). 100 ng hA3H hapI and 10 ng hA3H hapII plasmids displayed similar protein expression levels, and they showed similar strength of inhibition of SIVcpzPttMB897 (Red box in Fig 5A and 5B), which is consistent with previous studies [45,46]. These results indicate that the protein expression level of A3H is one of the key determinants for its antiviral activity. Human A3H has seven haplotypes and several splice variants, and the A3H protein stability determines the antiviral activity [29–31]. In the absence of Vif, cpzA3H, hA3H hapII, hA3H hapV, hA3H hapVII, and four hA3H hapII splice variants (SV182, SV183, and SV200) strongly inhibited SIVcpzPttMB897 (Fig 5C). However, Vif only counteracted the restriction of cpzA3H and was inactive against all the tested hA3H variants (Fig 5C). Corresponding immunoblotting results of lysates of the transfected cells confirmed that SIVcpzPttMB897 Vif only reduced the protein level of cpzA3H and protein levels of the hA3Hs were not changed by Vif co-expression (Fig 5D). To learn more about the strength of hA3H’s antiviral activity, the spreading replication of SIVcpz in human T cells (SupT11) that stably expressed hA3H hapII [47] was investigated. To facilitate replication of CCR5-tropic SIVcpz, we modified SupT11 cells to express human CCR5 (S3A and S3B Fig). The spreading replication was tested with full-length unmodified viruses (SIVcpzPttMB897, SIVcpzPttGab1, and SIVcpzPtsTAN1). SIVcpzPtsTAN1 did not replicate in the SupT11-hA3H hapII and SupT11-vector cells, regardless of the input of virus (1 ng reverse transcriptase (RT) activity or 50 ng RT) for the initial infection (Fig 5E and S3C Fig). Both SIVcpzPttMB897 and SIVcpzPttGab1 replicated efficiently in SupT11-vector cells, while no virus spreading was observed in SupT11-hA3H hapII cells (Fig 5E). These data indicate that hA3H hapII is a strong inhibitor of infection of SIVcpz in human T cells. We conclude, therefore, that stably expressed hA3H variants are Vif-resistant restriction factors of SIVcpz. Both cpzA3H and hA3H hapII displayed strong anti-SIVcpz activity, while they had different sensitivities to SIVcpz Vif counteraction (Figs 2C, 2D, 3C and 3D). One recent study demonstrated that residue 97 of cpzA3H and hA3H hapII determines the sensitivity to HIV-1 clone NL4-3 Vif [48]. Thus, we tested whether residue 97 would also be important for SIVcpz Vif inhibition of A3H. The Q97K and K97Q mutations were introduced into cpzA3H and hA3H hapII, respectively. The results showed that cpzA3H Q97K and hA3H hapII K97Q retained their anti-SIVcpz activity in the absence of Vif (Fig 6A and 6C). While SIVcpzPttMB897 Vif almost fully overcame the inhibition of wild-type cpzA3H, it only partially antagonized cpzA3H Q97K and, similarly, SIVcpzPtsTAN1 Vif did not counteract cpzA3H Q97K (Fig 6A and 6C). Additionally, hA3H hapII showed resistance to SIVcpzPttMB897 Vif, but this resistance was partially lost when the K97Q mutation was introduced (Fig 6A). In contrast to SIVcpzPttMB897 Vif, both wild-type hA3H hapII and its K97Q mutant showed resistance to SIVcpzPtsTAN1 Vif (Fig 6C). Furthermore, we analyzed the protein expression level of these A3H mutants in the presence of SIVcpz Vifs. hA3H E121K was included as a control mutant that could not be degraded by HIV-1 Vif [48,49]. SIVcpzPttMB897 Vif slightly reduced the protein level of cpzA3H Q97K compared to the no-Vif control, which is consistent with the infectivity data (Fig 6A and 6B). SIVcpzPtsTAN1 Vif did not affect the expression of cpzA3H Q97K (Fig 6D). hA3H hapII K97Q was depleted by co-expression of SIVcpzPttMB897 Vif, while the presence of SIVcpzPtsTAN1 Vif did not affect hA3H protein levels (Fig 6B and 6D). We conclude that Vifs from SIVcpzPttMB897 and SIVcpzPtsTAN1 have distinct interaction properties with hA3H hapII. To find out how diverse A3H is in chimpanzees, we analyzed the deep-sequencing reads from the recent Great Ape Genome Project [50]. We mapped reads to the hA3H region (hg19, chr22: 39496284–39498576) and the exons of A3H were isolated. The coding regions of A3H from 61 chimpanzees (10 Pan troglodytes ellioti, Pte; 16 Pan troglodytes schweinfurthii, Pts; 22 Pan troglodytes troglodytes, Ptt; 13 Pan troglodytes verus, Ptv) were analyzed. We found four single-nucleotide polymorphisms (SNPs) of cpzA3H (nucleotide positions 50,359,402, and 481; Table 1, Fig 7A and S4 Fig). Two of them (SNP_50 and SNP_359) were only present in Ptv with an overall frequency of 6. 5% and 9. 8%, respectively. SNP_402 was only found in Pts with a frequency of 9%. However, SNP_481 was detected in Pte, Pts, and Ptt with a frequency of 8. 2%, 19. 6%, and 34. 4%, respectively. The detailed SNP and zygosity information is described in Table 1. These four SNPs including the reference cpzA3H were named from haplotype I (hapI) to haplotype V (hapV) (Fig 7A). In addition, we performed a phylogenetic analysis of A3H from apes (rhesus macaque A3H was also included). The results showed that gibbon, rhesus macaque, and orangutan A3H were classified into one clade. Gorilla A3H formed a separate clade, and human and chimpanzee A3H were classified into two clades, respectively (Fig 7B). Bonobo A3H was classified into the clade of chimpanzee A3H. The protein stability differs in human A3H haplotypes and it is one of the determinants of its antiviral activity [29–31]. Thus, the expression of five cpzA3H haplotypes in 293T cells was tested by immunoblotting. All cpzA3H haplotypes produce stable proteins and had similar expression levels (Fig 7C). Moreover, these five cpzA3H haplotypes displayed similar anti-SIVcpz activities and were all sensitive to Vifs from SIVcpz lineages (Fig 7D and 7E). These data indicate that the polymorphism of cpzA3H does not affect its protein stability or antiviral activity. There have been four independent transmissions from different SIVcpz/gor strains to the human population, which caused HIV-1 groups M, N, O, and P, respectively [11,14]. Thus, we tested the sensitivity of cpzA3H and hA3H hapII to Vifs from several SIVcpz/HIV-1 lineages. The immunoblots of co-expressing cells indicated that cpzA3H was depleted by all the tested SIVcpz Vifs, and it was also depleted by Vifs from HIV-1 B-LAI (M group), N-116, and O-127, but was not degraded by HIV-1 F-1 Vif (Fig 8A). hA3H hapII was resistant to depletion of all SIVcpz Vifs tested, including SIVgor Vif (Fig 8A). However, HIV-1 B-LAI, F-1, and N-116 Vifs induced the degradation of hA3H hapII (Fig 8A). Unexpectedly, HIV-1 O-127 Vif, which protein expression was not detectable was inactive against hA3H hapII (Fig 8A). By testing chimeras of SIVcpzPttMB897 and HIV-1 LAI Vif, we identified that the Vif N-terminal region (residues 40–70) is essential for hA3H hapII depletion (Fig 8B and 8C). A previous study described the importance of HIV-1 Vif residues F39 and H48 for antagonism of hA3H hapII [33]. F39 is present in SIVcpzPttMB897 Vif, but at the 48 position, an asparagine (N) is found (Fig 8D). However, introducing an N48H mutation (construct M1) in SIVcpzPttMB897 Vif did not promote degradation of hA3H hapII (Fig 8E). However, the local area of residue 48 of SIVcpzPttMB897 Vif was important as an additional mutation revealed that changing residues 47EN48 to 47PH48 (construct M2) facilitated hA3H hapII depletion (Fig 8E). Furthermore, a replication-competent SIVcpzPttMB897_EN-PH with this substitution showed spreading replication in hA3H hapII-containing SupT11 cells (Fig 8F). A previous study showed that HIV-1 Vif from a homozygous hA3H haplotype II patient had greater activity against hA3H hapII compared to other laboratory HIV-1 Vifs, which correlated with the presence of four amino acid substitutions (60GDAK63 to 60EKGE63) [32]. This substitution was introduced into SIVcpzPttMB897 Vif and led to enhanced hA3H hapII depletion (Fig 8E, Vif M2 compared to M3 and M4). Based on a recent HIV-1 Vif-hA3H hapII co-structure model [49], the co-structure of SIVcpzPttMB897 Vif-hA3H hapII was modeled. From the structure, we found that residues 47EN48 and 60GDAK63 of SIVcpzPttMB897 Vif were in close contact with hA3H hapII (Fig 8G). Both regions are diverse in Vifs from distinct SIVcpz and HIV-1 lineages (Fig 8H). Next, we tested the replication of SIVcpz in human PBMCs from donors with different hA3H genotypes. We identified three donors who were homozygous for A3H hapI, hapIV and hapII, respectively. The protein expression of A3H in stimulated PBMCs was detected by immunoblotting, demonstrating highest protein levels in the PBMCs of hapII, moderate levels in PBMCs of hapI and weak levels in cells of hapIV (Fig 9A). However, the A3G expression in these PBMC was identical (Fig 9A). The viral replication experiments indicated that SIVcpzPttMB897 replicated fastest in PBMCs from the donor with haplotype IV, and moderately in PBMCs from the donor with haplotype I (Fig 9B and 9C). However, the replication of SIVcpzPttMB897 was inhibited in PBMCs from donor with haplotype II (Fig 9B and 9C). In summary, we speculate that stable hA3H forms a barrier for zoonotic transmission of SIVcpz to humans and Vif adaptation to stable hA3H would be needed for high-level infection of humans with this haplotype (Fig 10). SIVcpz originated from the cross-species transmission and recombination of three different SIVs [5,6]. After lentiviral transmission to a new host that differs in one or many A3 proteins, Vif adaptation is expected at the interface of both proteins [25,51]. In our study, all tested SIVcpz Vifs had the ability to counteract cpzA3Hs (Figs 2 and 7). Lucie Etienne et al. found that SIVrcm Vif acts like SIVcpz Vifs and can neutralize cpzA3H, while SIVmus Vif could not antagonize the restriction of cpzA3H [40]. Overcoming the restriction of cpzA3H may be one explanation for SIVcpz selectively acquiring the 5’ region (including vif) from SIVrcm during recombination, and acquiring the 3’ region (including vpu, env, and nef) from SIVgsn/SIVmus/SIVmon may have facilitated the counteraction of other restriction factors, such as Tetherin or Serinc3/5 [52–54]. Here, we also found that cpzA3D, F, and G were resistant to SIVagm Vif and similarly, cpzA3D was resistant to SIVmac Vif, confirming a previous report [40]. This observation indicates that cpzA3D and cpzA3G can protect chimpanzees from infection with SIVs of rhesus macaques and African green monkeys. On the other hand, SIVcpz Vifs could counteract all the tested cpzA3s. However, cpzA3F showed a moderate level of resistance to degradation induced by SIVcpz Vif (Fig 2C and 2D), possible suggesting that cpzA3F may provide some repression of SIVcpz infection. Human and chimpanzee A3D, F, and G display a similar sensitivity to SIVcpz Vif, indicating that the inhibitory activity against cpzA3s by SIVcpz may be a prerequisite for the cross-species transmission of SIVcpz to the human population. Here, we found that hA3C and hA3H hapI display a strong restriction against SIVmacΔvif and SIVagmΔvif; however, no antiviral activity was observed against SIVcpzΔvif (Fig 3) or HIV-1Δvif [26,41,55]. These data suggest that the viral sensitivity to hA3C and hA3H hapI was lost in the evolution of SIV lineages and not during the evolution of HIV-1. We cannot determine whether this happens during the creation of SIVcpz due to the lack of information regarding the antiviral activity of hA3C and hA3H hapI against SIVrcm/SIVgsn/SIVmus/SIVmon. We speculate that some SIVs similar to HIV-1 have the ability to escape hA3C and hA3H hapI restriction by a Vif-independent mechanism [55]. cpzA3H appears to be much less polymorphic than hA3H. However, A3F and A3G in chimpanzee are more diverse than the human orthologs [40]. Although our chimpanzee sample number was limited (61 chimpanzees), the results suggest that cpzA3H is relatively conserved among chimpanzees. Residues 15 and 105 of hA3H determine the protein stability and anti-viral activity [28]. However, no variability was identified at these two positions in cpzA3H, which is in agreement with the comparable protein stability and anti-viral activity of the currently recognized five cpzA3H haplotypes (Fig 7). Vifs of different SIVcpz isolates degrade all haplotypes of cpzA3H indicating that cpzA3H is not a restriction factor for inter-subspecies transmission of SIVcpz. Compared to cpzA3H, human A3H is more diverse and includes seven haplotypes and several splice variants [28,30,31]. In our study, the stably expressed hA3H haplotypes were identified as Vif-resistant inhibitors against SIVcpz, indicating that these active hA3Hs are strong barriers to prevent SIVcpz infection of humans. After the zoonotic transmission of SIVcpz to humans expressing unstable A3H haplotypes, the very early human-to-human transmission was likely to be severely affected by humans expressing the A3H haplotypes with a stable protein (Figs 9 and 10). A possible mutation that would enhance SIVcpz Vif adaptation was investigated by replacing residues 47EN48 of SIVcpzPttMB897 Vif with 47PH48 (Fig 8). It is possible—but unlikely—that there are currently not identified viruses circulating in chimpanzees with vif genes encoding 47PH48 residues enhancing SIVcpz cross-species transmission to humans. In fact, the 47PH48 motif is also found in HIV-1 patients who harbor hA3H hapII [32,34]. The frequency of active hA3H varies significantly between populations, with the highest frequency in Africans (around 50% harbor stable A3H) [28,30]. This observation may be the result of a selective sweep caused by exposure to a retrovirus such as SIV or HTLV or other A3H-sensitive pathogens [56,57]. Several previous studies described a positive and balancing selection of human and chimpanzee MHC loci, caused by HIV-1/SIVcpz infections [56,58–60]. In addition to hA3H hapII, human tetherin is also a strong barrier against SIVcpz transmission to humans. SIVcpz Nef recognizes the cytoplasmic domain of chimpanzee tetherin and inhibits its restriction, but it cannot overcome the restriction of human tetherin due to a deletion in this domain [54]. However, the virus adapts to this restriction by regaining Vpu-mediated inhibition of tetherin after transmission of SIVcpz to humans [54]. In fact, other unknown restriction factors may exist to control the cross-transmission of SIVcpz to humans. For example, a recent study found that introducing a M30R/K mutation in the Gag matrix could enhance SIVcpz replication fitness in human tonsil explant cultures [38]. Overall, our study suggests that the stable active human A3Hs can protect humans against the spillover of SIVcpz, and SIVcpz cross-species transmission to humans may have started in those that harbored unstable A3H proteins. Chimpanzee APOBEC3 (A3) expression plasmids (A3D, A3F, A3G and A3H) were provided by Michael Emerman [40], chimpanzee A3C plasmid was described recently [61]. Human A3s (A3A-A3H) were expressed by PTR600 vector with a carboxyl-terminal triple hemaggutinin (HA) tag [33]. Human A3H haplotype V, VII, splice variants and E56A of haplotype II expression plasmids with a carboxyl-terminal flag tag were provided by Viviana Simon [31]. Human A3H haplotype II with an N terminal HA tag was re-cloned into PTR600 vector by using standard PCR. All human and chimpanzee A3H mutants were generated by site direct mutagenesis and confirmed by sequencing. The MLV packaging construct pHIT60 was kindly provided by Jonathan Stoye, which encodes the gag-pol of MoMLV [62]. The Plasmid of pBABE. CCR5 that encodes human CCR5 was obtained from NIH AIDSREPOSITORY [63]. SIVmac-Luc (R-E-), SIVmac-Luc (R-E-) Δvif and SIVagm-Luc (R-E-) and SIVagm-Luc (R-E-) Δvif were provided by N. R. Landau [64]. The replication competent SIVcpzPtt clones MB897, EK505, Gab1 were kindly provided by Frank Kirchhoff [38,65]. SIVcpzPts clones TAN1. 910 and TAN2. 69 and SIVgor clone CP2139 were obtained from NIH AIDSREPOSITORY [10,66]. To generate the Nanoluciferase reporter virus of SIVcpzPttMB897, the nef gene was replaced (the first 7 amino acids of Nef remained) by nanoluciferase gene by overlapping PCR using NheI and XhoI restriction sites. Additionally, two stop codons were inserted amino-terminal of Vif (amino acid position 40 and 44) by overlapping extension PCR using PshAI and NheI restriction sites. The same method was performed to create nanoluciferase reporter virus of SIVcpzPtsTAN1, and the restriction sites are shown in S1 Fig. Simply, the nef gene was replaced (the first 7 amino acids of Nef remained) by nanoluciferase gene by overlapping PCR using AclIII and XbaI restriction sites. Additionally, two stop codons were inserted at the amino-terminal of Vif (amino acid position 40 and 44) by overlapping extension PCR using PshAI and AclI restriction sites. All constructs were verified by sequencing analysis. To generate the SIV Vif expression plasmids, Vif fragments from the following molecular clones: SIVcpzPtt EK505 (DQ373065), Gab1 (X52154), MB897 (EF535994) and SIVcpzPts TAN1 (AF447763), TAN2 (DQ374657) and SIVgor CP2139 (FJ424866) were amplified and inserted into pCRV1 by EcoRI and NotI. Vif expression plasmids of HIV-1 LAI, F-1, N-116 and O-127 were provided by Viviana Simon [33,61]. All SIVcpzPttMB897 Vif mutants were generated by overlapping PCR and cloned into pCRV1 without any tag, verified by sequencing. HEK293T (293T, ATCC CRL-3216) cells were maintained in Dulbecco’s high-glucose modified Eagle’s medium (DMEM, Biochrom, Berlin, Germany) supplemented with 10% fetal bovine serum (FBS), 2 mM L-glutamine, penicillin (100 U/ml), and streptomycin (100 μg/ml). SupT11 cells containing empty control and hA3H hapII were kindly provided by Reuben S. Harris and cultured in RPMI supplemented with 10% fetal bovine serum (FBS), 2 mM L-glutamine, penicillin (100 U/ml), and streptomycin (100 μg/ml) [47]. SupT11 cells with expression of hCCR5 were generated by MLV transduction. Simply, 1x106 SupT11 cells were transduced by MLV vector (produced by transfecting pBABE. CCR5, pHIT60 and VSV-G expression plasmid into 293T cells). 3 days after transduction, the SupT11 cells were selected for 3 weeks by using 1 μg/ml puromycin. For producing the single round infection of SIV reporter virus, 3×105 293T cells in 24-well plates were co-transfected with 300 ng SIVmac-Luc (R-E-), or SIVagm-Luc, or SIVcpzPttMB897-NLuc, or SIVcpzPtsTAN1-NLuc; or the corresponding delta Vif versions, 30 ng human A3s or 200 ng chimpanzee A3s expression plasmids and 50 ng VSV-G (pMD. G), and pcDNA3. 1 (+) (Thermo Fisher Scientific) was used instead of A3 expression plasmids. Human A3s were expressed in plasmid PTR600, while chimpanzee A3s were expressed in plasmid pcDNA3. 1 (+). 30 ng of PTR600-human A3s constructs had comparable expression levels with 200 ng pcDNA3. 1 (+) -chimpanzee A3s plasmids. Transfections were performed by using Lipofectamine LTX (Thermo Fisher Scientific) according to manufacturer’s instruction. The viral supernatants were collected 48 h post transfection. The reverse transcriptase (RT) activities of viruses were quantified by using the Cavidi HS lenti RT kit (Cavidi Tech, Uppsala, Sweden). For SIVmac and SIVagm infections, 5×104 293T cells were seeded in 96-well plates one day before transduction, and 50 pg RT of viruses were used for infection. After 48 h, firefly luciferase activity was measured with Steady-Glo Luciferase system (Promega) according to the manufacturer’s instructions on a MicroLumat Plus luminometer (Berthold Detection Systems, Pforzheim, Germany). For SIVcpz-NLuc, we observed high nanoluciferase enzyme activity in cell supernatant of transfected cells. 293T cells in 96-well plates were infected with 20 pg of SIVcpzPttMB897-NLuc or SIVcpzPtsTAN1-NLuc. To eliminate the effect of contaminating nanoluciferase in the supernatant of virus producer cells, we changed the medium 8 h post infection. 48 h after transduction, the cells were carefully washed by PBS once, and the nanoluciferase activity was measured with Nano-Glo Luciferase system (Promega) on a MicroLumat Plus luminometer (Berthold Detection Systems). Each sample was analyzed in triplicates; the error bar of each triplicate was shown. Infections in which the VSV-G glycoprotein was omitted served as control for nanoluciferase enzyme background enzyme activity. 1 x 106 293T cells were infected with DNase I (Thermo Fisher, Germany) treated SIVcpzΔVif-Nluc produced in 293T cells together with hA3G, hA3H hapII, hA3H hapII E56A or pcDNA3. 1 (+). At 12 h post-infection, cells were washed with PBS, and DNA was isolated using a DNasy DNA insolation kit (Qiagen, Germany). A 700-bp fragment of the SIVcpz-Nluc (200-bp C terminal of env plus 500-bp nanoluciferase gene) was amplified using DreamTaq DNA polymerase (Thermo Fisher, Germany) with primers: 5’-attctccagtattggggacaagag-3’ and 5’-ttacgccagaatgcgttcgcac-3’. The PCR parameters were: 95°C for 5 min; 30 cycles with 88°C for 30 s, 57°C for 30 s, 72°C for 1min; 10 min at 72°C. PCR products were cloned using CloneJET PCR cloning kit (Thermo Scientific). Seven to ten clones were sequenced for each sample. A3 induced hypermutations were analyzed with the Hypermute online tool (http: //www. hiv. lanl. gov/content/sequence/HYPERMUT/hypermut. html). The overall mutation rate was calculated by using the total number of G-A mutations divided by the total analyzed nucleotides. To analyze CD4 and CCR5 expression level of SupT11 cell lines, 5×105 cells were stained by α-hCD4 PE mouse IgG1k (Dako, Hamburg, Germany) and α-hCCR5 FITC (BD Bioscience, Heidelberg, Germany) separately according to the manufacturer’s instruction. The mouse IgG1/RPE isopeptidase was used as negative antibody control for CD4 staining. The measurement was carried out by BD FACSanto (BD Bioscience). Data analysis was done with the Software FlowJo version 7. 6 (FlowJo, Ashland, USA). Buffy-coats obtained from anonymous blood donors were obtained from University Hospital Düsseldorf blood bank. Whole blood was obtained from healthy and de-identified African donors that signed an informed consent. The research has been approved by the Ethics Committee of the Medical Faculty of the Heinrich-Heine-University Düsseldorf (Reference No 4767R - 2014072657) and performed according to the principles expressed in the Declaration of Helsinki. Cellular RNA from PHA stimulated human PBMCs was isolated by using QIAGEN RNA extraction kit (Qiagen). 1 μg of total cellular RNA was used for reverse transcription with the RevertAid H Minus First Strand cDNA synthesis kit (Thermo Scientific). Human A3H cDNA was amplified with Q5 High-Fidelity DNA Polymerase (New England Biolabs) using primers: 5’-atggctctgttaacagccgaaacattcc-3’ and 5’-ggactgctttatcctgtcaagccgtcgc-3’. PCR products were cloned using CloneJET PCR cloning kit (Thermo Scientific). Six to ten clones were sequenced for each donor. To produce SIVcpz, 1×106 293T cells in 6-well plate were transfected with 2 μg SIVcpz molecular clone plasmids (SIVcpzPtsTan1, SIVcpzPttMB897 and SIVcpzPttGab1). 2 days after transfection, the viral supernatants were collected and centrifuged at 12,000 rpm for 10 min to remove cell debris. Then the viral supernatants were concentrated through 20% sucrose cushion at 14,800 rpm 4 h, followed by resuspension in RPMI. The viral stock was quantified by using the Cavidi HS lenti RT kit (Cavidi Tech, Uppsala, Sweden). 5×105 cells of each SupT11 cell lines (SupT11-vector-hCCR5 and SupT11-hA3H hapII-hCCR5) were infected with 1 ng or 5 ng RT activity of SIVcpz in a 24-well plate (in 500 μl) and cells were washed with PBS 1 day post-infection. Each second day, 200 μl supernatant was collected, clarified, and stored at ­80°C, and cultures were supplemented with fresh media. The replications were performed in two independent experiments, and each infection was performed in duplicates. 3 x 105 PHA stimulated PBMC from three donors were infected overnight with SIVcpzPttMB897 representing either 1 ng RT activity or 5 ng RT activity in the presence of 30 U/ml Interleukin-2 (IL-2) in 96-well round-bottom plates (total volume 200 μl). After infection, cells were washed three times and maintained in complete RPMI with 30 U/ml of IL-2 for 15 days. 100 μl culture supernatant was collected every 2–3 days, and cultures were supplemented with fresh media. The RT activities of viruses in culture supernatant were quantified by using the Cavidi HS lenti RT kit (Cavidi Tech, Uppsala, Sweden). A total of 3×105 293T cells in 24-well plates were co-transfected with 200 ng chimpanzee A3H expression plasmid or 50 ng hA3H haplotype II in PTR600 expression vector and 300 ng pCRV1 Vif expression plasmids, pcDNA3. 1 (+) (Thermo Fisher Scientific) was used to fill up the total transfected plasmid DNA to 500 ng. Transfections were performed by using Lipofectamine LTX (Thermo Fisher Scientific). 48 h post transfection, cells were lysed and clarified by 14,000 rpm/30 mins centrifugation. The expression of A3H and Vif were analyzed by immunoblots. Transfected 293T cells were lysed in radioimmunoprecipitation assay (RIPA) buffer (25 mM Tris-HCl [pH 8. 0], 137 mM NaCl, 1% NP-40,1% glycerol, 0. 5% sodium deoxycholate, 0. 1% sodium dodecyl sulfate [SDS], 2 mM EDTA, and protease inhibitor cocktail set III [Calbiochem, Darmstadt, Germany]). To pellet virions, culture supernatant were centrifuged at 12,000 rpm for 10 min followed by centrifugation through 20% sucrose cushion at 14,500 rpm 4 h and resuspended in RIPA buffer, boiled at 95⁰C for 5 min with Roti load reducing loading buffer (Carl Roth, Karlsruhe, Germany) and resolved on a SDS-PAGE gel. The expression of A3s and SIV/HIV Vifs were detected by mouse anti-hemagglutinin (anti-HA) antibody (1: 7,500 dilution, MMS-101P; Covance, Münster, Germany), rabbit anti-HA antibody (1: 1,000 dilution, C29F4, cat. 3724, Cell Signaling, USA) and rabbit anti-Vif polyclonal antibody (1: 1,000 dilution, NIH AIDSREAGENTS, cat. 2221) [67]; tubulin and SIVcpzPttMB897 capsid protein was detected using mouse anti-α-tubulin antibody (1: 4,000 dilution, clone B5-1-2; Sigma-Aldrich, Taufkirchen, Germany) and mouse anti-capsid p24/p27 MAb AG3. 0 (1: 50 dilution) separately [68], followed by horseradish peroxidase-conjugated rabbit anti-mouse or donkey anti-rabbit antibodies (α-mouse or rabbit-IgG-HRP; GE Healthcare, Munich, Germany), and developed with ECL chemiluminescence reagents (GE Healthcare). The expression of A3H in SupT11 cell lines was detected by using anti-hA3H (1: 1,000) antibody [35] followed by horseradish peroxidase-conjugated rabbit anti-mouse and developed with ECL chemiluminescence reagents. 8 x 106 human PBMCs from three donors were lysed in 100 μl RIPA buffer with protease inhibitor cocktail set III [Calbiochem, Darmstadt, Germany]). The expression of A3G, A3H and tubulin were detected by using anti-hA3H (1: 1,000) [35], anti-hA3G (1: 10,000, NIH AIDSREAGENTS, cat. 9906) [69] and anti-tubulin (1: 4,000 dilution, clone B5-1-2; Sigma-Aldrich, Taufkirchen, Germany) antibodies, respectively. The primate A3H sequences were obtained from GenBank, the accession numbers are: EU861357, EU861358, EU861359, EU861360, EU861361, DQ408606 and DQ507277. CpzA3H SNPs were described in this study. The A3H sequences were aligned using the ClustalW in Mega 7 software. The phylogenetic analysis was performed in Mega 7 by using bootstrap neighbor joining method. Test parameters were estimated using 500 bootstrap replicates. To analyze the interaction surface between SIVcpz Vif and hA3H hapII, the structure of SIVcpz Vif was modeled using HIV-1 Vif (4N9F) [70] as template by using SWISS-MODEL online server (http: //www. swissmodel. expasy. org/). The recent crystal structure of hA3H hapII (6B0B) was also used to model the structure of cpzA3H. The SIVcpz Vif and hA3H hapII co-structure was modeled based on the recent HIV-1 Vif-A3H interaction surface analysis [49]. The graphical visualization was constructed using PyMOL (PyMOL Molecular Graphics System, version 1. 5. 0. 4; Schrödinger, Portland, OR). Data are represented as the mean with SD in all bar diagrams. Statistically significant differences between two groups were analyzed using the unpaired Student’s t-test with GraphPad Prism version 5 (GraphPad software, San Diego, CA, USA). A minimum p value of 0. 05 was considered as statistically significant: P value < 0. 001 extremely significant (***), 0. 001 to 0. 01 very significant (**), 0. 01 to 0. 05 significant (*), >0. 05 not significant (ns).

Cellular cytidine deaminases of the APOBEC3 (A3) family are potent restriction factors that are able to inhibit retroviruses, this A3 restriction is counteracted by lentivirus Vif proteins. Human APOBEC3H (A3H) represents the most evolutionarily divergent A3 gene; it includes seven haplotypes and several splice variants. The polymorphism of human A3H has relevance for HIV-1 infection and AIDS progression. HIV-1 originated from cross-transmission of SIVcpz to humans. However, little is known about how human A3s affect the replication or transmission of SIVcpz. In this study, we comprehensively analyzed the anti-SIVcpz activity of chimpanzee and human A3s. Human A3H haplotype II was identified as strong inhibitor against SIVcpz regardless of Vif. In addition, other stably expressed human A3H haplotypes and splice variants showed strong antiviral activity against SIVcpz. Moreover, based on the recent Great Ape Genome Project, we found that the polymorphism of chimpanzee A3H is lower compared with the diversity of human A3H. And chimpanzee A3H haplotypes identified in this study showed similar anti-SIVcpz activity and Vif sensitivity. Our results provide a model that stably expressed human A3H protects humans against the cross-species transmission of SIVcpz and that SIVcpz spillover to humans may have started in individuals that harbor haplotypes of unstable A3H proteins.

Abstract Introduction Results Discussion Methods

medicine and health sciences pathology and laboratory medicine 293t cells pathogens biological cultures microbiology vertebrates animals mammals genetic mapping plasmid construction retroviruses primates immunodeficiency viruses viruses rna viruses dna construction molecular biology techniques tubulins research and analysis methods proteins medical microbiology hiv microbial pathogens hiv-1 cell lines viral replication molecular biology siv biochemistry cytoskeletal proteins haplotypes eukaryota heredity viral pathogens virology genetics biology and life sciences apes lentivirus chimpanzees amniotes organisms

2017

Stably expressed APOBEC3H forms a barrier for cross-species transmission of simian immunodeficiency virus of chimpanzee to humans

Lung cancer, with its most prevalent form non-small-cell lung carcinoma (NSCLC), is one of the leading causes of cancer-related deaths worldwide, and is commonly treated with chemotherapeutic drugs such as cisplatin. Lung cancer patients frequently suffer from chemotherapy-induced anemia, which can be treated with erythropoietin (EPO). However, studies have indicated that EPO not only promotes erythropoiesis in hematopoietic cells, but may also enhance survival of NSCLC cells. Here, we verified that the NSCLC cell line H838 expresses functional erythropoietin receptors (EPOR) and that treatment with EPO reduces cisplatin-induced apoptosis. To pinpoint differences in EPO-induced survival signaling in erythroid progenitor cells (CFU-E, colony forming unit-erythroid) and H838 cells, we combined mathematical modeling with a method for feature selection, the L1 regularization. Utilizing an example model and simulated data, we demonstrated that this approach enables the accurate identification and quantification of cell type-specific parameters. We applied our strategy to quantitative time-resolved data of EPO-induced JAK/STAT signaling generated by quantitative immunoblotting, mass spectrometry and quantitative real-time PCR (qRT-PCR) in CFU-E and H838 cells as well as H838 cells overexpressing human EPOR (H838-HA-hEPOR). The established parsimonious mathematical model was able to simultaneously describe the data sets of CFU-E, H838 and H838-HA-hEPOR cells. Seven cell type-specific parameters were identified that included for example parameters for nuclear translocation of STAT5 and target gene induction. Cell type-specific differences in target gene induction were experimentally validated by qRT-PCR experiments. The systematic identification of pathway differences and sensitivities of EPOR signaling in CFU-E and H838 cells revealed potential targets for intervention to selectively inhibit EPO-induced signaling in the tumor cells but leave the responses in erythroid progenitor cells unaffected. Thus, the proposed modeling strategy can be employed as a general procedure to identify cell type-specific parameters and to recommend treatment strategies for the selective targeting of specific cell types. Lung carcinoma is one of the leading causes of cancer-related deaths worldwide. The main types of lung cancer are small-cell lung carcinoma (SCLC) and non-small-cell lung carcinoma (NSCLC). NSCLC is the most frequent form with a prevalence of around 85% and can be classified in squamous-cell carcinoma, large-cell carcinoma, and adenocarcinoma which is the most prevalent subgroup (40%) [1]. Because lung cancer metastasizes already at early stages independent of the tumor size [2], most of the patients receive chemotherapeutic agents such as cisplatin. As a side effect of chemotherapy, as well as due to tumor-related effects, anemia frequently occurs [3]. Anemia is treated either by blood transfusion or by erythropoiesis stimulating agents (ESAs) such as erythropoietin (EPO) alfa or beta [4]. EPO is the key regulator of red blood cell production and ensures survival, proliferation and differentiation of erythroid progenitors at the colony forming unit-erythroid (CFU-E) stage in the fetal liver, the adult bone marrow and spleen. Biosynthesis of EPO in the kidney is stimulated by reduced blood oxygen levels [5]. Unfortunately, recent studies suggested that EPO treatment could reduce the overall survival of NSCLC patients [6]. Furthermore, expression of the EPO receptor (EPOR) has been detected in some tumors and cancer cells including NSCLC cells [7–10]. Co-expression of EPO and the EPOR has been shown to be associated with poor survival of NSCLC patients, even at stage I [11]. Because of these ambivalent effects, the treatment of lung cancer patients with EPO in the context of cancer-related anemia is controversially discussed [12,13]. In addition to clinically administered recombinant EPO, endogenous EPO produced in the kidney may also influence cancer cells. It has been speculated that EPO might affect the survival of cancer cells [14]. The key signaling pathway that is activated by EPO binding to the EPOR and that is involved in survival signaling is the Janus kinase (JAK) 2 / signal transducer and activator of transcription (STAT) 5 pathway. EPOR is a member of the cytokine receptor superfamily and is present at the cell surface as a homodimer [15]. Upon EPO binding, the receptor undergoes conformational changes and activates the pre-bound tyrosine kinase JAK2 [16]. The activated JAK2 is transphosphorylated and phosphorylates (p) tyrosine residues on the EPOR cytoplasmic domain that serve as docking site for the interaction with Src-homology (SH) 2-domain containing proteins. The phosphorylated EPOR-JAK2 complex is able to bind and phosphorylate the latent transcription factor STAT5. Subsequently, pSTAT5 forms dimers that translocate to the nucleus, where they induce the transcription of target genes. JAK2/STAT5 target genes such as cytokine-inducible SH2-containing protein (CISH mRNA, translated to CIS) and suppressor of cytokine signaling 3 (SOCS3 mRNA, translated to SOCS3) are expressed that act as negative regulators of JAK2/STAT5 signaling at the EPOR/JAK2 level [17]. Specifically, CIS inhibits STAT5 activation by binding to the EPOR, whereas SOCS3 binds to the activated receptor and the kinase domain of JAK2, thereby inhibiting its tyrosine kinase activity [18–21]. In addition, protein tyrosine phosphatases (PTPs) negatively regulate the activated EPOR-JAK2 complex [21]. It has been shown that the PTP SHP1 negatively regulates EPOR signaling in hematopoietic cells [22]. EPO acts on the erythroid lineage, particularly on CFU-E cells residing in the bone marrow of human adults. It is difficult to obtain primary human CFU-E cells in adequate numbers to perform time-resolved biochemical experiments required for mathematical modeling. However, murine CFU-E cells can be readily purified from fetal mouse liver preparations subjected to negative depletion with lineage-specific antibodies [23] and thereby can be isolated in sufficient quantities for data-based mathematical modeling. Therefore, we employed murine CFU-E cells as a proxy for human CFU-E cells and as representative for healthy erythroid cells. Recently, an ordinary differential equation (ODE) model of the EPO-induced JAK2/STAT5 pathway in murine CFU-E cells was reported [19] that focused on the distinct roles of SOCS3 and CIS and linked the integrated nuclear pSTAT5 response to cell survival. To investigate differences of EPO-induced JAK2/STAT5 signaling in hematopoietic and non-hematopoietic cell types like cancer cells, such an ODE model can be generalized. The parameters of the generalized mathematical model can be estimated individually using time- and dose-resolved quantitative data for the studied cell types. A major challenge is to identify significant differences in parameter values that are characteristic of a respective cell type. In general, one could discriminate between cell type-specific parameters and cell type-independent parameters. Mathematically, all combinations of parameters being either cell type-specific or cell type-independent have to be tested. Finding the exact solution of this model selection task is challenging as the number of candidate models grows exponentially with the number of model parameters [24]. Established approaches to approximate the solution of such a selection task have been developed mainly in classical statistics and comprise approaches such as lasso (least absolute shrinkage and selection operator) [25], elastic net [26], forward selection and backward elimination [27], and combinations thereof [28]. These approaches favor small models by penalizing increasing numbers of parameters. Their main difference is the metric quantifying the model complexity. In the lasso approach, the L1 metric, i. e. , the absolute value of differences to zero is used. It has been shown that under certain conditions, the L1 metric produces similar results as the L0 metric that penalizes only the number of parameters [29]. However, due to convexity and continuity of the L1 metric, it is considered favorable over L0 [30]. The L2 metric, on the other hand, does not lead to a sparse (parsimonious) model and is therefore not suitable for our task. Because L1 regularization techniques have been mainly developed for linear models, only a subset of existing algorithms can be employed for models that are nonlinear, which is a prevalent property of mathematical models based on coupled ODEs. Additional issues arising from nonlinearity comprise local optima, non-identifiabilities, and decreased performance of numerical calculations in general. To tackle these challenges, state-of-the-art implementations for optimization in nonlinear ODE systems have been developed [31]. We utilize these established approaches and extend their functionality by identifying parameter differences with L1 regularization. For a successful implementation, strategies to consider discontinuities in the derivatives and adaptation of convergence criteria have to be taken into account [32]. Here, we show on the basis of simulated data that our implementation for modeling of nonlinear ODE systems in combination with an L1 approach is a suitable method to infer cell type-specific parameters of two cell types and demonstrate the interpretation of the results. This approach is then applied to assess parameter differences in EPO-induced JAK2/STAT5 signaling between the NSCLC cell line H838 that expresses the EPOR [9,10,33] and CFU-E cells. Furthermore, a sensitivity analysis of the biological readout is performed to discover potential targets for intervention that specifically reduce the effects of EPO on the cancer cells while leaving the EPO-induced survival of hematopoietic cells unaffected. We previously showed that the NSCLC cell line H838 expresses the EPOR at the mRNA and protein level, albeit at rather low levels [10]. To more robustly detect EPO-induced phosphorylation of signaling components and expression of target genes in this cell line, we stably expressed high levels of HA-tagged human EPOR (HA-hEPOR) in H838 cells by retroviral transduction (H838-HA-hEPOR). The amount of EPOR protein present in the human lung cancer cell line H838 and its derivative H838-HA-hEPOR was determined by immunoprecipitation followed by quantitative immunoblotting and compared to the amount present in mouse CFU-E cells (Fig 1A). A protein band in the immunoblot that corresponds to the EPOR was detected in the analyzed cell types. As shown in S1A and S1B Fig, the hEPOR levels in H838-HA-hEPOR cells were increased 170-fold compared to H838 cells. The shift in molecular weight of the EPOR in H838-HA-hEPOR cells is most likely due to the HA-tag introduced into the EPOR. For absolute quantification of receptor expression levels, signal intensities of purified protein standards of mouse or human GST-tagged EPOR [34] were used to establish calibration curves and to subsequently estimate the number of EPOR molecules per cell (S1D Fig). As shown in S1 Table, H838-HA-hEPOR cells harbor 620 000 ± 200 000 hEPOR molecules per cell, H838 cells 3 600 ± 1 200 hEPOR molecules per cell and CFU-E cells 4 300 ± 2 200 mEPOR molecules per cell. To assess the activation of signal transduction we examined ligand-induced phosphorylation of the EPOR (Fig 1A). H838 and H838-HA-hEPOR cells were treated for 10 min with 10 U/ml EPO beta and CFU-E cells for 10 min with 5 U/ml EPO alfa or were left untreated. In each case, a band in the immunoblot corresponding to the phosphorylated EPOR was detectable in the treated samples, but was absent in the unstimulated controls. The amount of pEPOR in H838-HA-hEPOR upon stimulation with 10 U/ml EPO beta was only approximately 12-fold increased compared to H838 (S1C Fig), indicating that in cells expressing elevated levels of the EPOR such as H838-HA-hEPOR possibly the transport of the receptor to the plasma membrane or phosphorylation of the EPOR by JAK2 might be limiting. S2A and S2B Fig show that EPO alfa and EPO beta activate EPO-induced signaling at a comparable extent. Further, we examined EPO-induced JAK2/STAT5 survival signaling in the NSCLC cell line H838 and its derivative H838-HA-hEPOR because it had been reported that EPO induces survival signaling via the JAK2/STAT5 pathway in CFU-E cells [19]. In H838 cells, pJAK2 was detected upon stimulation with 100 U/ml and in H838-HA-hEPOR an EPO dose-dependent increase in JAK2 phosphorylation was observed (Fig 1B). Because the latent transcription factor STAT5 is a key mediator of the JAK2/STAT5 signaling pathway, we determined the degree of STAT5 phosphorylation by mass spectrometry in H838 and H838-HA-hEPOR cells treated with different doses of EPO beta (Fig 1C). The highest degree of STAT5 phosphorylation was observed at 10 U/ml EPO beta, with approximately 60% of STAT5 in H838-HA-hEPOR cells being phosphorylated and 3% in H838 cells. As quantified in S2C and S2D Fig and summarized in S1 Table both H838 and H838-HA-hEPOR harbor approximately 1 200 JAK2 and 90 000 STAT5 molecules compared to 24 000 JAK2 and 20 000 STAT5 molecules in CFU-E. To also consider the cytoplasmic and nuclear volume of the cell types analyzed, trypsinized cells were analyzed by fluorescence microscopy and the diameter of the cell and of the nucleus was determined (S3A and S3B Fig). CFU-E cells with a cell volume of 700 μm3 and a nuclear volume of 300 μm3 are much smaller compared to H838 cells that have a cell volume of around 14 000 μm3 and a nuclear volume of around 2 000 μm3. The respective volumes were utilized to convert number of molecules per cell into cell type-specific initial concentrations, which are summarized in S1 Table. To test whether EPO influences survival and may reduce cisplatin-induced apoptosis in NSCLC cells, H838 and H838-HA-hEPOR cells were treated with 5 mg/l of the chemotherapeutic agent cisplatin in combination with 10 U/ml EPO beta or were left untreated for three days. Cell viability was measured every 24 hours by the CellTiter-Blue assay (Fig 1D, S4 Fig). In untreated proliferating H838 and H838-HA-hEPOR cells, a two- to three-fold increase in cell numbers was observed. No significant impact of EPO beta on cell viability was detected during this observation period. The treatment of H838 and H838-HA-hEPOR with 5 mg/l cisplatin reduced cell viability or at least decreased proliferation compared to the untreated control. Interestingly, co-treatment with 10 U/ml EPO beta decreased the impact of cisplatin in H838 and H838-HA-hEPOR at each time point. While there was some variability in the response of these cells to cisplatin, the rescue effect induced by EPO beta was consistent. To validate these findings, the Casper3-GR FRET (fluorescence resonance energy transfer) -based sensor was expressed in H838-HA-hEPOR (H838-HA-hEPOR-Casper3-GR) and caspase-3 activity, a key indicator for apoptotic responses, was measured by life-cell imaging in a time-resolved manner (Fig 1E, S5 Fig). As a positive control 10 μM staurosporine was applied that is known to rapidly induce caspase-3. Indeed, the effect of staurosporine was observed within 5 hours upon treatment whereas in untreated cells no peak of caspase-3 activation was detected. Upon treatment with 5 mg/l cisplatin the maximum peak of caspase-3 activity was reached at around 24 hours followed by a strong signal decrease due to cells undergoing apoptosis. Upon co-treatment with 10 U/ml or 500 U/ml EPO the amplitude of the cisplatin-induced maximal caspase-3 activation showed an EPO dose-dependent reduction, which is in line with the cell viability assay, and a peak shift to around 15 hours. Treatment with EPO alone did not induce caspase-3 activation. Concluding, we demonstrated that, similar to CFU-E, the EPOR expressed in H838 cells was capable of activating the JAK2/STAT5 signaling cascade. Further, in line with the observation that EPO-induced activation of JAK2/STAT5 signaling correlates with survival signaling [19], we showed in co-treatment experiments that EPO reduces the extent of cisplatin-induced apoptosis in H838 and H838-HA-hEPOR cells and thus exhibits a rescuing effect in the lung cancer cell line and its derivative. Due to the similarity of the core signaling components, it is difficult to predict cell type-specific differences in the signaling network based only on experimental observations. We developed a systematic mathematical modeling approach for the unbiased identification of cell type-specific differences with the aim to propose strategies to exclusively target the lung cancer cells but not the erythroid progenitor cells. To predict cell type-specific model parameters between two different cell types, we propose an ODE-based mathematical modeling strategy in combination with a L1 regularization method. To show that this approach is capable to identify these parameter differences, first an in silico model with simulated data was investigated. The exemplary mathematical model mimics a two-step phosphorylation cascade in two different cell types, in which a protein (Protein) is irreversibly converted into a phosphorylated protein (pProtein) with rate k1 followed by another phosphorylation step resulting in a doubly phosphorylated protein (ppProtein). This second step can be reversed by a dephosphorylation reaction (Fig 2). The ppProtein and its intermediate pProtein equilibrate depending on the ratio of the forward (k2) and backward reaction (k3). The corresponding differential equation system is given by: d[Protein]dt=−k1⋅ [Protein]d[pProtein]dt=k1⋅[Protein]−k2⋅[pProtein] + k3⋅[ppProtein]d[ppProtein]dt=k2⋅[pProtein] − k3⋅[ppProtein] The model structure and the parameters k1 and k2 were chosen to be identical for cell type 1 and cell type 2. In contrast, the following two parameters were set as cell type-specific: The rate k3 was set to 0. 5 min-1 in cell type 1 and 0. 1 min-1 in cell type 2 and the parameter [Protein]t = 0 was defined as 1 nM in cell type 1 and 2 nM in cell type 2. The initial concentration of pProtein and ppProtein at t = 0 min were assumed to be zero in both cell types. The ODE solution was calculated using the D2D software [31] and data (black dots) were simulated by adding normally distributed noise (σ = 0. 1) (Fig 2). The model trajectories are depicted as black lines and the shading represents the noise distribution. To approximate a realistic setting, only two proteins were assumed to be measurable and sampling time points were not equal. Hence, data was available for Protein (n = 21) and ppProtein (n = 10). To estimate parameters and identify cell type-specific differences given only by the model structure and the simulated data, the model was parameterized with parameter pi for cell type 1 and ri ∙ pi for cell type 2, with ri denoting fold-changes between both cell types. After calculating the maximum likelihood estimates individually for cell type 1 and 2, the regularization weight λ of the constrained likelihood C = Lcell type 1+Lcell type 2+λ∑i|log10ri| was scanned. Here, ℒcell type 1 and ℒcell type 2 denote the likelihood (i. e. the negative two-fold log-likelihood) of the respective cell type and the last term regularizes the fold-changes ri of parameters between the cell types. If the parameter pi is the same in both cell types, ri = 1 and the regularization term is zero. To obtain the regularization path for all parameters, we gradually increased λ from 10−4 to 104 and re-estimated the constrained likelihood at each step. With increasing regularization weight λ, the number of cell type-specific parameters decreased until all parameters were independent from the cell type (Fig 3A). To determine the optimal regularization weight λ, i. e. to statistically evaluate the minimum number of cell type-specific parameters that are necessary for the model to sufficiently describe the data (parsimonious model), the likelihood ratio test was utilized. The L1 regularization was used to select cell type-specific parameters. To reduce bias, the parameters of the non-regularized parsimonious model were then estimated in a second step. The likelihood ratio test statistic D = ℒλ − ℒfull was used, where ℒfull denotes the likelihood of the full model with only cell type-specific parameters (M) and ℒλ the likelihood of a model with N cell type-specific parameters that were selected for a given λ. If D < χ2dof, α, the smaller model cannot be rejected based on the χ2-distribution for a given confidence level α and degrees of freedom (dof = M − N). In Fig 3B, the test statistic D is shown by a solid blue line and the statistical threshold based on the likelihood ratio test is shown as a dashed red line. The crossing point of both lines defines the parsimonious model, in which only relevant parameters are defined as cell type-specific. The so-called regularization path shown in Fig 3C indicates the fold-changes between cell type 1 and cell type 2 for all parameters at a given regularization weight. The colors red or blue show whether a parameter is larger in cell type 1 or 2, respectively. The regularization path is also shown in S6A Fig as a line plot. In our example, the parsimonious model that is indicated by the vertical dashed line (Fig 3A, 3B and 3C) has two cell type-specific parameters. These two parameters were the rate k3 and the initial concentration of Protein [Protein]t = 0, which could be reconstructed from the example model as cell type-specific (indicated with asterisks in Fig 3C and S6A Fig). We determined the relative difference and by a profile likelihood approach [35] the corresponding confidence interval (S6B Fig): [Protein]t = 0 is larger in cell type 2 by a factor of 1. 88 ± 0. 17 (true value is 2) and k3 is larger in cell type 1 by a factor of 4. 71 ± 1. 20 (true value is 5). To illustrate how well individual models that differ in the number of cell type-specific parameters can describe the simulated data, model trajectories of three different scenarios were plotted with the simulated data (Fig 3D). The model with only cell type-specific parameters (green dashed line), which is the largest model and has the highest degree of freedom, was capable to describe all data sets. The smallest model with no cell type-specific parameters (blue dashed line) was not in statistical agreement with the simulated data, because e. g. for both cell types the species “ppProtein” was not correctly described. In line with the likelihood ratio test, the parsimonious model with only relevant cell type-specific parameters (solid red line) described the simulated data to a comparable extent as the model with only cell type-specific parameters. In sum, the L1 regularization was able to reconstruct cell type-specific parameters for simulated data using a single scan of the regularization weight λ, suggesting that our implementation for dynamic ODE-based mathematical modeling in combination with L1 regularization is appropriate to reveal differences between two different cell types. To pinpoint cell type-specific differences in EPO-induced JAK2/STAT5 signaling in CFU-E cells and H838 & H838-HA-hEPOR cells by the approach introduced above, the previously published mathematical model of EPO-induced JAK2/STAT5 signaling in murine CFU-E cells [19] was used as reference model structure: The JAK2/STAT5 model for mCFU-E cells described the EPO-induced JAK2 activation in the EPOR-JAK2 complex, the subsequent phosphorylation of the EPO receptor, attenuation by the phosphatase SHP1, the activation of the transcription factor STAT5 by the EPOR-JAK2 complex and the STAT5 transport from cytosol into and out of the nucleus, as well as the induced transcription and translation of the negative feedback regulators SOCS3 and CIS. To describe EPO-induced JAK2/STAT5 signaling in human H838 and H838-HA-hEPOR cells, a generalized model structure was developed by adding the following components to the reference model: (i) Basal RNA activation rates were included, because we observed that basal CISH mRNA and SOCS3 mRNA is also reduced by inhibiting transcription (S7A and S7B Fig). (ii) A dephosphorylation step of nuclear phosphorylated (np) STAT5 was included to be able to distinguish dephosphorylation and nuclear export of STAT5 [36]. (iii) The regulation of the SOCS3 promotor binding activity by npSTAT5 was described by a Hill coefficient to allow non-linear transcriptional activity. In addition, the receptor phosphatase SHP1 that is known to be restricted to hematopoietic cells was substituted by the general term PTP (Fig 4). The generalized model structure was parameterized based on cell type-specific quantitative data. First, the model was calibrated by directly implementing cell type-specific quantities for H838 and H838-HA-hEPOR cells such as cellular and nuclear volumes and the initial concentrations of the EPOR as well as JAK2 and STAT5 together with their respective experimental error (summarized in S1 Table). As the ratio of the EPOR and JAK2 was different for H838 and H838-HA-hEPOR cells possibly resulting in a saturation of the phosphatase in H838-HA-hEPOR cells, distinct properties were assumed for the phosphatase PTP in H838-HA-hEPOR cells. Except for this difference and the number of receptors on the surface as shown in S1 Fig, the JAK2/STAT5 signaling models for H838 and H838-HA-hEPOR cells were parameterized identically. Therefore, H838 and H838-HA-hEPOR cells were treated as the same cell type but with different conditions. To calibrate the model for CFU-E cells, quantitative data of the reference model [19] was reutilized (open circles in Fig 5). For the NSCLC cell line H838 and its derivative H838-HA-hEPOR, quantitative time-resolved data was generated (black closed circles in Fig 5). For H838 cells, the dynamics of pEPOR, pJAK2 and pSTAT5 and the expression of CISH mRNA upon stimulation with 10 U/ml EPO beta were determined. For H838-HA-hEPOR cells, the amount of total and phosphorylated STAT5, pEpoR, pJAK2 and the expression of CISH and SOCS3 mRNA were measured upon stimulation with 10 U/ml EPO beta. While the temporal changes of pEPOR were similar in CFU-E, H838-HA-hEPOR and H838 cells, the dynamics of CISH mRNA was more transient in CFU-E cells, while pSTAT5 was more sustained in H838-HA-hEPOR cells than in H838 cells. SOCS3 mRNA, on the other hand, was more sustained in CFU-E cells than in H838-HA-hEPOR cells. Thus, we observed substantial differences in the dynamics of signaling components in the three cell types analyzed. We performed a multitude of additional measurements with different conditions (S7 Fig). The basal expression levels of CISH mRNA and SOCS3 mRNA were determined for CFU-E and H838-HA-hEPOR cells by applying the transcription inhibitor actinomycin D. Time- and dose-resolved protein quantification was performed by quantitative immunoblotting and relative mRNA quantification by qRT-PCR. The degree of STAT5 phosphorylation (pSTAT5) was measured for both the lung cancer cell line and its derivative by quantitative mass spectrometry. The generalized model structure comprised 25 kinetic parameters and 3 initial concentrations (the EPOR-JAK2 complex, STAT5 and PTP at time point 0). Assuming that all kinetic parameters and initial concentrations were cell type-specific (i. e. different parameters for CFU-E and for H838 & H838-HA-hEPOR cells), this model was calibrated based on the experimental data of CFU-E (516 data points) using 83 additional observation parameters and of H838 and H838-HA-hEPOR cells (625 data points) using 77 additional observation parameters. For global optimization of the likelihood, a multi-start deterministic strategy was applied [37]. The model could describe all data sets of the cell types tested, shown representatively by the model trajectory of the model with only cell type-specific parameters (dashed green line in Fig 5). Concluding, we identified a model structure that was able to describe the data sets of EPO-induced JAK2/STAT5 signaling in CFU-E, H838 and H838-HA-hEPOR cells. Our experimental observations suggested that core reactions of the pathway operate comparable in the different cell systems. We therefore tested whether it is possible to identify cell type-independent and cell type-dependent parameters to establish a parsimonious model. The parsimonious model is defined as the model with the smallest number of cell type-specific parameters that is still compatible with the experimental data. To infer cell type-specific parameters, the L1 regularization strategy was applied while parameterizing the mathematical model based on the experimental data established for the different cell types. We tested the 25 kinetic parameters and the initial concentrations of PTP, while the initial concentrations of the EPOR-JAK2 complex and STAT5 were fixed to the measured values (S1 Table). As expected, by increasing the regularization weight the number of cell type-specific parameters was gradually decreased (Fig 6A). In analogy to the example (Fig 3), the parsimonious model was determined using the likelihood ratio test (blue line in Fig 6B). The statistical threshold corresponding to a significance level α = 0. 05 is shown by the red dashed line. At the crossing point of these two lines, the parsimonious model with relevant cell type-specific parameters is defined and depicted by the dashed black vertical line in all panels of Fig 6. The parsimonious model with the minimal number of cell type-specific parameters (red line, Fig 5) resulted in trajectories describing the data and showed similar performance as the model with only cell type-specific parameters (dashed green line, Fig 5). On the contrary, the model with only cell type-independent parameters (dashed blue line, Fig 5) was not capable of representing the dynamics of the data as evidenced e. g. by the difference of the model trajectory and the experimental data of SOCS3 mRNA. The parsimonious model simultaneously described all experimental data measured in CFU-E (S8 Fig), H838-HA-hEPOR (S9A and S9B Fig) and H838 (S9B and S9C Fig). Seven relevant cell type-specific parameters are highlighted with an asterisk in the regularization path of the 26 parameters in Fig 6C and in an alternative representation in S10A Fig. We ensured significance by calculating the confidence intervals of these parameter differences (S10B Fig). Table 1 provides a description of the 26 parameters and the model-predicted differences between CFU-E cells and H838 & H838-HA-hEPOR cells. The cell type-specific parameters predicted to have higher values in CFU-E cells (depicted in red, Fig 6C) were the CISH mRNA turnover rate (CISHRNAturn), the activation rate of the EPOR by JAK2 (EPORactJAK2) and the activation rate of JAK2 by EPO (JAK2actEPO). In contrast, the parameters that were predicted to have higher values in the H838 & H838-HA-hEPOR cells (depicted in blue, Fig 6C) were the parameter inversely defining the delay in SOCS3 mRNA production (SOCS3RNAdelay), the SOCS3 promoter activity (SOCS3prom) the import rate of pSTAT5 into the nucleus (STAT5imp) and the deactivation rate of STAT5 in the nucleus (nSTAT5deact). The parameter that showed strongest evidence for a difference between both cell types was the parameter regulating npSTAT5-induced SOCS3 promoter activity (SOCS3prom) (Fig 6C). The model predicted a linear correlation of the SOCS3 mRNA production rate and the concentration of npSTAT5 for CFU-E cells (S11A Fig). In contrast, the same prediction analysis for H838 & H838-HA-hEPOR cells indicated a threshold behavior of the SOCS3 mRNA production rate with respect to npSTAT5. The analysis showed that in these cells a certain amount of activated STAT5 has to be present in the nucleus to induce SOCS3 mRNA synthesis (S11B Fig). To test whether this difference between cell types can be explained by differences in promoter binding elements or epigenetic modifications, the human and murine SOCS3 promoter was analyzed for STAT5 binding sites. Indeed, the murine promoter contains four STAT5 binding sites, whereas the human harbors only two (S11C Fig). Additionally, we performed methylation analyses and demonstrated that the SOCS3 promoter is accessible in both CFU-E and H838 cells (S11C Fig). Taken together, the selected parsimonious model could simultaneously describe the data sets of CFU-E, H838 and H838-HA-hEPOR cells. Seven relevant cell type-specific parameters were identified and three of these parameters had higher values in CFU-E cells and four had higher values in H838 & H838-HA-hEPOR cells. The parsimonious model predicted several cell type-specific parameters comprising mRNA processing: the turnover rate of the CISH mRNA, the SOCS3 promoter activity and the SOCS mRNA delay parameter. Therefore, we selected mRNA processing as a biological process that can be measured experimentally and is connected to parameters identified by the model as cell type-specific for experimental validation of the parsimonious model. We hypothesized that by the application of the transcriptional inhibitor actinomycin D the dynamics of CISH mRNA and SOCS3 mRNA would be altered in a cell type-specific manner. First, the dynamics of CISH mRNA and SOCS3 mRNA were predicted for CFU-E and H838-HA-hEPOR cells for 300 min upon treatment with EPO (black) or EPO in combination with actinomycin D (blue) (Fig 7). We proposed by experimental design that the transcriptional inhibitor should be applied at the predicted peak of mRNA expression upon EPO stimulation, which was after 60 min for CFU-E cells treated with 5 U/ml EPO alfa and after 30 min for H838-HA-hEPOR stimulated with 10 U/ml EPO beta (blue arrows in Fig 7). To assess the uncertainty of the model dynamics, the prediction profile likelihood [38,39] was utilized. The shading with dotted lines indicates the 1σ confidence interval. The EPO-induced dynamic of CISH mRNA and SOCS3 mRNA was predicted to show a transient peak followed by a new steady state in both cell types. Addition of actinomycin D was predicted to result in a more dramatic decrease after the treatment and a steady state that is lower compared to EPO treatment alone. The SOCS3 mRNA in H838-HA-hEPOR cells was predicted to even decrease below the starting level after around two hours. To experimentally validate these predictions, H838-HA-hEPOR and CFU-E cells were treated either with 10 U/ml EPO beta or 5 U/ml EPO alfa alone (black dots) or in combination with 1 μg/ml actinomycin D (blue dots) at the predicted time points of either 60 min (CFU-E) or 30 min (H838-HA-hEPOR). The dynamics of CISH mRNA and SOCS3 mRNA were measured by qRT-PCR. Unknown experiment-specific offset and scaling parameters were estimated from this validation data. As shown in Fig 7, the experimentally measured CISH mRNA and SOCS3 mRNA (dots) were in line with the model predictions (lines with shadings). In summary, the model predicted cell type-specific differences in the dynamics of the SOCS3 and CISH mRNA could be experimentally validated. Therefore, the parsimonious model and the uncovered cell type-specific differences were used to recommend strategies to specifically target JAK2/STAT5 signaling in H838 cells. To identify potential targets in the EPO-induced JAK2/STAT5 signaling pathway, a sensitivity analysis was performed for CFU-E and H838 cells. The area-under-curve of npSTAT5 at 60 min was taken as read out (Κ), because we previously could correlate this to survival of CFU-E cells [19]. A control coefficient SpiK = piK⋅∂K∂pi was calculated for each model parameter (pi) in both cell types [23]. The larger the absolute value of the control coefficient, the larger is the influence of a parameter to the simulated biological readout. If a control coefficient is positive, a decrease of the parameter induces a decrease in the area-under-curve of npSTAT5. Interestingly, we observed not only that several parameters exerted major control over npSTAT5, but also that the control coefficients of these parameters differed between CFU-E and H838 cells (S12 Fig). We identified six parameters that had a larger control coefficient for JAK2/STAT5 signaling in H838 cells than on CFU-E cells and thus could represent suitable therapeutic targets (indicated with asterisks in S12 Fig): The activation rate of STAT5 by pEPOR, the activation rate of JAK2 by EPO, the activation rate of STAT5 by pJAK2, the activation rate of the EPOR by pJAK2, the SOCS3 mRNA turnover rate and the deactivation rate of PTP. The inhibition of these parameters would diminish the npSTAT5 level and thus the STAT5-mediated survival in the H838 cells more than in CFU-E cells. The cell type-specific parameters and the predicted therapeutic targets are summarized by a color code in the process diagram of the JAK2/STAT5 pathway (Fig 8). Cell type-specific differences in the parameter values are depicted in the upper part of each reaction square: Red indicates that the parameter had higher values in CFU-E cells, purple that the parameter value was higher in H838 and H838-HA-hEPOR cells and colorless that there is no difference between cell types. The predicted therapeutic targets are marked in blue in the lower part of each reaction square. Most of the differences were related to the receptor-kinase complex and the shuttling of STAT5. Interestingly, the potential drug targets identified by the sensitivity analysis did not entirely overlap with cell type-specific differences in parameter values. Parameters with a higher control coefficient in H838 cells than in CFU-E cells were either related to the activation of the JAK2-EPOR complex and its control of STAT5 phosphorylation or SOCS3 mRNA degradation. The shuttling of STAT5 was in fact cell type-specific, but the control coefficient had a higher value in CFU-E cells (S12 Fig), meaning that CFU-E cells would be particularly sensitive for inhibitors and npSTAT5 levels would decrease in the erythroid progenitor cells and not, as intended, in the lung cancer cells. Based on these results, the model suggests inhibitors for the receptor complex, e. g. a JAK2 inhibitor, as optimal drugs to diminish the EPO-induced JAK2/STAT5 signaling specifically in H838 cells while the CFU-E cells would continue to benefit from EPO treatment. In sum, we demonstrate that the approach of ODE-based mathematical modeling in combination with L1 regularization is an appropriate method to identify cell type-specific differences and suitable therapeutic targets. In this study, we present an approach that combines ODE-based mathematical modeling with L1 regularization to identify differences in the EPO-induced JAK2/STAT5 signaling pathway between the NSCLC cell line H838 and CFU-E cells. Based on these differences, targeted inhibitor treatments are predicted to reduce EPO-induced survival signaling in the lung cancer cells with only marginally affecting EPO-induced signaling in healthy erythroid progenitor cells. We demonstrated that the EPOR is expressed in the NSCLC cell line H838, is phosphorylated upon EPO stimulation and activates JAK2/STAT5 signaling associated with cell survival. We verified that EPO inhibits cisplatin-induced apoptosis in H838 cells. This indicates that EPO might not only have an effect on erythroid progenitor cells, but also could induce EPO-mediated survival signaling in other cells expressing the EPOR, including tumor cells. This effect has been previously observed in other cancer entities: It was shown that EPOR signaling affected survival of EPOR expressing melanoma cells in vitro and in xenograft mice [14] and that EPO activates survival pathways in breast cancer stem-like cells and increases the resistance to chemotherapeutic agents [40]. Having demonstrated that EPO potentially targets not only CFU-E cells but also lung cancer cells, it remained to be addressed if the signaling upon EPO stimulation is identical between the different cell types or if there is a therapeutic window to specifically target lung cancer cells. We quantitatively analyzed EPO-induced JAK2/STAT5 signaling in CFU-E, H838 and H838-HA-hEPOR cells and identified cell type-specific differences in the activation dynamics of the EPOR and STAT5 as well as the induction of CISH mRNA and SOCS3 mRNA. For CFU-E cells, we previously showed by mathematical modeling of the JAK2/STAT5 signaling pathway that CISH mRNA attenuates STAT5 activation primarily at low EPO levels, while SOCS3 mRNA reduces STAT5 phosphorylation at high EPO concentrations. Importantly, with this approach we quantitatively linked STAT5 phosphorylation to survival of CFU-E cells [41]. Previously, differences between cell types have been identified based on either differential gene expression or genomic mutations. The advantage of the mainly array- or sequencing-based methods is the simultaneous analysis of multiple genomic alterations. For example, a statistical deconvolution approach to derive the relative mRNA abundance for each cell type and infer the relative cell type frequencies from microarray data of mixed tissue samples has been reported [42]. A pioneering sequencing study analyzing samples of 17 NSCLC patients has revealed copy number alterations in the EPOR gene (one patient), mutations in the EPOR gene (two patients) and mutations in the JAK2 gene (two patients) [43]. These snapshot data yield static information on cellular differences, while our time- and dose-resolved measurements additionally provide insights into cell type-specific dynamic properties of the signaling pathways. Furthermore, it is unclear how mutations and copy number alterations translate into dynamic properties such as signaling behavior and cell fate decisions. An important difference affecting the dynamics of signaling pathways in individual cell types is the cellular concentrations of the involved proteins. Recently, by a mass spectrometric approach based on the proteomic ruler, protein abundances in primary human hepatocytes and the human hepatoma cell line HepG2 were compared, revealing that uptake transporters and phase I enzymes were either absent or expressed in very low amounts in HepG2 cells [44]. Also by mass spectrometry, differences in the protein amounts between primary mouse hepatocytes and the murine hepatoma cell line Hepa1-6 were identified [45]. Such results can be readily combined with our L1 approach. We have estimated the initial concentrations of the EPOR-JAK2 complex and STAT5 based on our quantitative immunoblotting results, while differences in the initial concentration of PTP were tested by L1 regularization. Our results revealed that the concentration of PTP was not decisive for the differences between CFU-E cells and H838 cells. In this study, we particularly considered the signaling components of the EPO-induced signaling pathway that are essential to describe the dynamics of the JAK2/STAT5 pathway by ODEs and identified parameters that are cell type-specific for either H838 or CFU-E cells by using L1-regularized optimization. After the cell type-specific parameters were identified, a non-regularized optimization was performed to calculate unbiased final parameter values. This procedure estimates parameter differences in analogy to classical backward elimination, which constitutes a major difference to lasso where regularization is used for parameter selection as well as to reduce variance by introducing bias (shrinkage) [25]. In such a statistical application, the main goal is typically to minimize prediction error, while the parameter values are not important per se. In our case, we were additionally interested in the relative differences of these parameter values. Here, we for example inferred that the import rate of STAT5 is five times larger in CFU-E cells than in H838 cells. Therefore, we used L1 regularization only to suggest a set of parameters that were cell type-specific and then calculated the final parameter estimates with a non-regularized optimization. The seven predicted cell type-specific parameters comprised the CISH mRNA turnover rate, the activation rate of the EPOR by JAK2, the activation rate of JAK2 by EPO, the delay in SOCS3 mRNA production, the SOCS3 promotor activity, the import rate of pSTAT5 into the nucleus, and the deactivation rate of nuclear pSTAT5. We identified faster activation rates of JAK2 by EPO and of the EPOR by pJAK2 in CFU-E than in H838 and H838-HA-hEPOR cells. This might be due to differences in the affinity of the human-derived EPO to the murine EPOR in CFU-E cells. Additionally, JAK2 is highly expressed in CFU-E cells, while the JAK2 concentration is limiting in H838 cells (S1 Table). It was proposed that JAK2 acts as a chaperone that binds to the EPOR in the endoplasmic reticulum and thereby enhances cell surface expression of the receptor [46]. The import rate of pSTAT5 to the nucleus and the deactivation rate of npSTAT5 were predicted to be higher in H838 and H838-HA-hEPOR cells. It was previously suggested that a cell with a smaller nucleus imports cargo faster [47]. While H838 and H838-HA-hEPOR have a larger nucleus than CFU-E cells, the nucleus to cell volumetric ratio is 0. 17 in these cells compared to a ratio of 0. 69 in CFU-E cells. The deactivation rate of nuclear pSTAT5 is controlled by nuclear phosphatases. The small dual-specificity phosphatase VHR has been identified to dephosphorylate IFNβ-induced pSTAT5 in the nucleus [48]. It remains to be shown if the same phosphatase dephosphorylates EPO-induced pSTAT5 and whether this phosphatase is differentially expressed in a cell type-specific manner. The model further identified that the parameter for the SOCS3 promoter activity had a higher value in H838 and H838-HA-hEPOR cells. This parameter resulted in a linear dependency between pSTAT5 concentration and SOCS3 mRNA expression in CFU-E cells. On the other hand, in H838 and H838-HA-hEPOR cells SOCS3 mRNA was only expressed if the pSTAT5 concentration in the nucleus exceeded a certain threshold. To analyze the underlying mechanism, we compared the structure of the human and murine SOCS3 promoters, revealing the presence of additional STAT5 binding sites in the murine promoter, which potentially strengthen the link between transcription factor concentration and mRNA expression. Previously, hypermethylation in CpG islands of the SOCS3 promoter in H838 correlating with transcriptional silencing was reported [49]. However, our analyses in CFU-E and H838 cells demonstrated low promoter methylation levels in both cell types, indicating that the SOCS3 promoter should be accessible for the transcriptional machinery in both cell types. The remaining two differential parameters concern mRNA processing: The CISH mRNA turnover rate has higher values in CFU-E and the delay parameter for SOCS3 mRNA production has higher values in H838 and H838-HA-hEPOR cells. Turnover rates of mRNAs can be controlled by the decay rates influenced by both mRNA sequence elements and cellular factors, as reviewed previously [50]. The delay parameter for SOCS3 mRNA production summarizes transcription, pre-mRNA processing and export to the cytoplasm. In line with our results, it was shown by global run-on sequencing that transcription rates not only vary between different genes, but that they can vary between identical genes in different cell types [51]. Of the seven identified cell type-specific differences, three were associated with mRNA induction and processing. We were able to predict the distinct cell type-specific dynamics of EPO-induced CISH mRNA and SOCS3 mRNA production and mRNA degradation in response to actinomycin D inhibition. Our experimental results confirmed these model predictions. To predict potential drug targets that primarily affect one cell type, we performed a sensitivity analysis. Since we aimed to diminish the EPO-induced STAT5-mediated survival signal in the lung cancer cells, we focused on potential targets for more effective inhibition in H838 cells. Interestingly, we identified that parameter differences and differential sensitivities do not entirely coincide. On the one hand, the activation rates of JAK2 by EPO and of the EPOR by pJAK2 are faster in CFU-E cells but can be targeted more efficiently in H838 cells. On the other hand, the activation rates of STAT5 by the EPOR-pJAK2 and by the pEPOR-pJAK2 complexes, while associated with the same parameter values in CFU-E and H838 cells, largely affect the pSTAT5 output in H838 cells and have only minor control in CFU-E cells. Other examples for cell type-independent rates but higher control coefficients in H838 cells include the SOCS3 mRNA turnover rate and the PTP deactivation rate. We previously performed mathematical modeling of the EPO-induced JAK2/STAT5 signaling in BaF3-mEPOR cells, a hematopoietic cell line that is a frequently used model system to study EPOR signaling, and we showed that the parameters of nuclear shuttling are most sensitive to perturbation [36]. Here, we identified STAT5 shuttling parameters to be cell type-specific. Additionally, the control coefficients of the pSTAT5 import and pSTAT5 export parameters have higher values in CFU-E cells than in H838 cells, indicating that the sensitivity of nuclear STAT5 shuttling seems to be restricted to cells with a larger nuclear to cytoplasmic ratio such as CFU-E cells. The observed incongruity of parameter differences and differential sensitivity has previously been predicted by theoretical approaches. It was proposed that kinases that are mutated tend to lose part of their control on signaling, while some of the non-mutated genes may become more important [52]. Concluding, it is not only important to identify which parameters are different, but also a mathematical model is necessary to understand how these differences affect cell type-specific intervention points. We identified possibilities for cell type-specific targets. Since we compared murine CFU-E cells to human H838 cells, we cannot entirely exclude that some of the predicted differences are species-related. Therefore, validation of the predicted drug targets in human CFU-E cells would be advantageous. Of the six predicted targets for more effective inhibition in H838 cells, the JAK2-mediated reactions are most promising. The model suggested that JAK2 inhibition in combination with EPO treatment affects lung cancer cells to a much higher extent than erythroid progenitor cells. Several JAK2 inhibitors have been developed, including Fedratinib and Ruxolitinib. Preclinically, it was observed by injection of MMTV-Wnt-1 tumor cells into mammary fat pads of mice that inhibition of EPO-induced JAK2 activation by Fedratinib was synergistic with chemotherapy for breast tumor-initiating cells [53]. Clinically, therapy of myelofibrosis with Fedratinib showed beneficial efficacy. However, severe toxic effects in some patients were observed and clinical development of Fedratinib was therefore discontinued [54]. On the other hand, the JAK2 inhibitor Ruxolitinib has been approved for the treatment of myelofibrosis in the United States and in the European Union [55]. Also for solid tumors, the benefit of several JAK inhibitors is currently investigated in clinical studies, including a phase II study with Ruxolitinib in combination with pemetrexed/cisplatin in NSCLC [56]. Concluding, we anticipate that both basic and translational research will benefit from the proposed strategies to identify cell type-specific differences and to predict drug targets that affect cancer cells without impairing healthy cells. Human lung adenocarcinoma cell line H838 was purchased from ATCC (CRL-5844) and cultivated in Dulbecco' s modified Eagle' s Medium (DMEM, Lonza) supplemented with 10% fetal calf serum (Gibco), 100 μg/ml streptomycin (Gibco) and 100 U/ml penicillin (Gibco). The Phoenix ampho packaging cell line [57] was cultured in DMEM (Gibco) supplemented with 10% fetal calf serum, 100 μg/ml streptomycin (Gibco) and 100 U/ml penicillin (Gibco). For the EPOR overexpressing cell line (H838-HA-hEPOR) 1. 5 μg/ml puromycin (Sigma) was added. As growth factor depletion medium DMEM without phenol-red (Lonza) supplemented with 1 mg/ml BSA (Sigma), 100 μg/ml streptomycin (Gibco), 100 U/ml penicillin (Gibco) and 2 mM L-glutamine (Gibco) was used. All cells were cultivated at 37°C, 5% CO2 and 95% relative humidity. Generation of the retroviral expression vectors pMOWS-GFP [58] and pMOWS-HA-hEPOR [59] have previously been described. To generate pMOWS-Casper3-GR, the Casper3-GR cassette of the vector pCasper3-GR (evrogen, #FP971) was cut out with the restriction enzymes BamHI and NotI and subcloned into the retroviral expression vector pMOWS [58], in which the puromycin resistance cassette was replaced with a neomycin resistance cassette. Transfection of Phoenix ampho cells was performed by calcium phosphate precipitation. Transducing supernatants were generated 24 hours after transfection by passing through a 0. 45 μm filter and supplemented with 8 μg/ml polybrene (Sigma). Stably transduced H838 cells were selected in the presence of 1. 5 μg/ml puromycin (Sigma) 48 hours after transduction for H838-GFP and H838-HA-hEPOR and for the H838-HA-hEPOR-Casper3-GR cell line in the presence of additional 400 μg/ml G418 (Sigma). Surface expression of the EPOR in H838-HA-hEPOR cells was verified by flow cytometry. H838-HA-hEPOR cells were detached with Cell Dissociation Solution (Sigma) according to the manufacturer’s instructions and stained with anti-HA antibody (Roche) diluted 1: 40 in 0. 3% PBS/BSA for 20 min at 4°C. Cells were washed with 0. 3% PBS/BSA and incubated with secondary Cy5-labeled antibody against rat (Jackson Immuno Research), diluted 1: 100 in 0. 3% PBS/BSA, for 20 min at 4°C in the dark. After washing samples with 0. 3% PBS/BSA, propidium iodide (BD Biosciences) was added to exclude dead cells. Canto II (BD Bioscience) was used for sample analysis. The expression of the Casper3-GR sensor and GFP was verified by life cell imaging. All animal experiments were approved by the governmental review committee on animal care of the state Baden-Württemberg, Germany (reference number DKFZ215). At E13. 5 Balb/c mouse embryos were dissected from the uteri of female mice euthanized by CO2 inhalation. Fetal livers were resuspended in PBS/0. 3% BSA and passed through a 40-μm cell strainer (BD Biosciences). Fetal liver cells (FLCs) were treated with 9 ml Red Blood Cell Lysis Buffer (Sigma-Aldrich) to remove erythrocytes. For sorting TER119− erythroid progenitors, FLCs were incubated with rat antibodies against the following surface markers: GR1, CD41, CD11b, CD14, CD45R/B220, CD4, CD8 and Ter119 (BD Pharmingen), and 42. 2. 2 for 30 min at 4°C. After washing, cells were incubated for 30 min at 4°C with anti-rat antibody-coupled magnetic beads and negatively sorted with MACS columns according to the manufacturer' s instructions (Miltenyi Biotech). Sorted CFU-E cells were cultivated for 12–14 hours in Panserin 401 (PAN-biotech) supplemented with 50 μM β-mercaptoethanol and 0. 5 U/ml EPO alfa. Before the experiments, cells were growth factor-depleted for 60 min. Total RNA was extracted using the miRNeasy Mini Kit (QIAGEN) according to the manufacturer’s instructions. cDNA was generated from 1 μg of total RNA using High Capacity cDNA Reverse Transcription Kit (Applied Biosystems) according to the manufacturer’s instructions. cDNA templates were analyzed by quantitative Real-Time-PCR (qRT-PCR) on a LightCycler 480 (Roche Applied Science) cycler using the LightCycler 480 Probes Master with final 0. 4 μM primer and 0. 2 μM FAM-labeled hydrolysis probes (Universal Probe Library, Roche Applied Science). Crossing point values were calculated using the second-derivative-maximum method of the LightCycler 480 Basic Software (Roche Applied Science). Quantitative RT-PCR efficiency correction was performed for each setup individually. Concentrations were normalized using the geometric mean of β-glucuronidase (GUSB) and esterase D (ESD) for the NSCLC cell line and its derivative and hypoxanthine-guanine phosphoribosyltransferase (HPRT) for the CFU-E cells. Primers were designed using the UniversalProbe Library Assay Design Center (Roche Applied Science). UPL Probes and primer sequences for murine samples were CISH_for 5′-gacatggtcctttgcgtaca-3′ CISH_rev 5′-atgccccagtgggtaagg-3′ probe#1, SOCS3_for 5′- gctggtactgagccgacct-3′ SOCS3_rev 5′-aacttgctgtgggtgaccat-3′ probe#83, and HPRT_for 5′-tcctcctcagaccgctttt-3′ HPRT_rev 5′-cctggttcatcatcgctaatc-3′ probe#95. UPL Probes and primer sequences for human samples were CISH_for 5′-agccaagaccttctcctacctt-3′ CISH_rev 5′-tggcatcttctgcaggtgt-3′ probe#20, SOCS3_for 5′-agacttcgattcgggacca-3′ SOCS3_rev 5′-aacttgctgtgggtgacca-3′ probe#36, ESD_for 5′-ttagatggacagttactccctgataa-3′ ESD_rev 5′-ggttgcaatgaagtagtagctatgat-3′ probe#27 and GUSB_for 5′-cgccctgcctatctgtattc-3′ GUSB_rev 5′-tccccacagggagtgtgtag-3′ probe#57. H838 and H838-HA-hEPOR cells were seeded in a 96-well plate at a density of 10 000 cells/well for three days. Before the experiment, cells were growth factor-depleted for 14–16 hours. To measure the viability of cells, CellTiter-Blue Viability Assay (Promega) was applied according to the manufacturer’s instructions. Incubation with the dye for 60 min was followed by measurement of the fluorescence with the infinite F200 pro Reader (TECAN). A blank well containing culture medium but no cells was measured as background. For the caspase-3 activity assay, H838 cells overexpressing HA-hEPOR and additionally expressing the FRET-based Casper3-GR sensor (H838-HA-hEPOR-Casper3-GR) were seeded in 8-well plates at a density of 20 000 cells/well. 24 hours after seeding, cells were growth factor-depleted for 16 hours and then treated with 5 mg/l cisplatin (Teva) or left untreated. Cells were imaged on an environment-controlled microscope (Zeiss LSM 710) and GFP and RFP intensity was determined. Images were acquired every 20 min for 64 hours. H838-GFP cells were amplified to a density of 80% and growth factor-depleted for 14–16 hours. The cells were trypsinized with 0. 025% Trypsin/EDTA/PBS (Invitrogen) for 5 min and resuspended in DMEM (Lonza) and Hoechst (H33342) (final concentration 1 μg/ml) was added. The cells were imaged on an environment-controlled microscope (Zeiss LSM 710) and the data was analyzed with Fiji software [60]. For the detection of the EPOR, JAK2 and STAT5,800 000 H838 and H838-HA-hEPOR cells, respectively, were seeded three days in advance in a 10-cm plate and washed three times with DMEM without additives and then kept for 3 hours in DMEM with 1% penicillin/streptomycin, 2 mM L-glutamine (Gibco) and 1 mg/ml BSA. The cells were stimulated with 10 U/ml EPO beta (Roche) and lysed with 500 μL 1. 25x NP-40 lysis buffer (1. 25% NP-40,187. 5 mM NaCl, 25 mM Tris pH 7. 4,12. 5 mM NaF, 1. 25 mM EDTA pH 8. 0,1. 25 mM ZnCl2 pH 4. 0,1. 25 mM MgCl2,1. 25 mM Na3VO4,12. 5% glycerol) supplemented with aprotinin and AEBSF (Sigma). The cell debris was removed by centrifugation and the supernatant was used for determination of protein concentration (BCA Protein Assays, Thermo Fisher). Immunoprecipitations (IP) were performed consecutively with first antibodies against hEPOR (MAB 307, R&D) and JAK2 (06–255, Merck Millipore) and then with antibodies against STAT5A/B (C-17, Santa Cruz) using protein A sepharose beads. The immunoprecipitates were loaded to a 10% polyacrylamide gel and transferred to a nitrocellulose membrane (Schleicher & Schuell). The membranes were blocked with 5% BSA for 1 hour and successively incubated with a phosphotyrosine antibody (4G10, Merck Millipore). To remove antibodies, membranes were treated with β-mercaptoethanol and SDS and subsequently incubated with an anti-hEPOR antibody (C-20, Santa Cruz), an anti-JAK2 antibody (06–255, Merck Millipore) or an anti-STAT5 antibody (C-17, Santa Cruz). Secondary horseradish peroxidase-coupled antibodies were obtained from GE Healthcare or Dianova. Detection was performed using ECL substrate (GE Healthcare) and acquired with the CCD camera-based ImageQuant LAS 4000 (GE Healthcare). For quantification, the ImageQuant TL version 7. 0 software (GE Healthcare) was used. For the detection of mEPOR in mCFU-E cells, the cells were growth factor-depleted for 3 hours and 1×107 cells shaking in 250 μl were stimulated for 10 min with 5 U/ml EPO alfa (Janssen-Cilag). The cells were lysed by addition of 250 μL 2x NP40 lysis buffer. IP was performed using an anti-mEPOR antibody (M-20, Santa Cruz) and protein A sepharose. For phosphorylated mEPOR, the phosphotyrosine antibody (4G10, Merck Millipore) was used. For total mEPOR anti-mEPOR antibody (M-20, Santa Cruz) was applied. For determination of protein abundances different recombinant fusion proteins were synthesized using a pGEX-2T vector (GE Healthcare). GSTΔhEPOR and GSTΔmEPOR consist of the complete cytoplasmic part of the respective receptor fused N-terminally to a GST tag leading to proteins of 52 214 Da mass for hEPOR and 52 309 Da for mEPOR. GSTΔJAK2 and GSTΔSTAT5 were constructed as described previously [61]. Briefly, in GSTΔJAK2 the tag is fused N-terminally to the kinase domain of murine JAK2 starting from L549 leading to a 94 572 Da fusion protein with 95% consensus to human JAK2. GSTΔSTAT5 consist of the N-terminal end of murine STAT5B starting from F332 which leads to a fusion protein of 78 432 Da with 97% consensus to hSTAT5B. The concentrations of the recombinant proteins were determined using a BSA standard curve on a Coomassie-stained SDS-PAGE gel (SimplyBlue SafeStain, Invitrogen). Different amounts of the respective calibrators were added to the cell lysate. IP and immunoblotting was performed with the indicated total antibodies. The linear calibration curve based on the intensities of the recombinant protein was estimated with SigmaPlot (V12. 5) and the endogenous signal was interpolated to calculate the corresponding number of molecules. The cell number was counted in parallel with a Neubauer improved counting chamber. For mass spectrometry experiments, at least 1. 5×107 H838 or H838-HA-hEPOR cells were used per time point. Treatment, lysis and IP conditions are described above. IPs were separated by 10% SDS-PAGE and gels were stained with SimplyBlue SafeStain (Invitrogen). STAT5A (90. 6 kDa) and STAT5B (89. 9 kDa) containing bands were excised between around 75 and 100 kDa (according to Precision Plus Protein marker, Bio-Rad) then destained, reduced (DTT, Sigma), alkylated (IAA, Sigma) and digested with trypsin gold (Promega). To ensure robust and accurate degree of phosphorylation analysis of Tyr694 (STAT5A) and Tyr699 (STAT5B) we applied an internal standard peptide mixture (One-source peptide-/phosphopeptide ratio standard) with defined ratio of labelled peptide and its phosphorylated counterpart. Starting from the in-house synthesized stable isotope labelled phosphopeptide A-[V+6Da]-D-G-pY-V-K-P-Q-I-K (identical for STAT5A and STAT5B) the standard mixture was generated by (i) dividing peptide dilution into two aliquots (ii) quantitative dephosphorylation of one of the aliquots (iii) remixing of dephosphorylated and untreated aliquot (described in detail [62]). The standard was spiked into the samples during tryptic digestion. All mass spectrometry sample preparation steps from lysis to standard addition have been previously described in more detail [63]. Following peptide extraction and concentration of eluates, samples were desalted and purified using C18 ZipTips (Merck Millipore). Samples were measured by nanoUPLC (nanoAcquity, Waters) coupled to an LTQ-Orbitrap XL mass spectrometer (Thermo Fisher). LC separations were performed on a 75 μm × 150 mm C18 column with 1. 3 μm particle size (Waters) using a water/acetonitrile based gradient up to 40% acetonitrile within 60 min. Intensities of native and labelled STAT5 peptide and phosphopeptide pairs were analyzed manually using Xcalibur 3. 0. 63 (Thermo Fisher). Genomatix Genome Analyzer was used for promoter analysis (www. genomatix. de). Promoter retrieval was performed by Genomatix Gene2Promoter (ElDorado 12–2013) and the transcription factor binding site analysis was performed using the Genomatix Overrepresented TFBS pipeline combining Genomatix MatBase and MatInspector based on Matrix Family Library Version 9. 1. Sequence alignment was performed using the Genomatix Multiple Alignment pipeline based on DiAlign professional TF Release 3. 1. 5 (June 2011). H838 and CFU-E cells were cultivated as described above. Genomic DNA was extracted with the DNeasy Tissue Kit (Qiagen) according to the manufacturer' s instructions and bisulfite converted using the EZ DNA Methylation kit (Zymo Research). Amplicons spanning the entire SOCS3 promoter region were designed and used to amplify this region from bisulfite-treated DNA (S2 Table). For MassARRAY EpiTYPER assay (Agena Bioscience), the PCR products were transcribed in vitro, cleaved by RNase A and subjected to matrix-assisted laser desorption ionization time-of-flight mass spectrometry to quantitatively assess methylation levels of CpG dinucleotides [64]. The ODE system consisted of 23 states, with 20 initial values that were implicitly dependent on kinetic parameters through steady-state assumptions, two initial values with prior knowledge available, and one initial value with prior knowledge only available for CFU-E, which was estimated for H838 & H838-HA-hEPOR. To obtain a numerical solution of the ODE system the solver CVODES was applied [65]. For CFU-E cells, 23 model variants representing experimental conditions were implemented, and 36 for H838 & H838-HA-hEPOR cells, respectively. For numerical efficiency, the ODE system and solver were compiled as C-executables, supplying calculation of states as well as sensitivities, and solved in a parallelized manner for all conditions. Data2Dynamics [31] was utilized to facilitate the automatic derivation of conditions and sensitivity equations. The ODE system was re-parameterized to disentangle internal concentrations from relative measurements and to decouple modules of the signaling network with different scales. Relative and absolute tolerances of the ODE solver were set to 10−6. For parameter estimation, all three cell types CFU-E, H838 and H838-HA-hEPOR were initially implemented separately. To achieve global optimization, a multi-start deterministic optimization strategy was used [37] with 1000 initial parameter vectors for each cell type. Each single optimization was performed using the MATLAB implementation of the trust-region method (lsqnonlin) [66]. The algorithm was set to terminate an optimization run if the proposed step size is smaller than 10−6. To account for flat regions in the parameter space, the termination criterion based on change of the objective function was omitted. A noise parameter was assumed for each experimental technique, observable and cell type. For intensity-based measurements, a log-normal error model was used, where σ is relative to the observation [67]. For degree of phosphorylation data obtained by mass spectrometry, a constant error model was assumed. Kinetic parameters, initial concentrations, observation parameters and error model parameters were estimated simultaneously based on the data [37]. Parameters were estimated in log-space and values between 10−5 and 103 were initially allowed. If an estimate was located at the boundary of the parameter space, the respective boundaries were enlarged. Efficiencies of inhibitors were limited to a maximum of one (100% efficiency), and the accuracy of degree of phosphorylation data obtained by mass spectrometry was restricted to a maximum of 5%. For CFU-E cells, 516 data points were used to estimate a total of 109 parameters. 1000 parameter estimation runs were started from random points of the parameter space. The global optimum was found in 9. 0% of parameter estimation runs. Likewise for H838,204 data points were used to estimate 54 parameters. Here, the global optimum was found in 5. 2%. For H838-HA-hEPOR, 407 data points were used to estimate 75 parameters with the global optimum reached in 0. 7% of all optimization runs. These results suggested that the global optimum was discovered for all cell types. The comprehensive model taking all cell types into account had a total of 216 parameters that were estimated based on 1141 data points. The parameters were estimated independently for CFU-E and H838 & H838-HA-hEPOR cells, and the corresponding fold-change was L1-regularized. Residuals and sensitivities were passed to the optimizer. To account for discontinuities of the L1 regularization gradient, optimization steps were truncated in the derivatives by preventing to cross zero at each step. To facilitate convergence, derivatives of the L1 regularization residuals were set to zero if their contribution was smaller than the respective gradient of the log-likelihood. The re-parameterization to decouple modules of the signaling network was performed with CFU-E as reference point for H838 & H838-HA-hEPOR. For generating the regularization path, the regularization strength was scanned in log-space from 10−4 to 104. Regularization was only used for selection of parameter differences, while for model selection the un-regularization solution was used for unbiased parameter estimates. The threshold for model selection was calculated using likelihood ratio test statistics (α = 0. 05), i. e. by the inverse cumulative density function of the χ2-distribution with the difference of number of parameters between full and reduced model as degrees of freedom. The profile likelihood was calculated to assess the standard deviation of the parameter ratios [35]. To conservatively estimate the standard deviation based on asymmetric confidence intervals, the maximum of the log10 difference between maximum likelihood estimate and lower/upper limit of the confidence interval was used to approximate σ. Modeling framework, examples and data are open source and publicly available at www. data2dynamics. org.

A major challenge in the development of therapeutic interventions is the selective inhibition of a signal transduction pathway in one cell type such as a cancer cell leaving the other cell type such as a healthy cell as unaffected as possible. Here, we propose a new approach that combines mathematical modeling based on quantitative experimental data with statistical methods. We demonstrate based on simulated data that our approach can determine which parameters are the same and which parameters differ in two exemplary cell types. We compare a lung cancer cell line to the precursor cells of red blood cells. We show that the same signal transduction network induced by erythropoietin (EPO), a hormone that is frequently employed to treat anemia in cancer patients, regulates survival of both cell types. Based on our experimental data in combination with our computational approach, we identify seven cell type-specific differences in this signaling pathway. Our strategy allows predicting therapeutic targets that could be inhibited to interfere with survival of lung cancer cells while leaving production of red blood cells unaffected.

Abstract Introduction Results Discussion Materials and Methods

phosphorylation medicine and health sciences molecular probe techniques messenger rna cancers and neoplasms immunoblotting simulation and modeling oncology immunoprecipitation molecular biology techniques research and analysis methods proteins lung and intrathoracic tumors stat signaling molecular biology precipitation techniques biochemistry rna signal transduction cell biology post-translational modification nucleic acids biology and life sciences non-small cell lung cancer cell signaling

2016

Identification of Cell Type-Specific Differences in Erythropoietin Receptor Signaling in Primary Erythroid and Lung Cancer Cells

Selenocysteine (Sec) is cotranslationally inserted into protein in response to UGA codons and is the 21st amino acid in the genetic code. However, the means by which Sec is synthesized in eukaryotes is not known. Herein, comparative genomics and experimental analyses revealed that the mammalian Sec synthase (SecS) is the previously identified pyridoxal phosphate-containing protein known as the soluble liver antigen. SecS required selenophosphate and O-phosphoseryl-tRNA[Ser]Sec as substrates to generate selenocysteyl-tRNA[Ser]Sec. Moreover, it was found that Sec was synthesized on the tRNA scaffold from selenide, ATP, and serine using tRNA[Ser]Sec, seryl-tRNA synthetase, O-phosphoseryl-tRNA[Ser]Sec kinase, selenophosphate synthetase, and SecS. By identifying the pathway of Sec biosynthesis in mammals, this study not only functionally characterized SecS but also assigned the function of the O-phosphoseryl-tRNA[Ser]Sec kinase. In addition, we found that selenophosphate synthetase 2 could synthesize monoselenophosphate in vitro but selenophosphate synthetase 1 could not. Conservation of the overall pathway of Sec biosynthesis suggests that this pathway is also active in other eukaryotes and archaea that synthesize selenoproteins. Selenocysteine (Sec) is a selenium-containing amino acid that is cotranslationally inserted into protein and is recognized as the 21st amino acid in the genetic code [1–3]. Sec is incorporated into protein in all three lines of descent, eukaryota, archaea, and eubacteria, but unlike other amino acids, Sec synthesis occurs on its transfer RNA (tRNA), designated tRNA[Ser]Sec [4,5]. tRNA[Ser]Sec is initially aminoacylated with serine by seryl-tRNA synthetase and the seryl moiety provides the backbone for Sec synthesis. The biosynthesis of Sec was established in Escherichia coli in the early 1990s [6–8]. Bacterial Sec synthase (SecS) (E. coli selenocysteine synthase [SelA]) is a pyridoxal phosphate (PLP) -dependent protein that converts the serine attached to tRNA[Ser]Sec to Sec by initially removing the hydroxyl group from serine to form an aminoacrylyl intermediate. This intermediate serves as the acceptor for activated selenium, and when selenium is donated, selenocysteyl-tRNA[Ser]Sec is formed. The active selenium donor in bacteria is synthesized from selenide and ATP by E. coli selenophosphate synthetase (SelD), and the product of the reaction has been identified as monoselenophosphate (SeP) [9]. A distant homolog of bacterial SelA (SelA-like) is present in some archaea but is not active as SecS [10], and it does not always co-occur in archaea with Sec insertion systems. In addition, no SelA sequences could be detected in eukaryotes. Although Sec insertion systems are different in bacteria from those in archaea and eukaryotes [11–13], several factors have been characterized in mammals that most certainly have a role in Sec biosynthesis. For example, the soluble liver antigen (SLA) was initially identified as a 48-kDa protein bound to Sec tRNA[Ser]Sec that was targeted by antibodies in patients with an autoimmune chronic hepatitis [14]. SLA was subsequently reported to exist as a separate family within a larger superfamily of diverse PLP-dependent transferases [15], and this protein has been proposed to function as the mammalian SecS (e. g. , see [3,15–17]). Further evidence that SLA is involved in selenium metabolism is that it was found to occur in a protein complex with other factors involved in the biosynthesis of Sec and/or its insertion into protein [17,18]. In addition, a kinase that phosphorylated a minor seryl-tRNA was reported in 1970 [19] that was subsequently isolated, characterized, and found to specifically phosphorylate the seryl moiety on seryl-tRNA[Ser]Sec [20]. The resulting phosphoseryl-tRNA[Ser]Sec was proposed either as a candidate substrate for SecS (see [3,20] and references therein) or it served as a storage form [21]. Furthermore, two genes initially thought to have a role in selenophosphate synthesis, sps1 and sps2, have been reported in mammals [22–25], and the product of sps2 is a selenoprotein, SPS2 [22,24]. The Sec-to-Cys mutant form of SPS2 has low enzyme activity [22,24,26] and can complement SelD in Escherichia coli cells transfected with the mammalian mutant form [26]. Complementation of SelD− E. coli cells with SPS1 or SPS2 has suggested that SPS1 may have a role in recycling Sec via a selenium salvage system and SPS2 may be involved in the de novo synthesis of selenophosphate from selenide [27]. However, it should be noted that, to our knowledge, selenophosphate has never been shown to serve directly as the active selenium donor in Sec biosynthesis in eukaryotes. Herein, we used a comparative genomics search and experimental analyses to show that SLA is the mammalian SecS. This protein belongs to a different family of PLP-containing enzymes and uses O-phosphoseryl-tRNA[Ser]Sec rather than seryl-tRNA[Ser]Sec as substrate. SecS dephosphorylates O-phosphoseryl-tRNA[Ser]Sec and accepts the active selenium donor to yield selenocysteyl-tRNA[Ser]Sec. We also demonstrated unequivocally that the selenium donor in eukaryotes is SeP by using this compound as a substrate in a reaction with SecS and phosphoseryl-tRNA[Ser]Sec. Selenophosphate is indeed synthesized in mammals by SPS2, whereas the distant homolog of SelD in mammals, SPS1, did not synthesize the active selenium donor. Conservation of the overall pathway of Sec biosynthesis suggests that it is also active in other eukaryotes and archaea. We analyzed completely sequenced genomes of eukaryotes and archaea for the occurrence of selenoproteins. Twenty-six eukaryotes and three archaea that had these proteins and 24 eukaryotes and 24 archaea that did not were identified. Comparative genomics studies were then carried out to identify genes that co-occur with selenoproteins in (1) eukaryotes (Table 1) and (2) archaea (Table 2). Each of the searches had known components of Sec insertion machinery as top candidates as well as an additional protein, herein designated as SecS. In mammals, SecS is also known as SLA. SLA was first detected as an autoimmune factor that coimmunoprecipitated tRNA[Ser]Sec from cell extracts in patients with autoimmune chronic hepatitis [14], and it also bound other Sec insertion components [17,18]. SecS formed a separate family within a larger superfamily of diverse PLP-dependent proteins and was previously suggested to convert a tRNA-bound serine to Sec [15]. We found that it occurred exclusively in both eukaryotes and archaea which had selenoproteins but was lacking in the other organisms examined (Figure 1 and Tables 1 and 2). These observations strongly suggested that SecS may be the missing SecS in eukaryotes and archaea. Based on the multiple sequence alignment and phylogenetic analysis of SecS and other PLP-dependent proteins, including SelA, SelA-like, and SepCysS, it was clear that bacterial SelA and archaeal SelA-like proteins [10], on one hand, and SecS, on the other, belonged to completely different families of PLP-containing proteins (Figures 1 and 2), suggesting that their similar functions arose by convergent evolution. SecS was also distantly homologous to SepCysS, a protein recently found to synthesize cysteine from phosphoserine in some archaea [28] (Figures 1 and 2). After identifying a likely SecS candidate by comparative genomics analysis, we experimentally verified its function as described below. To elucidate Sec biosynthesis in mammals, we initially examined the ability of tRNA[Ser]Sec, seryl-tRNA[Ser]Sec, and O-phosphoseryl-tRNA[Ser]Sec to bind to the recombinant mouse SecS (mSecS). The coprecipitated product was detected by Northern blotting (Figure 3A) and the amount of binding was quantitated (Figure 3B). O-Phosphoseryl-tRNA[Ser]Sec bound more efficiently to mSecS than the other tRNA[Ser]Sec forms, while seryl-tRNA[Ser]Sec bound least efficiently, suggesting that O-phosphoseryl-tRNA[Ser]Sec may be a substrate for mSecS. It is not clear why tRNA[Ser]Sec binds to mSecS (see also [18]), albeit less efficiently than O-phosphoseryl-tRNA[Ser]Sec. Seryl-tRNASer and tRNASer, however, did not manifest any binding to mSecS (unpublished data). To assess whether the phosphate moiety on O-phosphoseryl-tRNA may be removed by mSecS to generate an intermediate that serves as an acceptor for the active selenium donor, the 32P-labeled form of O-phosphoseryl-tRNA[Ser]Sec was incubated with mSecS (Figure 3C). mSecS removed the phosphoryl moiety from O-phosphoseryl-tRNA[Ser]Sec (see lane 2). Interestingly, SelA was also capable of dephosphorylating O-phosphoseryl-tRNA[Ser]Sec (lane 3). Neither mSPS2-Cys [mouse selenophosphate synthetase 2 containing an Sec (UGA) -to-Cys (UGC) mutation] nor SelD appeared to have any effect on O-phosphoseryl-tRNA[Ser]Sec (lanes 4 and 5). The dephosphorylation of O-phosphoseryl-tRNA[Ser]Sec by mSecS and SelA is further considered below. However, it should be noted that the data in Figure 3 strongly suggest that the dephosphorylated product is not seryl-tRNA[Ser]Sec as the product binds efficiently to mSecS but seryl-tRNA does not (Figure 3B). We next identified the active selenium donor by assessing whether mSPS1 and mSPS2-Cys synthesized SeP. 31P NMR spectroscopic analysis of the products of the mSPS2-Cys–catalyzed reaction manifested a signal at +23. 2 ppm, albeit weakly (Figure 4A1), that corresponded to SeP [9,29]. Since the mSPS2 used in this experiment was a Sec-to-Cys mutant and might not be expected to generate SeP efficiently, we cloned SPS2 from Caenorhabditis elegans, which naturally contains Cys in place of Sec at the presumed active site of SPS2 [13]. C. elegans SPS2 clearly generated a signal at +23. 2 ppm (Figure 4A2). As expected, SeP was also formed in the presence of E. coli SelD, selenide, and ATP (Figure 4A3) [9,29]. However, no signal at +23. 2 ppm was observed when mSPS1 replaced mSPS2-Cys or SelD in the reaction, indicating that mSPS1 did not synthesize SeP (Figure 4A4). As the peak at +23. 2 ppm was relatively weak in the product analysis of mSPS2-Cys (Figure 4A1), the ordinate and abscissa of the area between 15 and 30 ppm were expanded as shown in Figure 4B. Clearly, there was a peak at +23. 2 ppm corresponding to SeP, demonstrating that mSPS2-Cys produced SeP. The signal for SeP was also evident with C. elegans SPS2 and with SelD, but mSPS1 did not produce this signal. To further examine the hydrolysis of ATP by mSPS2-Cys, C. elegans SPS2, SelD, and mSPS1, each component was incubated with [α-32P]ATP with and without selenide (Figure 4C). Hydrolysis of ATP to AMP was largely dependent on the presence of selenide with the three enzymes, mSPS2-Cys, C. elegans SPS2, and SelD, that produced SeP (see above), and all three hydrolyzed ATP to ADP independently of selenide. Although mSPS1 hydrolyzed ATP to ADP and apparently only slightly to AMP, this degradation was independent of selenium. These data provide further evidence that mSPS1 cannot synthesize SeP from selenide. Previous studies analyzing Sec biosynthesis did not utilize SeP to assess whether this compound served directly as the active selenium donor. We therefore examined the ability of SeP to donate selenium directly in Sec biosynthesis. Sec was indeed synthesized when SeP [9,29] was added in the reaction with O-phosphoseryl-tRNA[Ser]Sec and mSecS (Figure 5A). This observation confirms unequivocally that SeP is the active donor of selenium in Sec biosynthesis and that SecS is the missing SecS. Control assays demonstrated that Sec was not formed when SeP was omitted from the reaction, when seryl-tRNA was used in place of O-phosphoseryl-tRNA[Ser]Sec, or when another protein, thioredoxin (Trx), was substituted for mSecS in the reaction. As expected, a reaction consisting of SelA, seryl-tRNA[Ser]Sec, and SeP also synthesized Sec (Figure 5A). O-Phosphoseryl-tRNA[Ser]Sec could also serve as a substrate and replace seryl-tRNA[Ser]Sec in reactions with SelA and SeP, thus using the dephosphorylated product as an acceptor for activated selenium to synthesize selenocysteyl-tRNA[Ser]Sec. Sec was also synthesized on tRNA[Ser]Sec when O-phosphoseryl-tRNA[Ser]Sec was incubated with mSecS, mSPS2-Cys, ATP, and selenide (Figure 5B). Control reactions demonstrated that Sec was not formed when selenide was omitted from the reaction, when seryl-tRNA[Ser]Sec was used in place of O-phosphoseryl-tRNA[Ser]Sec, or when Trx was substituted for mSecS (Figure 5B). SelA would substitute for mSecS when the substrate was seryl-tRNA[Ser]Sec or O-phosphoseryl-tRNA[Ser]Sec. As expected, SelD could substitute for mSPS2-Cys in synthesizing SeP and generating similar amounts of Sec as mSPS2-Cys in reactions with mSecS (unpublished data), and those reactions in Figure 5B were dependent on ATP as well as selenide wherein Sec was generated (unpublished data). Serine, alanine, and pyruvate were recovered from reactions with SelA using either seryl- or O-phosphoseryl-tRNA[Ser]Sec as substrates, wherein alanine and pyruvate were likely the deacylated, degraded products of the intermediate, aminoacrylyl-tRNA[Ser]Sec [7] (Figures 5A and 5B). In reactions with mSecS using O-phosphoseryl-tRNA[Ser]Sec as substrate, only a small amount of phosphoserine and a peak that comigrated with pyruvate were recovered as deacylated products, suggesting that pyruvate was, similar to the bacterial case, the deacylated, degraded product of the intermediate in Sec biosynthesis in eukaryotes. However, the amount of pyruvate recovered in reactions that were coupled with SeP synthesis by selenophosphate synthetase was lower (Figure 5B) than in reactions in which SeP was supplied directly as substrate (Figure 5A). Use of HSe− in these reactions required the addition of high levels of dithiothreitol (DTT) which were inhibitory to Sec synthesis (unpublished data), and apparently the intermediate that formed pyruvate as a deacylated product was unstable under these conditions. The intermediates in reactions with SelA [7] and mSecS are further considered in Discussion. We also examined the rate of Sec synthesis with O-phosphoseryl-tRNA[Ser]Sec and SeP as substrates in the presence of mSecS (Figure 5C). As the substrate, O-phosphoseryl-tRNA[Ser]Sec, was labeled with [3H]serine and the deacylated products, O-phosphoserine, Sec and the degraded intermediate, pyruvate, migrated separately in the chromatographic system used in Figure 5A; the amounts of each could be assessed during the course of the reaction. Dephosphorylation occurred rapidly and appeared to be near completion in about 10 min. Sec synthesis increased rapidly during the initial 10 min and then appeared to proceed more slowly until completion at about 40 min. The approximate initial rate was 0. 28 pmol Sec/min/pmol mSecS. Likewise, the intermediate formed rapidly during the initial stages of the reaction and then decreased over the remainder of the experiment. In this work, we defined the pathway of Sec biosynthesis in eukaryotes. In order to carry out Sec biosynthesis, we functionally characterized two previously known enzymes, selenophosphate synthetase [22,23,25] and O-phosphoseryl-tRNA[Ser]Sec kinase [20], as well as establishing the function of SLA [14] as the eukaryotic SecS. All these enzymes were found to be required for in vitro biosynthesis of Sec, and the implications of these findings are discussed below. The active selenium donor in bacteria is SeP [9,29] and it is synthesized from selenide and ATP by selenophosphate synthetase, also known as SelD [7,30]. Two homologs of bacterial SelD, designated SPS1 and SPS2, are present in mammals [22–25]. Interestingly, SPS2 is a selenoprotein. Direct roles of SPS1 and SPS2 in mammals have not been tested, but it was suggested that SPS2 supports the use of selenite, whereas SPS1 depends on a selenium salvage system when examined in E. coli [27]. Our results clearly demonstrate that SPS2 makes the active selenium donor, SeP, for the biosynthesis of Sec. We recently identified the gene that phosphorylates seryl-tRNA[Ser]Sec and characterized the gene product, O-phosphoseryl-tRNA[Ser]Sec [20]. However, the precise role of O-phosphoseryl-tRNA[Ser]Sec was not determined. The current results showed that O-phosphoseryl-tRNA[Ser]Sec is the substrate of SecS, and therefore O-phosphoseryl-tRNA[Ser]Sec kinase is involved in the Sec biosynthesis pathway. What is the intermediate produced by mSecS? Most certainly the dephosphorylated product of O-phosphoseryl-tRNA[Ser]Sec cannot be seryl-tRNA[Ser]Sec (Figure 3A). The intermediate could possibly be aminoacrylyl-tRNA[Ser]Sec (dehydroalanyl-tRNA[Ser]Sec) which could yield pyruvate on hydrolysis. The facts that the intermediate generated by SelA is aminoacrylyl-tRNA[Ser}Sec and that mSecS contains pyridoxal phosphate suggest that the Schiff base intermediate of aminoacrylyl-tRNA[Ser}Sec postulated for Sec synthesis in prokaryotes [7] is analogous to that formed in eukaryotes. However, it proved possible to trap the proposed aminoacrylyl-tRNA[Ser]Sec by reduction with KBH4 leading to the formation of alanine on hydrolysis in the prokaryotic, but not eukaryotic, case. This result may be due to differences between the enzyme to which the aminoacrylyl-tRNA[Ser]Sec is bound; that is, reduction can occur before hydrolysis in the prokaryotic, but not eukaryotic, case. Nevertheless, identification of the intermediate in Sec biosynthesis in eukaryotes must await further study. The biosynthesis of Sec in eukaryotes is shown in Figure 6. tRNA[Ser]Sec is aminoacylated by seryl-tRNA synthetase and the seryl moiety is phosphorylated by O-phosphoseryl-tRNA[Ser]Sec kinase to form O-phosphoseryl-tRNA[Ser]Sec [20]. O-Phosphoseryl-tRNA[Ser]Sec is a substrate for SecS which replaces the phosphoryl moiety of phosphoserine, derived from the selenium donor, SeP, to yield Sec. SeP is synthesized by SPS2 in the ATP-dependent reaction. SecS does not use seryl-tRNA[Ser]Sec as a substrate (Figure 5A and 5B). Although no enzyme comparable to O-phosphoseryl-tRNA[Ser]Sec kinase has been identified in E. coli, it is of interest to note that SelA can utilize O-phosphoseryl-tRNA[Ser]Sec as a substrate. The major difference between the Sec biosynthetic pathway characterized herein and that in eubacteria is the extra step in the synthesis of O-phosphoseryl-tRNA[Ser]Sec which serves as a substrate for SecS. In E. coli, seryl-tRNA[Ser]Sec serves directly as the substrate for SelA [7]. The occurrence of SecS exclusively in selenoprotein-containing organisms in eukaryotes and archaea (Tables 1 and 2) indicates that the SecS-based pathway also operates in other animals, lower eukaryotes, and archaea where the Sec machinery occurs [3]. Considering the difficulties with identification of other components of Sec biosynthesis and insertion machinery (e. g. , SBP2, EFsec), SecS might become the most characteristic feature of the Sec trait in eukaryotes and archaea. [α-32P]ATP and [γ-32P]ATP (specific activity, approximately 6,000 Ci/mmol) and Hybond N+ nylon membranes were purchased from Amersham (http: //www. amersham. com), 3H-serine (specific activity, 29. 5 Ci/mmol) and 14C-pyruvate (specific activity, 19 mCi/mM) were from Perkin Elmer (http: //www. perkinelmer. com), Ni-NTA agarose was from Qiagen (http: //www. stratagene. com), and pfu polymerase and pBluescript II were from Stratagene (http: //www. stratagene. com). pET32b vector (encoding the 109–amino acid thioredoxin with a His-tag) and BL21 (DE3) competent cells were obtained from Novagen (EMD Biosciences, http: //www. emdbiosciences. com), alkaline phosphatase from New England Biolabs (http: //www. neb. com), T7 RiboMAX Express Large Scale RNA Production System and 3M filter paper from Whatman (http: //www. whatman. com), and unlabeled amino acids, PEI TLC plates, and selenocystine from Sigma (http: //www. signaaldrich. com). [ (CH3) 3SiO]3PSe was chemically synthesized [31]. All other reagents were commercial products of the highest grade available. A total of 50 completely or nearly completely sequenced eukaryotic genomes (Table S1) were analyzed for occurrence of selenoproteins by TBLASTN using the set of all known selenoproteins. Twenty-six organisms were found to contain selenoproteins and 24 organisms lacked these proteins. In organisms lacking selenoproteins, the Sec insertion machinery was also missing. To identify proteins with phylogenetic profiles corresponding to selenoproteins, annotated genes in D. melanogaster were used as a query dataset. BLAST homology analyses were used to scan genomic databases using the following criteria: E-value less than 1e−06 and length of the conserved region greater than 50 amino acids. Genes present in any of the organisms lacking selenoproteins were dismissed. The remaining genes were searched against selenoprotein-containing organisms to determine their occurrences. Top candidate genes were included in Table 1; they were present in 80% of selenoprotein-containing eukaryotes. These candidates were further manually analyzed for possible function. A similar search strategy was carried out in archaea. A total of 27 archaeal genomes (Table S1), including three selenoprotein-containing organisms (M. jannaschii, M. maripaludis, and M. kandleri), were analyzed for gene occurrence using all annotated genes in M. jannaschii. Top candidate genes are included in Table 2; these proteins were present in all three selenoprotein-containing archaea and absent in other completed archaeal genomes. We used ClustalW to generate multiple sequence alignment. Phylogenetic trees were built with PHYLIP programs. The coding regions of E. coli selA and selD, mouse secS, sps2, sps1, and C. elegans sps2 genes were amplified from BL21 genomic DNA or mouse liver cDNA or C. elegans total cDNA using pfu polymerase, respectively [20]. The resulting product was cloned into the pET32b vector at the NdeI-XhoI cloning sites in which the vector contained a His-tag immediately downstream of, and in frame with, the open reading frame. The Sec TGA codon in sps2 was mutated to a Cys TGC codon using a site-directed mutagenesis kit (Stratagene), and the resulting gene product was designated mSPS2-Cys. The cDNA constructs were confirmed by sequencing and transformed into BL21 (DE3) cells. Expression and purification of each protein were carried out as described [20]. For mSecS and SelA expression and purification, 1 nM PLP was added in the LB medium during expression and 5 μM PLP was added in the elution buffer during purification. The proteins were dialyzed against 1× TBS for 2 h and stored at −20 °C in 50% glycerol before use. Native tRNA[Ser]Sec was purified and aminoacylated with serine and the seryl moiety phosphorylated as described [20]. Then 200 ng of purified mSecS containing a His-tag on its C-terminal was added in a total volume of 100 μl solution (20 mM Tris-HCl [pH 7. 4], 0. 01 mM EGTA, 1 mM DTT, 10 mM MgCl2, and 5 μg of yeast tRNA) and approximately 50 ng of purified tRNA[Ser]Sec (with either serine or phosphoserine attached, or no amino acid) added, and the reaction was incubated for 30 min at room temperature. Anti-His agarose (10 μl) was added to pull down mSecS. After washing three times with 1 ml of 1× TBS/0. 1% Tween, the agarose was suspended in 40 μl of TBE-urea loading buffer (90 mM Tris-HCl [pH 8. 3], 64. 6 mM boric acid, 2. 5 mM EDTA, 3. 5 M urea), and 5 μl of each sample was loaded onto a 15% TBE-urea gel. After electrophoresis and transfer of the RNA to a nylon membrane, RNA was detected by Northern blotting with the Sec tRNA probe [20]. Of each binding reaction, 2 μl had been removed immediately after the incubation period and electrophoresed along with the reaction samples for analysis by Northern blotting that served as a loading control. Native tRNA[Ser]Sec was aminoacylated with serine and phosphorylated with [γ-32P]ATP using O-phosphoseryl-tRNA[Ser]Sec kinase as described [20]. The 32P-labeled O-phosphoseryl-tRNA[Ser]Sec was added to a 10-μl reaction mixture containing 50 mM Tris-HCl (pH 7. 5), 20 mM DTT, 10 mM KCl, 10 mM MgCl2 and 1 μg of each purified protein. Reactions were carried out for 30 min at 37 °C. Following incubation, the tRNA was deacylated by adding an equal volume of 1 M Tris-HCl (pH 8. 0) and incubating at 37 °C for 45 min. Reactions were then spotted onto 3M paper (Whatman), placed in a TLC chamber, and chromatographed for 8 h using a mixture of butanol/acetic acid/water (12: 3: 5). The chromatogram was then exposed to a PhosphorImager screen. ATP hydrolysis assays were carried out in a volume of 10 μl with 40 mM HEPES (pH 7. 4), 20 mM KCl, 10 mM MgCl2,5 μCi of [α-32P]ATP, and 10 mM DTT and either with or without 0. 25 mM selenide. After adding 0. 3 mg/ml final concentration of each selenophosphate synthetase protein, reactions were incubated at 37 °C for 1 h under anaerobic conditions. Then 0. 5 μl of each reaction was loaded onto PEI TLC plates, the plates were run in 0. 8 M LiCl, and the developed TLC plates were exposed to a PhosphorImager screen. For NMR analysis, ATP hydrolysis reactions were carried out under anaerobic conditions in 3-mm NMR tubes in a total volume of 200 μl with 2 mM ATP instead of [α-32P]ATP. NMR tubes were sealed and incubated at 37 °C for 4 h prior to 31P NMR spectroscopic analysis [9]. Synthetic Sec tRNA was used in all biosynthetic reactions. Synthesis, purification, and aminoacylation of Sec tRNA were carried out as described [20]. All of the reactions were set up under anaerobic conditions before chromatographic analysis. For Sec biosynthesis, the selenium donor SeP was either generated from selenide by using mSPS2-Cys or hydrolyzed from chemically synthesized [ (CH3) 3SiO]3PSe [31]. For generating the selenium donor with mSPS2-Cys, a 10-μl reaction containing 50 mM NH4HCO3 (pH 7. 6), 10 mM DTT, 2 mM MgCl2,2 mM KCl, 2 mM ATP, and 2 μg of mSPS2-Cys with or without 1 mM selenide was preincubated at 37 °C for 1 h. The mSPS2-Cys reaction was added to 10 μl containing 50 mM Tris-HCl (pH 7. 0), 20 mM DTT, 10 mM MgCl2,2 μM PLP, 1. 0 μg of purified mSecS, and approximately 5 μg (about 30,000 cpm) of either O-phospho-[3H]seryl-tRNA[Ser]Sec or [3H]seryl-tRNA[Ser]Sec and, in a positive control reaction, SelA in place of mSecS, and in a negative control reaction, thioredoxin (with a His-tag) in place of mSecS. Reactions were incubated at 37 °C for 2 h and then heated at 75 °C for 5 min, aminoacyl-tRNAs were deacylated [20], and 1 μl of an unlabeled amino acid mix (containing 12. 5 mM concentration each of serine, O-phosphoserine, Sec, and alanine in 50 mM of KBH4) was added. Each reaction, along with several control lanes containing unlabeled amino acids and 14C-pyruvate, was chromatographed on Whatman 3M filter paper (45 × 60 cm) in ethanol/acetic acid/water (12: 3: 5) for 16 to 20 h. Then 1. 0-cm strips were cut out of the dried chromatogram and counted in a liquid scintillation counter. The locations of each amino acid were determined by staining the lanes with unlabeled amino acids in 0. 3% ninhydrin in acetone or by cutting out 1. 0-cm strips of lanes with 14C-pyruvate and counting in a liquid scintillation counter. For direct use of SeP as a selenium donor, reactions contained the same components as above except 1 mM SeP was used in place of mSPS2-Cys reaction solutions and DTT was omitted from the mSecS reactions, since we found that the activity of mSecS is higher without DTT. SeP was generated by hydrolysis of 20 mM chemically synthesized [ (CH3) 3SiO]3PSe [31]. Reactions were incubated and analyzed as above. For measurement of the SecS synthesis rate, reactions were carried out in a total volume of 10 μl of 50 mM Tris-HCl (pH 7. 0), with 10 mM MgCl2,10 mM KCl, 0. 2 mM SeP, and 4 μg (approximately 120 pmol) of O-phospho-[3H]seryl-tRNA[Ser]Sec. Reactions were initiated by adding 2 μg (approximately 35 pmol) of mSecS and were stopped at specific time points between 0 and 80 min by boiling for 2 min and then deacylating and counting as described above. GenBank (http: //www. ncbi. nlm. nih. gov/Genbank) accession numbers for the sequences used in this paper are Homo sapiens (SecS_HS), Q9HD40; Mus musculus (SecS_MM), Q6P6M7; Drosophila melanogaster (SecS_DM), NP_649556; C. elegans (SecS_CE), Q18953; Methanococcus jannaschii (SecS_MJ), Q58027; Methanopyrus kandleri (SecS_MK), Q8TXK0. SelA sequences: Escherichia coli (SelA_EC), BAE77702; Geobacter sulfurreducens (SelA_GS), P61736. SelA-like sequences: Methanococcus jannaschii (SelA-like_MJ), Q57622; Methanopyrus kandleri (SelA-like_MK), AAM01835; E. coli selA, M64177; selD, M30184; mouse secS, AL049338; sps2, NM_009266; sps1, NM_175400; and C. elegans sps2, NM_070203. The numbers for other PLP-containing proteins are Methanococcus jannaschii Sep-tRNA: Cys-tRNA synthase (SepCysS_MJ), Q59072; Medicago truncatula Orn/Lys/Arg decarboxylase (COG1982_MT), ABE83138. 1; Archaeoglobus fulgidus glutamate decarboxylase (COG0076_AF), O28275; and Thermococcus kodakarensis glycine/serine hydroxymethyltransferase (COG0112_TK), 05JF06.

Biosynthesis of the 20 canonical amino acids is well established in eukaryotes. However, many eukaryotes also have a rare selenium-containing amino acid, selenocysteine, which is the 21st amino acid in the genetic code. Selenium is essential for human health, and its health benefits, including preventing cancer and heart disease and delaying aging, have been attributed to the presence of selenocysteine in protein. How selenocysteine is made in eukaryotes has not been established. To gain insight into its biosynthesis, we used computational analyses to search completely sequenced genomes for proteins that occur exclusively in organisms that utilize selenocysteine. This approach revealed a putative selenocysteine synthase, which had been previously identified as a pyridoxal phosphate–containing protein dubbed soluble liver antigen. We were able to characterize the activity of this synthase using selenophosphate and a tRNA aminoacylated with phosphoserine as substrates to generate selenocysteine. Moreover, identification of selenocysteine synthase allowed us to delineate the entire pathway of selenocysteine biosynthesis in mammals. Interestingly, selenocysteine synthase is present only in those archaea and eukaryotes that make selenoproteins, indicating that the newly defined pathway of selenocysteine biosynthesis is active in these domains of life.

Abstract Introduction Results Discussion Materials and Methods Supporting Information

biochemistry eukaryotes molecular biology genetics and genomics

2007

Biosynthesis of Selenocysteine on Its tRNA in Eukaryotes

Critical aspects of HIV-1 infection occur in mucosal tissues, particularly in the gut, which contains large numbers of HIV-1 target cells that are depleted early in infection. We used electron tomography (ET) to image HIV-1 in gut-associated lymphoid tissue (GALT) of HIV-1–infected humanized mice, the first three-dimensional ultrastructural examination of HIV-1 infection in vivo. Human immune cells were successfully engrafted in the mice, and following infection with HIV-1, human T cells were reduced in GALT. Virions were found by ET at all stages of egress, including budding immature virions and free mature and immature viruses. Immuno-electron microscopy verified the virions were HIV-1 and showed CD4 sequestration in the endoplasmic reticulum of infected cells. Observation of HIV-1 in infected GALT tissue revealed that most HIV-1–infected cells, identified by immunolabeling and/or the presence of budding virions, were localized to intestinal crypts with pools of free virions concentrated in spaces between cells. Fewer infected cells were found in mucosal regions and the lamina propria. The preservation quality of reconstructed tissue volumes allowed details of budding virions, including structures interpreted as host-encoded scission machinery, to be resolved. Although HIV-1 virions released from infected cultured cells have been described as exclusively mature, we found pools of both immature and mature free virions within infected tissue. The pools could be classified as containing either mostly mature or mostly immature particles, and analyses of their proximities to the cell of origin supported a model of semi-synchronous waves of virion release. In addition to HIV-1 transmission by pools of free virus, we found evidence of transmission via virological synapses. Three-dimensional EM imaging of an active infection within tissue revealed important differences between cultured cell and tissue infection models and furthered the ultrastructural understanding of HIV-1 transmission within lymphoid tissue. HIV-1 remains a significant public health concern with over 33 million people infected world-wide [1]. Most HIV-1 transmissions occur across an epithelial barrier, resulting in generation of a founder population within the mucosa, viral dissemination to lymphatic tissue, and exponential viral replication throughout the lymphatic system [2]. These events result in depletion of most CD4-positive T cells in mucosal compartments, and establishment of a reservoir of resting cells with integrated provirus that is not susceptible to antiretroviral therapy. In the absence of therapy, progressive immune system collapse and progression towards AIDS ensue in most infected persons. Accumulating evidence indicates that both acute and chronic HIV-1 infection profoundly affect the gastrointestinal (GI) tract [3], [4]. Studies of SIV infection in non-human primates demonstrated that intestinal CD4 T cell depletion occurs within days, even before T cell depletion can be detected in the peripheral blood or lymph nodes [5]; similar events occur in HIV-1–infected humans [2], [6]. Several features of the GI tract facilitate its susceptibility to HIV-1 infection: (i) the GI mucosa includes high levels of pro-inflammatory, HIV-1–stimulatory cytokines produced by exposure to antigens in the external environment, (ii) a dense clustering of cells that facilitates cell-to-cell transmission, and (iii) a majority of the activated memory T cells expressing CD4 and CCR5 that serve as the preferred target cells for HIV-1 infection [7], [8]. Indeed, the gut-associated lymphoid tissue (GALT) harbors the greatest concentration of potential HIV-1 target cells in the human body [9]; >50% of CD4 T cells from the lamina propria in the lower GI tract are destroyed during acute HIV-1 infection, and early infection of the GALT is believed to be central to chronic HIV-1 infection and disease progression [10], [11]. Furthermore, the presence of CD4 and CD8 T cells, dendritic cells, and macrophages in the GALT make this tissue an integral site for HIV-mediated immune depletion. Mouse models with humanized immune systems are emerging as a tractable, cost-effective means by which to study HIV-1 infection in mucosal lymphoid tissue [12]. One such model, humanized bone marrow/liver/thymus (BLT) mice, are individually created by transferring human fetal thymic and liver organoid tissues, along with CD34-positive human stem cells, into immunocompromised mice. BLT mice reconstitute significant levels of human lymphoid immune cells; e. g. , T and B cells, monocytes, dendritic cells and macrophages in peripheral blood and organs including the GI tract [13], [14]. Important aspects of human HIV-1 infection are recapitulated in this system, including T cell depletion in the gut and peripheral blood, and both systemic and mucosal virus transmission during the course of the disease [15], [16]. Furthermore, BLT mice exhibit high levels of human immune cell engraftment at mucosal sites and significant antigen specific immune responses by multiple cell types [17], [18]. Electron microscopy (EM) was instrumental in the original identification of HIV-1 [19], [20]. Subsequently, diagnostic EM analyses of biopsies from infected patients revealed important aspects of HIV-1 transmission in humans at varying stages of infection, from early acute disease to AIDS progression [21]. More recently, 3-D EM, specifically electron tomography (ET), cryoelectron tomography (cryoET) and ion-abrasion scanning electron microscopy, have been applied at increasingly higher resolutions, facilitating improved understanding of HIV-1 virion structure [22]–[24], virus budding [25], [26], and virus transmission between immune cells [27], [28]. 3-D EM of isolated virions and infected cells can provide a detailed understanding of HIV-1 ultrastructure and transmission between cultured cells, but does not address the complex cellular environment found in mucosal tissues within an organism experiencing an active infection. Here we used ET to analyze GALT from humanized HIV-1–infected BLT mice in order to visualize HIV-1 infection in mucosal tissues in 3-D at ultrastructural resolution. These analyses allowed us to localize infected substructures within intestinal tissue, classify virions as mature or immature, identify infected cells, visualize structures we interpreted as components of the host cell machinery involved in viral budding, and assess the propensity for viral spread by cell-to-cell versus free virus routes of infection. In parallel studies, we used immunofluorescence (IF) and immuno-electron microscopy (immunoEM) to verify the identities of viral particles, locations of infected tissue, and to distinguish human from murine and infected from uninfected cells. Human hematopoietic cells derived from transplanted human stem cells have been shown to repopulate the GALT of BLT mice, and HIV-1 infection of these mice results in CD4 T cell depletion, initially in GALT and then systemically [13], [16]. Following established protocols [13], BLT mice were infected with HIV-1 approximately 20 weeks after transfer of human immune tissues and cells, using only mice that met the following criteria for adequate human immune reconstitution: >25% of peripheral blood cells were within a lymphocyte gate on forward-versus-side scatter plots; >50% of cells in the lymphocyte gate were human (human CD45+/mouse CD45−); and >40% of human cells in the lymphocyte gate were T cells (human CD3+). Ten to twenty weeks post infection, mice were sacrificed and segments of small intestine and colon were excised. IF was used to survey locations of HIV-1–infected cells in GALT (Figure 1A, B). Following infection with HIV-1, human CD4 T cells were depleted from the lamina propria (Figure 1B), as previously reported [13], [16]. Staining for the p24 capsid protein of HIV-1 localized primarily in CD4+ cells in regions near the crypts (Figure 1B, inset), which harbor significant populations of immune cells and multipotent stem cells [29]. No evidence of human cells or HIV-1 infection was found in non-humanized infected controls (data not shown). We next analyzed GALT samples in parallel by ET and immunoEM/ET. Tomography of frozen hydrated tissue samples by cryoET was not possible because the samples were too thick for imaging without sectioning and were infectious biohazards. We therefore imaged fixed and sectioned samples, either positively-stained plastic-embedded or negatively-stained methylcellulose-embedded sections. For ET alone, preservation quality was improved by lightly fixing HIV-1–infected tissue with aldehydes and then further processing them by high-pressure freezing and freeze substitution fixation [30]. This “hybrid” fixation method allowed for safe handling of infectious material and obviated the most structurally damaging steps of traditional chemical fixation [31], yielding well-preserved positively-stained samples. Tomograms were reconstructed from 200 nm or 300 nm sections, often in montaged serial sections of volumes up to 6. 1 µm×6. 1 µm×1. 2 µm. Although these samples could not be used for immunoEM because antibody epitopes are rarely accessible in epoxy-embedded, positively-stained samples [32], analogous GALT samples generated from the same animal were prepared for immunoEM/ET as negatively-stained methylcellulose-embedded sections [33]. Measurements of virions and other structures reflected proportional thinning typical of plastic-embedded and negatively-stained samples [34]. Consequently most structures were ∼30% smaller than counterparts from cryoEM studies or virions in solution or in cultured cells [22]–[24], [35], [36]. ET surveys of HIV-1–infected BLT mouse GALT revealed budding virions (Figure 2A, B; Figure S1A) and free mature and immature particles (Figure 1C, Figure 2C–D, Figure S1B). Virions were detected in all HIV-1–infected mice, while none were found in mock-infected controls (data not shown). The virions were verified as HIV-1 using antibodies against HIV-1 p24 and the envelope spike (Figure 2E, F). Virions were imaged in tissue at all stages of egress, from early plasma membrane Gag assembly to nearly completed buds and fully mature, free HIV-1 (Figure S2). Budding profiles and immature free virions were distinguished by core structures that exhibited radial layers and often appeared as an incomplete internal sphere (a “C” shape in projection) [24], [26]. Mature HIV-1 particles were distinguished from immature particles by the collapse of their cores into a variety of conical shapes, typically “bullet-shaped” cones but often cylinders or ellipsoids [23], [37] (Figure S2). Although envelope spikes on HIV-1 and SIV can be distinguished in positively-stained samples [38], we observed few projections emanating from virion surfaces, consistent with biochemical and cryoET studies of purified HIV-1 virions that demonstrated a low number of envelope spikes: an average of ∼14 (ranging from 4–35) per virus particle [39], [40]. After establishing that HIV-1 could be identified in infected BLT GALT by ET and immunoEM, we surveyed GALT samples to determine locations of infection. Plastic-embedded sections of small intestine (jejunum and ileum) and large intestine (colon) were examined to find HIV-1 and infected cells, which were identified by budding profiles at their surfaces. Within a given animal, the extent of infection and the distribution of virions were similar between the small and large intestine. However, virions were found in differing amounts amongst sub-structures in the intestinal mucosa. The largest populations of HIV-1 virions and infected cells identified by EM were located in crypts (Figure 1A, C), consistent with IF (Figure 1B). Approximately one in ten crypts showed evidence of HIV-1 infection. The mucosal region surrounding the villus base and the crypts contained few free virions or infected cells (∼1 in >100); when present, infected cells were often near a capillary or venule (Figure S1A). The numbers of free virions and infected cells in the lamina propria were less than in the crypts (Figure S1B). Typically, infected lamina propria were in villi continuous with infected crypts. Few infected cells or virions were found in the smooth muscle layer surrounding the intestine. In addition, free virions were rarely found in blood vessels because even the high viral loads of the HIV-1–infected BLT mice from which the samples were derived (up to 126,000/mL in peripheral blood) translated to only ∼1×10−7 virions/µm3. Thus at the scale of individual EM images or even large-format tomograms, HIV-1 virions would be rarely seen, and our imaging of >50 blood vessels contained within tomograms yielded only two examples of free virions (data not shown). To identify potential human target cells of HIV-1 infection, we conducted immunoEM (Figure S3A–D) using antibodies specific for human proteins. Human CD4 localized primarily to the plasma membrane in uninfected cells (Figure S3A), but we found extensive CD4 labeling in the endoplasmic reticulum (ER) of CD4-positive cells with budding virions or nearby free virions (Figure S3B), correlating with the finding that HIV-1 Vpu induces cell surface CD4 to redistribute to the ER to avoid surface retention of newly-forming virions [41]. Double labeling with antibodies against HIV-1 Nef and human CD4 (Figure S3C) or class I human leukocyte antigen (HLA) and human CD4 (Figure S3D) confirmed that cells exhibiting a predominantly ER localization of CD4 were human cells infected with HIV-1. No instances of Nef expression were found in uninfected or non-human cells (data not shown), which served as an internal control for the specific of the antibodies and further validated the BLT model of HIV-1 infection. Tomograms of immature virions derived from negatively-stained infected tissue revealed detailed structural information. With the exception of the widening of lipid bilayer membranes, presumably caused by obligatory light fixation associated with this method, the overall architecture of the Gag shell in immature virions conformed to known properties of HIV-1 determined from studies of viruses isolated from cultured cells [22], [24], [35], [36], [42] (Figure 3; Figure S4). Indeed, the immature virions in our tissue samples (Figure 3, S4A, B) exhibited features observed in cryoET analyses of purified frozen hydrated HIV-1 [22], [24] (Figure S4C); e. g. , individual layers of the Gag shell, including the hexagonal lattice of the capsid (CA) portion (Figure 3). The symmetry of the CA layer was confirmed by hexagonal features in the Fourier transforms of immature virions, but not in transforms of adjacent cytoplasm (Figure 3B; Figure S4A). More than 50 crypts of Lieberkühn were imaged in the course of this study. In the ∼10% of crypts that were infected, HIV-1 virions were found primarily in pools within dilated regions of intercellular spaces (Figure 1C; Figure 4; Figure S5; Movie S1). Pools were defined as a population of virions within an intercellular space that was continuous within a given 3-D volume. Multiple intercellular spaces could be present within the volume, but unless the spaces were visually continuous, virions within them were regarded as separate pools (Figure 4B, C). The numbers of free virions in intercellular pools ranged from 5 to >200. In single-frame tomograms (3. 2 µm×3. 2 µm×200 nm), most pools contained 10–40 particles. Larger pools were observed in serial-section reconstructions encompassing greater tissue volumes. In longitudinal sections of crypts, most pools were found between the base and middle. Infected human immune cells, identified by the presence of budding virions, were often found near virion pools. Virions within a given pool were distinguished as mature or immature based on the presence of a cone-shaped core in mature particles and radial Gag layers in immature particles (Figure S2). The numbers of mature and immature particles in intercellular pools were quantified within reconstructed volumes of infected crypts. Pools could be classified as either “mostly mature” or “mostly immature” (Figure S5A). Of >100 pools containing many hundreds of virions, approximately 90% of pools were classified as mostly mature and 10% were mostly immature. Potential HIV-1 target cells and pools of virions were plentiful in GALT, particularly in crypts, thus it was not always possible to determine from which cell a particular virion population originated. In order to quantify virions from a particular cell and infer temporal data with respect to virion pools, we imaged regions of the intestinal smooth muscle layer (Figure 1A), which contains few HIV-1 target cells. Figure S5B shows an HIV-1-infected cell in the smooth muscle. The surface of this cell exhibited several HIV-1 budding profiles, and groups of free virions were located both in close proximity to and at varying distances from it. There were no other infected cells within several microns, thus we could be confident that nearby free virions had originated from that cell. We found that 62% of virions (n = 16) in immediate proximity (≤0. 5 µm) to the cell were immature, while 73–75% of virions in groups located 0. 8 µm (n = 15) and 1. 3 µm (n = 32) away were mature. Of >100 virion pools that were imaged, most were in obvious extracellular spaces. Some pools (∼5%) appeared to be intracellular, but were revealed by ET to be connected to the extracellular space by narrow channels that averaged ∼27 nm in width (range = 23–32 nm; n = 6) (Figure 4D–E) and contained 2–20 mature virions. A few of the budding regions were large enough that potential continuities with the plasma membrane were outside of the reconstructed volume. The presence of seemingly intracellular virion pools connected to microchannels could identify the cell as an infected macrophage, a cell type in which internal virus-containing compartments were proposed to represent specialized domains of the plasma membrane that were sequestered intracellularly [43], [44] and/or endosomal compartments [45], [46]. ET surveys of HIV-1 infected GALT showed evidence of virological synapses for direct cell-to-cell virus transmission, a route of HIV-1 transmission within tissues whereby a virus buds from an infected cell and directly contacts and infects an adjacent uninfected cell [47]. Formation of a virological synapse results from interaction of gp120 on an infected cell with its receptors on a target and also involves other host proteins such as LFA-1 and ICAM proteins on the surfaces of both the donor and target cells [48], [49]. A large format reconstruction (2×3-frame montage) of GALT revealed an HIV-1–infected cell, likely a dendritic cell or macrophage based on the convoluted processes intercalating between neighboring cells (Figure 5A; Movie S2). A presumptive virological synapse was visualized as a region of contact between a budding virion and an adjacent cell (Figure 5B; Movie S2). Although this positively-stained sample could not be examined by immunoEM, we found similar features in negatively-stained samples that labeled with antibodies against LFA-1 and ICAM-1 (Figure 5C, D), supporting the identification of these regions as virological synapses. In another example, an infected cell that showed numerous budding profiles included one that closely approached the surface of an adjacent cell although still attached to its host cell via a ∼50 nm neck (Figure S6A). The surface region of the cell proximal to the approaching bud was denser than surrounding surface regions and extended toward the bud. In a third example, a budding profile from an infected cell appeared to project into an invagination in the plasma membrane of an adjacent cell (Figure S6B, C). Tomographic views through the volume containing this region showed the boundaries of the invagination followed the contours of the budding profile (Figure S6C), suggesting a dynamic response to the approaching nascent virion. By reconstructing a large 3-D volume of infected tissue, we could address whether direct cell-to-cell transmission was an obligatory means of virion transfer between two adjacent cells. Movie S1 shows a 1. 4 µm×2. 9 µm×1. 2 µm tomogram in which the outlines of two adjacent cells were distinguished. Both cells were identified as infected by the presence of budding virions and were therefore HIV-1 targets. A region resembling a virological synapse was not observed in the reconstructed volume, however a large accumulation of free mature virions were present in the space between the cells, suggesting that direct cell transfer is not a required mechanism of HIV-1 transmission between closely apposed infected cells. The lack of an observed virological synapse in such cases could be the consequence of CD4 down-regulation in the infected cells. However the existence of natural recombinant HIV-1 strains, which could result from infection by one HIV-1 strain of a cell already infected with a different viral strain [50], suggests that residual CD4 remaining at an infected cell surface can allow for infection via free virus or direct cell-to-cell transfer. The large number of budding virions within BLT GALT tomograms offered the opportunity to characterize structural aspects of HIV-1 budding in infected tissue (Figure S7). Actin filaments were often found near forming buds (Figure S7A) similar to those previously observed at HIV-1 budding sites in cultured cells [25]. Budding profiles exhibited varying lengths of necks, including some with no neck (Figure 3C, D; Figure S7B). In the colon, early budding virions without necks were often observed forming from surfaces that were not obviously plasma membrane. However, serial-section tomography revealed that these domains were usually continuous with the plasma membrane proper, indicating that they were convoluted regions of the cell surface and not distinct cytoplasmic compartments. Some budding virions exhibited necks with 50–80 nm lengths and varying widths (Figure S7C), with narrower necks likely representing those approaching scission. Virions were also observed budding at the ends of extremely long cellular projections (Figure 2A, B) that were likely filopodia extending from dendritic cells, as observed in culture [27], [51]. ET analyses of HIV-1 budding in cultured cells revealed a subset of RNA-free immature virions with a novel “thinner” Gag lattice lacking the nucleocapsid-RNA layer, which were suggested to represent aberrant, noninfectious virions resulting from premature activation of HIV-1 protease [25]. Using our measuring convention, the previously-described thin Gag lattice [25] measured 9–10 nm. Analysis of 100 free or budding immature virions from tissue samples yielded no examples with a thin (9–10 nm) Gag lattice that lacked discernable RNA densities. Instead, we found that the Gag lattice widths in all of the immature virions we surveyed (n = 100) within infected tissue was 14. 6±0. 8 nm (Figure S4B, Figure S7D, E); significantly different than the thin 9–10 nm Gag lattices previously described [25]. In addition, there were no systematic structural differences in Gag lattices correlating with the type of budding profile: the Gag shell thicknesses measured in 30 long-necked and 30 neck-free buds were similar and presumptive RNA densities were present in all cases (Figure S7D). Release of HIV-1 virions from infected cells involves recruitment of the host endosomal sorting complexes required for transport (ESCRT) machinery to sites of virus assembly by the Gag polyprotein [52]. These interactions culminate with the polymerization of ESCRT-III proteins, recruitment of vacuolar protein sorting-associated protein 4 (VPS4) ATPase oligomers, fission of the cellular membrane attaching the virion to the host cell, and disassembly of the ESCRT machinery. We used antibodies against ESCRT-III proteins, human charged multivesicular body proteins (hCHMPs) 1B and 2A, and ALG2-interacting protein X (ALIX), an ESCRT adaptor protein that facilitates the transport of Gag to the cell membrane [53] and can mediate interactions between ESCRT-I and ESCRT-III complexes [54], to detect components of the ESCRT pathway in infected tissue by immunoEM. We found that hCHMP1B, hCHMP2A and hALIX localized predominantly to the neck regions of budding HIV-1 virions (Figure 6A-C). The labeling was specific, but sparse due to the small number of epitopes and their availability only at section surfaces. At scission regions of budding virions in which the neck of the bud was greater than half the diameter of the bud, clusters of 4–6 spoke-like projections nearly 20 nm in length radiating from a centralized origin at the base of the budding virion were sometimes observed (Figure 3C, D; Figure 6D; Figure S8A; Movie S3). As the larger neck diameter may define these buds as being at an initial stage of egress, these radial projections could represent components of the early portions of the ESCRT pathway such ESCRT-I or ALIX recruited by assembling HIV-1 Gag molecules. Indeed, the size and shape of the structures approximate models for the ESCRT-I-II supercomplex determined by a combination of spectral techniques [55]. By contrast, in tomograms of budding virions with narrower necks (less than half the diameter of the bud itself), we observed parallel electron dense striations circumscribing the neck of the bud in both positively- and negatively-stained sections (Figure 7A–E; Figure S8B, C; Movie S4) suggestive of ESCRT-III components polymerizing at membranes [56], [57]. Similar electron dense striations were detected at the necks of budding virions arrested at a late stage by expression of dominant-negative ESCRT-III or VPS4 proteins [58]. In addition, budding profiles in positively-stained samples often showed 1–5 electron-dense “spots” in the neck or base of a bud (Figure 7F, G; Movie S5). The spots were observed in over half of ∼50 budding profiles in which the diameter of the neck was half or less of the diameter of the budding virion; presumably a late stage of budding. Available antibodies against VPS4 did not stain efficiently by immunoEM, however their interpretation as VPS4 oligomers was consistent with fluorescence imaging showing recruitment of 2–5 VPS4 dodecamers to the sites of viral budding just prior to virion abscission [59], [60]. In addition, the size and relative shape of the putative VPS4 densities (Figure 7G) correlated with cryoEM reconstructions of VPS4 [61]. Many aspects of the pathologies related to HIV-1 infection, including immune cell death and tissue destruction, occur in GALT. However, 3-D ultrastructural details of a natural GALT infection were unknown because ET had not been applied to in vivo infection in GALT or other lymphatic tissues. BLT humanized mice are an emerging model for studying HIV-1 infection, and BLT GALT maintains cellular architecture, cell-cell interactions, immune cell populations and signaling more accurately than cell culture infection models [12]. As such, the BLT mouse system is a reliable model for structural studies of HIV-1 infection in a tissue environment. In addition, the inclusion of human thymic tissue in BLT mice allows for T cell maturation in the context of human, rather than murine, MHC proteins; an aspect that is not present in humanized mouse model systems produced with human hematopoietic stem cells but without thymic tissue. Dense areas of HIV-1–infected cells, including CD4 T cells, macrophages and dendritic cells, and free HIV-1 virions were found in crypts within BLT GALT by IF, ET and immunoEM (Figure 1B, C). Blood vessels were imaged in mice with a wide range of viral loads; however, we were unable to correlate the relative abundance of virions detected in GALT with the viral load measured in the blood. In fact, only two examples of virions within blood vessels of BLT mice were detected as compared with hundreds of virions within mucosal tissue. This finding is consistent with reports highlighting a discrepancy between blood viral load and HIV-1 levels in tissues [62], [63]. Thus analysis of HIV-1–infected tissues by methods such as ET may provide valuable information in addition to blood viral load measurements when evaluating treatment regimens. Potentially relevant to infection and immune cell recognition mechanisms, large pools of free HIV-1 were found within infected GALT (Figures 1C, Figure 4, Figure S5). Although most pools contained mainly mature virions, some pools contained a majority of immature virions (Figure S5A), a phenomenon not observed in EM studies of HIV-1 infection of cultured cells. Pools of virions were usually found between cells, but also in compartments that appeared to reside within cells. These compartments were often connected to the cell surface by microchannels 20–30 µm in width (Figure 4D). These narrow channels likely undergo dynamic changes in morphology, as their width would be too narrow to accommodate passage of HIV-1 to the extracellular space. We interpreted such channels as invaginations of the plasma membrane, consistent with reports that macrophages can assemble HIV-1 in intracellular virus-containing compartments created by internally sequestered plasma membrane [43], [44], [64]. In infected tissue, we found that pools of HIV-1 virions located between two cells could contain mature or immature virions (Figure S5A), whereas the intracellular pools connected by microchannels contained only mature virions (Figure 4D). One possibility for the difference in maturation states of inter- versus intracellular pools of HIV-1 is that intracellular virions connected to the extracellular space by microchannels are not subject to movement by interstitial fluid through intestinal tissue and could remain in a single location long enough to complete maturation, perhaps representing viral reservoirs that allow low levels of de novo infection to proceed in the presence of anti-retroviral therapy and/or antibodies [65]. Although the discovery of virion pools suggested that infection by free virus could occur within infected tissue, we also found evidence of direct cell-to-cell transmission of HIV-1 in infected GALT (Figure 5; Movie S2). The virological synapse is a mechanism of cell-to-cell transmission in which juxtaposition of an infected and uninfected cell promotes infection by directing viral assembly, budding, maturation, and fusion machinery to discrete locations of cellular contact between cells [47]. In a large 3-D reconstruction of two adjacent HIV-1–infectable target cells (Movie S1), we found a large pool of mature virions but no evidence for a virological synapse, suggesting that formation of virion pools and infection by free virus can occur even when adjacent cells are both infectable by HIV-1, or had been infectable prior to down-regulation of CD4. In addition, this result validated our frequent finding of large pools of free virions in HIV-1–infected tissue, demonstrating that this phenomenon was not necessarily the consequence of the juxtaposition of a human infected cell and a murine cell, as may occur in BLT GALT. EM studies of HIV-1 virions produced in cultured cells suggested that maturation is a rapid process, because intermediate maturation states were not detected and because virions found near cells were predominantly mature [66]. However, our finding of pools containing immature virions in proximity to infected cells in tissue suggested maturation dynamics and/or virion diffusion properties differ between cells organized within tissue versus those cultured in vitro. In addition, we never found examples of RNA-negative budding virions with a thin Gag lattice in tissue samples, as had been observed in ∼18% of immature particles in cryoET analyses of HIV-1 produced in cultured cells [25]. Thus, higher numbers of aberrant particles and of exclusively mature virions in close proximity to producer cells could be artifacts of producing virions in cultured cells, suggesting that the BLT model of in vivo infection more accurately recapitulates the HIV-1 lifecycle than cell culture models. Although ET relies on fixed tissue and cannot directly recapitulate virion dynamics in live cells, our studies provided a glimpse into temporal aspects of HIV-1 maturation. We determined that an isolated infected cell within a large tissue volume was the sole producer of several populations of imaged virions located at varying distances from the cell. This allowed us to determine that a single infected cell can produce at least 63 viruses (the number of virions in the three pools in Figure S5B). The total number of virions produced per cell is likely far larger, as regions above and below the cell were not represented in the reconstruction. Using a predicted rate of interstitial fluid movement in intestinal tissue of 0. 1–2 µm/sec [67], a virion would travel 2 µm in 1–20 sec, indicating that maturation could occur just seconds after release from an infected cell. This argues that, in tissue, virions found ∼2 µm away from a producer cell budded only seconds earlier, supporting an assumption of rapid virus maturation. Furthermore, our finding of mostly immature virion pools in close proximity to the infected cell and mostly mature virion pools further away from the cell (Figure S5B) is consistent with synchronous release and subsequent maturation of HIV-1. The trigger (s) for and/or block (s) to maturation that could promote synchronized virus maturation in tissue could include proximity to an infected producer cell, lack of an adjacent target cell to form a virus synapse, and/or contact with a non-infectable cell. Late events in HIV-1 budding had been visualized by fluorescence microscopy [59], [60] and ET of cultured cells [25], [26], [66], [68] but not yet in infected tissue. In our infected tissue samples, we detected distinct electron dense structures near virions at various stages of budding that may represent aspects of the host cell ESCRT machinery at sites of HIV-1 egress (Figures 6,7, S8). Although we could not identify the structures conclusively, our assignments of their possible identities are consistent with what is known temporally about the involvement of host cell machinery in HIV-1 budding and release from infected cells. Tomograms revealed that virions in the initial stages of budding contained 4–6 spoke-like projections emanating from the center of the forming neck of the budding virion (Figure 3C–D, Figure 6D, Figure S8A), potentially representing components of host ESCRT-I or ALIX recruited by assembled HIV-1 Gag. The shape, size, and temporal occurrence of these structures agree with a proposed model for vesicle budding and fission based on biophysical analyses of the ESCRT-I-II supercomplex in solution [55]. Virions at later stages of budding that were connected to the host cell membrane by thinner (<50 nm) elongated necks showed parallel, electron dense striations along the membrane surface of the neck (Figure 7A–E, Figure S8B) that we interpreted as features of polymerized ESCRT-III proteins [56], [57]. These late budding profiles often displayed dense spots along the center of the neck (Figure 7F, G) that we suggest were VPS4 oligomers recruited immediately prior to fission of the new virion from the cell membrane, consistent with fluorescence microscopy studies [59], [60]. ET of budding virions within tissue allowed a spatial and temporal interpretation of HIV-1 budding. First, the Gag lattice reached a sufficient point of closure, which allowed formation of a spoke-like structure at the base of the early budding virion. Next, the virion formed an elongated neck; concomitant with polymerization of host cell factors in a spiral around the inside of the membrane [56], [57]. In tomographic slices of budding profiles, these presumptive spirals appeared as two or more parallel lines bisecting the neck region (Figure 7A–E). Finally, the recruitment of large oligomers, possibly VPS4, coincided with the separation of the virion from the infected cell [59], [60], completing the budding process (Figure 7F, G). In summary, our 3-D ultrastructural characterization of HIV-1–infected GALT identified dense regions of virus transmission, provided insights into the temporal nature of virus maturation, revealed HIV-1 transmission occurring by both free virus and direct cell-to-cell mechanisms, and demonstrated important differences between cultured cell and tissue HIV-1 infection models. Differences included the identification of free immature virions and the scarcity of aberrantly formed viral particles during an active infection. The high resolution of our positively- and negatively-stained tissue samples allowed 3-D visualization of HIV-1 transmission within lymphoid tissue, providing a new approach for understanding HIV-1 infection in vivo. Humanized mice were prepared and cared for in an AAALAC-certified animal care facility at the Massachusetts General Hospital (OLAW Assurance #A3596-01), in accordance with a protocol approved by the MGH IACUC (Protocol #2009N000136/25). The protocol as submitted and reviewed conforms to the USDA Animal Welfare Act, PHS Policy on Humane Care and Use of Laboratory Animals, the “ILAR Guide for the Care and Use of Laboratory Animals” and other applicable laws and regulations. Every effort was made to minimize animal suffering throughout all experiments. Human tissue for preparing the humanized mice was procured and used in accordance with a protocol approved by the local Institutional Review Board (Partners Human Research Committee, Protocol #2012-P-000409/5). NOD/SCID/IL2Rγ−/− mice (The Jackson Laboratory) were reconstituted with human tissue as described [13]. Approximately 20 weeks after transfer of human immune tissues and cells, mice were infected intraperitoneally with 1×105 TCID50 of JR-CSF HIV-1. Every 2 weeks after infection, ∼200 µl of blood was obtained through puncture of the retro-orbital sinus or submandibular vein for determination of HIV-1 plasma viral load. Viral RNA was isolated using the QIAamp Viral RNA Mini Kit (Qiagen) and viral loads were determined by quantitative RT-PCR using primers for HIV-1 Gag [69]. Immunofluorescence experiments were conducted using tissues from a mouse with a blood viral load of 940,000 copies/mL. ImmunoEM experiments were conducted using tissues from a mouse with a viral load of 100,000/mL. The remaining mice had blood viral loads as follows: 0 (control) 9,800,8, 400,18,500 and 126,000 copies/mL. As previously shown, the range of blood viral loads did not correlate with virus populations found in tissue samples [62], [63]. Infected mice were sacrificed 10–20 weeks post infection, and then necropsied with segments of small intestine and colon excised and fixed. Immunofluorescence (IF) studies were conducted as described in the Supplementary Methods (Text S1). For positively-stained samples, HIV-1–infected tissue was prepared by a hybrid method that employed primary chemical fixation followed by high-pressure freezing/freeze substitution fixation (see Text S1). Negatively-stained samples were prepared as described [33] and in the Supplementary Methods (Text S1). 200 nm positively-stained sections and 90 nm negatively-stained sections were imaged in a Tecnai-12 G2 transmission electron microscope at 120 KeV, and 300 nm sections were imaged in a Tecnai G2 TF30-FEG microscope at 300 KeV (FEI Company, Holland) in a dual-axis tomography holder (2040; Fischione Instruments, Export, PA). Dual axis tilt series (+/−60°; 1° intervals), including multi-frame montaged datasets, were acquired automatically using the SerialEM software package [70]. Tomographic data were aligned, backprojected, analyzed and segmented using IMOD [71]. The Gag lattice in the tomographic slice closest to the equator of each virion or budding profile slice was measured in five randomly selected areas as a line from the base of the innermost layer to the outside of the outermost layer (green lines in Figures S3B and S6D) using IMOD [71]. The values were combined to give an average Gag thickness for each virion. The symmetry of the Gag lattice was evaluated by Fourier transformation of Gag regions in negatively-stained tomograms. Budding profiles were viewed in tomographic slices taken near the surfaces of their Gag layers (Figure 3B) and images were displayed using the Slicer tool in IMOD, which allowed for 3-D rotation. When the Gag structure was optimally oriented, the image was transformed to Fourier space using an FFT algorithm within IMOD. Selected tomographic datasets are available at http: //www. br. caltech. edu/bjorker/ladinsky, on the Electron Microscopy Databank (http: //www. emdatabank. org) under submission number 28207, or will be provided upon request.

HIV/AIDS remains a global public health problem with over 33 million people infected worldwide. High-resolution imaging of infected tissues by three-dimensional electron microscopy can reveal details of the structure of HIV-1, the virus that causes AIDS, how it infects cells, and how and where the virus accumulates within different tissue sub-structures. Three-dimensional electron microscopy had previously only been performed to image infected cultured cells or purified virus. Here we used three-dimensional electron microscopy to examine an active infection in the gastrointestinal tract of HIV-1–infected mice with humanized immune systems, allowing visualization of the interplay between the virus and host immune cells. Recapitulating the course of infection in humans, immune cells were depleted in infected humanized mouse gut-associated lymphoid tissue, and individual HIV-1 particles were detected as they budded from host cells and accumulated in pools between cells. HIV-1 was mapped to different substructures and cell types within the gut, and free virions were found to accumulate in pools between cells and also to infect adjacent cells via regions of cell-to-cell contact called virological synapses. Our three-dimensional imaging of an HIV-1 infection in tissue uncovered differences between cultured cell and tissue models of HIV-1 infection and therefore furthered our understanding of HIV-1/AIDS as a disease of mucosal tissues.

Abstract Introduction Results Discussion Materials and Methods

medicine immune cells viral classification immunology host-pathogen interaction microbiology lymphoid organs retroviruses animal models immunodeficiency viruses model organisms animal models of infection infectious diseases viral immune evasion hiv t cells biology mouse immune system virology hiv clinical manifestations viral diseases

2014

Electron Tomography of HIV-1 Infection in Gut-Associated Lymphoid Tissue

Chromatoid bodies (CBs) are spermiogenesis-specific organelles of largely unknown function. CBs harbor various RNA species, RNA-associated proteins and proteins of the tudor domain family like TDRD6, which is required for a proper CB architecture. Proteome analysis of purified CBs revealed components of the nonsense-mediated mRNA decay (NMD) machinery including UPF1. TDRD6 is essential for UPF1 localization to CBs, for UPF1-UPF2 and UPF1-MVH interactions. Upon removal of TDRD6, the association of several mRNAs with UPF1 and UPF2 is disturbed, and the long 3’ UTR-stimulated but not the downstream exon-exon junction triggered pathway of NMD is impaired. Reduced association of the long 3’ UTR mRNAs with UPF1 and UPF2 correlates with increased stability and enhanced translational activity. Thus, we identified TDRD6 within CBs as required for mRNA degradation, specifically the extended 3’ UTR-triggered NMD pathway, and provide evidence for the requirement of NMD in spermiogenesis. This function depends on TDRD6-promoted assembly of mRNA and decay enzymes in CBs. During mammalian gametogenesis, substantial changes in chromosome structure and in gene expression profiles occur. Male germ cell transcription and transcriptomes have been studied quite extensively. Some transcriptome studies concern the whole testis, and thus represent combined somatic and germ cell data, which limits germ cell-specific conclusions. Other transcriptome reports describe meiosis- and/or postmeiosis-specific gene expression patterns (reviewed in [1]), some reports deal with the silencing of unsynapsed axial elements in early meiosis (reviewed in [2]), with the silencing associated with histone replacement by specific histone variants or by protamines (reviewed in [3]), with specific RNA species such as small RNA families (reviewed in [4]), with alternative splicing in germ cells (reviewed in [5]) or with transcription factors acting specifically in male germ cells (reviewed in [6]). Various RNA binding proteins were studied (reviewed in [7] and are mostly involved in processing of RNA species. Proper post-transcriptional processing of RNA molecules is essential for germ cell development and thus to producing gametes. For example, in late spermiogenesis, transcripts with short 3’ untranslated regions (UTRs) become preferred and are more stable [8] due to changes in 3’ UTR processing factors [9]. An example of such regulation is the Rnf4 transcript of which a long isoform is present in spermatocytes and a shorter isoform, due to truncation of the 3’ UTR at an upstream polyadenylation site, is present in spermatids [10]. The localization and regulation of mRNA translation is also specifically regulated. mRNAs coding for proteins important for the late stages of spermiogenesis, such as Prm1 and Prm2, are synthesized in the early round spermatid stage, but reside in translationally repressed mRNP complexes. The transcripts become competent for active translation only during the later stages [11]. The apparent reason for such regulation is the transcriptional inactivity of the elongating spermatids due to the extensive nuclear condensation, i. e. transcripts are produced early and stored for later translation. RNA-rich structures termed “nuage’” or “granules” appear in spermatogenic cells and seem to orchestrate this peculiar post-transcriptional program. The cytoplasmic presence of “germ granules” is a unique feature of germ cells is. Germ granules are RNA-rich, non-membranous cytoplasmic structures that have been proposed to play important roles in RNA post-transcriptional regulation [12]. On the basis of structural features and protein composition, the different germ granules of mammalian spermatogenic cells are designated, for example, chromatoid bodies (CBs) in spermatids and intermitochondrial cement (ICM) in spermatocytes [13]. The CBs appear first as fibrous and granulated material in the interstices of mitochondrial clusters and in the perinuclear area of pachytene spermatocytes. After meiosis, the CBs condense into one single lobulated, perinuclear granule in round spermatids and disassemble later during spermatid elongation. It has been shown that CBs contain mRNA, miRNA and piRNA species [14,15]. Proteins that participate in RNA transport such as KIF17b, miRISC (miRNA induced silencing complex) proteins such as the Argonaute family proteins AGO2, AGO3 and Dicer, proteins implicated in piRNA biogenesis and function such as MIWI and MILI [14,16], RNA helicases/binding proteins such as MVH (DDX4), UPF1, GRTH (DDX25) and PABP and RNA decaying factors such as SMG6 [17] were recently found in CBs. According to their molecular composition, CBs were proposed to function in translational repression, RNA silencing and mRNA storage. Thus, CBs may provide a platform for various RNA processing enzymes and processes, which are, however, little described. The CBs are enriched also in TUDOR domain (TDRD) containing proteins. It was proposed that the TUDOR domains of these proteins provide interaction interfaces and create a scaffold to organize the CB structure [18]. The association of TDRD5, TDRD9 and TDRD7 with piRNA biogenesis enzymes such as MILI and MIWI is essential for piRNA biogenesis and retrotransposon mRNA silencing, and established a role of CBs in piRNA processing [19–21]. A role of CBs in splicing was suggested based on the observation that TDRD1 participates in complexes with snRNAs in the context of CBs [22]. Previous work by us [23] showed that TDRD6 is major component of CBs and is required for its architecture. Ablation of TDRD6 disrupts the CB structure and leads to developmental arrest at the round-to-elongated spermatid stage. Altered presence of miRNAs was observed in Tdrd6-/- spermatocytes, but piRNA biogenesis and retrotransposon silencing were not affected. However, whether TDRD6 is implicated in other mRNA metabolic processes that may occur within the CB was unknown. To gain insights into functions of TDRD6 and thus likely of the CB, we performed proteomics of purified CBs. Having identified UPF1 and UPF2 in the CB, which are key proteins in the nonsense mediated mRNA decay (NMD) pathway, we aimed at determining the contribution of TDRD6 to mRNA decay. Previous reports on processes associated with the 3’ end of mRNAs in spermatogenesis describe specific signals embedded in the 3’ UTR sequences or with individual proteins binding there (reviewed in [24,25]. However, the NMD pathway has not been described for mouse or human spermatocytes or spermatids. Processes and complexes that serve mRNA stability and function in germ cells are not sufficiently understood. We show here that TDRD6 is essential for UPF1 localization to CBs and is critical for UPF1-UPF2 and UPF-MVH interactions. We report that a specific branch of NMD, the 3’ UTR length-triggered pathway, but not the downstream exon-exon junction dependent mode of NMD, is affected by absence of TDRD6 and thus CB distortion. We further show that association of some mRNAs with UPF1 is impaired in Tdrd6-/- spermatids, perturbing mRNA processing. We suggest that in spermatids TDRD6 is required for the specific long 3’ UTR dependent NMD pathway, which most likely acts within the CB. TDRD6 was proposed to play an architectural role in the assembly of CBs such as a scaffold protein. Morphological studies showed that in Tdrd6 -/- spermatids, the CBs were found less compacted and of lower density [23]. We investigated the contribution of TDRD6 to CB composition by determining the protein constituents of CBs from Tdrd6+/- and Tdrd6 -/- spermatids. Based on a method described previously [15] we isolated CBs from adult Tdrd6+/- and Tdrd6 -/- testes. Testicular cell suspensions were chemically fixed to preserve the CB structures during the subsequent step of cell lysis in high stringency buffer. The lysates were centrifuged at low speed to acquire a CB-rich pellet. By immunostaining of MVH we could observe the presence of large, ring-like CB structures in Tdrd6+/- samples, and in Tdrd6 -/- samples smaller structures that may represent less compacted CBs as expected for in Tdrd6 -/- cells or precursor building blocks of CBs (S1Aii Fig). Next, Tdrd6+/- and Tdrd6 -/- CBs were attached to anti-MVH Dynabeads and immunoprecipitated (Fig 1A). Similar efficiency of immunoprecipitation from Tdrd6+/- and Tdrd6 -/- samples was obtained as seen by immunoblotting of the preparations for MVH (S1B Fig). The immunoprecipitated Tdrd6+/- and Tdrd6 -/- CB samples were resolved in SDS-PAGE gels and subjected to mass spectrometric analysis. Substantial differences in protein content were observed in TDRD6-deficient CBs compared to controls. We found 286 proteins in Tdrd6-/- CB preparations and 254 proteins in Tdrd6+/- CBs (S3 Table). We reckon that TDRD6 is key to supporting a normal protein composition of CBs. Only 96 proteins are present in both samples and thus do not require TDRD6 for their assembly in CBs. Those 96 proteins were excluded form the further analysis and we focused on the 158 proteins enriched exclusively in Tdrd6+/- CB samples (Fig 1B). The 190 proteins not present in unperturbed CBs but present in absence of TDRD6 may result from aberrant associations with MVH made possible by the removal of TDRD6. We further analyzed the proteome data by focussing on proteins whose presence in the CB depends upon TDRD6. We performed Domain and GO term analysis through the DAVID platform [26] and QIAGEN’s Ingenuity Pathway Analysis (IPA, QIAGEN Redwood City). Analysis of enriched domains revealed that proteins with TUDOR domains, RNA binding domains, or helicase domains are enriched in the Tdrd6+/- CBs (Fig 1C) confirming previous studies [15]. The most represented GO terms and thus molecular and cellular functions are Cell Death and Survival, Cell Development, Protein Synthesis and RNA metabolism such as RNA trafficking, RNA Damage and Repair, RNA Post-Transcriptional Modification (Fig 1D). Thus, the CB localization of proteins bearing RNA binding domains and of proteins that participate in RNA post-transcriptional modification mechanisms was confirmed and is prominently affected by the loss of TDRD6. Among the proteins identified in the “RNA Post-Transcriptional Modification” group are DEAD box RNA helicases (DDX21, DDX25), snoRNP related proteins (IMP3, DKC1, NOP56, NOP58), rRNA metabolism related proteins (EBNA1BP2, FBL, NCL), pre mRNA binding protein (HNRNPH3), exon junction complex proteins (EIF4A3, CASC3, RBM8A) and RNA decaying enzymes (UPF1, SMG6) (Fig 1E). UPF1 is a key factor of nonsense mediated mRNA decay (NMD). Initially the NMD pathway was considered as a quality control system that recognizes and degrades aberrant mRNAs with truncated open reading frames (ORF) due to the presence of a premature termination codon (PTC) [27]. However, recent studies demonstrated a general role of NMD in post-transcriptional regulation of non-aberrant mRNAs. Upstream ORF (uORF), introns in 3’ UTR and long 3’ UTRs have been identified as features that activate NMD [28]. UPF1 was found to be enriched at long 3’ UTR sequences [29,30] and increased association of UPF1 with the 3’ UTR triggers the decay of the mRNA [31]. Along with UPF1, UPF2 supports the decay of mRNAs with long 3’ UTR [32]. UPF1 was previously shown to be a component of the CB [17] and the fact that we identified UPF1 in intact, but not in disturbed CBs, motivated us to further analyze UPF1 and its partner UPF2 in spermatids. Generally, very little is known about the presence and function of UPF complexes in germ cell development. To address whether the expression of Upf genes is developmentally regulated in the testis, we isolated and analyzed RNA from testis of successive days post partum, i. e. during the first wave of spermatogenesis and spermiogenesis (S2A Fig). Upf1 and Upf2 are weakly expressed in neonatal testes and their expression increases during the development of spermatocytes and spermatids. Upf1 and Upf2 expression increases moderately in meiotic cells, which were identified by Hormad1 expression at day 10 postpartum (pp) [33]. Later Upf1 and Upf2 expression peaks and coincides with high Tdrd6 expression at day 22 pp, which marks the appearance of early round spermatids [23]. Upf1 and Upf2 expression remain at high levels as round spermatids differentiate to elongated spermatids marked by Prm2 expression form 26 pp onwards. These data suggest that UPF complexes may play an hitherto undescribed role in the late meiotic and postmeiotic stages of spermatogenesis. The parallel expression of Upf1, Upf2 and Tdrd6 led us to test whether the levels of Upfs are affected by TDRD6 deficiency. We isolated mRNA (S2B Fig) and protein extracts (S2C Fig) from Tdrd6+/- and Tdrd6-/- spermatids and compared mRNA and protein expression of UPFs. The absence of TDRD6 and thus the disruption of CBs did not affect levels of UPFs protein or mRNA. We investigated the associations—direct or indirect—of TDRD6, MVH, UPF1 and UPF2 in Tdrd6+/- and Tdrd6-/- by co-immunoprecipitation in the presence or absence of RNAse A treatment. RNAse A treatment efficiency was assessed by RNA electrophoresis of the flow-through sample of the IPs (S2D Fig). We investigated the interaction between MVH, UPF1 and UPF2 by performing MVH immuno-precipitation (IP) (Fig 2Ai). Vinculin (VINC) a membrane-cytoskeletal protein was used as a loading and negative co-IP control (Fig 2Aii). The previously reported in vitro interaction between TDRD6 and MVH [23] was recapitulated by co-IP from spermatids and was independent of RNAseA inclusion (Fig 2Aiii). Reverse IP of TDRD6 (Fig 2Bi) showed also that TDRD6 associated with MVH (Fig 2Biii) irrespectively of RNAse A treatment. UPF1 co-IP with MVH (Fig 2Aiv) was observed specifically only in Tdrd6+/- spermatids but not in the Tdrd6-/- spermatids or in IgG control IPs and was dependent on RNA. Reverse IP of UPF1 (Fig 2Ci) showed also its association with MVH (Fig 2Cv), but only when TDRD6 and intact RNA were present. UPF2 co-immunoprecipitated with MVH (Fig 2Av) irrespectively of the genotype and RNAse A treatment and the reverse IP of UPF2 (Fig 2Di) demonstrated its association with MVH (Fig 2Dv). Thus, only UPF1, but not UPF2, requires a TDRD6-supported, intact CB for its association with the key CB component MVH. Since we observed differential association of UPF1 and MVH in Tdrd6+/- and Tdrd6-/- spermatids, we investigated the interaction between TDRD6, UPF1 and UPF2 by performing TDRD6 IP (Fig 2Bi). UPF1 co-immunoprecipitated with TDRD6 (Fig 2Biv) specifically only in Tdrd6+/- spermatids but not in the Tdrd6-/- spermatids or in IgG control IPs. Reverse IP of UPF1 (Fig 2Ci) showed also its association with TDRD6 (Fig 2Civ). UPF2 co-immunoprecipitated with TDRD6 (Fig 2Bv) and the reverse IP of UPF2 (Fig 2Di) demonstrated its association with TDRD6 (Fig 2Div). The interactions of TDRD6 and UPF1 or UPF2 are resistant to RNAse A treatment. This data suggested the involvement of TDRD6 in complexes containing MVH, UPF1 and UPF2. UPF1 binds directly to UPF2 via an UPF2-interacting domain [34], but upon IP of UPF1 from Tdrd6+/- and Tdrd6-/- round spermatids (Fig 2Ci), UPF2 was found to co-immunoprecipitate only in the Tdrd6+/- samples, UPF1 interaction with UFP2 is almost entirely abrogated upon loss of TDRD6 (Fig 2Ciii). Confirming these results, IP of UPF2 from Tdrd6+/- and Tdrd6-/- round spermatids (Fig 2Di) showed that UPF2 associated with UPF1 in the Tdrd6+/- sample but hardly in absence of TDRD6 (Fig 2Diii). UPF1-UPF2 association in Tdrd6+/- samples was not affected by the presence of RNAse A as expected. In conclusion, the absence of TDRD6, accompanied by distortion of CB structure, prevented UPF1-MVH and UPF1-UPF2 interactions. Given the distinctly TDRD6-dependent associations of UPF1 and UPF2 shown above, the localization of UPFs in germ cells was determined by staining meiotic and postmeiotic cells with antibodies against UPF1 and UPF2. The localization of UPF proteins has been extensively investigated in mammalian cell lines where UPF1 is mainly cytoplasmic [35], but a fraction of UPF1 resides in the nucleus where it promotes DNA replication, S phase progression [36] and telomere stability [37]. More recently it was shown that UPF proteins localize to P-bodies in mammalian cells [38]. UPF2 is a cytoplasmic protein [35]. In meiosis I spermatocytes, positive for SYCP3, UPF1 localized to the perinuclear space of the cytoplasm and there was hardly any staining observed in the nucleus (Fig 3A). UPF2 was distributed in some clusters throughout the cytoplasm (Fig 3B). No apparent co-localization with TDRD6 was detected suggesting no participation in the precursor structures of CB in meiotic cells. The localization pattern of UPFs in meiotic cells remained unaffected by the loss of TDRD6 (Fig 3A and 3B). In Tdrd6+/- mice, UPF1 was absent form the cytoplasm of round spermatids and was exclusively concentrated in CBs where it co-localized with MVH and TDRD6 (Fig 4A and 4Ci). However, in Tdrd6-/- round spermatids UPF1 failed to co-localize with MVH positive foci, i. e. with the distorted CBs found in Tdrd6-/- round spermatids, and remained diffuse in the perinuclear cytoplasm (Fig 4A and 4Cii). This suggested that TDRD6-positive, undistorted CBs are required for UPF1 re-localization from the cytoplasm of meiotic cells to the CBs of round spermatids. On the other hand, UPF2 (Fig 4B, 4Ciii and 4Civ) primarily co-localized with MVH in Tdrd6+/- and Tdrd6-/- CBs. UPF2 is a newly identified component of CBs. 100% (n = 67) and 97% (n = 76) of Tdrd6+/- CBs scored contain UPF1 and UPF2, respectively. In Tdrd6-/- round spermatids 0% (n = 41) of CBs contained mUPF1, while mUPF2 localized to 86% (n = 72) of Tdrd6-/- CBs (S3 Fig). This indicated a TDRD6 independent manner of localization of UPF2 to CBs, although CB presence of these proteins was slightly affected probably by the distorted architecture of the Tdrd6-/- CBs. UPF1 is a key factor of nonsense mediated mRNA decay (NMD). Initially the NMD pathway was considered as a quality control system that recognizes and degrades aberrant mRNAs with truncated open reading frames (ORF) due to the presence of a premature termination codon (PTC) [27]. PTCs can arise from aberrant splicing events, 5’ UTR upstream open reading frames (uORFs) or by mutations. In principle, a termination codon residing more than ~55 nucleotides upstream of an exon-exon junction complex is considered a PTC and the transcript is a likely target for the so called downstream exon-exon junction stimulated (dEJ) NMD [27]. We hypothesized that mis-localization of UPF1 and failing interaction of UPF1 with UPF2 in the Tdrd6-/- strain would lead to accumulation of NMD sensitive transcripts in Tdrd6-/- round spermatids. To test whether the dEJ mode of NMD was affected by loss of TDRD6, we generated the mRNA profiles of germ cell populations enriched for round spermatids of Tdrd6+/- and Tdrd6-/- mice by deep sequencing. The MACS-purified population was more than 95% positive for expression of the marker hCD4 in both genotypes (S4A Fig), and the hCD4 is expressed at the same levels in Tdrd6+/- and Tdrd6-/- cells (S4B Fig). The preparations enriched for round spermatids contained approximately 70% round spermatids in both Tdrd6+/- and Tdrd6-/- samples (S4C Fig), and the fraction of primary and secondary spermatocytes was about the same. We used 4 biological replicates, i. e. round spermatid samples of four individual animals per genotype and acquired over 250 million RNA seq reads (S5 Fig). We aligned the RNA sequence reads with TopHat, assembled transcripts with Cufflinks and annotated them using Ensemble v67 [39–41]. Expression analysis was performed with Cuffdiff with a FDR of 0. 1. We used 2 different approaches to classify transcripts with PTCs, which are putative dEJ NMD targets to be further analyzed. In the first approach, the mouse annotation of Ensembl v67 was used for the classification of the transcripts. Transcripts which had the biotype “Nonsense Mediated Decay” were extracted from the complete data set and used for the subsequent analysis. Here, if the coding sequence of a transcript finishes >50 bp from a downstream splice site, it is tagged as a putative NMD sensitive transcript. In the second approach, SpliceR version 1. 12. 0 was used for the annotation of transcripts with PTC [42]. We used the Cufflinks results files for SpliceR and filtered the isoforms with the setting" expressedIso" and" isoOK" within SpliceR. Furthermore SpliceR requires CDS information, which was retrieved with SpliceR internal function from UCSC. Annotation of transcripts was done with “annotatePTC” and transcripts were extracted, which were set to PTC equals TRUE. These transcripts were used for the comparison. There are 564 dEJ NMD sensitive transcripts identified only by Ensemble v67 database, 2004 transcripts identified only by the SpliceR pipeline and 1268 identified by both ways (S6A Fig). 1,520 (82%) out of 1,832 dEJ NMD sensitive transcripts by Ensemble v67 have a log2 fold change between -1 and 1 and 1,089 (59%) transcripts have a log2 fold change between -0. 5 and 0. 5 (Fig 5A). Similarly 2780 (85%) out of 3272 dEJ NMD sensitive transcripts by SpliceR analysis have a log2 fold change between -1 and 1 and 2031 (62%) transcripts have a log2 fold change between -0. 5 and 0. 5 (Fig 5A) indicating normal regulation of dEJ NMD transcripts between Tdrd6+/- and Tdrd6-/- round spermatids. Next, we looked at the expression values measured in FPKM for the dEJ NMD sensitive transcripts in both genotypes (Fig 5B). A Wilcoxon-Mann-Whitney test (p-value = 0. 4789 for the dEJ NMD sensitive transcripts by Ensemble v67 and p-value = 0. 8998 for the dEJ NMD sensitive transcripts by SpliceR analysis) showed that there is no difference in the FPKM values of the dEJ NMD sensitive transcripts between the TDRD6-proficient and -deficient samples. We further examined a number of known NMD substrates to validate the high throughput analysis. Abnormal splicing events such as intron inclusion, exon skipping and splicing downstream of a normal termination codon can induce the dEJ mode of NMD. During an intron inclusion event, a PTC can be introduced either because it resides in the included intron or is generated due to frameshift of the physiological ORF. In an exon skipping event a frameshift of the ORF might produce a PTC. We tested PTC generation by intron inclusion and exon skipping events that characterized for specific transcripts in other murine tissues [43]. Performing RT-PCR using specific primers (arrows), which span intron inclusion events for Pkm2, Srsf2, Srsf3, Hnrpl, Brd2 and exon skipping events for Hnrnph3 and Mdm2, we investigated NMD sensitive transcript variants (marked by arrowheads). NMD sensitive or NMD resistant variants showed the same levels in Tdrd6+/- and Tdrd6-/- samples (Fig 5C). Auf1 mRNA can also be used as a marker of NMD efficiency due to its unusual 3’ UTR architecture [44]. Splicing of exon 9 and exon 10 generates an exon junction more than 50 nt downstream of the normal termination codon, producing NMD sensitive transcript variants II and III (S6B Fig and [44]). We designed specific primers to map different splicing events and found that splicing events producing the NMD sensitive transcripts II and III occur largely the same in Tdrd6+/- and Tdrd6-/- samples (S6C Fig). Finally, uORFs of a transcript would lead to premature translational termination and subsequent NMD. We compared the expression of transcripts with uORFs such as Atf5, Map3k14, Arfp1 and Dusp10, which were previously shown to be recognized by NMD in other cell types [45,46] by RT-qPCR. We found no difference in their expression levels in Tdrd6+/- and Tdrd6-/- round spermatids (Fig 5D). Together these data showed normal function of the downstream exon-exon junction dependent mode of NMD in TDRD6 deficient spermatids. Although NMD was initially characterized in PTC dependent mRNA degradation as a quality control mechanism, there is evidence that NMD is implicated in the metabolism of normal mRNAs. A well studied feature of mRNAs which can elicit NMD is the long 3‘ UTR. UPF1 was found to be enriched at long 3’ UTR sequences [29,30] and increased association of UPF1 with the 3’ UTR triggers the decay of the mRNA [31] in an UPF2 and SMG6 dependent way [32]. Since we found TDRD6 to associate with UPF1 and UPF2, we assessed the effect of TDRD6 deficiency on the general mRNA transcriptome. We analyzed the expression of normal mRNAs in the transcriptome data described above, derived from germ cell populations enriched for round spermatids from Tdrd6+/- and Tdrd6-/- mice. We aligned the RNA seq reads with TopHat, assembled transcripts with Cufflinks and annotated them using Ensemble v67 [39–41]. Expression analysis was performed with Cuffdiff with a FDR of 0. 1 and we found 2704 transcripts to be significantly (p-value <0. 05) mis-regulated in absence TDRD6 and thus of intact CBs. More specifically, 1375 were up-regulated and 1329 down-regulated in Tdrd6-/- round spermatids (Fig 6A and S1 Table). Thus, TDRD6 is required for the presence of a proper mRNA repertoire in spermatids. To further characterize the changes of the mRNA content of CB-disrupted spermatids, we grouped mis-regulated transcripts with p-values <0. 05 into 3 groups according to the length of their 3’ untranslated region (UTR): short 3’ UTR <350 nt, medium 3’ UTR >350 nt and <1500 nt and long 3’ UTR >1500 nt. We analyzed the log2 fold distribution of these groups of transcripts. We observed that the majority of mis-regulated transcripts (514,82% out of 628) with a long 3' UTR had a positive log2 fold change, i. e. they are present at higher levels in the Tdrd6-/- compared to the Tdrd6+/- round spermatids. The distribution of positive and negative log2 fold change of mis-regulated transcripts was not significantly altered for short and medium 3‘UTR length groups, but the log2 fold distribution of transcripts with long 3’ UTR, showing enrichement of upregulated transcripts in Tdrd6-/- samples was statistically different from the others (Wilcoxon-Mann-Whitney test p-value <2. 2−16) (Fig 6B and S2 Table). We conclude that the mis-regulation of mRNAs in TDRD6 deficient, CB-disrupted spermatids correlates with an accumulation of transcripts carrying long 3’ UTRs. Significantly mis-regulated transcripts with long 3’ UTRs >1500 nt correspond to 628 genes. 288 (46%) genes have a single transcript with long 3 UTR and 340 (54%) genes have multiple transcripts and transcripts with long 3’ UTRs among them (S6D Fig). These 340 genes code for 1176 putative transcript isoforms. These putative isoforms include the 340 long 3’ UTR transcripts tested previously, but in addition there are 167 isoforms that have 3’ UTRs shorter than 1500 nt and for the rest there is no reliable information on the 3’ UTR length. From the 167 short 3 ‘UTR isoforms of the genes with mis-regulated long 3’ UTR transcript isoforms, there are 15 transcripts, corresponding to 14 genes, showing a significant mis-regulation, while the large majority of 152 transcripts remained unchanged. Thus loss of TDRD6 affects specifically the long 3’ UTR isoforms of genes with multiple isoforms with different 3’ UTR lengths. Further, we used the database of EMBL-EBI Expression Atlas and looked the expression analysis of different murine tissues. There are 10760 genes expressed in testis above a standard expression cutoff value of 0. 5. We consider a transcript to be testis-specific when its expression is 10 times higher in testis than in any other tissue examined. There are 1851 genes that fall in this group. Of the 628 mis-regulated transcripts with long 3’ UTR there are 35 which can be considered testis specific (6%). 514 transcripts with long 3‘ UTRs are higher in the Tdrd6-/- round spermatids and 15 of them are testis-specific (3%). 113 transcripts with long 3’ UTRs are lower in the Tdrd6-/- round spermatids and 20 of them are testis-specific (17%). UPF1 is an RNA helicase that can bind to all transcripts, although it preferentially associates with transcripts carrying long 3’ UTRs [30,47]. Our initial observation that mRNAs with long 3’ UTR tend to be up-regulated in CB distorted round spermatids led us to investigate particular mRNAs with 3’ UTR length >1000 nt with respect to UPF1 binding, mRNA levels and translational potential. We assessed the in vivo binding of UPF1 to selected mRNAs by performing anti-UPF1 RNA immunoprecipitation (RIP) from Tdrd6+/- round spermatids, followed by RT PCR. We examined 9 transcripts that carried long 3’ UTRs >1000 nucleotides and 2 transcripts with 3’ UTRs <350 nucleotides as a negative control. Positive signals in the anti-UPF1 RIP RT-qPCR were obtained for transcripts with long 3’ UTR as Spen (3’ UTR length = 1016 nt), Diap1 (3’ UTR length = 2244 nt), Mdc1 (3’ UTR length = 2040 nt), Ube2c (3’ UTR length = 1598 nt), Twsg1 (3’ UTR length = 3218 nt), Dixdc1 (3’ UTR length = 3438 nt), Daam1 (3’ UTR length = 2478 nt), Yap1 (3’ UTR length = 2497 nt) (Fig 7A) and Wnt3 (3’ UTR length = 1889 nt) (S7A Fig) and for transcripts with short 3’ UTR as Ecsit (3’ UTR length = 78 nt), Prss51 (3’ UTR length = 119 nt) (S7A Fig), suggesting that UPF1 binds in vivo to all these transcripts. Next, we expanded the analysis of mRNA to UPF1 binding in the Tdrd6+/- versus Tdrd6-/- round spermatids. Significantly decreased binding to UPF1 in Tdrd6-/- round spermatids was observed for transcripts with long 3’ UTR such as Spen, Mdc1, Diap1, Ube2c, Twsg1, Dixdc1, Daam1 and Yap1 (Fig 7Bi–7Bviii). To assess the effect of impaired UPF1-mRNA binding on the mRNA levels, we performed RT-qPCR. The presence of mature mRNAs of Spen, Mdc1, Diap1, Ube2c, Twsg1, Dixdc1, Daam1 and Yap1, i. e. the transcripts with long 3’ UTR showing decreased association with UPF1 in Tdrd6-/- samples, was increased 2- to 3-fold in Tdrd6-/- round spermatids (Fig 7C). The pre-mRNA levels of these genes remained unchanged (Fig 7D), showing that the higher levels were not caused by increased transcription, but by increased stability. These results suggested that UPF1 binding to mRNAs carrying long 3’ UTR is perturbed upon TDRD6 deletion—and thus distortion of the CB and mis-localization of UPF1 –and correlates with increased mRNA stability possibly through decreased degradation. We identified short 3’ UTR such as Ecsit and Prss51 (S7Bi and S7Bii Fig) and long 3‘UTR such Wnt3 (S7Bii Fig) to associate with UPF1 equally between Tdrd6+/- and Tdrd6-/- round spermatids. The expression levels of transcripts that show unaffected association with UPF1 among the different genotypes were mis-regulated both on the mature mRNA (S7C Fig) and nascent pre mRNA levels (S7D Fig) possibly through indirect mechanisms. In RIP experiments using anti UPF2 antibody, we found that most of the mRNAs with long 3’ UTR that associate with UPF1 also associate with UPF2 in round spermatids (Fig 7E). The analysis of binding of UPF2 to long 3’ UTR mRNAs in the Tdrd6+/- versus Tdrd6-/- round spermatids showed significantly decreased binding of Spen, Mdc1, Diap1, Ube2c, Dixdc1 and Yap1 to UPF2 in absence of TDRD6 (Fig 7Fi–7Fvi). Thus, both UPF1 and UPF2 associations with long 3’ UTR mRNAs are affected by loss of TDRD6. To assess the effect of impaired UPF1-mRNA binding on the translation potential of UPF1 bound mRNAs, we performed sucrose gradient fractionation to isolate translationally active fractions, which are those rich in polysomes (fractions #1–7), translationally inactive fractions that are rich in ribosomal subunits (fractions #8–10), and ribosome-free mRNPs (fractions # 11–12) (Fig 8A). Although the majority of UPF1 protein from yeast and human cell line cultures was shown to associate with polysomes [48], we found that in 26 dpp murine testis, enriched for round spermatids, UPF1 was underrepresented in polysome/ribosome fractions #1–10, confirmed by the presence of RPS6. The majority of UPF1 was detected in fractions #11–12 containing ribosome-free mRNPs, indicated by GAPDH. The same distribution was observed for MVH. The UPF1 and MVH association with ribosome-free mRNPs was not compromised by absence of TDRD6 (Fig 8B). To test the translational capacity of UPF1-associated mRNA species we extracted RNA from each fraction and performed RT-PCR. The transcripts with long 3’ UTR, which associated less with UPF1 in Tdrd6-/- samples such as Spen (Fig 8C and 8D), Diap1 (S8A and S8B Fig) Mdc1 (S8A and S8C Fig), showed relatively equal distribution in translationally active fractions #1–7 (52% for Spen, 43% for Diap1 and 53% for Mdc1) and translationally inactive ribosome fractions #8–12 (48% for Spen, 57% for Diap1 and 47% for Mdc1) in Tdrd6+/- samples. In contrast, in Tdrd6-/- samples, Spen (Fig 8C and 8D), Diap1 (S8A and S8B Fig) and Mdc1 (S8A and S8C Fig) showed increased abundance in the translationally active fractions #1–7 (75% for Spen, 68% for Diap1 and 59% for Mdc1) and decreased abundance in translationally inactive ribosome and ribosome-free fractions #8–12 (25% for Spen, 32% for Diap1 and 41% for Mdc1). Other transcripts with long 3’-UTRs, i. e Twsg1 (S8A and S8D Fig) and Yap1 (S8A and S8E Fig), are underrepresented in translationally active fractions #1–7 (37% for Twsg1 and 38% for Yap1) and overrepresented in translationally inactive fractions #8–12 (63% for Twsg1 and 62% for Yap1) in Tdrd6+/- samples. On the other hand, in Tdrd6-/- samples, Twsg1 (S8A and S8D Fig) and Yap1 (S8A and S8E Fig) mRNAs are more abundant in the translationally active fractions #1–7 (49% for Twsg1 and 48% for Yap1) and almost equals the abundance in translationally inactive fractions #8–12 (51% for Twsg1 and 52% for Yap1). Together this shows that mRNAs with reduced UPF1 binding in Tdrd6-/- spermatids such as Spen, Diap1, Mdc1, Twsg1 and Yap1 associated to a larger extent with polysomal fractions compared to the controls, suggesting these mRNAs were more actively translated in Tdrd6-/- spermatids. On the other hand, mRNAs which showed no difference in UPF1 binding between Tdrd6+/- and Tdrd6-/- mice, such as Ecsit (S7Ei and S7Eii Fig), Prss51 (S7Ei and S7Eii Fig) and Wnt3 (S7Ei and S7Eii Fig) displayed a similar distribution pattern across all fractions in Tdrd6+/- and Tdrd6-/- samples (Ecsit: 45% in translationally active fractions #1–7 and 55% in translationally inactive fractions #8–12 in Tdrd6+/- and 52% in translationally active fractions #1–7 and 48% in translationally inactive fractions #8–12 in Tdrd6-/-; Prss51: 47% in translationally active fractions #1–7 and 53% in translationally inactive fractions #8–12 in Tdrd6+/- and 41% in translationally active fractions #1–7 and 59% in translationally inactive fractions #8–12 in Tdrd6-/- Wnt3: 43% in translationally active fractions #1–7 and 57% in translationally inactive fractions #8–12 in Tdrd6+/- and 35% in translationally active fractions #1–7 and 65% in translationally inactive fractions #8–12 in Tdrd6-/-). Overall, this data set shows that decreased binding of mRNA to UPF1 in Tdrd6-/- round spermatids correlates with increased mRNA stability and translational potential. The aim of the present study was to define the role of a germ cell-specific protein, TDRD6, which in spermatids resides in the CB and is a main structural component of this cell organelle whose functions remained hitherto largely unknown. CBs are considered large RNP complexes in the cytoplasm close to the nuclei of round spermatids. CBs were proposed to be sites of accumulation of mRNPs exported from the nuclei [49]. It was postulated that these mRNPs are translationally repressed through piRNAs or miRNAs or by translational regulators such as Nanos, Pum and Gemin3 [49,50]. These mRNPs would be stored or targeted to other cytoplasmic sites [51]. The dominance of Tudor domain proteins in the CB and their interactions with PIWI and other proteins was suggested to provide the molecular scaffold for CB [18]. TDRD6 and TDRD7 were shown to be indispensable for CB architecture [20,23], and in our study we used TDRD6 deficient mice. The CB has been implicated in piRNA biogenesis and retrotransposon silencing [19]. The loss of TDRD6 results in male infertility and disruption of CB architecture of which a remnant “ghost body” is left. Genome methylation remains normal as did retrotransposon silencing, which depends on MIWI and MILI, suggesting that the proper architecture of the CB is required for other functions. To decipher some of these functions, we used proteome analysis to determine differences between Tdrd6+/- and Tdrd6-/- CB compositions and determined distinct perturbations of the CB proteome in TDRD6-deficient samples. In CB preparations we identified 158 proteins that depend on TDRD6 for their enrichment in CBs. To compare these with the transcriptomics data, the 158 IPI protein IDs were converted to 139 Ensemble transcript IDs (88%). Next we looked at the expression values of these transcripts in our RNA deep sequencing analysis of the Tdrd6+/- and Tdrd6-/- round spermatid transcriptomes. Out of the total of 139 transcripts, 96 remained unchanged between the genotypes (70%). 27 transcripts expressed lower in the Tdrd6-/- round spermatids (19%) and 16 transcripts expressed higher in the Tdrd6-/- round spermatids (11%). The failure to identify the 24 transcripts, that expressed lower in Tdrd6-/- round spermatids, in the Tdrd6-/- CB may be due to the very low expression levels. However, the the vast majority of the proteins (82%) are normally or higher expressed in Tdrd6-/- round spermatids, so the failure to identify them in Tdrd6-/- CB is likely a consequence of CB distortion in this mutant. Many of the proteins absent in TDRD6-deficient CBs bear RNA binding domains pointing to a critical role of CBs in RNA metabolism, i. e. post transcriptional regulation. Among the RNA-related proteins found within the CB proteome were components of the RNA degrading machinery, which we further investigated. Our data are in agreement with a recent study [17] that provided first insights into the molecular composition of CBs. A major subset of proteins that localizes to CBs are those implicated in RNA degradation processes e. g. mRNA decapping enzyme DCP1a and RNA endonuclease SMG6. Indeed CBs share common features with P-bodies and the NMD core factor UPF1 was found in both structures [14,38]. Here we demonstrate that key NMD factors UPF1 and UPF2 are highly expressed in post-meiotic male germ cells and accumulate in CBs implying a key role of NMD for the completion of spermatogenesis. We provide evidence that the localization of UPF1 to CBs depends on a TDRD6-supported CB structure, while UPF2 is targeted to CBs via different mechanism (s). We analyzed protein-protein interactions in the presence and absence of TDRD6. In the wild-type situation, MVH and UPF1 associated with UPF2 and TDRD6. In absence of TDRD6, MVH and UPF2 interacted with each other localizing to the CB “ghost body”, but UPF1 failed to associate with them. Thus, TDRD6 supports the formation of UPF1-containing mRNPs in the CBs. It is very unlikely that TDRD6 itself binds directly to RNA, since there are no RNA binding domains identified in this protein. TUDOR domains bind to methylated arginines or lysines [18]. One may speculate that such methylated residues in UPF1 would enable UPF1-TDRD6 interaction, which is subject of future investigations. In any case, TDRD6 likely provides a protein scaffold, where RNA binding proteins are brought into proximity so that correctly assembled mRNPs can be formed and stabilized. Accumulation of UPF proteins in CBs indicated that CBs support NMD, for example CBs may serve as storage sites for NMD proteins or even as sites of active NMD. The loss of TDRD6 and subsequent perturbation of UPF1 interactions did not affect the levels of PTC containing transcripts, thus did not affect PTC induced, downstream exon-exon junction dependent NMD. In Tdrd6-/- round spermatids dEJ-triggered NMD is functional despite the compromised interaction of UPF1 and UPF2, suggesting an alternative pathway of UPF1 activation on a PTC containing transcript. However, we observed increased levels of transcripts with long 3’ UTR in Tdrd6-/- sample, suggesting that TDRD6 supports the long 3’ UTR triggered pathway of NMD. To our knowledge, this is the first mutant that discriminates between different modes of stimulating NMD. We also demonstrate that specific mRNAs with long 3‘ UTR associate with UPF1 and UPF2 in vivo in round spermatids, but this association is much reduced in Tdrd6-/- cells. The reduced association with UPF1 correlated with increased levels of these mRNAs and their increased translational potential in the Tdrd6-/- background. The presence of a few mRNAs with either long or short 3‘ UTR that bind to UPF1 in a TDRD6-independent manner but are nevertheless altered in levels in absence of TDRD6 suggests that TDRD6 regulates the levels of some mRNAs independently of UPF1 through a distinct pathway. It has been shown that the average 3‘ UTR length of transcripts required for spermiogenesis is shorter compared to transcripts required for pre-meiotic, meiotic or testicular cell development [8]. Transcripts with shorter 3’ UTR may be more stably stored for longer periods and thus may be particularly competent for efficient translation during the last stages of spermiogenesis. NMD is important for many developmental processes as systemic depletion of the murine Upf1 gene results in complete loss of NMD and leads to post implantation embryonic death [52]. NMD is essential for hematopoietic stem cells and for B and T lymphocyte maturation, since conditional ablation of murine UPF2 in the hematopoietic system is detrimental to proliferation of progenitor cells and leads to up regulation of aberrant TCR and Ig locus recombination products [53]. On the other hand, NMD activity is down-regulated in neural stem cell upon neurogenic signaling to allow differentiation [54]. Thus, tissue- and cell-type specific roles of NMD exist, but are known in only a few instances. We provide the first evidence of NMD functioning in the regulation of transcripts during spermiogenesis. Successful completion of the spermiogenic program depends strongly on post-transcriptional regulation as the transcriptional production of RNA ceases from the mid to later stages because of the extensive nuclear compaction. The use of mice was approved by the State of Saxony animal welfare officials, Az DD24-5131/339/6 and was performed according to the national and EU guidelines. Construction of TDRD6‐deficient mice was described previously [23]. In all experiments, except otherwise noted, testes from postnatal day 26 (P26) Tdrd6+/- and Tdrd6-/- mice were dissected to be enriched in round spermatid cells. Tdrd6+/- mice used as control for the experiments do not exhibit any phenotype and provide the targeting vector with the hCD4 gene in frame with the Tdrd6 5′ UTR and ATG (start) codon, that allows isolation of TDRD6 expressing cells through an anti hCD4-MACS approach [23]. For cell preparations enriched in round spermatids, the Tunica albuginea was removed and seminiferous tubules resuspended in 10 ml PBS and passed subsequently through 100μm and 40μm stainers. Cells were washed once with PBS and hCD4-positive cells were magnetically labeled with CD4MicroBeads (Miltenyi Biotec) and MACS isolated (Miltenyi Biotec) according to manufacturer instructions. Testes for immunostaining were fixed in freshly prepared 4% PFA for 1h on ice, briefly washed with PBS, and incubated O/N in 30% sucrose. Testes were embedded in OCT blocks, frozen on dry ice, and cryo-sectioned at 7 μm thickness. CBs were isolated according to Meikar et al. (2010) with some modifications. hCD4 positive cells from Tdrd6+/- and Tdrd6-/- adult mice were fixed in 1% PFA (Sigma) solution for 10 min at RT. The reaction was stopped by adding glycine (Roth) pH 7 to a final concentration of 0. 25 M. The fixed cells were lysed by sonication in 0. 5 mL of RIPA buffer (50 mM Tris-HCl at pH 7. 4 (Roth), 150 mM NaCl (Roth), 1% NP-40 (Sigma), 0. 5% sodium deoxycholate (Sigma), 0. 1% SDS (Roth), 1 mM EDTA (Sigma), 1 mM DTT (Roth), 5mM NaF (Sigma), 1mM Na2VO3 (Sigma), 1mM PMSF (Sigma), 1x protease inhibitor cocktail complete mini (Roche) ) supplemented with 100U RNAse inhibitor (Invitrogen). The lysate was centrifuged at 300g for 10 min and the CB enriched pellet resuspended in 0. 5 mL of RIPA buffer. The CBs were immunoprecipitated using Dynabead Protein G (Invitrogen) coupled to rabbit polyclonal anti-MVH (Abcam) O/N at 4°C. Dynabeads were washed 4 times with RIPA buffer and the crosslinks of the isolated CBs were reversed by incubation at 70°C for 45 min in 1x Laemmli buffer. CB samples were separated in mini-protean TGX pre-cast gradient gels (BioRad) and stained with SimplyBlue SafeStain (Life Technologies). Gel pieces were excised from the sample lanes, followed by in-gel digestion with trypsin (Promega) and extraction of the peptides. The peptides were analyzed using LC-MS/MS with an Ultimate 3000 (Dionex Corp, Sunnyvale CA) nanoLC system connected to a LTQ Orbitrap mass-spectrometer (ThermoScientific Corp. , San Jose CA) equipped with an automated nanoelectrospray ion source TriVersa (Advion BioSciences, Ithaca NJ). All MS/MS samples were analyzed using Mascot (Matrix Science, London, UK; version 2. 2. 04). Mascot was set up to search the ipi. MOUSE_V3. 76_20110304 database assuming the digestion enzyme trypsin. Mascot was searched with a fragment ion mass tolerance of 0. 50 Da and a parent ion tolerance of 5. 0 PPM. Oxidation of methionine and propionamide of cysteine were specified in Mascot as variable modifications. Scaffold (version Scaffold_3. 6. 4, Proteome Software Inc. , Portland, OR) was used to validate MS/MS based peptide and protein identifications. Peptide identifications were accepted if they could be established at greater than 95% probability as specified by the Peptide Prophet algorithm [55]. Protein identifications were accepted if they could be established at greater than 99. 0% probability and contained at least 2 identified peptides. Protein probabilities were assigned by the Protein Prophet algorithm [56]. Scaffold normalizes the MS/MS data between samples. Normalization is done on the MS sample level, which is the total sample run through the mass spectrometer. The normalization method that Scaffold uses is to sum the “Unweighted Spectrum Counts” for each MS sample. For the purposes of protein identification, Scaffold uses a ProteinProphet model, assigning the peptide exclusively to the protein with the most evidence. The result is that the peptide has a weight of 1 in one protein and a weight of zero in all other proteins. However, if there are two proteins, and each protein has the same peptide, then each spectrum for this peptide has ions contributed from both proteins. The “Unweighted Spectrum Count” option on Scaffold' s Samples page will count this spectrum twice, once in the first protein and once in the second protein. This count is “unweighted” in the sense that the spectrum counts the same in each of the shared proteins. Scaffold counts unweighted spectra for determining protein abundance. These sums are then scaled so that they are all the same. The scaling factor for each sample is then applied to each protein group and adjusts its “Unweighted Spectrum Count” to a normalized “Quantitative Value”. International Protein Index (IPI) accession numbers of proteins identified more than 2 fold enriched in Tdrd6+/- CB (S3 Table) were uploaded to DAVID platform [26], functional annotation for protein domains from PFAM database was performed with threshold count 3 and threshold EASE 0. 1. The same list was uploaded to QIAGEN’s Ingenuity Pathway Analysis (IPA, QIAGEN Redwood City) and functional analysis was performed with custom parameters. Immunofluorescence labeling of frozen sections of mouse testis was performed using rabbit polyclonal anti UPF1, rabbit polyclonal anti UPF2 [35], rabbit polyclonal anti MVH (Abcam), mouse monoclonal anti SYCP3 [57], guinea pig polyclonal anti C-term TDRD6 (this study). Sections were fixed using 4% PFA for 20 minutes, blocked and permeabilized with 2% BSA, 0. 1% Triton-X100 in PBS and incubated overnight with primary antibodies. Slides were washed with PBST and probed for 2 h with secondary antibodies Alexa-566-labeled goat anti guinea pig, Alexa-488-labeled goat anti rabbit, or Alexa-647-goat anti mouse (Molecular probes, Invitrogen). For double immunostaining rabbit polyclonal antibodies were labeled using the Zenon Rabbit IgG Labeling Kits (Molecular Probes, Invitrogen). Slides washed again with PBST and nuclei were visualized with DAPI. Images acquired with a Zeiss LSM 510 confocal microscope and quantification of signal intensity was done with ImageJ. Tdrd6+/- and Tdrd6-/- round spermatid cell suspension fixed in 1% PFA solution for 10 min at RT to capture RNA-protein and protein-protein interactions. Cells were lysed in 0. 5 mL RIPA buffer supplemented with 100U RNAse inhibitor (Invitrogen) for 20 min on ice. The lysate was centrifuged 1000 rpm for 10’ at 4°C and the protein concentration of the supernatant was quantified. 150 μg of protein extract diluted in IP buffer (50 mM Tris-HCl at pH 7. 4 (Roth), 150 mM NaCl (Roth), 0. 25% Triton-X100 (Sigma), 1 mM EDTA (Sigma), 5mM NaF (Sigma), 1mM Na2VO3 (Sigma), 1mM PMSF (Sigma), 1x protease inhibitor cocktail complete mini (Roche) ) to a final volume of 250 μL and antibodies coupled to Dynabeads used for immunoprecipitation: goat polyclonal anti UPF1 (Bethyl), rabbit polyclonal anti MVH (Abcam), rabbit polyclonal anti TDRD6 (Antibody Verify), rabbit polyclonal anti UPF2 [35], goat IgG (Invitrogen) and rabbit IgG (Invitrogen). The beads were then washed 4 times with IP buffer. The retained proteins were resolved by SDS-PAGE and immunoblotted with with the aforementioned antibodies and mouse monoclonal anti Vinculin (Sigma). 3 μg of total RNA per sample were used for library preparation. Ribosomal RNA was depleted by using the GeneRead rRNA Depletion Kit (Qiagen). RNA fragmentation, cDNA synthesis and further RNA-Seq library preparation was done with the NEBNext Ultra Directional RNA Library Prep Kit (New England Biolabs). After enrichment and XP bead (Agencourt AMPure Kit; Beckman Coulter, Inc.) purification, quality control was done using Fragment Analyzer (Advanced Analytical). The bar-coded libraries were equimolarly pooled and subjected to 75 bp single-end sequencing on Illumina HiSeq 2000, resulting in an average of 33 million reads per sample. Sequencing raw data were deposited in GEO database under the GSE63948 accession number. The “Tuxedo Suite” of Bowtie, TopHat, Cufflinks and Cuffdiff [39,40,58,59] was used for the alignment and expression analysis. We aligned the samples separately to the mm9 genome using the splice junction mapper Tophat (version 2. 0. 9), which used Bowtie 2 (version 2. 1. 0) for mapping. The Ensembl version 67 [41] was used as a support for the annotation during the alignment. Total RNA was extracted using the TRIZOL Reagent (Invitrogen), according to the manufacturer’s instructions. The concentration and purity of the RNA samples were determined using spectrophotometer scan in the ultraviolet (UV) region. Total RNA (1 μg) was reverse transcribed (RT) with SuperScript II Reverse Transcriptase (Invitrogen) using random primer mix (NEB) according to manufacturer’s instruction. RT PCR amplification was carried out as follows with specific primers (S4 Table): 30” at 95°C, 20” at 60°C, and 30” at 72°C, for 30 cycles using DreamTaq Green DNA Polymerase (Fermentas). RT PCR products were visualized on 1% agarose gels by ethidium bromide staining. RT-qPCR amplification was carried out as follows with specific primers (S4 Table): 5” at 95°C and 30” at 60°C for 40 cycles using GoTaq qPCR Master Mix (Promega). Data analyses was performed with the ddCT method and the unpaired, one tail t-test was implemented. We performed anti UPF1 RNA immunoprecipitation according to [30] with modifications. Briefly, testicular cell suspension was prepared in 20 ml ice-cold PBS and subjected three times to 150 mJ/cm2 UV-C light (Stratagene Stratalinker 1800). After irradiation, hCD4 positive cells were selected as described above and lysed in 0. 75 ml RIPA buffer supplemented with 100U RNAse inhibitor (Invitrogen) for 20 min on ice. The cell lysate was centrifuged at 13,000g for 10 min (4°C). The supernatant was split in 3 samples: 0. 25 ml for input, 0. 25 ml for anti UPF1 RIP and 0. 25 ml for control IgG RIP. RIP samples were diluted with IP buffer to a final volume of 2 ml and pre-cleared with 15 μl Dynalbeads Protein G (Life Technologies). Then, 5 μL of anti UPF1 antibody (Bethyl) or normal goat IgG (Santa Cruz) were added and rotated at 4°C for 4 h. Afterwards, 15 μl Dynalbeads Protein G (Life Technologies) were added and incubated at 4°C for 1 h. After IP, the beads were washed four times with IP buffer and incubated with 1 mg/ml Proteinase K (Roth). Then, RNA extraction, RT and RT-qPCR were performed as described above. RIP RT-qPCR data analysis was performed with fold enrichment method. Briefly, each RIP RNA fractions’ CT value was normalized to the Input RNA fraction Ct value for the same RT-qPCR assay (ΔCt) to account for RNA sample specific differences as ΔCt [normalized RIP] = (Ct [RIP]— (Ct [Input]—Log2 (Input Dilution Factor) ) ). Then, the normalized [RIP] fraction Ct value was adjusted to the normalized background [control Ab RIP] fraction Ct value (ΔΔCt) as ΔΔCt[RIP/control RIP] = ΔCt [normalized RIP]—ΔCt [normalized control RIP]. Fold enrichment above the sample specific control was calculated as linear conversion of ΔΔCt: Fold enrichment = 2 (-ΔΔCt[RIP/control RIP]). RIP assays were conducted in 3 biological replicates and unpaired one-tailed t-test was implemented. For UPF2 RIP, Tdrd6+/- and Tdrd6-/- MACS enriched round spermatid cell suspension fixed in 1% PFA solution for 10 min. After fixation, the cells were lysed in 0. 45 ml RIPA buffer supplemented with 100U RNAse inhibitor (Invitrogen) for 20 min on ice, followed by 2x15 s sonication. Cell lysate was centrifuged at 1,000g for 10 min (4°C). The supernatant was split in 3 samples: 0. 15 ml for input, 0. 15 ml for anti UPF2 RIP and 0. 15 ml for control IgG RIP. RIP samples were diluted with IP buffer to a final volume of 0. 5 ml and 15 μl of serum containing rabbit polyclonal anti UPF2 or normal rabbit IgG (Santa Cruz) were added and rotated at 4°C for 16 h. Afterwards 30 μl Dynalbeads Protein G (Life Technologies) were added and incubated at 4°C for 2 h. After IP, the beads were washed six times with IP buffer and incubated with 1 mg/ml Proteinase K (Roth). RNA extraction, RT, RT qPCR and analysis were performed as described for UPF1 RIP. Testicular extracts from Tdrd6+/- and Tdrd6-/- mice (P26) were subjected to sucrose gradient fractionation as described previously [60]. Briefly, testicular lysates (100 mM NaCl, 10 mM MgCl2,20 mM HEPES, pH 7. 6,0. 5% Triton X-100,200U RNAseOUT) were centrifuged at 13,000 × g at 4°C for 2 min, and the supernatant was applied to the top of a 15–40% linear sucrose gradient. The gradient was centrifuged at 115,000 × g for 200 min (Beckman Coulter). Absorbance tracing at A254 was obtained with 759A Absorbance Detector (Applied Biosystems) and twelve fractions (1 mL) were collected manually. RNAs were extracted from 0. 5 ml of each fraction using the TRIZOL Reagent (Invitrogen). Reverse transcription and RT PCR reactions performed as described above. Proteins were separated by SDS/PAGE, and Western blots were probed with rabbit polyclonal anti RPS6 (Antibody Verify) mouse monoclonal anti GAPDH (Santa Cruz), goat polyclonal anti UPF1 (Bethyl) and rabbit polyclonal anti MVH (Abcam). Cell preparations from total testis or MACS purified hCD4+ were stained with FITC-anti-Human CD4 for 20 min at 4°C and subsequently with 1 μg/ml Hoechst 33342 for 30 min at 32°C. Cells were washed with PBS and resuspended in FACS buffer (PBS, 1% BSA and 1mM EDTA). Before the analysis, 1 μg/ml PI was added to exclude dead cells. Stained cells were analyzed on a BD LSRII (BD Biosciences) using FACSDiva software (BD Biosciences). Data were analyzed using FlowJo software (TreeStar).

Tudor-domain containing protein 6 (TDRD6) is a central component of the chromatoid body (CB) in male germ cells. Chromatoid bodies, which are present in spermatids, contain RNA and protein, are not enclosed by membranes, and typically reside close to the nucleus. Without TDRD6, a much distorted CB structure is observed, and this work asked for the functional contribution of TDRD6 to spermatids. We found that TDRD6 is required for localization of an RNA degradation machinery to the CB. This so-called nonsense mediated decay (NMD) machinery, known from somatic cells, destroys mRNAs that feature premature stop codons. Absence of TDRD6 significantly impairs one specific mechanism of NMD, which depends on long 3’ untranslated regions of the transcripts. Thus, the CB component TDRD6 acts in the assembly of the NMD machinery in the CB.

Abstract Introduction Results Discussion Materials and Methods

rna-binding proteins 3' utr messenger rna rna extraction germ cells untranslated regions extraction techniques sperm research and analysis methods spermatids animal cells proteins gene expression biochemistry rna cell biology nucleic acids protein translation genetics protein domains biology and life sciences cellular types

2016

Chromatoid Body Protein TDRD6 Supports Long 3’ UTR Triggered Nonsense Mediated mRNA Decay

Cell membranes are complex multicomponent systems, which are highly heterogeneous in the lipid distribution and composition. To date, most molecular simulations have focussed on relatively simple lipid compositions, helping to inform our understanding of in vitro experimental studies. Here we describe on simulations of complex asymmetric plasma membrane model, which contains seven different lipids species including the glycolipid GM3 in the outer leaflet and the anionic lipid, phosphatidylinositol 4,5-bisphophate (PIP2), in the inner leaflet. Plasma membrane models consisting of 1500 lipids and resembling the in vivo composition were constructed and simulations were run for 5 µs. In these simulations the most striking feature was the formation of nano-clusters of GM3 within the outer leaflet. In simulations of protein interactions within a plasma membrane model, GM3, PIP2, and cholesterol all formed favorable interactions with the model α-helical protein. A larger scale simulation of a model plasma membrane containing 6000 lipid molecules revealed correlations between curvature of the bilayer surface and clustering of lipid molecules. In particular, the concave (when viewed from the extracellular side) regions of the bilayer surface were locally enriched in GM3. In summary, these simulations explore the nanoscale dynamics of model bilayers which mimic the in vivo lipid composition of mammalian plasma membranes, revealing emergent nanoscale membrane organization which may be coupled both to fluctuations in local membrane geometry and to interactions with proteins. Our growing knowledge of lipid-lipid interactions [1], [2] and of lipid involvement in the control of membrane protein function [3]–[8] highlights the importance of the complexity of composition, structure and dynamics of cell membranes. The large number of different lipid species in vivo has led to an understanding of the cell membrane as a multicomponent system, which is highly heterogeneous in the lipid distribution and composition [9]–[12]. Studies of lipids influencing protein function have revealed that the lipid components of cell membranes play key functional roles in cells [3]–[8]. The lipid composition of membranes differs between species, and also between the plasma membrane of mammalian cells and intracellular membranes such as those of the endoplasmic reticulum, nucleus, and mitochondria [9], [10]. This spatial dependency of membrane lipid composition further highlights its complexity and potential important role in cell function. It is also related to the observed spatial complexities of distribution of proteins within living cell membranes [13], [14]. The lipid compositions of the extracellular and the intracellular leaflets of plasma membranes are highly asymmetric [9]–[16]. Mammalian plasma membranes are composed of approximately 65% glycerolipids, 10% sphingolipids and 25% sterols such as cholesterol (Chol) [10]. The extracellular leaflet is enriched in phosphatidylcholine (PC) such as 1-palmitoyl-2-oleoyl-sn-glycero-3-phosphocholine (POPC) and 1,2-dioleoyl-sn-glycero-3-phosphocholine (DOPC), and in sphingolipids such as sphingomyelin (Sph) and glycosphingolipids. In contrast, the intracellular leaflet is enriched in phosphatidylethanolamine (PE) such as 1-palmitoyl-2-oleoyl-sn-glycero-3-phosphoethanolamine (POPE) and 1,2-dioleoyl-sn-glycero-3-phosphoethanolamine (DOPE), in phospatidylserine (PS) such as 1-palmitoyl-2-oleoyl-sn-glycero-3-phospho-L-serine (POPS) and 1,2-dioleoyl-sn-glycero-3-phospho-L-serine (DOPS), and in phosphatidylinositol (PI) including the di-phosphorylated lipid phosphatidylinositol-4,5-bisphosphate (PIP2). One consequence of this composition is that the inner leaflet of the plasma membrane is anionic in nature [7], [9], [10]. This compositional complexity is likely to result in a corresponding spatial and dynamic complexity, based on e. g. in vitro biophysical studies of lipid vesicles containing three or more lipid types [17], [18]. Molecular simulations provide a ‘computational microscope’ whereby the nanoscale details of the dynamic spatial distributions of lipids may be examined [19]. To date such simulations have largely focussed on relatively simple lipid compositions, thus informing our understanding of in vitro experimental studies [20]–[24]. In contrast, relatively few simulation studies have addressed the lipid asymmetry present in vivo, and these have generally focused on membranes containing only a few different lipid types [25]–[28]. Here we exploit a novel approach for modeling compositionally complex lipid bilayer membranes, which is fast, automated and allows for full control over lipid composition within both the outer and inner leaflets. This has enabled us to construct physiologically relevant membrane models which form the starting point for microsecond duration coarse grained (CG) molecular dynamics simulations [29]. In particular we focus on a complex asymmetric plasma membrane model, which contains the glycolipid GM3 (monosialodihexosylganglioside), within its outer leaflet. This plasma membrane model was simulated both alone and together with model proteins, revealing localised nano-domain formations of the GM3 within the outer leaflet and also of the key anionic lipid, phosphatidylinositol 4,5-bisphophate (PIP2), within the inner leaflet. In order to explore the behavior of mixed lipid bilayers with a composition mimicking that of mammalian plasma membranes, a number of CG bilayer models containing 1500 lipid molecules were generated (see Supporting Information Table S1). These were derived from an initial model bilayer containing 1500 POPC molecules and with dimensions of ca. 20×20 nm which was generated via self-assembly simulations [30]. POPC molecules within either the upper or the lower leaflet were then randomly exchanged for other lipid species (see Methods for details). This yielded an asymmetric plasma membrane model composed of the lipid types abundant within the mammalian plasma membrane in vivo. Thus the overall lipid composition of the model plasma membrane (PM; Supporting Information Table S1) was POPC (25%), POPE (25%), POPS (7. 5%), GM3 (5%), Sph (7. 5%), Chol (25%) and PIP2 (5%). We also explored the effect of increasing the degree of lipid tail unsaturation by including DOPC, DOPE and DOPS lipids (PMUnsat; Supporting Information Table S1). The behavior of the asymmetric PM model has also been compared to that of symmetric lipid bilayer with compositions equivalent to either the upper (i. e. extracellular) or lower (intracellular) leaflets of PM in the PMUpper and PMLower simulations respectively. Plasma membrane models consisting of 1500 lipids and resembling the in vivo composition were constructed and the CG simulations were run for 5 µs. (PM; Fig. 1A). The plasma membrane composition was POPC: POPE: Sph: GM3: Chol (40∶10∶15∶10∶25) within the outer leaflet and POPC: POPE: POPS: PIP2: Chol (10∶40∶15∶10∶25) within the inner leaflet. This membrane composition and distribution mimics a human plasma membrane (9). Two symmetric membrane models were additionally constructed and simulated as a reference for the asymmetric simulation; one system consisting of a symmetric bilayer with the same composition as the outer leaflet of the plasma membrane (PMUpper; Supporting Information Table S1) within both leaflets and one system with the same composition as the inner leaflet of the plasma membrane (PMLower; Supporting Information Table S1) within both leaflets. Cholesterol is able to flip-flop between the leaflets during the simulations. The cholesterol composition of both leaflets was the same in the initial setup. Even though many flip-flops occur during the simulation the overall percentage composition of the upper and lower leaflets remains almost the same throughout the simulations, with ∼49% in each leaflets and the remaining cholesterol located within the membrane core (see Supporting Information Table S2). The flip-flop rate for e. g. the 1500 lipid simulations is between 0. 11–0. 17 flip-flops/ns. The flip-flop rate in the PMProtein system is slightly slower than that observed for the protein-free membranes. As will be discussed later this is largely a result of favorable interactions between the protein and cholesterol within the bilayer. The initially equal distribution of cholesterol between leaflets remains the same during simulations of both the asymmetric and the symmetric lipid bilayers. The change in the membrane organization within the asymmetric plasma membrane during the 5 µs simulation is shown in Fig. 1B–C. Most striking is the apparent formation of nano-domains of GM3 within the outer leaflet (Fig 1B). Large GM3 clusters up to 60 nm have also experimentally been proposed to occur in living cell membranes [31] and on the basis of X-ray scattering studies of GM3 bilayers this has been suggested to be due to strong and cooperative head group interactions [32]. In contrast, the inner leaflet lipids seem to retain more random distribution of the lipids. Similar patterns of behavior were observed within the symmetric membranes resembling the outer and inner leaflets (PMUpper & PMLower). This local clustering of GM3 is further supported by analysis of fractional interactions between different lipid types (Fig. 2A). Thus, approximately 45% of the lipid-lipid interactions of GM3 are with another GM3 molecule. All other lipid types within the outer leaflet are nearly randomly distributed, as indicated by approximately 25% fractional interaction with all other lipids. We did not observe any preferential Sph-Sph interactions within the simulations, even though this has previously been suggested [2], [33], [34]. This co-clustering of Sph is thought to be driven by the ability to form a hydrogen-bonding network through the hydroxyl group of the tails [34]. Our failure to observe this may therefore reflect a limitation of the current CG model for Sph. As mentioned above, cholesterol flip-flops during the simulation and was therefore not assigned to a unique leaflet and not included within the fractional interaction analysis. The simulations of the PMUpper symmetric bilayer illustrate the same behavior with significant GM3 interactions while the other lipids were almost randomly distributed (Supporting Information Fig S1A, B). Similar analysis of the inner leaflet interactions revealed that the anionic PIP2 molecules form interaction networks with each other. However they do not form as large clusters as observed for GM3 Thus the fractional interaction between two PIP2 lipids is approximately 30% (Figure 2B). Interestingly, a similar degree of preferential interaction is observed for the POPC-POPC interactions within the inner leaflet in the asymmetric PM and in the symmetric PMLower membrane simulations (Fig. 2 and SI Fig. S1C, D) but is absent from the outer leaflet of PM and from PMUpper despite the higher content of POPC in the upper leaflet. The distribution of the GM3 cluster sizes was assessed for the PM simulations. The clustering of GM3 was calculated over time utilizing a density-based clustering approach with a cutoff radius of 15 Å and a density requirement of 3 lipids (see Methods for details). The distribution of GM3 clusters over time is shown in the Supporting Information Figure S2A. There are a total of 75 GM3 lipids within the outer leaflet of the PM system. Based on the cluster distribution we see that the number of ‘free’ lipids (defined as a cluster size 1–3) equilibrates to a value of approximately 20% of the total lipids after 1 µs. This indicates that the remaining 80% of GM3 co-clusters into domains of 4–20 lipids, 21–40 lipids or bigger. This analysis also demonstrates that we observe convergence of the GM3 clustering i. e. the GM3 molecules are not evolving towards a single large cluster. As observed from the fractional interaction analysis, PIP2 molecules also cluster but into smaller domains, such that more than ∼50% of PIP2 molecules are distributed into clusters of between 4 and 20 lipids, which remain the same throughout the entire simulation (Supporting Information Fig S3A). To evaluate the robustness of the clustering of GM3, the PM simulations were repeated with the terminal N-acetyl-neuraminic acid group of the headgroup in either a protonated (i. e. neutral) or a deprotonated (anionic) state (and with a minor difference in headgroup restraint parameters). Similar to what others have observed for simulations of cardiolipin-containing bilayers [35], we did not observe any significant differences in the behavior of the lipids within the membrane systems dependent on the presence vs. absence of the negative charge on the GM3 headgroup. When the GM3 headgroup is charged the clustering is accompanied by Na+ counter ions present in the system. Similar behavior is seen for PIP2. This reflects charge neutralization rather than specific lipid-ion interactions, as might be expected given the inherent approximations in the CG model of ions. This suggests that sodium ions and water may facilitate the stabilization of GM3 nano-domains. Indeed sodium ions and water are observed to form stabilizing interactions with the anionic head groups of GM3. We also explored possible effects of lipid tail length and saturation by introducing lipids with di-unsaturated tails instead of just mono-unsaturated lipids. This yielded the PMUnsat model (see SI Table S1) with a composition of POPC: DOPC: POPE: DOPE: Sph: GM3: Chol (20∶20∶5∶5∶15∶10∶25) within the upper leaflet, and of POPC: DOPC: POPE: DOPE: POPS: DOPS: PIP2: Chol (5∶5∶20∶20∶8∶7∶10∶25) within the lower leaflet. The fractional interactions between the lipids in this simulation illustrate a similar behavior to that seen for the simpler PM simulation containing just mono-unsaturated lipids (Fig. 2 and SI Fig. S1E, F). Again, significant inter-GM3 interactions were observed within the outer leaflet alongside some degree of inter-PIP2 interactions within the inner leaflet. This suggests that the head group functionality is most likely the most dominating factor for the localized nano-domain interactions. No separation between the mono- and di-unsaturated lipids was observed in these simulations. This is not surprising since previous studies of symmetric lipid bilayer and monolayers have revealed that modification of lipid tail particle types is a necessity for phase separation of the lipids into raft-like domains [22], [24], [36], [37], indicating that the current CG model is not able to fully capture all such differences in lipid saturation. To explore the possible influence of a simple transmembrane protein [38] on the properties of the model plasma membrane, a system containing sixteen single α-helical transmembrane domains (TMDs) was studied. The TMD chose was from a signaling protein, the cytokine receptor gp130. This was selected on the basis of extending earlier studies of anionic lipid clustering [38], and more importantly because signaling receptors (including e. g. receptor tyrosine kinases such as the EGFR), are thought to interact with a number of lipid types [6] and to co-localize with lipid nano-domains [39]. The TMDs were initially placed on a regular grid, with 60 Å between each protein to ensure no bias was introduced into the initial protein-protein interactions. This membrane system after 5 µs of simulations clearly illustrates that the proteins do not prevent GM3 cluster formations. Indeed, the proteins were observed to co-cluster together with GM3 (Fig 3A). The proteins associate into mainly dimers, and a single stable trimer throughout most of the simulation. The largest protein cluster observed being a trimer of dimers. The distribution and stability of protein cluster sizes can be seen in Supporting Information Fig. S4. The protein clusters were observed to be surrounded by GM3 molecules (Fig. 3A). Similarly to what we previously observed in a simpler system setup [38], the basic C-terminal of the gp130 TMD is able to attract anionic lipids, which can be illustrated by the interaction networks observed between protein and PIP2 within the inner leaflet of the plasma membrane (Fig. 3B). The PIP2 clusters do however not increase significantly in size as a result of the TMD proteins (Supporting Information Fig. S1J) and both the GM3 and PIP2 clusters converge to a similar size as seen in the PM system (Supporting Information Fig. S2B and S3B). GM3 and PIP2 have previously been described to be important in regulation of protein function [5], [6]. In addition to PIP2 and GM3, cholesterol is also observed to form favorable interactions with the protein as judged from radial distribution functions of the different lipid types around the protein (Fig. 3C). The interaction between the lipids and proteins were further assessed by analyzing the average number of interactions between all proteins and the head group of the different lipid species. The average number of interactions of each lipid species has been mapped onto the amino acid sequence of the protein (Fig 4A). Similar to what was seen from the radial distribution function, it was observed that in particular cholesterol and PIP2 interacted strongly with the protein. As mentioned above, we noted that cholesterol is able to flip between the leaflets during the simulation [40] and consequently forms interactions along the TMD sequence within the membrane center. PIP2 form favorable interactions with the basic amino acid enriched C-terminus of the protein similar to previous observations for POPS in a simpler membrane composition [38]. Furthermore, POPC formed many interactions with the protein within the extracellular leaflet. This is nevertheless not surprising as POPC accounts for 40% of the lipids within the outer leaflet. From the radial distribution function analysis, GM3 appeared to interact relatively favorably with the protein. This is not mirrored in the interactions of the anionic head group of GM3 with the protein TMD. This suggests that GM3 interactions with the TMDs are mainly mediated via the lipid tails and/or at the interface between the tails and the head groups. The interactions between cholesterol, PIP2 and the protein are illustrated in greater detail in Figure 4B–D. Favorable interactions between the PIP2 and protein molecules form quite quickly and remain stable during most of the simulations (Fig 4B). A similar pattern is observed from the cholesterol interactions, and as also seen in Fig. 4A, we observe some interactions of the head group of cholesterol within the membrane embedded parts of the protein as a result of the ability of cholesterol to flip into the membrane core. The average number of interactions has been mapped onto the protein structures in Fig. 4D, highlighting attraction of anionic lipids by the basic C-terminus of gp130 and strong interactions both within the N- and C-termini of the proteins with cholesterol. It is evident from the results mentioned above that GM3, and to a lesser extent PIP2, form favorable interaction networks independently of membrane asymmetry, lipid tail saturation, bilayer size, and the presence or absence of proteins. It would be expected that the diffusion of lipid within nano-domains would be slower than freely moving lipids. To assess this, the mean square displacement (MSD) of the different lipids components was calculated during the 5 µs of simulations in the six different plasma membrane simulations (Supporting Information table S3). Within the plasma membrane the diffusion of GM3 (D = 1. 5×10−7 cm2/s) is reduced by ∼40% compared to most other lipid types (D = 2. 6×10−7 cm2/s). Also, PIP2 is observed to have a slightly slower diffusion (D = 2. 0×10−7 cm2/s). The diffusion of the lipids we observe in the plasma membrane models is slower than that previously described in CG simulations of membranes. In simpler membrane models the diffusion of the lipids in pure membranes or with very few proteins is around 9−10×10−7 cm2/s [38] while it is decreased down to 4×10−7 cm2/s in highly protein crowded membranes [41]. The slower diffusion in complex membrane models, and especially for GM3, is most likely caused by the formation of lipid nano-domains and its larger head group able to interact with the surrounding water. The same approximately 40% reduced diffusion of GM3 relative to the other lipids is observed in all the simulations containing this glycolipid. One of the slowest diffusion constant of the anionic PIP2 lipid is observed in the simulations containing the membrane-spanning region of sixteen gp130 receptors (D = 1. 5×10−7 cm2/s). Interactions between the basic C-terminal and anionic lipids were previously described in details [6], and this type of interaction is most likely the reason for the observed decrease in diffusion of PIP2. The single molecule diffusion constant of the glycolipid GM1 within live plasma membrane cells has been determined to 5×10−9 cm2/s [42]. The difference observed in our simulations and the experiments is likely to be due to both crowding effects of proteins and interactions with cytoskeletal proteins in vivo. The evolution of the lipid diffusion was explored by calculating the diffusion coefficients for consecutive time intervals of 1 µs (Supporting Information Fig. S5). The diffusion of GM3 and PIP2 are slower than for the other lipids throughout all of the simulations. This suggests that both the more complex head groups (and hence more extensive interactions), the clustering of GM3 and PIP2 and in the case of PIP2 protein interactions lower their diffusion rates. To better understand the effects of the size of the simulated bilayer patch on membrane behavior and lipid clustering, we performed a simulation of a substantially larger membrane patch consisting of 6000 lipids. This larger membrane system contained more than 150,000 particles with an area dimension of 39 nm ×39 nm and was run for 5 µs. The same overall behavior (including diffusion coefficients) was observed for this larger system compared to the 1500 lipid plasma membrane (Fig. 5). Inter-GM3 interactions were again shown to account for more than 40% of all interactions revealing similar clustering behavior as seen for the 1500 lipid simulations (Supporting Information Fig. S1G). Again large nano-domains of GM3 were observed. The interactions between the head groups are tightly mediated by water and sodium ions (Fig. 5B). As seen for the PM system, the number of non-clustered GM3 lipids converges to approximately 20% with the rest of the GM3 lipids participating in clusters (Supporting Information Fig. S2C). This simulation allows us to investigate larger-scale emergent properties of the PM model. In our simulations of the PM6000 system we observe curvature of the membrane bilayer, as has been observed in other simulations of large (but simpler) lipid bilayers [43]–[45]. Local curvature of the membranes occurs within the first 0. 1 µs of simulation time and continues to fluctuate dynamically over the course of the simulation. This curvature is unlikely to be a tension bias arising from the initial configuration of the complex membrane, as the area per lipid is expected to be the same for both the inner and outer leaflets with the lipid composition employed (see Methods and Table S1). Furthermore, the local curvature observed is irregular, rather than a global uniform deformation in one direction, again suggesting local dynamic fluctuations rather than an overall curvature bias in the system as a whole. Visualization of the simulation (Fig. 6A) suggested a correlation between the curvature in the PM6000 bilayer surface and the clustering of lipid molecules. Viewed from the extracellular (upper) surface, the bottom of the waves (i. e. the concave surfaces) were enriched in GM3 whilst if one views from the intracellular (lower) side the concave surfaces are enriched in PIP2 and cholesterol. We quantified this by calculations of the cross correlation between the local displacement of the membrane lipids from their average position along the bilayer normal (z) and the local composition of the bilayer (see Fig. 6B and Methods for details). This analysis revealed clear correlations between local bilayer geometry and local lipid composition. Thus the concave (downwards) deflections of the bilayer from the extracellular side were locally enriched in GM3 and to a lesser extent PE in the outer leaflet of the bilayer, whilst the concave (upwards) deflections from the intracellular side were enriched in PIP2, cholesterol and PE (Fig. 6C). Thus we would expect the GM3 and PIP2 clusters to be anti-correlated, which can be seen from the correlation matrix between the lipids within each leaflet (Supporting Information Fig. S6). The pattern of local enrichment of PE corresponds to the known preference of this lipid for an inverted hexagonal (HII) phase [46]. Furthermore, GM3 micro domains have been observed in mixed lipid bilayers by AFM imaging [47], and PIP2 has been shown to have direct effects on bilayer properties [48]. Cholesterol is believed to be involved in nano-domain formation [2], [33] and the clear correlation between curvature and composition observed may indicate that local cholesterol enrichment has an impact on the geometry of the membrane. The behavior of GM3 and PIP2 in our simulations is presumed to reflect the local clustering seen in the smaller scale simulations, mediated in part by Na+ and water (Fig 5B). However, note that we observe similar behavior in the simulations where GM3 is modeled as neutral and with slightly different head group parameters (data not shown). The local membrane curvature does not lead to local thinning or thickening of the bilayer. In our simulations of complex plasma membrane-like lipid bilayers we observe asymmetric formation of lipid nano-clusters. The location of these nano-clusters within the bilayer correlates with local bilayer distortion/deflection. It seems that the formation of glycolipid nano-clusters in the outer leaflet is (largely) independent of the less pronounced clustering of PIP2 lipids within the inner leaflet mediated by sodium ions and water. Lipids within one leaflet are able to form local clusters independently of the composition of the opposite leaflet. Our simulations also illustrate this, since GM3 forms a tight network in both symmetric and asymmetric bilayers, and additionally in membranes containing protein. A similar type of glycolipid clustering (of GM1 [49]) was recently observed by others [24]. Similarly, PIP2 has been shown from experiments and simulations to cluster dependent on the presence of interacting membrane proteins [50]. Our results extend these previous studies in that we are for the first time able to study clustering of glycolipids and PIP2 simultaneously in a complex asymmetric membrane model. We are not only able to show local clustering as previously observed, but also the effect of membrane proteins and the anti-correlation between these types of nano-domains and their spatial organization with respect to membrane curvature. Asymmetric membranes have been investigated in a number of previous computational studies [25]–[28], but not in bilayers of comparable complexity to those studied here. Through the simulations described here, we have been able to capture the clustering of GM3 comparable to that observed in cell membranes [31], although the cluster sizes of GM3 in our simulations are of the order of ca. 50–100 Å, compared to ca. 600 Å in the experimental studies. The differences may reflect the greater complexity of the cell membrane and its interactions but also the resolution limit of the experimental method applied. When we introduce model proteins into the PM bilayer model we are able to see the co-clustering of proteins and lipids. We also see that the large number of different lipid types slows down the diffusion of the lipids compared to that in membranes of simpler lipid composition [38], [41], and that the local clustering of GM3 leaves this lipid less dynamic than other lipid types. Residues of the model TMD are both able to favorably cluster with glycolipids on the extracellular side of the membrane mainly through the tails in addition to forming favorable interactions with anionic lipids within the inner leaflet. Cholesterol is observed to form persistent interactions with the proteins on both the extracellular and the intracellular side. This suggests that there may be a cholesterol interaction site on the surface of the TMD of the signaling protein gp130. Of course, the limitations of the CG force field apply to cholesterol and its interactions (see e. g. [51]), so one might more cautiously conclude that the gp130 TMD exhibits a possible sterol interaction site, the specificity of which may be explored in more detail in future atomistic simulations. Thus the effects of local and asymmetric clustering of lipids on protein function may correlate both via specific protein-lipid interactions but maybe also because of the slower diffusion once captured in nano-domains. This is clearly related to on-going discussion of both direct (protein binding) and indirect (bilayer mediated) effects of PIP2 on membrane protein function [48] and organization [50]. The MARTINI force field has previously been shown to too ‘sticky’ for the interactions of globular proteins in aqueous solution [52]. However, simulated protein-lipid interactions [53], [54] and protein-protein interactions within the lipid bilayer [55]–[57] seem to reproduce experimental results quite well. We are therefore reasonably confident that the lipid-lipid clustering we observe within these simulations are not simply artefacts of the CG model. We observe that the clustering of GM3 is independent of lipid head group charge and is mediated by both water particles and sodium ions (Fig 5B). Similarly the clustering of PIP2 is bridged by sodium ions and sensitive to introduction of proteins within the system (Fig 3B). We therefore suggest that the lipid clustering observed in this study a result of favorable interaction between lipids of certain shape and charge, as has also been shown from experiments. The degree of clustering of GM3 that we observe seems to have converged in that that the number of unclustered lipids equilibrates towards 20% during the simulations. However, a limitation of the current CG representation is the difficulty in modelling phase separations when changing the tail saturation of the lipids. For example DOPC and DPPC phase separate in vitro but not in MARTINI. Additionally we are not able to capture the experimentally observed sphingomyelin co-clustering within our models. Both of these limitations indicate that important but subtle difference in lipid tails may not be sufficiently captured in CG simulations using the MARTINI and related force fields and hence the electrostatics, size and charge of the lipid head groups are main driving force in our observations of lipid clustering. A further possible limitation is that of convergence of these complex bilayer simulations. A number of recent studies have discussed some of the difficulties of assessing convergence in complex membrane simulations (e. g. [58]): to some extent it is a question of assessing “unknown unknowns” [59]. However, as discussed above, the number of free (i. e. unclustered) lipid molecules plateaus with respect to time for both GM3 and PIP2. This suggests that the clustering we do observe is unlikely to be bias from the initial configuration of the simulation, but rather is a genuine local property of a complex bilayer, especially as our simulation systems do not seem to simply drift towards a single large cluster. In summary, simulations now allow us to explore the nanoscale dynamics of model bilayers, which mimic the in vivo lipid composition of cell membranes. In these simulations we can see indications of emergent larger scale membrane organization which may be coupled both to fluctuations in local membrane geometry and to interactions with proteins. It will be of interest to extend these studies to higher levels of protein crowding [41] to better understand the interplay of compositional complexity and local spatial clustering of both lipids and proteins on aspects of membrane protein function. The complex membranes were generated by starting from a self-assembled POPC bilayer which was then ‘edited’. Thus, for lipids with the same or shorter chain length, the required coordinates were simply transferred from a randomly selected POPC molecule and the lipid was relabeled to e. g. POPE, POPS, PPCS, or POPE and the particle types altered accordingly. For larger lipids (e. g. PIP2, GM3, DOPC, DOPE, DOPS) the new lipid molecule was superimposed on the first two tail beads of the randomly selected POPC molecule and the two lipid molecules (new and POPC) were exchanged. Cholesterol was superimposed on the head group and 3 of the hydrophobic tail beads of a randomly selected POPC. In the asymmetric bilayers the overall percentage of lipids (other than cholesterol) in each leaflet of the bilayer is the same to avoid potential issues with differences in area per lipids. Table S1 shows that the area per lipid after equilibration of the systems. The standard deviation is less than 1% of the average area per lipid for the symmetric systems with similar compositions (PM, PMUpper, PMLower, PM6000). Thus no strain is added to either leaflet of an asymmetric bilayer as a result of non-matching areas per lipid. All simulations were run using the MARTINI CG force field for lipids [60] and proteins [61] and all membranes except one were built by an initial self-assembly of either 1500 or 1759 POPC lipids. For a list and details of simulations see the Supporting Information, Table S1. The 6000 lipid system was created from the 1500 lipid system using the genconf module within gromacs [62] (www. gromacs. org). All systems were solvated by standard MARTINI water beads and neutralized by NaCl to a concentration of 0. 15 M. The 1500 lipid plasma membrane simulations (PM, PMUpper, PMLower) consisted of between 35,000 to 47,000 particles, the 6000 lipid plasma membrane simulation (PM6000) consisted of 151,431 particles and the simulations containing proteins were composed of 60,751 particles All simulations except PM6000 were initiated directly from a minimized system. The PM6000 system needed further equilibration to remove steric issues. This was done utilizing built-in free energy functions within gromacs. Initially GM3 and water were removed from the system. The lipids were then gradually equilibrated using the free energy perturbation method, where the presence of all lipids were increased to full van der Waals over 1000 steps using 1 fs. Afterwards the GM3 lipids were included at the positions determined from the exchange lipid protocol (described above) and they were also gradually included using free energy perturbation over 1000 steps using 1 fs time steps. The system was subsequently solvated, neutralized and NaCl was added to a concentration of 0. 15 M. All simulations were run utilizing gromacs 4. 5. x or 4. 6. x [62] (www. gromacs. org). The CG MD simulations were performed utilizing the MARTINI version 2. 0 lipid CG force field except for PIP2, GM3, and with PIP2 described in [63]. Parameters for Sph were from the MARTINI force field itp file (http: //md. chem. rug. nl/cgmartini/images/parameters/ITP/martini_v2. 0_lipids. itp). The ceramide tail from Sph was utilized as the tail for GM3. The head group of GM3 was newly parameterized (see Supporting Information and Fig S7 and S8 for details). Sph and GM3 contains two 4 bead tails with the first bead of the sphingosine tail being unsaturated, while PIP2 contains a 4 and a 5 bead unsaturated tail. For an overview of the CG models of all lipids see Fig S9. In all simulations the pressure was maintained at 1 bar using the Berendsen barostat with a 1 ps coupling constant. In all simulations the temperature was maintained at 310 K and the temperature was controlled by a Berendsen thermostat [64] using a coupling constant of 1 ps. Semi-isotropic pressure coupling was used in all simulations with compressibility of 3×10−4 bar−1. A time step of 20 fs was used in all the simulation and the van der Waals and coulomb interactions were shifted to zero between 9 Å and 12 Å and 0 and 12 Å respectively. All the simulations were run for 5 µs. The built-in g_msd function of gromacs was used to calculate the lipid diffusion. The normalized fractional interactions were calculated as the relative number of contacts of a lipid species with each of the other lipid types with a correction for the total number of lipids in the system. Others have previously used this type of calculations to characterize the degree of phase formation in CG simulations [24]. For a two and four component system a fraction of 0. 5 and 0. 25 respectively correlate to a randomly mixed bilayer. Even though more than one bead was located with the cutoff distance of 11 Å only one contact was registered. To allow for sufficient equilibration only the last 1 µs of the simulation was used from the calculation (4–5 µs) with a 1 ns interval. The nature of such interactions is not necessary symmetric since the density and clustering of specific lipid species will make contacts between lipid A and lipid B different from lipid B interactions with lipid A as a result of number of nearest neighbors and the lipid sizes. Lipid cluster sizes of GM3 and PIP2 within the PM, PMProtein and PM6000 systems during the 5 µs of simulation were calculated using a python implementation of the density based algorithm DBSCAN [65]. Clusters were identified according to the following parameters: the cutoff distance between neighbors was set to 15 Å and the minimum numbers of elements was set to 3. The results was plotted as a function of 4 groups, with 1–3 lipids classified as non-clustered, 4–20 lipids as small clusters, 21–40 as medium clusters and>40 as large clusters (Fig S2 and S3). The protein cluster sizes were determined during 5 µs using the connectivity networkx python module [66] whereby a point is considered connected to another group of points if within the cutoff distance from at least one the points within the group. The cutoff distance was set to 8 Å for minimum distance between proteins. Visualization of the PM6000 simulation suggested a correlation between the curvature in the PM6000 bilayer surface and the clustering of lipid molecules. We quantified this by calculating the cross correlation between the local displacement of the membrane lipids from their average position along the bilayer normal (z) and the local composition of the bilayer within grid boxes evenly distributed across the membrane. The PM6000 lipid was split into 8×8 grid boxes yielding grid boxes of approximately 50 Å2. The normalized cross correlation (RL, z) at a given snapshot was calculated as: where Ln is the number of lipids of a given species in a grid box, zn is the z coordinate of the interface between the head groups of the lipids (excluding the current species being calculated) and tails in that box, and the averaging is across all grid boxes. The average RL, z and the standard deviation over the 5 µs simulation were displayed (Fig. 6) for all lipid species. Figures were generated in VMD [67].

Cell membranes play important roles in vivo both in shielding the cell interior from the surrounding environment and in cell function through lipid components of the membrane having roles in controlling protein function, cell signaling etc. We employ molecular dynamics simulations to explore the behavior of biologically realistic membrane models. Our simulations reveal nano-domain clustering of the glycolipid GM3 and to a lesser extent of the anionic lipid phosphatidylinositol 4,5-bisphophate (PIP2). When including transmembrane proteins we are able to observe preferential interactions of known regulatory lipids (e. g. GM3, PIP2 and cholesterol) with the proteins. Membrane curvature is shown to be coupled to the local lipid composition, suggestive of a link between lipid nano-domains and membrane geometry.

Abstract Introduction Results Discussion Methods Cluster distribution Correlation between bilayer surface curvature and the clustering of lipid molecules

cell physiology biochemistry lipids cell biology biology and life sciences computational biology molecular cell biology biophysics membrane trafficking biophysical simulations

2014

Lipid Clustering Correlates with Membrane Curvature as Revealed by Molecular Simulations of Complex Lipid Bilayers

Drosophila melanogaster, like other invertebrates, relies solely on its innate immune response to fight invading microbes; by definition, innate immunity lacks adaptive characteristics. However, we show here that priming Drosophila with a sublethal dose of Streptococcus pneumoniae protects against an otherwise-lethal second challenge of S. pneumoniae. This protective effect exhibits coarse specificity for S. pneumoniae and persists for the life of the fly. Although not all microbial challenges induced this specific primed response, we find that a similar specific protection can be elicited by Beauveria bassiana, a natural fly pathogen. To characterize this primed response, we focused on S. pneumoniae–induced protection. The mechanism underlying this protective effect requires phagocytes and the Toll pathway. However, activation of the Toll pathway is not sufficient for priming-induced protection. This work contradicts the paradigm that insect immune responses cannot adapt and will promote the search for similar responses overlooked in organisms with an adaptive immune response. Immune responses are typically characterized as being either adaptive or innate. Adaptive immunity, which requires T and B cells, is specific, has memory, and is generally considered to be restricted to vertebrates. In contrast, the innate immune response is thought to act naïvely to each encounter with a pathogen [1,2]. Innate immunity depends on the recognition of broadly conserved molecular moieties and exhibits only weak specificity, such as the ability to distinguish between different structural classes of peptidoglycan [3]. However, recent work suggests that the invertebrate innate immune response may exhibit adaptive characteristics (reviewed in [4] and [5]). Functional immune adaptation can be defined most broadly as any case where an immune response differs between a first and second challenge. The simplest form involves the immune system remaining activated after an initial challenge. This sort of response has long been known in invertebrates as shown by Hans Boman and coworkers [6]. They found that antibacterial activity in Drosophila hemolymph persists after bacterial challenge and can provide protection against subsequent challenges. More recently, Moret et al. [7] found a similar persistence of humoral antibacterial activity in mealworms. More complex adaptive phenomena have also been observed in arthropods; for example, flour moths [8] and Daphnia [9] possess strain-specific immunity that is passed from a mother to her offspring. The molecular mechanisms underlying the maternal transfer of strain-specific protection have not been characterized. Specific memory has also been examined in cockroaches [10] and bumblebees [11]. In both insects, the initial immune activation is nonspecific and confers protection against many types of challenges. However, both cockroaches and bees are also able to mount long-term specific protection: a priming dose of a particular species of bacteria only protects against that species (or class of species in the case of bumblebees). From this work, it seems that innate immunity possesses the adaptive characteristics of specificity and memory; unfortunately, the animals used in past studies have not been amenable to deeper analysis. We examined the well-understood Drosophila innate immune response for specificity and memory because this model organism would give us genetic and physiological assays to dissect adaptive aspects of innate immunity. Drosophila has been proved to be a powerful model organism to study innate immunity [1,2]. The Drosophila innate immune response has three effector mechanisms: the humoral response, melanization, and the cellular response [1,2]. The humoral immune response involves the secretion of soluble factors, such as antimicrobial peptides (AMPs), into the hemolymph following immune activation. Melanization is the process whereby melanin is deposited at wound sites and parasite surfaces, resulting in the release of toxic reactive oxygen species. The cellular immune response consists of hemocytes that phagocytose, encapsulate, and kill invading microbes, much like vertebrate macrophages. These mechanisms depend in various ways on pathogen detection via the Toll or imd signaling pathways [1,2, 12,13]. We found that S. pneumoniae–primed flies are protected against a subsequent lethal challenge with S. pneumoniae. This response is specific for S. pneumoniae and persists for the life of the fly. In this paper, we demonstrate that the Toll pathway, but not the imd pathway, is required for this protective effect. Notably, activation of the Toll pathway is not sufficient to elicit a primed response. We show that AMPs are not involved and that phagocytes are the critical effectors of the primed response. Taken together, we demonstrate that the Drosophila primed response is specific and persists for the life of the fly. We identified a signaling pathway required for the process and ascertained which branch of the fly immune response is responsible for the primed response. We found that previous exposure to S. pneumoniae permanently alters the fly' s response to this bacterium. S. pneumoniae is a Gram-positive encapsulated bacterium that is the causative agent of otitis media, pneumonia, and meningitis [14]. The injection of 3,000 colony-forming units (CFU) directly into Drosophila hemolymph is normally lethal, killing the fly within 2 d (Figure 1A), and death is correlated with bacterial proliferation (Figure 1B). However, flies primed with a sublethal dose of bacteria were protected against a lethal challenge of S. pneumoniae administered 1 wk later (Figure 1C). Flies primed with S. pneumoniae and challenged with a lethal dose died at the same rate or slower than wounded controls. A priming dose of dead heat-killed S. pneumoniae was also sufficient to protect flies against a subsequent lethal challenge. Priming flies with S. pneumoniae thus induces long-term changes in the fly immune response, conferring protection against an otherwise-lethal challenge. We considered two scenarios that could explain enhanced survival in S. pneumoniae–primed flies. First, bacterial numbers may not differ between naïve versus primed flies, and primed flies might survive the stress of the infection better than naïve flies. Second, S. pneumoniae–primed flies could kill the bacteria faster, and bacterial clearance would correlate with enhanced survival. To distinguish between these two possibilities, we examined bacterial load in naïve versus primed flies. Flies were injected with a priming dose of either dead S. pneumoniae or phosphate-buffered saline (PBS) 1 wk prior to a challenge of 400 CFU S. pneumoniae. This dose was chosen to emphasize the difference between naïve and primed flies; 400 CFU is the lowest dose that is lethal to naïve flies but not S. pneumoniae–primed flies (unpublished data). Within 1 d, S. pneumoniae–primed flies had killed almost all of the S. pneumoniae, whereas naïve flies still contained bacteria (Figure 1D). This result indicates that the survival difference between naïve and S. pneumoniae–primed flies results from different rates of S. pneumoniae killing. Functionally, immunological memory is characterized by a more effective immune response upon repeat exposure that persists for the life of the animal. Although the best-described model for immune memory involves T and B cells and recombination-derived variation of receptors, we note that the definition of memory is independent of mechanism and immune memory could arise in a variety of ways. We demonstrated that preinoculation with dead S. pneumoniae alters the fly immune response such that it is more effective against subsequent challenges with S. pneumoniae. To determine how long these immune changes persist in the fly, we next varied the length of time between the priming and challenge dose. Flies primed with dead S. pneumoniae on day 0 were challenged between 1 and 14 d later with a lethal dose of S. pneumoniae. S. pneumoniae–primed flies always died significantly more slowly than naïve PBS-injected flies (Figures 1E and S1). An interval of 2 wk between the priming dose and challenge dose was the longest time we could assay; at 3 wk postpriming, the flies are actually 5 wk old and die from the stress of wounding alone (unpublished data). In summary, protection due to a priming dose of S. pneumoniae was detectable within 24 h and persisted for the life of the fly, or as long as we could assay survival differences. To explore the specificity of this immune response, we asked whether other pathogens can induce a protective response against themselves. We chose a broad range of microbes pathogenic to wild-type Drosophila, including a Gram-negative bacterium, a Gram-positive bacterium, a Mycobacterium, and a natural fungal pathogen [15–18]. Priming doses of heat-killed Salmonella typhimurium, Listeria monocytogenes, and Mycobacterium marinum did not elicit a protective effect against a subsequent lethal challenge of the same bacteria used in the priming dose (Figures 2A and S2). Dead bacteria were used as priming doses in these experiments because all of these bacteria have an LD50 of one bacterium. Protection by priming is thus not a general characteristic of all microbial challenges in the fly. However, a priming dose of the natural fungal pathogen, Beauveria bassiana [17], conferred a protective effect against a subsequent lethal challenge (Figures 2A and S2). Perhaps S. pneumoniae is a uniquely powerful immune activator; priming with S. pneumoniae might protect against challenges with other pathogens. To test this hypothesis, we challenged S. pneumoniae–primed flies with lethal doses of the same panel of microbes as above. S. pneumoniae–primed flies were not protected against lethal challenges of other pathogens (Figures 2A and S2). The protective effect of the primed response thus is not due to general activation of the Drosophila immune response. Having seen that the priming-induced protective response was specific in the sense that S. pneumoniae was incapable of protecting against other immune challenges, we next determined whether other immune activators were capable of inducing a protective response against S. pneumoniae. To test this, we primed flies with a mixture of strong immune activators (dead Escherichia coli, dead Micrococcus luteus, and dead Beauveria bassiana) [17,19–21]. Flies injected with this mixture were not protected against a lethal challenge of S. pneumoniae (Figure 2B). Furthermore, injection with this mixture does not interfere with the ability to induce a protective response because the addition of dead S. pneumoniae to the mixture protected the flies from a second lethal challenge (Figure 2B). These results also demonstrate that a priming dose of B. bassiana does not protect against a lethal dose of S. pneumoniae. We conclude that protection conferred by a priming dose of S. pneumoniae specifically protects against lethal doses of S. pneumoniae and persists for the life of the fly. Toll and imd are the best-characterized fly immunity pathways that control the majority of genes found to be induced by bacterial and fungal infections, including AMPs [1,2, 17,19–21]. The mixture of microbes used in Figure 2B was chosen to strongly activate both the Toll and imd pathways [19,20]. Although activation of Toll and imd signaling is insufficient to induce a protective response, it remained possible that the pathways are necessary for protection. We tested loss-of-function mutants in each pathway to determine whether they were necessary for a protective response. Because both Toll and imd pathway mutants are immunocompromised with respect to S. pneumoniae (Figure S3), we reduced the lethal challenge dose of S. pneumoniae (20 CFU for Toll mutants, 100 CFU for imd mutants). These doses are normally sublethal to wild-type flies, but higher doses killed the mutant flies too quickly to detect a difference in survival. Loss-of-function mutants of imd (Figure 3) and dTak1 [22] (unpublished data) were protected by a priming dose of dead S. pneumoniae. Thus, although the imd pathway normally contributes to killing S. pneumoniae, it is not necessary to elicit a protective response. In contrast, flies homozygous for partial loss-of-function mutations that disrupt the Toll pathway, PGRP-SA [23,24] (Figure 3) and Dif [25] (unpublished data), were not protected by a priming dose of dead S. pneumoniae. The Toll pathway is therefore necessary for the primed response, but the imd pathway is not required. Next, we wanted to determine the contribution of melanization, the humoral response, and the cellular response to priming (Figure 4A) [1,2]. We do not observe melanization in response to S. pneumoniae infections and therefore do not expect that it plays a strong role in the protective response (unpublished data) [26]. Because the Toll pathway is required for the priming-induced protection, we first examined the contribution of the inducible humoral response. We found four lines of evidence suggesting that AMPs are not responsible for protecting flies against a second lethal challenge of S. pneumoniae. First, S. pneumoniae was not a strong AMP inducer because the peak of AMP transcription in response to S. pneumoniae was not as high as the peak induced by the positive control elicitor E. coli [27]. Flies were injected with a priming dose of dead S. pneumoniae, media (wounding control), or E. coli (positive control), and quantitative real-time reverse transcription–PCR (qRT-PCR) was used to assess transcript levels of three different AMPs: defensin (Figure 4B), attacin (Figure S4A), and diptericin (Figure S4B) [27]. These AMPs were chosen because they are strongly induced by Gram-positive bacterial infections. Second, we found that not only was S. pneumoniae a poor inducer of AMPs but also AMP transcript levels did not remain elevated 1 wk later (Figure 4B). Thus, AMP transcription should be back to the ground state by the time the challenge dose is administered, 1 wk after the priming dose. It has been reported that AMPs can persist in the hemolymph [28]. However, we have shown that simultaneous activation of Toll and imd, and thus AMP induction, is not sufficient to protect the fly (Figure 2B). Finally, we asked if this ground state is “sensitized”—that is, whether AMP induction is enhanced in flies that have been primed with S. pneumoniae compared to media. Using qRT-PCR, we measured defensin (Figure 4C), attacin (Figure S5A), and diptericin (Figure S5B) transcript levels after a second challenge of S. pneumoniae, media (wounding control), or E. coli (positive control). None of the AMPs were differentially induced in S. pneumoniae–primed flies compared to naïve media-injected flies. Thus, we find no evidence to support the involvement of AMP induction in the primed immune response. In light of these data, and the fact that hemocytes are altered in Toll pathway mutants [12,13], we examined whether the cellular immune response is the main effector of the primed response—that is, a priming dose of S. pneumoniae might specifically increase S. pneumoniae clearance by phagocytes upon a second exposure. We first assessed the contribution of phagocytosis to S. pneumoniae killing in naïve flies by injecting flies with polystyrene beads to block phagocytosis prior to infection [29]. Bead-inhibited flies were extremely sensitive to S. pneumoniae; 3,000 bacteria were required to kill a wild-type fly, whereas 20 CFU was sufficient to kill a wild-type fly that lacks phagocytosis (Figure 4D). We demonstrated above that the survival difference between S. pneumoniae–primed flies and naïve flies is linked to enhanced clearance (Figure 1D); here we show that phagocytosis is required to kill S. pneumoniae in naïve flies. We then asked if the enhanced clearance in primed flies is due to increased killing by phagocytes and not a second, mechanistically, different method of killing. To test this, we inhibited phagocytosis in both primed and naïve flies and then challenged with a lethal dose of S. pneumoniae 1 wk later, at which time point phagocytosis remained inhibited [29]. Primed flies died at the same time as naïve flies and were therefore not protected by a priming dose of S. pneumoniae, regardless of whether they were injected with beads before (Figure 4E) or after (unpublished data) the priming dose. Fly phagocytes are therefore an essential effector of the primed response. These data suggested that a priming dose of S. pneumoniae activates fly phagocytes to kill S. pneumoniae more efficiently. Is this enhanced killing specific to S. pneumoniae, or does a priming dose of S. pneumoniae simply cause general phagocyte activation? If phagocytes are generally more activated in S. pneumoniae–primed flies, these flies should be able to clear other bacteria more rapidly. To test this hypothesis, S. pneumoniae–primed flies were tested for their ability to kill E. coli (Figure 4F). There was no difference between naïve PBS-injected and S. pneumoniae–primed flies in their ability to clear E. coli. Combined with the fact that a priming dose of S. pneumoniae does not offer protection against any other lethal challenges of bacteria (Figure 2A) and the fact that the primed response is specific for S. pneumoniae (Figure 2), we conclude that a priming dose of S. pneumoniae alters the fly immune system in a persistent manner that specifically allows phagocytes to recognize and kill S. pneumoniae more efficiently. We have presented evidence that the fly modulates its immune response as a result of multiple challenges: a priming dose of S. pneumoniae is sufficient to protect the fly against a subsequent lethal dose of S. pneumoniae. Using a functional immune assay, we have shown that the fly immune system exhibits the adaptive characteristics of specificity and persistence. The mechanism underlying this protective response requires the Toll pathway, although its contribution is not via activation of AMPs. We have eliminated contributions from the imd pathway and AMPs and identified phagocytes as the critical effectors of the primed response. This system is uniquely positioned to further characterize the molecular basis underlying specific phagocyte activation and other adaptive aspects of innate immunity. Flies were maintained on standard dextrose medium at 25 °C and 65% humidity. All experiments were performed with 5- to 7-d-old male wild-type Oregon R flies. All mutant flies were back-crossed onto the Oregon R background to limit background effects. In particular, white mutant flies are very sensitive to S. pneumoniae and do not elicit a primed response (unpublished data). Mutant lines used in this study include PGRP-SAseml (from P. Ligoxygakis), Dif, imd10191, and Tak12527. Molecular information for imd10191 is included below. The imd10191 line has a 26-nucleotide deletion that frameshifts the protein at amino acid 179, which is the beginning of the death domain. Microbial strains used in this study include S. pneumoniae strain SP1, E. coli DH5α, M. luteus, B. bassiana, L. monocytogenes strain 10403S, S. typhimurium strain SL1344, and M. marinum strain M. S. pneumoniae cultures were grown standing at 37 °C 5% CO2 in brain heart infusion broth (BHI) (BD Bioscience, http: //www. bdbioscience. com) to an OD600 of 0. 15, and aliquots were frozen at −80 °C in 10% glycerol. For infection, an aliquot of S. pneumoniae was thawed, diluted 1: 3 in fresh BHI, and allowed to adjust for 2 h at 37 °C 5% CO2. E. coli, S. typhimurium, and L. monocytogenes cultures were grown standing overnight in BHI at 37 °C. M. luteus was cultured standing at 29 °C in BHI for 1 wk or until a sufficient density was reached. M. marinum was cultured standing at 29 °C in Middlebrook 7H9 broth (BD Bioscience) supplemented with Middlebrook OADC (BD Bioscience) and 0. 2% Tween. B. bassiana spores were grown on malt agar (BD Bioscience) at 29 °C for 2 wk or until a sufficient density was reached. For injection, flies were anesthetized with CO2 and injected with a total volume of 50 nl using individually calibrated pulled glass needles attached to a Picospritzer III injector (Parker Hannifin, http: //www. parker. com). Flies were always injected in the abdomen, close to the junction with the thorax and just ventral to the junction between the ventral and dorsal cuticles. Flies were never anesthetized for longer than 10 min. After each injection, all flies were transferred to a new vial and maintained at 29 °C and 65% humidity. To prepare priming doses of microbes, concentrated cultures were boiled for 30 min, centrifuged 5 min at 2,000g, and washed three times in PBS. Bacterial cultures were diluted in PBS to an OD600 of 0. 1 and stored at −80 °C. Heat-killed B. bassiana spores were counted on a hemocytometer, adjusted to a concentration of 1 × 107/ml in PBS, and stored at −80 °C. Aliquots were plated on the appropriate media to verify that the microbes had been heat-killed. To prime flies, 50 nl of these solutions was injected into the fly. Flies were incubated at 29 °C and 65% humidity until they received their second challenge. Fresh cultures were washed three times in PBS and diluted to the appropriate OD600 in PBS. For the different concentrations of S. pneumoniae, appropriate bacterial load corresponding to different optical densities was experimentally determined. For reference, an OD600 of 0. 1 corresponds to 3,000 CFU. Lethal doses of other bacterial species are as follows: S. typhimurium, OD 0. 1 (10,000 CFU); L. monocytogenes. OD 0. 01 (6,500 CFU); and M. marinum, OD 0. 05 (500 CFU). Bacterial load after injection was verified for all strains except M. marinum by plating on the appropriate media (blood agar for S. pneumoniae, BHI agar for all other strains). For lethal B. bassiana challenges, flies were anesthetized in groups of 20 and shaken on a plate of spores for exactly 30 s. All infections were carried out at 29 °C. Individual flies were homogenized in 100 μl of PBS, diluted serially, and spotted onto appropriate plates. S. pneumoniae were grown on blood agar supplemented with 500 μg/ml streptomycin (Sigma, http: //www. sigmaaldrich. com), 10 μg/ml colistin (Sigma), and 5 μg/ml oxolinic acid (Sigma) to eliminate the growth of bacterial contaminants from the fly. Plates were incubated overnight at 37 °C 5% CO2. E. coli colonies were grown on LB and incubated overnight at 37 °C. Flies were challenged as described above and incubated at 29 °C for the indicated time points. At the given times, triplicates of three flies were anesthetized, placed in 1. 5-ml tubes, and homogenized in 100 μl of Trizol-LS (Invitrogen, http: //www. invitrogen. com). RNA was extracted using the standard Trizol-LS protocol, and remaining genomic DNA was degraded with DNase I treatment. RT-PCR was carried out with a Bio-Rad iCycler (http: //www. bio-rad. com) using TaqMan probes and rTth polymerase (Perkin-Elmer, http: //www. perkinelmer. com) as directed by the manufacturer. The following primers below were used. Relative RNA quantities were determined with respect to Drosophila ribosomal protein 15a, and all levels were normalized with respect to the zero time point for media injection: defensin TTCTCGTGGCTATCGCTTTT (left primer), GGAGAGTAGGTCGCATGTGG (right primer), AGGATCATGTCCTGGTGCATGAGGA (Taqman probe); attacin CAATGGCAGACACAATCTGG (left primer), ATTCCTGGGAAGTTGCTGTG (right primer), AATGGTTTCGAGTTCCAGCGGAATG (Taqman probe); diptericin ACCGCAGTACCCACTCAATC (left primer), CCCAAGTGCTGTCCATATCC (right primer), CAGTCCAGGGTCACCAGAAGGTGTG (Taqman probe); and ribosomal protein 15a TGGACCACGAGGAGGCTAGG (left primer), GTTGGTTGCATGGTCGGTGA (right primer), TGGGAGGCAAAATTCTCGGCTTC (Taqman probe). Carboxylate-modified blue fluorescent 0. 2-μm-diameter polystyrene beads (Molecular Probes, http: //www. invitrogen. com) were injected to block phagocytosis essentially as previously described [29]. Briefly, beads were washed twice in sterile water and resuspended in one fourth of the original volume. Flies were injected with 50 nl of bead solution or water as an injection control. To confirm that phagocytosis was inhibited, the in vivo phagocytosis assay was performed as described previously with FITC-conjugated E. coli or FITC-conjugated Staphylococcus aureus. Phagocytic inhibition was confirmed each time bead-injected flies were manipulated. All experiments were performed at least three times. For survival analysis, a minimum of 45 flies were injected for each condition. Dead flies were counted daily, and survival data were graphed and analyzed using GraphPad Prism (GraphPad Software, http: //www. graphpad. com). Mean survival time with standard error was calculated using R (http: //www. r-project. org).

Due to the common practice of vaccination and prominence of AIDS, people are already aware of the distinction between adaptive and innate immunity without realizing it. All organisms have an innate immune response, but only vertebrates possess T cells and the ability to produce antibodies. It has been a long-standing assumption that invertebrate immune systems are not adaptive and respond identically to multiple challenges. In this study, we demonstrate that the fly innate immune response adapts to repeated challenges; flies preinoculated with dead Streptococcus pneumoniae are protected against a second, otherwise-lethal dose. Although the underlying mechanisms are likely to be very different, this primed response is reminiscent to vaccine-induced protection in that it exhibits coarse specificity (dead S. pneumoniae only protects against itself), persists for the life of the fly and is dependent on phagocytic cells. This result prompts the obvious question of whether the innate immune system of vertebrates shares a similar biology. Such a finding is of particular interest since immunocompromised individuals only possess an innate immune system.

Abstract Introduction Results/Discussion Materials and Methods

drosophila melanogaster beauveria bassiana streptococcus pneumoniae immunology microbiology

2007

A Specific Primed Immune Response in Drosophila Is Dependent on Phagocytes

Cutaneous leishmaniasis (CL) represents a range of skin diseases caused by infection with Leishmania parasites and associated with tissue inflammation and skin ulceration. CL is clinically widespread in both the Old and New World but lacks treatments that are well tolerated, effective and inexpensive. Oleylphosphocholine (OlPC) is a new orally bioavailable drug of the alkylphosphocholine family with potent antileishmanial activity against a broad range of Leishmania species/strains. The potential of OlPC against Old World CL was evaluated in a mouse model of Leishmania (L.) major infection in BALB/c mice. Initial dose-response experiments showed that an oral daily dose of 40 mg/kg of OlPC was needed to impact time to cure and lesion sizes. This dose was then used to directly compare the efficacy of OlPC to the efficacy of the antileishmanial drugs miltefosine (40 mg/kg/day), fluconazole (160 mg/kg/day) and amphotericin B (25 mg/kg/day). OlPC, miltefosine and fluconazole were given orally for 21 days while amphotericin B was administered intraperitoneally for 10 days. Ulcer sizes and animal weights were followed up on a weekly basis and parasitemia was determined by means of a real-time in vivo imaging system which detects luminescence emitted from luciferase-expressing infecting L. major parasites. Amphotericin B and OlPC showed excellent efficacy against L. major lesions in terms of reduction of parasitic loads and by inducing complete healing of established lesions. In contrast, treatment with miltefosine did not significantly affect parasitemia and lesion sizes, while fluconazole was completely ineffective at the dose regimen tested. Given the data showing the outstanding efficacy and tolerability of OlPC, our results suggest that OlPC is a promising new drug candidate to improve and simplify current clinical management of L. major CL. Leishmaniasis describes a range of visceral and cutaneous disease forms caused by infection with protozoal parasites of the Leishmania genus, transmitted to humans by phlebotomine sandflies [1], [2]. Cutaneous leishmaniasis (CL) is characterized by primary localized skin infections that sometimes resolve without treatment, but can also evolve into disseminated, diffuse, or mucocutaneous lesions. In the Old World, CL is caused mainly by L. major, L. tropica and L. aethiopica, whereas in the New World L. braziliensis, L. panamensis, L. amazonensis, L. guyanensis and L. mexicana are the main causative agents [3]. Based on most recent estimates, about 0. 7 to 1. 2 million new CL cases occur annually [4]. Treatment of leishmaniasis in most endemic regions relies on multiple intralesional, intramuscular or intravenous injections of pentavalent antimonials, old generation drugs that cause considerable toxicity and have unacceptably long treatment schedules which undermine adherence to therapy and contribute to resistance development [2], [5]. Although in the past decade significant progress was made in the field of antileishmanial drug development with the approval of amphotericin B, paromomycin and miltefosine, considerable disadvantages remain [6]. In particular for CL, treatment regimens are poorly justified and have sub-optimal efficacy. Although local therapy can be used to treat certain forms of CL, procedures such as intralesional injection, cryo- or thermotherapy can be painful and may require local anesthesia [3]. Ointments or creams such as those containing paromomycin (WR279,396) are more suitable for uncomplicated CL cases, and their efficacy for treatment of New World CL and complicated CL (multiple lesions) is still under study in well controlled clinical trials in Panama and Peru. Whether administered topically or systemically, treatment efficacy against CL is highly variable and depends both on the infecting Leishmania strain and on the geographic region [2]. As CL is not a life-threatening disease, treatment recommendation is based on a risk-benefit ratio for every case [2]. In view of the considerable drawbacks of current therapies, in particular the long treatment times and associated side effects, moderate clinical manifestations of CL are likely to be undertreated which increases the chance of patients developing debilitating scars or more severe forms of the disease [3]. Orally bioavailable and well-tolerated agents that are effective against a wide range of clinical CL manifestations are needed, especially against complicated CL. So far the only oral drug with acceptable efficacy against leishmaniasis is miltefosine, an alkylphosphocholine generally used in a long 28-day treatment regimen that associates with dose-limiting gastro-intestinal toxicity [3], [6], [7]. Miltefosine has been tested for CL treatment showing acceptable but variable clinical efficacy [8]. Despite these variable results miltefosine (brand name Impavido) was recently approved by the United States Food and Drug Administration. Oleylphosphocholine (OlPC) is a new chemical entity belonging to the alkylphosphocholine family showing antileishmanial activity against a broad range of Old and New World Leishmania species/strains. While OlPC and miltefosine demonstrate comparable activity in vitro, OlPC revealed to be of higher efficacy in vivo when tested in a predictive hamster model of visceral leishmaniasis [9]. This study evaluates the value of OlPC for the treatment of Old World CL (OWCL) by testing it in laboratory models of L. major infected-mice. These models have undergone internal validation and are reproducible according to industry standards [10]. Female BALB/c mice weighing 20–25 grams were purchased from Charles River (Wilmington, MA). The animal protocol was approved by the Walter Reed Army Institute of Research (Silver Spring, MD) institutional animal ethics committee in accordance with national guidelines (protocol number 13-ET-26). Research was conducted in compliance with the Animal Welfare Act, other federal statutes, and regulations that relate to animals and experiments involving animals, and principles stated in the Guide for the Care and Use of Laboratory Animals [11]. The authors abide to the reductionist approach of using animal models in drug development. Luciferase-labeled or standard L. major promastigotes were cultured in Schneider' s medium (Lonza) supplemented with 20% heat-inactivated fetal bovine serum at 25°C. Animals were infected at the base of the tail with 1×107 stationary phase promastigotes. The ulcer areas were measured with a calibrated digital caliper once a week. The average diameter of each tail lesion was calculated as the mean of the horizontal and vertical diameters, and this value was used to calculate the ulcer size area in mm2. The following parameters were examined to determine toxicology inequity of the study drugs: cure or distress/death, body weight, general physical and coat appearance. Crystalline OlPC was supplied by Dafra Pharma Research & Development (Turnhout, Belgium) while miltefosine was purchased from Panslavia Chemicals LLC and provided by the WRAIR depository (Rockville, USA). Fluconazole was purchased from Sigma-Aldrich (St-Louis, USA) and amphotericin B (Ambisome) from Astellas Pharma US Inc. (Northbrook, USA). Miltefosine and OlPC stock solutions were prepared in 1× PBS and stored at room temperature in the dark for a maximum of 7 days. Fluconazole was dissolved in HECT (in 0. 5% (w/v) hydroxyethyl cellulose and 0. 2% (0. 5% HECT, v/v) Tween-80 in distilled water), then homogenized using a PRO Scientific Inc. Monroe, CT homogenizer. AmBisome was dissolved in double distilled sterile water. Efficacy was assessed by comparing the suppression of lesion size after 28 days in the drug treated group to that in negative vehicle control as previously described [10]. Percent suppression is defined as {[ (LS (-) C) −LS (drug) ]/LS (-) C}×100, where LS (-) C = lesion size in negative control and LS (drug) = lesion size in drug group. The threshold for success is a percent suppression which is at least 50% of the positive control amphotericin B [10]. Luciferin (D-Luciferin potassium salt, Xenogen Corporation, Almeda, CA /Goldbio, St Louis, MO), the luciferase substrate, was intra-peritoneally injected into mice at a concentration of 200 mg/kg 18 minutes before bioluminescence analysis. Mice were anaesthetized with isoflurane (MWI veterinary Supply, Harrisburg, PA) and maintained in the imaging chamber for analysis. Emitted photons were collected by auto acquisition with a charge couple device (CCD) camera (IVIS Imaging System 100 Series) using the medium resolution (medium binning) mode. Analysis was performed after defining a region of interest (ROI) that delimited the surface of the affected area. Total photon emission from each infected tail base area was quantified with Living Image software (Xenogen Corporation, Almeda, CA), and results were expressed in photons/sec. A first set of dose-response experiments was used to assess the capacity of OlPC to cure L. major cutaneous lesions in BALB/c mice when given orally. L. major promastigotes were injected at the tail base of the mice and local lesions were allowed to develop until they reached optimal lesion size of ∼50 mm2. Mice were then grouped (n = 5 per group) based on equivalent average lesion sizes and daily oral treatment with OlPC was initiated. Based on previous data generated in L. infantum infected hamsters 9, doses of 10,20, and 40 mg/kg of OlPC were selected to be given for 5 or 10 consecutive days (total doses of 50,100,200 and 400 mg/kg) (Table 1). Ulcer sizes were measured from the first treatment day (Day 0) up to Day 28 post treatment start, and compared to those of vehicle treated animals. In this mouse treatment model, dosing of 10 and 20 mg/kg daily for 5 or 10 days had little to no impact on lesion growth, while the dose of 40 mg/kg was able to significantly reduce their sizes. For the 5-day and 10-day regimens, lesion sizes were reduced by 34. 0% and 93. 5%, respectively (Table 1). The 10-day regimen at 40 mg/kg was independently validated by intraperitoneal (IP) treatment. In this experiment the reduction of lesion sizes was also significant (66. 8%, Table 1), although lower than what had been seen with oral treatment. No sign of drug toxicity (as defined in Materials and Methods) was observed in any of the treatment groups. Taken together, these data pointed that an oral daily dose of 40 mg/kg was needed for effective treatment of L. major lesions in BALB/c mice, although total disappearance of lesions was not observed with 10 days of treatment. Building on the previous dataset, the efficacy of OlPC to cure L. major induced lesions in BALB/c mice was directly compared to those of the clinically used antileishmanial drugs miltefosine, fluconazole and amphotericin B. For this experiment, mice were infected with luciferase-labeled promastigotes and lesions were allowed to develop until they reached optimal lesion size of ∼50 mm2. Mice were then grouped (n = 6 per group) based on equivalent average lesion sizes. To allow direct comparison between treatments, OlPC (40 mg/kg/day), miltefosine (40 mg/kg/day) and fluconazole (160 mg/kg/day [10]) were used orally for 21 days alongside PBS-treated control animals (considering first treatment day as Day 0). As amphotericin B is not orally bioavailable, this drug was administered IP at 25 mg/kg/day for 10 days based on previous experience [10] and served as a positive control. Parasitemia (IVIS), ulcer sizes, and animal weights were followed in each group on a weekly basis. As expected, treatment with the reference drug, Amphotericin B, led to a rapid reduction of the parasite loads (visible on Day 7) correlating with lesion size reduction as of Day 19. By Day 27, lesions had healed/cured (defined as 100% re-epithelialization – normal skin) and parasites could not be detected in the mice of this group (Figure 1A, 1B ◊). Although occurring more slowly, response to oral OlPC also led to gradual but complete clearance of parasitemia (seen on Day 12) followed by lesion regression/healing as of Day 27 (Figure 1A, 1B •). In the OlPC-treated group the lesions had completely re-epithelialized/healed by Day 34. In contrast, parasitemia in both miltefosine and fluconazole treated groups never significantly differed from those of the control group, and no lesion regression was observed (Figure 1A, 1B). On Day 34, the control group had an average lesion size of 118. 9 ± SEM 32. 1 mm2, the fluconazole-treated group 129. 2 ± SEM 25 mm2 and the miltefosine-treated group 55 ± SEM 23. 7 mm2. A detailed analysis of the Day 19 post treatment start time point is presented on Figure 2, with pictures of the luminescent signal in individual mice (Figure 2A), individual group luminescence values (Figure 2B) and lesion sizes (Figure 2C). The IVIS analysis of OlPC-treated mice clearly shows that OlPC is highly effective at clearing parasites at the lesion site despite the fact that the lesions have not yet started to regress in size. Although the OlPC- and miltefosine- treated groups show similar average lesion sizes at this time point, the difference in the activities of both drugs is nevertheless unambiguous. Mice remaining beyond Day 34 were closely monitored until the lesion grew to a size >200 mm2 or until recrudescence of the ulcers, which were considered as clinical end points. Mice of the control group were euthanized as of Day 35 due to excessive lesion sizes together with the fluconazole-treated mice, indicating that fluconazole was ineffective at the selected dose regimen. Miltefosine treatment appeared to slow down the progression of the ulcers up to Day 34 (Figure 1B) suggesting partial efficacy at 40 mg/kg/day×21 days. However the lesions showed enlargement as of Day 40 (end point 96. 4 ± SEM 30. 8 mm2). In contrast, both OlPC and amphotericin B-treated animals remained lesion free up to Day 54. On day 74, which was the final time point evaluated, both the amphotericin B and OlPC mice had relapsed, showing average ulcer sizes of 31. 3 ± SEM 14. 9 mm2 and 28. 4 ± SEM 14. 6 mm2, respectively (not shown). In conclusion, although treatment with oral OlPC had a slower action than IP amphotericin B, the overall capacity of both drugs to clear the infection in the studied model appeared to be similar and far superior to the one of oral miltefosine at equivalent dose. A moderate weight loss was observed in all treatment groups during the 35-day follow-up period, which generally correlated well with disease progression in terms of lesion size. The average weight loss reached a maximum of 8. 9% in the vehicle control group and 13. 4% in the fluconazole-treated group on Day 34 (Table 2). The miltefosine-treated mice experienced higher weight loss during the treatment period, with an average of 20. 6% weight loss on Day 13 (i. e. mid-treatment), indicating potential drug safety issues at the dose used. Amphotericin B- and OlPC- treated mice both experienced a ∼7% weight loss during treatment (peaking on Day 10 and Day 13, respectively), followed by overall weight gain compared to baseline by Day 34 post treatment start. Animals euthanized or found dead during the first 35-day follow-up period are reported in Table 3. For two of the mice (1 in control group, 1 in OlPC group), any association with drug toxicity is formally excluded. As for the other found dead animals (3 in the miltefosine group, 2 in the OlPC group and 1 in the fluconazole group), the possibility of cumulative toxicity or complications due to daily gavage (or a combination or the two) could not be excluded. Of those, it is interesting to note that the 3 deaths in the miltefosine group occurred earlier (Day 12,13 and 19) compared to the ones in the OlPC group (Day 22 and 23), and were associated with piloerection, a recognized sign of sickness in mice, and important weight losses (based on last weight measurement before death; Table 3). Gross necropsy in the two found dead animals of the OlPC-treated group (mouse # 581 and #575) revealed no specific pathological findings, and only mouse #575 underwent weight loss during treatment. In conclusion, both OlPC and amphotericin B showed excellent efficacy against L. major lesions in mice by reducing parasitemia and inducing healing of established lesions. In contrast, treatment with miltefosine at the same dosing regimen as OlPC did not significantly affect parasitemia or induce lesion regression, while fluconazole was completely ineffective at the dose tested. OlPC also appeared better tolerated than miltefosine at equivalent dosing regimen. As human leishmaniasis comprises several clinical syndromes caused by dozens of Leishmania species across the globe, it is unlikely that one drug or drug combination will be effective for all clinical forms of the disease [12]. Therefore, the development of new antileishmanial drugs is needed, preferably with a low side effect profile, oral bioavailability, efficacy in a short treatment regimen, and which can be manufactured at low-cost and adapted for use in rural areas [10], [13]. Currently the only orally bioavailable drug for leishmaniasis is miltefosine, an alkylphosphocholine with a narrow therapeutic window mainly due to its gastrointestinal toxicity. Vomiting and/or diarrhea have been reported in every clinical trial performed with miltefosine [8], and although clinical evidence has suggested efficacy against CL, there is a large variation in clinical response and, in particular for OWCL, more data is needed [8], [14], [15]. Its main limitations are treatment compliance and hence potential for selection of drug resistant parasites and teratogenicity (pregnancy must be avoided during treatment and during the following two months). In this study, the two alkylphosphocholines, miltefosine and oleylphosphocholine (OlPC), were compared side by side for efficacy and safety in a mouse model of OWCL. Of note, the daily dose used approximates the human equivalent dose at which miltefosine is generally used in clinical practice against CL, namely 2. 5–3. 3 mg/kg (corresponding to 30–40 mg/kg/day in mice) [3], but for 21 days instead of the recommended 28 day regimen. Based on data accumulated so far in two independent rodent models of leishmaniasis, namely L. infantum visceral infection in Golden hamsters [9] and L. major cutaneous infection here in BALB/c mice, OlPC has greater in vivo efficacy and superior safety profile compared to miltefosine when compared at equivalent dose regimen. In addition, although no direct comparison with miltefosine was performed, the clinical efficacy of OlPC against L. infantum canine leishmaniasis (CanL) was also demonstrated in naturally infected dogs using a 14-day regimen of 4 mg/kg/day [16], a daily dose exceeding the maximum tolerated dose of miltefosine in that species (recommended miltefosine regimen in dogs: 2 mg/kg for 28 days). Therefore overall, OlPC is better tolerated than miltefosine and has better efficacy (i. e. wider therapeutic window). Since the antileishmanial activity of OlPC and miltefosine is similar in vitro [9], the difference in therapeutic windows between the two drugs could result from differences in oral bioavailability, tissue distribution, or in the affinity of the drugs for the parasites. The detailed PK/PD analysis of OlPC vs. miltefosine in animal models is interesting and deserves further attention. These studies will allow efficient translation of the knowledge accumulated in animal models in future clinical studies in humans aiming at comparing the clinical efficacy of OlPC and miltefosine. The other two comparative drugs used in our study were fluconazole and amphotericin B. Regarding fluconazole, despite the fact that clinical efficacy in CL patients has been reported in the literature against L. major CL at 200 mg daily for six weeks in Saudi Arabia [17] and L. braziliensis CL with 8 mg/kg (highest dose tested; corresponding to about 100 mg/kg as an equivalent dose for mice) [18], this drug appeared to be ineffective in our study when given at 160 mg/kg×21 days. However it cannot be excluded that fluconazole would still be effective in the context of a longer treatment period in Balb/c mice. As for amphotericin B, this drug came out as the “best” overall (considering speed of recovery, average group weight loss and group mortality). However as this drug was given IP, it most likely had an earlier Tmax compared to the other drugs given orally. In addition, IP injections require different types of manipulations than oral gavage, which might influence overall group mortality, aside from the fact that the treatment was only for 10 days as opposed to 21 days for the other groups. Nevertheless, despite the differences in the routes of administration, OlPC achieved similar absolute efficacy than amphotericin B in terms of reduction of parasite loads and lesion remission. The fact that OlPC is orally bioavailable represents a huge practical advantage over amphotericin B considering similar efficacy. Taken together, our study suggests that even though the optimal oral regimen of OlPC against CL still requires further study and optimization, this new alkylphosphocholine opens the possibility of future improvement of CL patient management in terms of having a well-tolerated oral treatment associated with good patient compliance. In this regard the full FDA/EMA-compliant toxicology and safety pharmacology analysis of oleylphophoscholine is being assembled to pave way to further human clinical development in CL/MCL patients. Having a new oral treatment available will also reduce treatment costs, a factor extremely important in remote settings where cold chain distribution and parenteral drug administration remain challenging.

Cutaneous leishmaniasis (CL) is a vector-borne parasitic disease transmitted to humans by sandflies and characterized by local ulcerative skin lesions. The disease is linked to poverty in the Middle-East, North and East Africa, South-Central Asia and South America, with 0. 7 to 1. 2 million new annual cases estimated. In most endemic regions CL treatment relies on injections with pentavalent antimonials, old generation drugs with considerable side effects and long treatment regimens. CL is therefore a highly undertreated disease in need of easy-to-administer, orally bioavailable and well-tolerated agents with broad clinical activity. To date, the only oral drug with acceptable efficacy against leishmaniasis is miltefosine, an alkylphosphocholine with a narrow therapeutic window that limits its use. Given the existing clinical need for CL, we tested the efficacy of oleylphosphocholine (OlPC) in a validated mouse model of Old World (Leishmania major) CL. OlPC is a new orally bioavailable drug of the same family as miltefosine with potent and broad leishmanicidal activity. In direct comparison with miltefosine, our results indicate that OlPC induces higher parasite clearance and lesion healing with measurable improved tolerance. These promising observations warrant further research on OlPC as a new drug to improve clinical management of CL.

Abstract Introduction Materials and Methods Results Discussion

dermatology infectious diseases drug therapy veterinary diseases zoonoses medicine and health sciences skin infections pharmaceutics population modeling leishmaniasis protozoan infections biology and life sciences tropical diseases infectious disease modeling computational biology parasitic diseases veterinary science

2014

Direct Comparison of the Efficacy and Safety of Oral Treatments with Oleylphosphocholine (OlPC) and Miltefosine in a Mouse Model of L. major Cutaneous Leishmaniasis

We developed new methods for parameter estimation-in-context and, with the help of 125 authors, built the AmP (Add-my-Pet) database of Dynamic Energy Budget (DEB) models, parameters and referenced underlying data for animals, where each species constitutes one database entry. The combination of DEB parameters covers all aspects of energetics throughout the full organism’s life cycle, from the start of embryo development to death by aging. The species-specific parameter values capture biodiversity and can now, for the first time, be compared between animals species. An important insight brought by the AmP project is the classification of animal energetics according to a family of related DEB models that is structured on the basis of the mode of metabolic acceleration, which links up with the development of larval stages. We discuss the evolution of metabolism in this context, among animals in general, and ray-finned fish, mollusks and crustaceans in particular. New DEBtool code for estimating DEB parameters from data has been written. AmPtool code for analyzing patterns in parameter values has also been created. A new web-interface supports multiple ways to visualize data, parameters, and implied properties from the entire collection as well as on an entry by entry basis. The DEB models proved to fit data well, the median relative error is only 0. 07, for the 1035 animal species at 2018/03/12, including some extinct ones, from all large phyla and all chordate orders, spanning a range of body masses of 16 orders of magnitude. This study is a first step to include evolutionary aspects into parameter estimation, allowing to infer properties of species for which very little is known. The role of biodiversity in ecosystem structure and functioning is central for conservation and environmental quality management, as well as biospherics and earth system studies. Biodiversity is not only about the number of species present, but also the number and nature of the different characteristics and functions which make up a community or an ecosystem, often referred to as traits. Scientists and managers are turning towards such trait-based approaches to measure the health and vitality of ecosystems. In this context of apprehending biodiversity on the basis of diversity of characteristics and functionalities we have been developing the AmP (Add-my-Pet) project. AmP is a database of referenced data on animal energetics, parameter values of models based on Dynamic Energy Budget (DEB) theory [1–4], and properties derived from these parameters. Some 125 authors contributed to the database at 2018/03/12. The AmP project aims: (i) to find the simplest organization principles for metabolism upon which all life is based and (ii) to understand taxon-specific patterns as variations on this common organization. The development of DEB theory started in 1979 and meanwhile over 700 papers have been published on DEB theory, see www. zotero. org/groups/500643/deb_library/. Partly based on the fact that a large number of popular empirical models turned out to be special cases of DEB models [4], we claim that DEB theory is presently the best tested quantitative theory in biology. The comparison of species on the basis of parameter values is an important aspect of the AmP project. Species-comparisons based on measured quantities suffer from the problem that these quantities typically have contributions from many underlying interacting processes, and were not measured for all species of interest. Parameters of mechanistic models, however, have much simpler links with such processes, which makes it easier to find explanations for differences between species. Moreover, the complete parameter set is available for each entry, allowing to predict, e. g. respiration, without any measured data on respiration. Comparison of species is, however, not the only important application of the AmP website. Prediction of effects of global change [5], understanding the geographic distribution of species [6–8], the effects of (toxic) chemical compounds [9–12], the optimization of bio-production (e. g. aquaculture and agriculture [13–15]), stock management, the best re-introduction of endangered species or the control of invading species [16], are just examples of applications where detailed knowledge of energetics of species in a DEB context is very useful. Like many ecologists, we see energetics as the key to understand the ecological behavior of species [17], and as the root of population [18] and ecosystem dynamics [19,20], with consequences at the planetary level [21]. This is the context that motivated the development of DEB theory, of which the AmP project is an application. In view of the rapid build-up of ecological stress all over the world, we think that the field is in urgent need of an online database like AmP. AmP started in 2009 as an educational initiative, to teach researchers how to estimate DEB parameters from their data and animals have the simplest metabolisms (if we compare with say plants, bacteria or microalgae). The database grew and in 2013, (at about 300 entries) we teamed up, formed the first AmP curator board, and together developed the code and web-platform underlying AmP. We wanted to get an overview on: (i) how well the standard DEB model worked for describing animal metabolism, (ii) standardize and improve the parameter estimation procedure and (iii) improve our capacity to judge the realism of parameter values. We start with a brief introduction of DEB theory. All applications of models, including testing of the model against data, start with knowledge of parameter values. The parameters of a DEB model must be estimated from a collection of data sets on the various aspects of energy budgets and life history, using all this information in combination. This involves new features in the methodology of parameter estimation. The next section of this paper describes improvements we implemented, based on the co-variation method [22,23], which was used in an early phase of the database project. Moreover we needed more administrative rigor and improved methods for detecting patterns in parameter values. Development of new routines and re-organization of the previous estimation procedure allowed us to include these new extensions. So the following section describes the new web-interface and structure of the database. Finally we present and discuss the results obtained after implementing the new method and reaching 1035 species in the collection at 2018/03/12. Given that these entries employ together 270 different types of data, in 585 combinations, the estimation of 14 parameters of each species not only illustrates the scope of the data reduction, but also the step-up in comparison potential. Dynamic Energy Budget (DEB) theory, aims to specify commonalities underlying metabolic organization for all life. It does this by delimiting a small set of assumptions from which mathematical formulae for metabolism are derived, covering the start of embryo development to death by aging through a range of life stages [24–27]. DEB models are meant to apply to all life on earth and allow species comparisons on the basis of (functions of) parameters of that model. The standard DEB model (’std’) is the simplest non-degenerated DEB model implied by the theory and it applies to heterotrophic animals (see Fig 1 for the summary). We refer the reader to [28] for an accessible summary of the principles of DEB theory and a description of the standard DEB model. A detailed derivation of the model on the basis of underlying assumptions is presented in [4], chap. 2. Each DEB parameter of the std-DEB model (see Table 1) has a clear link with one underlying physiological process (specified on the arrows of Fig 1). The combination of parameters covers all aspects of energetics throughout the full life cycle of organisms, include feeding, digestion, storage, maintenance, growth, development, reproduction, aging. Parameter values are individual-specific in the context of DEB theory, but the difference between individuals are typically small enough to average for a species in a meaningful way. Parameter values determine how state variables of an individual (reserve, structure, maturity and reproduction buffer) change in time through all life stages (embryo, juvenile, adult). Life stages have specific definitions: Embryos do not assimilate; Juveniles assimilate and allocate to maturity but not to reproduction; Adults feed, no longer allocate to maturity but store energy/ mass allocated to reproduction into a reproduction buffer which is converted to offspring. To complete the background information needed for this study, we must point to the co-variation rules [29]. If there was no selection pressure of parameter values then particular parameters would co-vary in a simple way with the dimensionless zoom factor z (Table 1), purely based on parameters being either intensive or extensive. The estimation of parameters for each species allows comparing parameters between species and seeing to what extent each parameter is in fact either dependent or independent of body size for different taxonomic groups, thus revealing important patterns in how environmental pressures exert selection pressure on parameter values. Prior to this work, DEB parameters for the AmP entries were obtained with the covariation-method for parameter estimation [22,23]. All the code for parameter estimation is developed in the DEBtool_M package, which is frequently updated and freely available at https: //github. com/add-my-pet/DEBtool_M/. In this section we describe improvements to the method which are implemented in DEBtool. The improvements comprise: The overview of the parameter estimation procedure is presented in Fig 2. In the next section we describe the important new components. One way to judge how good the parameter estimates are is to compute a goodness of fit measure which assesses how close the model predictions are to all of the data. Goodness of fit is not enough, one also needs to check for biological realism. The previous system used goodness of fit measures defined in section 2 of [23]. The new system uses the Mean Relative Error (MRE) and the Symmetric Mean Squared Error (SMSE) to quantify the goodness of fit. MRE = 1 n ′ ∑ i = 1 n RE i, where RE i = ∑ j = 1 n i w i j w i | p i j - d i j | | d i | (w i = ∑ j = 1 n i w i j > 0), that simplifies to R E i = | p i 1 d i 1 - 1 | for zero-variate data. n′ is the number of data sets with wi greater than 0. SMSE = 1 n ′ ∑ i = 1 n SSE i where SSE i = ∑ j = 1 n i w i j w i (p i j - d i j) 2 d i 2 + p i 2, that simplifies to SSE i = (p i 1 / d i 1 - 1) 2 1 + (p i 1 / d i 1) 2 for zero-variate data. See section on loss functions for the definition of the other symbols. MRE can have values in the interval [0, ∞], while SMSE has values in the interval [0,1]. In both cases, 0 means predictions match data exactly. MRE assesses the differences between data and predictions additively, judging equally an overestimation and underestimation of the same relative size (e. g, +20% or −20% have the same contribution), while SMSE assesses the difference multiplicatively, judging overestimation and underestimation by the same factor equally (e. g. × 2 or ×1/2 have the same contribution). Notice that the result of the minimization of loss functions does not, generally, correspond with the minimum of MRE or SMSE (unless the fit is perfect). AmPtool is a software package that is designed to analyse patterns in (functions of) parameter values in selected entries. It is available via github. com/add-my-pet/AmPtool/, and changes frequently since the AmP collection is rapidly expanding. Meta-data, parameters and implied properties (biologically relevant quantities which are functions of parameters as well as food and temperature) are collected in a single Matlab structure (allStat. mat). Advanced plot-routines were developed to plot (functions of) parameters against other (functions of) parameters, for selected species. These selections make use of the taxonomic tree. The tree actually consists of lists-of-lists, based on the newest insights in taxonomy, as presented in the Catalog of Life, the Encyclopedia of Life and Wikipedia. Legends can be created where different taxomonic groups in the tree are attributed user-defined markers. This legend functionality allows selecting taxa at a large number of taxonomic levels. We mentioned in the previous section that goodness of fit is not enough to judge a set of DEB parameters and that one must also check for biological realism. AmPtool is one way to do this. The user can visualize how parameters of closely and less closely related taxonomic groups relate to each other and see if the new parameter estimates are extreme outliers or not. Careful examination of the coherence of the implied properties within an entry is the other way. AmPtool can also be used to find entries on the basis of data types that have been used, or print values of parameters and statistics of selected species. The AmP web-interface was developed as a result of this work. The interface allows examining each entry as well as obtaining on overview on how entries relate to each other. From 2013 till 2017 the number of AmP entries (Fig 3, left) was dramatically increased from ca 300 in 2013 to 1035 at 2018/03/12 to find out whether DEB models do apply to all animals and determine how problematic it is to have limited amounts of data for estimating DEB parameters. AmP receives entries which are submitted by the international scientific community and to date some 125 authors have contributed to the collection. Every author and their associated entry (with links to the entries) are listed on the AmP web-interface. Over the course of this study, we included a bit over 700 extra species to the collection since 2013 as well as converted the previous 300 entries to this new format. We were careful to select species such that the collection had a broad taxonomic scope, as well as included species of commercial relevance and species relevant for conservation and toxicity testing. We aimed at maintaining a good balance between the different taxonomic groups (Fig 3, right) We were on the look-out for exceptional species (in terms of size or aging) that might show that the DEB model was not applicable. We did not yet find such an animal species. The increasing number of DEB applications on animal species motivated the continuing amelioration of the method, making it more robust, more efficient and easier to apply. As the collection of species grew, it became evident that the standard DEB model (Fig 1) required simple extensions for particular taxa, to accommodate larval life stages, fetal development, various forms of metabolic acceleration [36], substantial programmed shrinking (as sported by Elopocephalai), etc. DEB models are classified as s-models, a-models and h-models according to the mode of metabolic acceleration. Since we see the structured model collection that resulted from the AmP project as an important insight into animal metabolism, we come back to it in the discussion section. We here present the typified models. They must be understood as variations on the standard DEB model (std), see Methods section on DEB theory. The general idea is that the choice of typified model depends more on higher-level classifications than the species-level. Delayed stage transitions are also accounted for in the different model families. Most mammals delay the start of fetal development during gestation. Some bivalves delay the start of metabolic acceleration; this phenomenon can prove to be more common with the increase of available data. The three sets of models are detailed below. The AmP database not only allows us to test model against data and evaluate implications, but also to identify evolutionary and ecological patterns in parameter values. The number of papers on patterns in parameter values is now increasing rapidly [10,15,30,31,33,36–41]. Apart from being of direct scientific relevance, future improvements of DEB parameter estimation methods might exploit these patterns, since lack of data is the rule rather than the exception and having some testable prediction is better than no prediction at all. Where, for instance, (measured) maximum body weight is treated as an independent variable in the eco-physiological literature, DEB theory sees this as a property resulting from underlying processes that are quantified via parameter values. So maximum body weight is a function of parameter values (and food availability). Respiration (e. g. the use of dioxygen) is another function of parameter values, which has contributions from various underlying processes, such as maintenance, development, growth and assimilation overheads, etc. While respiration has been measured for only a small part of the species in the AmP collection, it is available for all species as a prediction (i. e. a function of parameter values). This is just an illustration of the power of comparison on the basis of parameter values, rather than on the basis of measurements. The Patterns page on the AmP website illustrates some patterns in parameter values, such as (predicted) respiration as function of (predicted) maximum body weight, confirming Kleiber’s empirical law stating that (measured) respiration is about proportional to (measured) maximum weight to the power 3/4. Two other patterns illustrate the explanation provided by DEB theory: relative reserve capacity is increasing and specific somatic maintenance is decreasing with body size. Reserve does not require maintenance, so does not contribute to respiration. The increase of reserve capacity [ E m ] = { p ˙ A m } / v ˙ with maximum (structural) length follows from co-variation rules [29], where energy conductance v ˙ (which quantifies reserve mobilization) is independent of maximum length, and specific maximum assimilation rate { p ˙ A m } is proportional to it. The increase of specific somatic maintenance with decreasing maximum (structural) length is seen as an ecological adaptation to exploit short-lasting peaks in food abundance [36]. Another pattern shows that, contrary to popular believe, the maximum growth rate of dinosaurs is just in line with other taxa, given their body size. The full analysis of all patterns in parameter values is beyond the scope of this work, and is an ongoing activity [33,41]. The classification of DEB models into s–, a– and h–models has more or less clear links with the evolution of metabolism. This will be discussed here using a feature that sets a– and h–models apart from s–models: metabolic type M acceleration [36], where surface area temporarily scales with volume, while before and after this period it scales with volume to the power 2/3. The amount of type M acceleration is quantified by the ratio of body length at the end and the start of the acceleration period. The effect of this change in a DEB context is that both the specific assimilation and the energy conductance increase with length, while outside the acceleration period they stay the same. The ratio of the two, the specific reserve capacity, is not affected by this type of acceleration. The h–models differ mainly from the a–models by the fact that acceleration extends into the adult stage. Fig 7 presents evolutionary relationships among animal taxa with a color code to indicate the amount of acceleration in their species. Since the oldest animal group, the Radiata, and the oldest deuterostomes, the echinoderms, accelerate, it might well be that acceleration became suppressed in several other groups and this suppression evolved several times in evolution [36]. The Ecdysozoa (n = 99 at 2018/01/01) beautifully illustrate the link between model type and taxonomic relationship. Fig 7 shows that Chaetognatha, Tardigrada, Nematoda and Entognatha hardly accelerate metabolism, just a factor 2 or less. The basic insects, the ephemeropterans and odonata accelerate a bit more, while the crown groups, the holometabolic insects, accelerate very much. All of the h–models are found in insecta, but very interestingly, springtails Entognatha (n = 6), which are no longer classified as insects, follow the abj model. Most insects seem to skip the juvenile phase and allocate to reproduction as larvae, which classifies them as adult by definition in DEB terms, while the imago neither grows, nor eats (frequently). Holometabolic insects insert a pupal phase between the larval and imago phases that behaves like an embryo with a reproduction buffer, where most of the larval structure is first converted to reserve [42] and imago structure is build from reserve. Crustacea (n = 62) sport a mix of s– and a–models Branchiopoda (n = 18) are described by std. Copepoda (n = 6) are hardly resolved and require more research. Calanaus does not accelerate (sbp), the others do (abp). Copepods are special with respect to the other crustaceans in that the κ-rule no longer applies to the adult stage. The 1 species of ostracod is described by abj. Malacostraca are better represented in the collection (n = 36) and are described by abj. While abp now applies to most copepods, it may be that it also applies to ostracods, arachnids and scorpions. The future will teach us. The results further show that among the spiralia only some mollusk species accelerate metabolically, see Fig 8. Ray-finned fish (n = 206) sport a wide range of acceleration factors, see Fig 9, but extreme forms of acceleration are confined to the Otomorpha, the Paracanthomorphacea and the crown groups of the Percomorphaceaei only. The coupling between the amount of acceleration in mediterranean perches with the spawning season, as reported in [39], shows that, apart from evolutionary aspects, ecological ones are involved as well and the two aspects cannot be fully separated. These examples beautifully show that the occurrence of acceleration is far from random. Energy conductance, one of the two parameters that are affected by metabolic acceleration, controls reserve mobilisation, so dominates the incubation (or gestation time), since eggs start their development as a lump of reserve. We have several cases with data for embryo development in combination with post natal development, which show that energy conductance remains constant before and after birth. We met one convincing case, the Asian freshwater leech, Barbronia weberi [43], where energy conductance makes a jump up at birth. There are quite a few cases, however, where incubation time is under-estimated. These cases do not have data on embryo development, so we cannot be sure if energy conductance also makes a jump here, or that the start embryo development is delayed. The latter might be due to a variety of reasons. Most mammalian embryos, for instance, have a period to prepare for growth during which the fetus does not increase in size. The onset of growth is typically quite clear, since consistent with DEB expectations, structural length starts to increase linearly [44,45]. Data on embryo development is relatively scarce. The colors in the taxa names of Figs 7,8 and 9 reflect the range of values of metabolic acceleration that were found in the various taxa. Although over 1000 entries is very large for a database of this type, compared to the 10 million existing animal species, it is close to nothing. Thus we cannot assume that the species in the collection are fully representative for the species in nature. Moreover, the number of species in each taxon, both in nature and in the database, varies enormously; some have just a single species. It is already an accomplishment to indicate the ranges this way, and we will not be surprised if insights on which taxa accelerate change somewhat as the collection grows. Since the database is online and freely accessible, the reader can easily check the values of metabolic acceleration for each species in each taxonomic group using the free package AmPtool (we refer the reader to the online manual for how to do this). This study represents a large scale application of a general theory for metabolic organization of living organisms: the Dynamic Energy Budget theory. Although DEB theory applies to all organisms, the AmP collection only deals with animals. The reason is that animals eat other organisms, which do not vary much in chemical composition. As a first approximation, their environment can be characterized by the variables food availability and temperature. This characterization is hard to make “complete” for other organisms, which hampers comparison. And comparison is the most useful asset of this collection. We contend that animal species can be compared on the basis of DEB parameters and that this offers a tractable means to study animal biodiversity in an ecological and evolutionary context. Moreover, by being mechanistic (= based on first principles), DEB models interpret data, rather than just describe it. They can therfore reveal inconsistencies in data and predict un-measured properties of species as functions of parameters. We present and discuss how DEB parameters can be extracted from eco-physiological data: the AmP approach. The two associated software packages, DEBtool and AmPtool, are freely available via GitHub and have online user manuals. We demonstrate the applicability of DEB theory, by showing that it is possible to extract DEB parameters for animal species even when there is little data. We evaluate goodness of fit with respect to data completeness per species and overall the models fit data well. We found that a family of related DEB models, which share the same 14 DEB parameters, are needed to capture the diversity of life-cycles in the animal kingdom. A main metabolic feature which distinguishes the life-cycles is that some groups have ‘metabolic acceleration’, which has links with larval stages. We present the latest evolutionary overview on which groups were found to have metabolic acceleration. Knowledge gaps are highlighted. The AmP project was initiated at 2009/02/12 and meanwhile 125 authors contributed with entries. It has 1035 entries at 2018/03/12, including all larger phyla and all chordate orders, and is both the smallest as well as the largest database of this kind, since it is unique. We expect that it will remain unique for a long time to come, in view of the huge amount of effort to arrive at the state we are presently in and because we think that DEB models will not have alternatives with matching generality, simplicity and realism. We hope that this study motivates the scientific community to contribute and use the AmP collection.

We discovered that parameters of Dynamic Energy Budget (DEB) models can be estimated from a set of simple data on animal life history aspects, growth and reproduction, if treated in combination. Apart from goodness-of-fit as an estimation criterion, relations with parameter values of other species are important, since DEB parameters have a clear physiological interpretation and a good fit for the wrong reasons is always a risk to consider. We developed and optimized methods for this type of parameter estimation-in-context and organized the results of over 1000 animal species in the open-access Add-my-Pet (AmP) database, to which 125 authors contributed so far. We also developed software package AmPtool to compare parameter values in the collection, that builds on DEBtool to assist applications of DEB theory. A family of related DEB models, structured with respect to the modes of metabolic acceleration, captures biodiversity, including various life stages. We discuss some features of the family structure of DEB models in an evolutionary context. The AmP collection has a great potential for research on the role of biodiversity in ecosystem structure and functioning, which will grow with the size of the database.

Abstract Introduction Methods Results Discussion

taxonomy invertebrates organismal evolution medicine and health sciences ecology and environmental sciences evolutionary biology biodiversity animals endocrine physiology developmental biology data management physiological parameters embryos imagos embryology computer and information sciences endocrinology life cycles insects arthropoda eukaryota ecology physiology animal evolution biology and life sciences puberty organisms

2018

The AmP project: Comparing species on the basis of dynamic energy budget parameters

Dengue dynamics are driven by complex interactions between human-hosts, mosquito-vectors and viruses that are influenced by environmental and climatic factors. The objectives of this study were to analyze and model the relationships between climate, Aedes aegypti vectors and dengue outbreaks in Noumea (New Caledonia), and to provide an early warning system. Epidemiological and meteorological data were analyzed from 1971 to 2010 in Noumea. Entomological surveillance indices were available from March 2000 to December 2009. During epidemic years, the distribution of dengue cases was highly seasonal. The epidemic peak (March–April) lagged the warmest temperature by 1–2 months and was in phase with maximum precipitations, relative humidity and entomological indices. Significant inter-annual correlations were observed between the risk of outbreak and summertime temperature, precipitations or relative humidity but not ENSO. Climate-based multivariate non-linear models were developed to estimate the yearly risk of dengue outbreak in Noumea. The best explicative meteorological variables were the number of days with maximal temperature exceeding 32°C during January–February–March and the number of days with maximal relative humidity exceeding 95% during January. The best predictive variables were the maximal temperature in December and maximal relative humidity during October–November–December of the previous year. For a probability of dengue outbreak above 65% in leave-one-out cross validation, the explicative model predicted 94% of the epidemic years and 79% of the non epidemic years, and the predictive model 79% and 65%, respectively. The epidemic dynamics of dengue in Noumea were essentially driven by climate during the last forty years. Specific conditions based on maximal temperature and relative humidity thresholds were determinant in outbreaks occurrence. Their persistence was also crucial. An operational model that will enable health authorities to anticipate the outbreak risk was successfully developed. Similar models may be developed to improve dengue management in other countries. Dengue viruses are the most important arthropod-borne viruses affecting humans. During the past century, the four serotypes (DENV 1 - DENV 4) have spread to about a hundred countries in the tropical and subtropical world including Asia, Africa, the Americas and the Pacific. Each year, an estimated 50 million people contract dengue fever with at least 500,000 cases of dengue haemorrhagic fever or dengue shock syndrome leading to 25,000 deaths [1]. The spatial distribution of this emerging infectious disease largely reflects the distribution of its primary urban mosquito vector, Aedes aegypti [2]. As no effective vaccine and specific treatment exist, vector control currently represents the only resource to mitigate dengue outbreaks. Epidemic dynamics of dengue, like those of other vector-borne diseases, are driven by complex interactions between hosts, vectors and viruses that are influenced by environmental and climatic factors. Several determinants in dengue fever emergence have been identified including human population growth, accelerated urbanization, increased international transport, weakened public health infrastructure as well as a lack of effective vector control and disease surveillance [3]–[6]. On the other hand, there is growing interest in the impact of climate change on the emergence or re-emergence of vector-borne infectious diseases such as dengue [7]–[10]. It has been shown that climate-induced variations in modelled A. aegypti populations were strongly correlated to reported historical dengue cases (1958–1995) at the global scale [11], and a potential increase in the latitudinal and altitudinal distribution of A. aegypti and dengue are expected under global warming [5], [12]. In a specific ecosystem, the required conditions for the occurrence of a dengue outbreak include i) the presence of a dengue virus, ii) the presence and a sufficient density of competent vectors, iii) a sufficient number of susceptible humans that is serotype-specific, and iv) favorable environmental and climatic conditions for dengue transmission. Despite evidence that climate can influence dengue like other vector-borne diseases (i. e. vector population size and distribution, vector-pathogen-host interactions, and pathogen replication [7],), the relationships between climate, Aedes mosquitoes density and behaviour, human populations and dengue incidence are not well understood. Previous studies have shown that temperature influences the lengths of the mosquito gonotrophic cycle and the extrinsic incubation period of the virus within the mosquito, the survival rate of adults, the mosquitoes population size and feeding behaviours and the speed of virus replication [7], [13], [15]–[19]. Water is necessary for eggs and larva development, mosquito breeding, and humidity affects adult mortality [16]–[17], [20]–[22]. Temperatures and precipitations have been identified as influencing incidence rates of dengue in several endemic areas in the world (i. e. Thailand [23]–[24], Taiwan [25]–[27], Singapore [28], and Puerto Rico [24], [29]). On a broader scale, it is plausible that El Niño-Southern Oscillation (ENSO) also influences patterns of dengue transmission [23]–[24], [30]–[31]. This coupled ocean-atmosphere phenomena results in warm waters displacement and changes in sea surface temperatures (SST) across the Pacific Ocean, and has a strong influence on regional climates, particularly in the Pacific. ENSO can induce large temperature, humidity and precipitation changes for months (see the websites of the International Research Institute for Climate and Society (IRI, www. iri. org), and the National Oceanic and Atmospheric Administration (NOAA, www. noaa. gov) for more details). Importantly, previous studies revealed a positive correlation between ENSO, as measured by the Southern Oscillation Index (SOI), and dengue outbreaks in the South Pacific islands [30]–[31]. Our study was conducted in New Caledonia where dengue represents a major public health problem like in many Pacific Islands Countries and Territories [32]. The first dengue outbreak in New Caledonia occurred in 1884–1885 [33]. Disease transmission increased after World War II, and successive waves of epidemics involving all four serotypes were reported. Since 2000, serotype 1 has been predominant [34] causing more than 6,000 cases during the 2003–2004 epidemics [35] and about one thousand of cases in 2008. Although the serotype 4 [36] was involved in a major outbreak in 2009 (8,456 cases), the serotype 1 is still circulating. New Caledonia has had an effective surveillance system for dengue and access to high quality meteorological data for many years. Since 2000, regular entomological surveillance is performed. This provides an opportunity to study the influence of climate variations on dengue dynamics. We analyzed the epidemiology of dengue fever in Noumea, the capital of New Caledonia, from 1971 to 2010 together with local and remote climate influences. The objectives of this study were i) to improve our knowledge of the relationships between meteorological variables, entomological surveillance indices and dengue fever dynamics at seasonal to inter-annual time scales, ii) to identify suitable conditions for an epidemic occurrence, and iii) to develop a predictive model for dengue outbreaks that can be integrated in an early warning system in New Caledonia. New Caledonia is a French overseas territory located in the subregion of Melanesia in the southwest Pacific, about 1,200 kilometres east of Australia and 1,500 kilometres northwest of New Zealand. It lies astride the Tropic of Capricorn, between 19° and 23° south latitude. Its climate is tropical. This archipelago of 18,575 square kilometres is made up of a main mountainous island elongated northwest-southeast 400 kilometres in length and 50–70 kilometres wide, the Loyalty Islands (Mare, Lifou, and Ouvea), and several smaller islands (e. g. Isle of Pines). The population was estimated in January 2009 to be 245,580 [37]. Approximately half of inhabitants are concentrated in the southeast region of the main island around Noumea, the capital. A. aegypti is the only mosquito vector of dengue in New Caledonia. The two others vectors of dengue present in the Pacific region, A. albopictus and A. polynesiensis, have never been detected in this archipelago [38]–[40]. In Noumea, most of A. aegypti breeding sites are outdoors and therefore rainfall dependent. Bivariate and multivariate analyses were conducted using the R software package (R development Core Team version 2. 9. 1 [42]). During the 1971–2010 period, a significant correlation was found between dengue incidence rates and mean annual mean Temp in Noumea (Spearman' s coefficient rho = 0. 426, p-value = 0. 007, Figure 1) but there was no significant correlation with annual mean RH and Precip. Similar results were obtained with conserved trends and detrended data. Anomalies of annual means of mean Temp, Precip and mean RH were significantly correlated with ENSO, as measured by Niño 3. 4 (rho = −0. 365, p-value = 0. 029; rho = −0. 481, p-value = 0. 003; rho −0. 486, p-value = 0. 003, respectively). During El Niño (positive value of Niño 3. 4), the weather was cooler and drier. During La Niña (negative value of Niño 3. 4), the weather was warmer and wetter. However, no direct correlation was found between ENSO and dengue incidence rates at the inter-annual scale (rho = −0. 106, p-value = 0. 539). Dengue outbreaks occurred during either El Niño, La Niña or neutral phases of ENSO. During the 2000–2009 period, dengue incidence rates, meteorological and entomological data were analyzed in Noumea at a monthly scale. A strong seasonal distribution of HI, BI and API was observed (Figure 3), and significant correlations were found between monthly entomological surveillance indices and climate variables (data not shown). Although the highest dengue incidence rates and the highest values of HI, BI and API were observed during the same period of the year (from January to July), no significant time-lagged correlation has been found between monthly entomological indices and dengue incidence rates reported in Noumea over the 2000–2009 period (supporting Figure S1). We did not find relevant entomological patterns during dengue outbreaks. Accordingly, entomological surveillance indices were not used for the modelling of dengue outbreak risk. Based on the tercile method, there were 13 epidemic years (dengue incidence rate in the upper tercile, i. e. >19. 48 cases/10 000 inhabitants) and 13 non epidemic years (dengue incidence rate in the lower tercile, i. e. <4. 13 cases/10 000 inhabitants). A detailed analysis was performed based on monthly and quarterly meteorological data measured from September (year y-1) to April (year y), i. e. four months before and after the outbreak onset. Temperatures (min Temp, mean Temp and max Temp) were higher during epidemic years than during non epidemic years. The peak of max Temp, observed usually in February, preceded the epidemic peak of dengue with a lag of 1–2 months (Figure 4a). Analysis of daily data allowed identifying important temperature thresholds. It revealed that the number of days with max Temp exceeding 32°C, mean Temp exceeding 27°C, and min Temp exceeding 22°C were significantly higher during epidemic years than during non epidemic years. The most important and significant differences were observed during the first quarter of the year, principally in February for max Temp (p-value<0. 01 using a t-test, Figure 4b). By contrast, the relationships between Precip, mean RH and dengue dynamics were not clear, as shown in supporting Figure S2. Highest Precip and mean RH were observed in February–March–April during the epidemic phase of dengue. Using a t-test, Precip and mean RH were significantly lower in February during epidemic years than during non epidemic years (p-value<0. 01 and = 0. 04, respectively). Inversely, the ETP was significantly higher in February (p-value = 0. 02). WF, HB, ENSO indices and entomological surveillance indices were not significantly different between epidemic and non epidemic years. Meteorological variables showing strongest correlations with the epidemic years series, as defined in the Methods section, are presented for each family of variables in Table 1. Significant correlations were identified with several local meteorological variables (particularly Temp, Precip, RH, and ETP) but not with ENSO indices. No or poor correlation was found with WF and HB. In accordance with Figure 4 and supporting Figure S2, Temp were positively correlated with dengue outbreaks in Noumea, whereas Precip and RH measured in February were negatively correlated with dengue outbreaks. A positive correlation was found between the ETP measured in February and the occurrence of dengue outbreaks. First, in order to produce an explicative model of dengue outbreak, we selected meteorological variables observed within the period of dengue outbreak onset, i. e. from January to April (Figure 2). The best SVM model based on the minimum AICc (−79. 21) was obtained using two meteorological variables, i. e. the number of days with maximal temperature exceeding 32°C during the first quarter of the year (NOD_max Temp_32_JFM), and the number of days with maximal relative humidity exceeding 95% during January (NOD_max RH_95_January). The addition of a third meteorological variable did not improve the performance of the model. Results obtained in leave-one-out cross validation (Figure 5) were close to those obtained with the complete dataset (Figure S3) and were characterized by a high ROC-AUC value reaching 0. 80 and 0. 85, respectively. As indicated by the ROC curves, most of epidemic years were predicted correctly with high probability and few false alarms. Importantly, with bivariate analysis, NOD_max Temp_32_JFM was positively correlated with the occurrence of dengue outbreak (rho = 0. 57, p-value = 0. 002) whereas NOD_max RH_95_January did not appear to be a discriminatory meteorological variable (rho = −0. 11, p-value = 0. 58). With multivariate analysis, these two variables were highly informative and discriminatory. Scatter plots of epidemic and non epidemic years as a function of these two variables allowed the identification of three distinct groups (Figure 6): group A including years characterized by low NOD_max Temp_32_JFM (<12 days) and low NOD_max RH_95_January (<12 days), group B including years characterized by high NOD_max Temp_32_JFM (>12 days) and low NOD_max RH_95_January, and group C including years characterized by low NOD_max Temp_32_JFM and high NOD_max RH_95_January (>12 days). According to the tercile method of years classification, all non epidemic years belonged to group A whereas all epidemic years, except 1973 and 2003, belonged to either group B or group C. Similar results were obtained using the median method ensuring the inclusion of all years, preferable for the development of SVM models. Only four years (1978,1979,1985, and 2002) belonging to the middle tercile (dengue incidence rate ranging from 4. 13 to 19. 48 cases/10 000 inhabitants/year) were incorrectly classified using the median method. In 2002, although favorable climatic conditions for dengue outbreak were observed, the incidence rate (5. 24 dengue cases/10 000 inhabitants/year) was close to the median (7. 65 dengue cases/10 000 inhabitants/year). In 1978,1979 and 1985, the low values of NOD_max Temp_32_JFM and NOD_max RH_95_January were not favorable for dengue outbreak. However, incidence rates (7. 74,10. 63, and 11. 24 dengue cases/10 000 inhabitants/year, respectively) were close to the median. Two years (1973 and 2003) belonging to epidemic years using either a tercile or a median method of classification were characterized by low NOD_max RH_95_January and intermediate NOD_max Temp_32_JFM, as members of group A (non epidemic years). However, dengue outbreaks occurred with high incidence rates (23. 64 and 213. 58 dengue cases/10 000 inhabitants/year in 1973 and 2003, respectively). These mismatches indicate that i) the model fails for years that are difficult to classify as their dengue incidence rates were close to the median and in the middle tercile and, ii) NOD_max Temp_32_JFM and NOD_max RH_95_January alone cannot account for all dengue outbreaks (Figure 6). It is likely that other climate events and other factors influencing dengue dynamics contribute to the epidemic spread of dengue viruses during these peculiar years. We were thus able to build an efficient explicative model of dengue epidemics based on meteorological variables contemporaneous to the outbreak. Another challenge was to construct a predictive model for dengue epidemics using variables available prior to the outbreak onset, i. e. from September (year y-1) to December (year y-1). Accurate predictive skill (AICc = −66. 64) was achieved with the SVM model built from the value of the two following variables: the quarterly mean of maximal relative humidity during October–November–December (max RH_OND), and the monthly mean of maximal temperature in December (max Temp_December) of the year y-1 with a ROC-AUC value of 0. 83 (supporting Figure S4). Probabilities obtained in leave-one-out cross validation (Figure 7) and the corresponding ROC-AUC value reaching 0. 69 illustrate the robustness of this predictive model. Importantly, max RH_OND and max Temp_December were not significantly correlated with the risk of dengue outbreak with bivariate analysis (rho = 0. 24, p-value = 0. 14; and rho = 0. 25, p-value = 0. 14, respectively). Scatter plots of epidemic years and non epidemic years built from the combination of meteorological variables used for the SVM explicative model (Figure 8) and for the SVM predictive model development (Figure 9) show that dengue outbreaks occurred in distinct climatic conditions in Noumea. With the SVM predictive model, as noted with the SVM explicative model, epidemic years belonged to two different groups of data according to the value of max RH_OND and max Temp_December (see the two red kernels corresponding to high risk of dengue outbreak in Figure 9). Dengue outbreaks occurred following either years characterized by high max Temp_December and relatively low max RH_OND, or years characterized by high max RH_OND_December, and max Temp_December. To note, the high value of max Temp_December (31. 2°C) and the relatively low value of max RH_OND (86. 8%) measured in 2010 indicate a high risk (74%) of dengue outbreak for 2011. A comparison of the results obtained with the explicative model and the predictive model was performed together with a detailed analysis of the relationships between meteorological variables used to build the explicative model (NOD_max Temp_32_JFM and NOD_max RH_95_January) and those used to build the predictive model (max RH_OND and max Temp_December). As shown in Figure S5, strong relationships exist between the values of max Temp and max RH measured at the end of the year y-1, and those measured at the beginning of the year y. Low max RH_OND and max Temp_December (year y-1) were predictive of low NOD_max Temp_32_JFM and NOD_max RH_95_January (years y, group A). High max RH_OND and max Temp_December (year y-1) were predictive of either high NOD_max Temp_32_JFM and low NOD_max RH_95_January (years y, group B), or low NOD_max Temp_32_JFM and high NOD_max RH_95_January (years y, group C). Results obtained with the predictive model were highly consistent with those obtained with the explicative model with similar probabilities of dengue outbreak risk obtained for 30 of the 40 studied years. Failures of the predictive model can be explained by a lack of correlation between these meteorological variables on a few occasions (e. g. 1982,1983,1995). For example, although the predictive model estimated a risk of dengue outbreak close to 5% in 1995, the explicative model estimated a risk over 90%, and a major outbreak occurred. The value of max RH_OND and max Temp_December measured in 1994 (87% and 27. 6°C, respectively) were relatively low and therefore not predictive of outbreak risk. However, climatic conditions were favorable for a dengue outbreak occurrence (NOD_max Temp_32_JFM = 20 days, NOD_max RH_95_January = 0 day, group B). This suggests that other climate variables or meteorological processes may impact on the local value of NOD_max Temp_32_JFM and NOD_max RH_95_January. Overall, the high performance of the climate-based models of dengue outbreak risk developed in our study suggest that dengue dynamics were essentially driven by climate during this 1971–2010 period in Noumea. The explicative model provides important and new information. We have shown that maximal values of temperature and relative humidity were determinant in dengue outbreaks occurrence and precise thresholds of their value were identified. Importantly, we found that the most relevant meteorological variables explaining dengue outbreaks were built using the number of days for which the variable was greater than a threshold value introducing the importance of the persistence of suitable climatic conditions. Our findings are compatible with the mosquito biology and viral transmission cycle. The length of Aedes gonotrophic cycle is shorter at temperatures above 32°C and feeding frequency is more than twofold at 32°C as compared to 24°C; pupae development period reduced from four days at 22°C to less than one day at 32–34°C [16]–[17], [47]. Additionally, the experimental infection of A. aegypti with DENV-2 viruses showed that the extrinsic incubation period shortens from 12 days at 30°C to seven days at 32–35°C leading to an increasing risk of viral transmission from an infected mosquito to a susceptible host [15]. The influence of temperature on the rate of virus replication inside mosquitoes was also evidenced in the study of Watts et al. Temperatures may also influence the vector size and its biting rate [19], [21]. Consequently, it is likely that the increased level of viral transmission characterizing dengue outbreaks in Noumea at temperatures exceeding 32°C may be a consequence of shortening of the A. aegypti gonotrophic cycle and extrinsic incubation period, and of increased vector feeding frequency. Mortality rate of larvae, pupae and adult mosquitoes as a function of temperature between 10 and 40°C can be represented by a wide-base ‘U’ graphical shape with lower mortality rate at temperature ranging from 15 to 30°C [16]–[20], [22]. Hence, A. aegypti mortality rate may be relatively constant at temperatures observed usually in Noumea, and the increasing mortality rate expected above 32°C is not likely to be an important limiting parameter in the spread of dengue viruses in this specific ecosystem. Larval breeding places are mostly outdoors in Noumea and mosquito abundance increases during the rainy and humid season. Moreover, relative humidity may be determinant in A. aegypti egg development and adult population size that may itself be correlated with vectorial capacity [48]. High humidity shortens incubation and blood-feeding intervals; it favours adult mosquito longevity [20] and thus dengue transmission. This may explain why a sustained high RH during January is associated with a higher risk of dengue outbreak in Noumea. On a broader scale, a growing number of studies have shown that ENSO may be associated with changes in the risk of mosquito borne diseases such as dengue [23]–[24]. By contrast, Hales et al. [31] further analyzed the relationships between the annual number of dengue cases in New Caledonia, ENSO, temperature and rainfall using global atmospheric reanalyses climate based data, and they did not find any significant correlation between SOI and dengue (Pearson' s coefficient = 0. 20). In accordance with this study, and with the advantage of observational and long term data, we found significant inter-annual correlations between ENSO and our local climate but not between ENSO and dengue (Table 1). Moreover, the selection process of multivariate models did not select any ENSO index neither in explicative mode nor in predictive mode. These findings suggest that, in New Caledonia, large-scale climate indices such as ENSO cannot account for the complexity of the local meteorological inter-annual situations. However, at a larger scale, Hales et al. showed that the number of dengue outbreaks in the South Pacific islands (aggregated data, 1970–1995) were positively correlated with the SOI [30], suggesting that La Niña may favour dengue outbreaks in this region of the world. The impact of ENSO on local weather in the South Pacific may strongly vary from one place to another. New Caledonia, located around 20° south latitude in the western Pacific is relatively far from the main centre of action of ENSO located in the equatorial central/eastern equatorial Pacific and its local weather is thus not only influenced by ENSO, but also by other climate modes such as the Madden-Julian Oscillation which strongly influences local meteorological parameters at intra-seasonal (30 to 90 days) time scales [49]. In contrast, ENSO influence may be stronger in islands located closer to the equator, the relationship between ENSO and dengue epidemics being therefore more straightforward [29]. Our long-term study also suggests an increasing risk of dengue outbreaks in New Caledonia in the context of global warming (Figure 1). Even though a global upward trend of dengue incidence rates was noted along the 1971–2010 period, and as surveillance methods and laboratory tests have evolved, it is difficult to know if the amplitude of dengue outbreaks is significantly growing. Even though climate influenced the disease epidemiology in Noumea during this forty-year period, the reasons of dengue emergence in New Caledonia are multiple, including population growth (119,710 inhabitants in 1973 to 245,580 in 2009), accelerated urbanization particularly around Noumea, tourism development and increasing international and inter-islands traffic [50]. The emergence of dengue fever in other parts of the world, particularly South East Asia where dengue is endemic with a co-circulation of the four serotypes, represents an increasing source of virus introduction into New Caledonia. Indeed, multiple and repeated introductions of dengue viruses have been detected from several countries in Asia [34]. Moreover, the geographical distribution of A. aegypti has expanded during recent decades in New Caledonia (Paupy and Guillaumot, unpublished data). Well known factors may have contributed to the epidemic dynamics such as the size of susceptible human hosts and vectors populations. In the absence of seroprevalence data, and due to the lack of long term entomological data, these variables were not included in the input dataset of the models. Nevertheless, as dengue is known to confer a prolonged serotype-specific immunity in the long term, herd immunity represents an important factor in understanding dengue dynamics [51]–[54]. In New Caledonia, successive waves of dengue outbreaks involving the same serotype were reported in 1980 and 1986 (DENV-4), 1989 and 1995 (DENV-3), 2003 and 2008 (DENV-1). This constant interval time between two epidemics involving the same serotype has already been observed in other South Pacific Islands [55]–[57]. Recently, a large molecular characterization of DENV-1 viruses collected regularly in French Polynesia between the 2001 and 2006 outbreaks revealed that the virus responsible for the severe 2001 outbreak was introduced from South-East Asia, and evolved under an endemic mode until its re-emergence under an epidemic mode five years later [56]. These findings suggest that 5–6 years may be necessary for the renewal of the susceptible population in these islands. In New Caledonia, at four occasions, dengue outbreaks were detected between January and July during two successive years: in 1976–1977 (DENV-1), 1995–1996 (DENV-3), 2003–2004 (DENV-1), and 2008–2009 (DENV-1 and DENV-4). This suggests that environmental conditions may be not favorable for dengue transmission all through the epidemic year, particularly during the second semester of the year characterized by lower values of entomological indices. It is likely that dengue re-emerged the following year when climatic conditions were favorable for dengue transmission (as suggested by the results of our explicative model in 1977,1996,2004 and 2009) and the size of the mosquito-vector and susceptible human populations were still sufficient for a large spread of dengue viruses. In these four examples of recurrent outbreaks during two consecutive years, it is more likely that the end of the epidemic was driven by limiting climatic factors and intricate entomological factors rather than by the depletion of the susceptible population. The relationship between Aedes density and the intensity of dengue transmission remains unclear [47], [58]–[60]. Although dengue viruses cannot circulate if mosquito vectors are not present, the vector density of adult female A. aegypti necessary for dengue viruses to become endemic or epidemic remains unknown. In Noumea, entomological indices (HI, BI and API) were not correlated with the incidence rate of dengue, they were sometimes lower during epidemic than during non epidemic periods and lowest values were measured during the largest outbreak in 2009. The fact that these usual entomological surveillance indices (particularly API) are good indicators of adult density in Noumea suggests that the mosquito density threshold under which dengue viruses cannot spread widely may be very low and has never been reached up to now. Moreover, mosquito populations are influenced by human behaviours and meteorological variables alone cannot account for their geographical distribution and abundance [14], [61]. At the domestic level, A. aegypti populations are also influenced by global trends in urbanization, socioeconomic conditions, and vector control efforts. For instance, the outbreak predicted in 2002 with a probability close to 90% did not occur. A possible explanation is that strong vector control policies (e. g. increased efforts to reduce mosquito breeding sites and undertake human population education, development of perifocal spraying of insecticides) were undertaken in New Caledonia at the time of large dengue outbreaks in the other Pacific French overseas territories (French Polynesia in 2001, Wallis and Futuna in 2002). A relaxation in vector control efforts at the end of 2002 may have allowed the resurgence of dengue in the East coast and the spread of the virus through the archipelago during the next year. Overall, our results suggest that the local climate had a major effect on dengue dynamics in Noumea during the last forty years. It is likely that other factors, not included in the input dataset of the models, had a lower influence on dengue epidemic dynamics. The introduction of dengue viruses may have been relatively constant, and the number of human hosts susceptible to a given serotype and of mosquito-vectors may have been always sufficient for an epidemic to occur when suitable climate conditions were met. It is likely that the susceptibility of human populations influenced the serotype involved in the outbreak and the epidemic magnitude. The variability of the length of the gonotrophic cycle, the extrinsic incubation period, and the life span of infected mosquitoes under climate change rather than the overall vector density may play a major role on the epidemic dynamics of dengue at the seasonal scale. Although the meteorological variables contemporaneous to the epidemic season provide crucial information on local dengue dynamics as discussed above, prediction models are needed to anticipate the risk before the dengue outbreak onset and to make the model useful for health authorities in New Caledonia. In this study, we were able to build such a predictive model relying on maximal temperature and relative humidity measured in Noumea at the end of the previous year. Biological interpretations about statistical associations between specific climatic conditions and the yearly risk of dengue outbreak in Noumea can be made in the frame of the explicative model as it uses relevant climatic variables that occur within the period of outbreak onset. The meteorological variables selected in the frame of the predictive model are tightly connected with the explicative meteorological variables (Figure S5). As Noumea concentrates the majority of inhabitants and of dengue cases, as this city has been affected by all dengue outbreaks that occurred in New Caledonia during the last 40 years, and as dengue epidemics usually begin in Noumea, our predictive model is useful to anticipate the risk of dengue outbreak in New Caledonia. However, climatic conditions in Noumea can not account for dengue epidemics in other localities in New Caledonia that would not involve Noumea, even if this situation has never been observed in 40 years. Depending on the user' s objectives, different detection thresholds corresponding to a probability of dengue outbreak can be used. In the case of dengue, it is likely that decision makers would prefer to choose a detection threshold with high true positive rate and low false positive rate, as obtained with a detection threshold of 65% (Figure 7b). The model initialized in December 2009 indicated no risk of dengue outbreak for 2010 that was in accordance with the current epidemiological situation. To note, a high risk of dengue outbreak is predicted for 2011 (74%, Figure 9). Up to now, only a few cases of dengue fever have been reported. Only one case imported from the Philippines was possible to type and belonged to the serotype 1. It is likely that a significant part of the human population is immunized against the serotypes 1 and 4 involved in the largest dengue outbreaks reported in New Caledonia in 2008 and 2009 but the introduction of a new serotype (DENV-2 or DENV-3) may lead to another epidemic. However, several important confusing factors may interfere with dengue dynamics this year such as the massive rainfalls brought by the tropical cyclone Vania in middle January 2011 with its unknown effects on vector populations, the introduction and worrying local diffusion of Chikungunya viruses transmitted by the same mosquito and the subsequent enhancement of vector control policies. In conclusion, the epidemic dynamics of dengue fever were strongly influenced by climate variability in Noumea during the 1971–2010 period. Local thresholds of maximal temperature and relative humidity have been identified with precision allowing the development of explicative and predictive climate-based models of dengue outbreak risk. The health authorities of New Caledonia have now integrated these models into their new decision making process in order to improve their management of dengue, in combination with clinical, laboratory (e. g. serotype determination), and entomological surveillance data. This work provides an example of the practical utility of research projects in operational public health fields and reinforces the need for a multidisciplinary approach in the understanding and management of vector-borne diseases. Our results provide also new insights for future experimental studies. It seems important now to study the impact of maximal temperatures exceeding 32°C and maximal relative humidity exceeding 95%, and the influence of their duration (more or less than 12 days) on the length of the extrinsic incubation period, feeding frequency and longevity of A. aegypti from New Caledonia. The epidemic dynamics of dengue are driven by complex interactions between human-hosts, mosquito-vectors and viruses. These interactions are influenced by environmental and climatic factors that may have more or less burden according to the geographical localisation, the local climatic conditions, the vector characteristics (e. g. Aedes species and strains), the size and movements of human populations and the epidemiology of dengue. Consequently, our results can not be applied to other ecosystems. However, the methodology of analysis used in this study could be extended to other localities highly threatened by the emergence of dengue in the South Pacific, like in other tropical and subtropical countries. As global atmospheric reanalyses climate based data exist, there is hope for the development of local predictive models of dengue outbreak in countries where no reliable weather data are available.

Dengue fever is a major public health problem in the tropics and subtropics. Since no vaccine exists, understanding and predicting outbreaks remain of crucial interest. Climate influences the mosquito-vector biology and the viral transmission cycle. Its impact on dengue dynamics is of growing interest. We analyzed the epidemiology of dengue in Noumea (New Caledonia) from 1971 to 2010 and its relationships with local and remote climate conditions using an original approach combining a comparison of epidemic and non epidemic years, bivariate and multivariate analyses. We found that the occurrence of outbreaks in Noumea was strongly influenced by climate during the last forty years. Efficient models were developed to estimate the yearly risk of outbreak as a function of two meteorological variables that were contemporaneous (explicative model) or prior (predictive model) to the outbreak onset. Local threshold values of maximal temperature and relative humidity were identified. Our results provide new insights to understand the link between climate and dengue outbreaks, and have a substantial impact on dengue management in New Caledonia since the health authorities have integrated these models into their decision making process and vector control policies. This raises the possibility to provide similar early warning systems in other countries.

Abstract Introduction Methods Results Discussion

medicine

2012

Climate-Based Models for Understanding and Forecasting Dengue Epidemics

Malaria parasites (Plasmodium spp.) and related apicomplexan pathogens contain a nonphotosynthetic plastid called the apicoplast. Derived from an unusual secondary eukaryote–eukaryote endosymbiosis, the apicoplast is a fascinating organelle whose function and biogenesis rely on a complex amalgamation of bacterial and algal pathways. Because these pathways are distinct from the human host, the apicoplast is an excellent source of novel antimalarial targets. Despite its biomedical importance and evolutionary significance, the absence of a reliable apicoplast proteome has limited most studies to the handful of pathways identified by homology to bacteria or primary chloroplasts, precluding our ability to study the most novel apicoplast pathways. Here, we combine proximity biotinylation-based proteomics (BioID) and a new machine learning algorithm to generate a high-confidence apicoplast proteome consisting of 346 proteins. Critically, the high accuracy of this proteome significantly outperforms previous prediction-based methods and extends beyond other BioID studies of unique parasite compartments. Half of identified proteins have unknown function, and 77% are predicted to be important for normal blood-stage growth. We validate the apicoplast localization of a subset of novel proteins and show that an ATP-binding cassette protein ABCF1 is essential for blood-stage survival and plays a previously unknown role in apicoplast biogenesis. These findings indicate critical organellar functions for newly discovered apicoplast proteins. The apicoplast proteome will be an important resource for elucidating unique pathways derived from secondary endosymbiosis and prioritizing antimalarial drug targets. Identification of new antimalarial drug targets is urgently needed to address emerging resistance to all currently available therapies. However, nearly half of the Plasmodium falciparum genome encodes conserved proteins of unknown function [1], obscuring critical pathways required for malaria pathogenesis. The apicoplast is an essential, nonphotosynthetic plastid found in Plasmodium spp. and related apicomplexan pathogens [2,3]. This unusual organelle is an enriched source of both novel cellular pathways and parasite-specific drug targets [4]. It was acquired by secondary (i. e. , eukaryote–eukaryote) endosymbiosis and has evolutionarily diverged from the primary endosymbiotic organelles found in model organisms. While some aspects of apicoplast biology are shared with bacteria, mitochondria, and primary chloroplasts, many are unique to the secondary plastid in this parasite lineage. For example, novel translocons import apicoplast proteins through specialized membranes derived from secondary endosymbiosis [5–8], while the parasite’s pared-down metabolism necessitates export of key metabolites from the apicoplast using as-yet unidentified small molecule transporters [9,10]. These novel cellular pathways, which are also distinct from human host cells, can be exploited for antimalarial drug discovery. Indeed, antimalarials that target apicoplast pathways are currently in use as prophylactics or partner drugs (doxycycline, clindamycin) or have been tested in clinical trials (fosmidomycin) [11–15]. However, known apicoplast drug targets have been limited to the handful of pathways identified by homology to plastid-localized pathways in model organisms. Meanwhile, the number of druggable apicoplast targets, including those in unique secondary plastid pathways, is likely more extensive [16]. A major hurdle to identifying novel, parasite-specific pathways and prioritizing new apicoplast targets is the lack of a well-defined organellar proteome. So far, the apicoplast has not been isolated in sufficient yield or purity for traditional organellar proteomics. Instead, large-scale, unbiased identification of apicoplast proteins has relied on bioinformatic prediction of apicoplast targeting sequences [17–19]. These prediction algorithms identify hundreds of putative apicoplast proteins but contain numerous false positives. Confirmation of these low-confidence candidate apicoplast proteins is slow due to the genetic intractability of P. falciparum parasites. Unbiased identification of apicoplast proteins in an accurate and high-throughput manner would significantly enhance our ability to study novel apicoplast pathways and validate new antimalarial drug targets. Proximity-dependent biotin identification (BioID) and other proximity-based proteomics methods are attractive techniques for identification of organellar proteins [20,21]. In BioID, a promiscuous biotin ligase from Escherichia coli (BirA*) is fused to a bait protein and catalyzes biotinylation of neighbor proteins in intact cells. Proximity labeling methods have been used for unbiased proteomic profiling of subcellular compartments in diverse parasitic protozoa, including Plasmodium spp. [22–29]. Here, we used BioID to perform large-scale identification of P. falciparum apicoplast proteins during asexual blood-stage growth. Extending beyond previous BioID studies of unique parasite compartments, we achieved high positive predictive value of true apicoplast proteins by implementing an endoplasmic reticulum (ER) negative control to remove frequent contaminants expected based on the trafficking route of apicoplast proteins. Furthermore, higher coverage was achieved by using the proteomic data set to develop an improved neural network prediction algorithm, ApicoPLAST Neural Network (PlastNN). We now report a high-confidence apicoplast proteome of 346 proteins rich in novel and essential functions. To target the promiscuous biotin ligase BirA* to the apicoplast, the N-terminus of a green fluorescent protein (GFP) –BirA* fusion protein was modified with the apicoplast-targeting leader sequence from acyl carrier protein (ACP) (Fig 1A). Since apicoplast proteins transit the parasite ER en route to the apicoplast [30], we also generated a negative control in which GFP–BirA* was targeted to the ER via an N-terminal signal peptide and a C-terminal ER-retention motif (Fig 1A). Each of these constructs was integrated into an ectopic locus in Dd2attB parasites [31] to generate BioID–Ap and BioID–ER parasites (S1A Fig). Immunofluorescence colocalization and live imaging of these parasites confirmed GFP–BirA* localization to either the apicoplast or the ER (Fig 1B and S1B Fig). To test the functionality of the GFP–BirA* fusions in the apicoplast and ER, we labeled either untransfected Dd2attB, BioID–Ap, or BioID–ER parasites with DMSO or 50 μM biotin and assessed biotinylation by western blotting and fixed-cell fluorescent imaging. As has been reported [28], significant labeling of GFP–BirA*-expressing parasites above background was achieved even in the absence of biotin supplementation, suggesting that the 0. 8 μM biotin in standard parasite growth medium is sufficient for labeling (Fig 1C). Addition of 50 μM biotin further increased protein biotinylation. Fluorescent imaging of biotinylated proteins revealed staining that colocalized with the respective apicoplast- or ER-targeted GFP–BirA* fusion proteins (Fig 1D). These results confirm that GFP–BirA* fusions are active in the P. falciparum apicoplast and ER and can be used for compartment-specific biotinylation of proteins. For large-scale identification of apicoplast proteins, biotinylated proteins from late-stage BioID–Ap and BioID–ER parasites were purified using streptavidin-conjugated beads and identified by mass spectrometry. A total of 728 unique P. falciparum proteins were detected in the apicoplast and/or ER based on presence in at least 2 of 4 biological replicates and at least 2 unique spectral matches in any single mass spectrometry run (Fig 2A and S1 Table). The abundance of each protein in apicoplast and ER samples was calculated by summing the total MS1 area of all matched peptides and normalizing to the total MS1 area of all detected P. falciparum peptides within each mass spectrometry run. To assess the ability of our data set to distinguish between true positives and negatives, we generated control lists of 96 known apicoplast and 451 signal peptide-containing nonapicoplast proteins based on published localizations and pathways (S2 Table). Consistent with an enrichment of apicoplast proteins in BioID–Ap samples, we observed a clear separation of known apicoplast and nonapicoplast proteins based on the apicoplast: ER abundance ratio (Fig 2A). Using receiver operating characteristic (ROC) curve analysis (Fig 2B), we set a threshold of apicoplast: ER abundance ratio ≥5-fold for inclusion of 187 proteins in the BioID apicoplast proteome, which maximized sensitivity while minimizing false positives (Fig 2A, dotted line; S1 Table). This data set included 50 of the 96 positive control proteins for a sensitivity of 52% (95% confidence interval (CI): 42%–62%). None of the original 451 negative controls were present above the ≥5-fold enrichment threshold, but manual inspection of this list identified 5 likely false positives not present on our initial list (S1 Table) for a positive predictive value (PPV; the estimated fraction of proteins on the list that are true positives) of 91% (95% CI: 80%–96%). To benchmark our data set against the current standard for large-scale identification of apicoplast proteins, we compared the apicoplast BioID proteome to the predicted apicoplast proteomes from 3 published bioinformatic algorithms: Predict Apicoplast-Targeted Sequences (PATS) [17], Plasmodium falciparum Apicoplast-targeted Proteins (PlasmoAP) [18], and Apicomplexan Apicoplast Proteins (ApicoAP) [19] (S3 Table). At 52% sensitivity, apicoplast BioID identified fewer known apicoplast proteins than PATS or PlasmoAP, which had sensitivities of 89% and 84%, respectively, but outperformed the 40% sensitivity of ApicoAP (Fig 2C). All 3 algorithms as well as apicoplast BioID achieved high negative predictive values (NPV), since NPV is influenced by the larger number of true negatives (known nonapicoplast proteins) than true positives (known apicoplast) from literature data (S2A Fig). We expected that the advantages of apicoplast BioID would be its improved discrimination between true and false positives (Fig 2A) and the ability to detect proteins without classical targeting presequences. Indeed, bioinformatic algorithms had poor PPVs, ranging from 19%–36% compared to the 91% PPV of BioID (Fig 2D). Even a data set consisting only of proteins predicted by all 3 algorithms achieved a PPV of just 25%. Similarly, the specificity of BioID outperformed that of the bioinformatic algorithms (S2B Fig). Consistent with the low PPVs of the bioinformatic algorithms, many proteins predicted by these algorithms are not enriched in BioID–Ap samples and are likely to be false positives (S3 Fig). Altogether, identification of apicoplast proteins using BioID provided a dramatic improvement in prediction performance over bioinformatic algorithms. To determine whether lumenally targeted GFP–BirA* exhibited any labeling preferences, we assessed proteins identified based on the presence of transmembrane domains, their suborganellar localization, and their functions. First, we determined the proportion of the 187 proteins identified by apicoplast BioID that are membrane proteins. To ensure that proteins were not classified as membrane proteins solely due to misclassification of a signal peptide as a transmembrane domain, we considered a protein to be in a membrane only if it contained at least 1 predicted transmembrane domain more than 80 amino acids from the protein’s N-terminus (as determined by annotation in the Plasmodium genome database [PlasmoDB]). These criteria suggested that 11% of identified proteins (20/187) were likely membrane proteins (Fig 3A), indicating that lumenal GFP–BirA* can label apicoplast membrane proteins. Second, apicoplast proteins may localize to 1 or multiple subcompartments defined by the 4 apicoplast membranes. It was unclear whether BirA* targeted to the lumen would label proteins in nonlumenal compartments. Based on literature descriptions, we classified the 96 known apicoplast proteins on our positive control list as either lumenal (present in lumenal space or on the innermost apicoplast membrane) or nonlumenal (all other subcompartments) and determined the proportion that were identified in our data set. Apicoplast BioID identified 56% (45/81) of the classified lumenal proteins and 33% (5/15) of the nonlumenal proteins (Fig 3B), suggesting that the GFP–BirA* bait used can label both lumenal and nonlumenal proteins but may have a preference for lumenal proteins (though this difference did not reach statistical significance). Finally, we characterized the functions of proteins identified by apicoplast BioID. We grouped positive control apicoplast proteins into functional categories and assessed the proportion of proteins identified from each functional group (Fig 3C). BioID identified a substantial proportion (67%–100%) of proteins in 4 apicoplast pathways that are essential in blood-stage and localize to the apicoplast lumen, specifically DNA replication, protein translation, isoprenoid biosynthesis, and iron–sulfur cluster biosynthesis. Conversely, BioID identified few proteins involved in heme or fatty acid biosynthesis (0% and 17%, respectively), which are lumenal pathways that are nonessential in the blood-stage and which are likely to be more abundant in other lifecycle stages [32–36]. We achieved moderate coverage of proteins involved in protein quality control (44%) and redox regulation (38%). Consistent with the reduced labeling of nonlumenal apicoplast proteins, only a small subset (29%) of proteins involved in import of nuclear-encoded apicoplast proteins were identified. Overall, apicoplast BioID identified soluble and membrane proteins of diverse functions in multiple apicoplast compartments, with higher coverage for lumenal proteins required during blood-stage infection. Apicoplast BioID provided the first experimental profile of the blood-stage apicoplast proteome but is potentially limited in sensitivity due to 1) difficulty in detecting low-abundance peptides in complex mixtures, 2) inability of the promiscuous biotin ligase to access target proteins that are buried in membranes or protein complexes, or 3) stage-specific protein expression. Currently available bioinformatic predictions of apicoplast proteins circumvent these limitations, albeit at the expense of a low PPV (Fig 2D). We reasoned that increasing the number of high-confidence apicoplast proteins used to train algorithms could improve the accuracy of a prediction algorithm while maintaining high sensitivity. In addition, inclusion of exported proteins that traffic through the ER, which are common false positives in previous prediction algorithms, would also improve our negative training set. We used our list of previously known apicoplast proteins (S2 Table) as well as newly identified apicoplast proteins from BioID (S1 Table) to construct a positive training set of 205 apicoplast proteins (S4 Table). As a negative training set, we used our previous list of 451 signal peptide-containing nonapicoplast proteins (S2 Table). For each of the 656 proteins in the training set, we calculated the frequencies of all 20 canonical amino acids in a 50-amino acid region immediately following the predicted signal peptide cleavage site. In addition, given that apicoplast proteins have a characteristic transcriptional profile in blood-stage parasites [37] and that analysis of transcriptional profiles has previously enabled identification of apicoplast proteins in the related apicomplexan Toxoplasma gondii [38], we obtained transcript levels at 8 time points during intraerythrocytic development from previous RNA-Seq data [39]. Altogether, each protein was represented by a vector of dimension 28 (20 amino acid frequencies plus 8 transcript levels). These 28-dimensional vectors were used as inputs to train a neural network with 3 hidden layers (Fig 4A and S5 Table). Six-fold cross-validation was used for training, wherein the training set was divided into 6 equal parts (folds) to train 6 separate models. Each time, 5 folds were used to train the model and 1 fold to measure the performance of the trained model. We named this model PlastNN. PlastNN recognized apicoplast proteins with a cross-validation accuracy of 96 ± 2% (mean ± standard deviation (SD) across 6 models), along with sensitivity of 95 ± 5% and PPV of 95 ± 4% (Fig 4B). This performance was higher than logistic regression on the same data set (average accuracy = 91%; S6 Table). Combining the transcriptome features and the amino acid frequencies improves performance; the same neural network architecture with amino acid frequencies alone as input resulted in a lower average accuracy of 91%, while using transcriptome data alone resulted in an average accuracy of 90% (S6 Table). Comparison of the performance of PlastNN to existing prediction algorithms indicates that PlastNN distinguishes apicoplast and nonapicoplast proteins with higher accuracy than any previous prediction method (Fig 4C). To identify new apicoplast proteins, PlastNN was used to predict the apicoplast status of 450 predicted signal peptide-containing proteins that were not in our positive or negative training sets. Since PlastNN is composed of 6 models, we designated proteins as “apicoplast” if plastid localization was predicted by ≥4 of the 6 models. PlastNN predicts 118 out of the 450 proteins to be targeted to the apicoplast (S7 Table). Combining these results with those from apicoplast BioID (S1 Table) and with experimental localization of proteins from the literature (S2 Table) yielded a compiled proteome of 346 putative nuclear-encoded apicoplast proteins (S8 Table). To determine whether candidate apicoplast proteins from this study have the potential to reveal unexplored parasite biology or are candidate antimalarial drug targets, we assessed the novelty and essentiality of the identified proteins. We found that substantial fractions of the BioID and PlastNN proteomes (49% and 71%, respectively) and 50% of the compiled apicoplast proteome represented proteins that could not be assigned to an established apicoplast pathway and therefore might be involved in novel organellar processes (Fig 5A). Furthermore, we identified orthologs of identified genes in the 150 genomes present in the Ortholog Groups of Protein Sequences database (OrthoMCL-DB) [40]: 39% of the compiled apicoplast proteome were unique to apicomplexan parasites, with 58% of these proteins found only in Plasmodium spp. (Fig 5B). Of the 61% of proteins that were conserved outside of the Apicomplexa, we note that many of these contain conserved domains or are components of well-established pathways, such as DNA replication, translation, and metabolic pathways (S8 Table). This analysis indicates that many of the newly identified proteins are significantly divergent from proteins in their metazoan hosts. Consistent with the critical role of the apicoplast in parasite biology, a genome-scale functional analysis of genes in the rodent malaria parasite P. berghei showed that numerous apicoplast proteins are essential for blood-stage survival [41]. Using this data set, we found that 77% of proteins in the compiled apicoplast proteome that had P. berghei homologs analyzed by the Plasmodium Genetic Modification project (PlasmoGEM) were important for normal blood-stage parasite growth (Fig 5C). Notably, of 49 proteins that were annotated explicitly with “unknown function” in their gene description and for which essentiality data are available, 38 are important for normal parasite growth, indicating that the high rate of essentiality for apicoplast proteins is true of both previously known and newly discovered proteins. In concordance with the PlasmoGEM data, recent genome-scale transposon mutagenesis in P. falciparum [42] identified 75% of proteins in the compiled apicoplast proteome as nonmutable (Fig 5D), suggesting essential functions in the blood stage. Overall, these data suggest that we have identified dozens of novel proteins that are likely critical for apicoplast biology. Our analyses of the apicoplast BioID and PlastNN data sets suggested that these approaches enabled accurate, large-scale identification of apicoplast proteins (Figs 2D and 4C) and included many proteins of potential biological interest due to their novelty or their essentiality in the blood stage (Fig 5). As proof of concept of the utility of these data sets, several newly identified apicoplast proteins were experimentally validated. Fortuitously, while this manuscript was in preparation, 7 new apicoplast membrane proteins in P. berghei were validated by Sayers and colleagues [43]. Of these, apicoplast BioID identified the P. falciparum homologs of 3 proteins (PF3D7_1145500/ABCB3, PF3D7_0302600/ABCB4, and PF3D7_1021300), and PlastNN identified 1 (PF3D7_0908100). In addition to these, we also selected 4 candidates from apicoplast BioID and 2 from PlastNN to validate. From the BioID list (S1 Table), we chose a rhomboid protease homolog ROM7 (PF3D7_1358300) and 3 conserved Plasmodium proteins of unknown function (PF3D7_1472800, PF3D7_0521400, and PF3D7_0721100) and generated cell lines expressing C-terminal GFP fusions from an ectopic locus in Dd2attB parasites. With the exception of ROM7, which was chosen because of the biological interest of rhomboid proteases, we focused on proteins of unknown function to begin characterizing the large number of unannotated proteins in the Plasmodium genome (see Materials and methods section for additional candidate selection criteria). To assess the apicoplast localization of each candidate, we first detected the apicoplast-dependent cleavage of each protein as a marker of its import. Most nuclear-encoded apicoplast proteins are proteolytically processed to remove N-terminal targeting sequences following successful import into the apicoplast [44,45]. This processing is abolished in parasites rendered “apicoplast-minus” by treatment with an inhibitor (actinonin) to cause apicoplast loss [16,46]. Comparison of protein molecular weight in apicoplast-intact and -minus parasites showed that ROM7, PF3D7_1472800, and PF3D7_0521400 (but not PF3D7_0721100) were cleaved in an apicoplast-dependent manner (Fig 6A). Next, we demonstrated colocalization of these 3 proteins with the apicoplast marker ACP by co-immunofluorescence analysis (co-IFA; Fig 6B). ROM7, PF3D7_1472800, and PF3D7_0521400 clearly colocalized with ACP. PF3D7_0721100 localized to few large puncta not previously described for any apicoplast protein, which partly colocalized with the apicoplast marker ACP (Fig 6B and S4 Fig) but also appeared adjacent to ACP staining (Fig 6B and S4 Fig, arrowheads). Finally, we localized the candidate–GFP fusions by live fluorescence microscopy and assessed the apicoplast dependence of their localization. ROM7-GFP, PF3D7_1472800-GFP, and PF3D7_0521400-GFP localized to branched structures characteristic of the apicoplast (S5 Fig). Upon actinonin treatment to render parasites “apicoplast-minus, ” these proteins mislocalized to diffuse puncta (S5 Fig) previously observed for known apicoplast proteins [46]. Interestingly, while in untreated live parasites PF3D7_0721100-GFP again localized to a few large bright puncta, this protein also relocalized to the typical numerous diffuse puncta seen for genuine apicoplast proteins in apicoplast-minus parasites (S5 Fig). Taken together, these data validate the apicoplast localization of ROM7, PF3D7_1472800, and PF3D7_0521400. Though transit peptide cleavage and the characteristic branched structure were not detected for PF3D7_0721100, partial colocalization with ACP and the mislocalization of PF3D7_0721100-GFP to puncta characteristic of apicoplast-minus parasites indicates that this protein may also be a true apicoplast protein. Further studies using either endogenously tagged protein or antibody raised against endogenous protein will be necessary to better characterize this localization. From the PlastNN list (S7 Table), we selected 2 proteins of unknown function, PF3D7_1349900 and PF3D7_1330100. As above, each protein was appended with a C-terminal GFP tag and expressed as a second copy in Dd2attB parasites. In agreement with apicoplast localization for each of these proteins, actinonin-mediated apicoplast loss caused loss of transit peptide processing (Fig 7A) and redistribution from a branched structure to diffuse puncta (S6 Fig). Furthermore, both proteins colocalized with the apicoplast marker ACP (Fig 7B). Altogether, we confirmed the apicoplast localization of 5 novel apicoplast proteins, with a sixth protein (PF3D7_0721100) having potential apicoplast localization. These results, combined with validation of 4 apicoplast membrane proteins predicted in our data sets by Sayers and colleagues, show that the apicoplast BioID and PlastNN data sets can successfully be used to prioritize apicoplast proteins of biological interest. Given the potential of ATP-binding cassette (ABC) proteins as drug targets, we sought to experimentally validate the essentiality of newly discovered apicoplast ABC proteins and assess their roles in metabolism or organelle biogenesis. Apicoplast BioID identified 4 ABC proteins: 3 ABCB-family proteins (ABCB3, ABCB4, and ABCB7) and an ABCF-family protein (ABCF1). We expected that these proteins might be important for apicoplast biology, as ABCB-family proteins are integral membrane proteins that typically act as small molecule transporters, and ABCF-family proteins, which do not contain transmembrane domains, are typically involved in translation regulation [47,48]. We pursued reverse genetic characterization of ABCB7 (PF3D7_1209900) and ABCF1 (PF3D7_0813700), as the essentiality of ABCB3 and ABCB4 has been previously studied [43,49]. To assess localization and function of ABCB7 and ABCF1, we modified their endogenous loci to contain a C-terminal triple hemagglutinin (HA) tag and tandem copies of a tetracycline repressor (TetR) -binding RNA aptamer in the 3′ UTR of either gene (S7 Fig) [50,51]. Co-IFA confirmed ABCF1-3xHA colocalization with the apicoplast marker ACP (Fig 8A). ABCB7-3xHA localized to elongated structures that may be indicative of an intracellular organelle but rarely colocalized with ACP, indicating that it has a primarily nonapicoplast localization and is likely a false positive from the BioID data set (S8A Fig). Taking advantage of the TetR-binding aptamers in the 3′ UTR of ABCF1, we determined the essentiality and knockdown phenotype of this protein. In the presence of anhydrotetracycline (ATc), a repressor comprised of TetR fused to the P. falciparum development of zygote inhibited (DOZI) protein cannot bind the aptamer and ABCF1 is expressed. Upon removal of ATc, binding of the TetR–DOZI repressor binding blocks gene expression [50,51]. Knockdown of ABCF1 caused robust parasite growth inhibition (Fig 8B and 8C). Growth inhibition of ABCF1-deficient parasites was reversed in the presence of isopentenyl pyrophosphate (IPP) (Fig 8C), which bypasses the need for a functional apicoplast [46], indicating that ABCF1 has an essential apicoplast function. Essential apicoplast functions can be placed into 2 broad categories: those involved in organelle biogenesis and those involved solely in IPP production. Disruption of proteins required for organelle biogenesis causes apicoplast loss, while disruption of proteins involved in IPP production does not [16,46,52]. We determined whether knockdown of ABCF1 caused apicoplast loss by assessing 1) absence of the apicoplast genome, 2) loss of transit peptide processing of nuclear-encoded apicoplast proteins, and 3) relocalization of apicoplast proteins to puncta. Indeed, the apicoplast: nuclear genome ratio drastically decreased in ABCF1 knockdown parasites beginning 1 cycle after knockdown (Fig 8D), and western blot showed that the apicoplast protein caseinolytic protease subunit P (ClpP) was not processed in ABCF1 knockdown parasites (Fig 8E). Furthermore, IFA of the apicoplast marker ACP confirmed redistribution from an intact plastid to diffuse cytosolic puncta (Fig 8F). In contrast to ABCF1, a similar knockdown of ABCB7 caused no observable growth defect after 4 growth cycles despite significant reduction in protein levels (S8B and S8C Fig). Together, these results show that ABCF1 is a novel and essential apicoplast protein with a previously unknown function in organelle biogenesis. Since the discovery of the apicoplast, identification of its proteome has been a pressing priority. We report the first large-scale proteomic analysis of the apicoplast in blood-stage malaria parasites, which identified 187 candidate proteins with 52% sensitivity and 91% PPV. A number of groups have also profiled parasite-specific membrane compartments using proximity biotinylation but observed contamination with proteins in or trafficking through the ER, preventing accurate identification of these proteomes without substantial manual curation and validation [23,24,26–29]. This background labeling is expected since proteins traffic through the ER en route to several parasite-specific compartments, including the parasitophorous vacuole, host cytoplasm, food vacuole, and invasion organelles. The high specificity of our apicoplast BioID proteome depended on 1) the use of a control cell line expressing ER-localized GFP–BirA* to detect enrichment of apicoplast proteins from background ER labeling and 2) strong positive and negative controls to set an accurate threshold. We suspect a similar strategy to detect nonspecific ER background may also improve the specificity of proteomic data sets for other parasite-specific, endomembrane-derived compartments. Leveraging our successful proteomic analysis, we used these empirical data as an updated training set to also improve computational predictions of apicoplast proteins. PlastNN identified an additional 118 proteins with 95% sensitivity and 95% PPV. Although 2 previous prediction algorithms, PATS and ApicoAP, also applied machine learning to the problem of transit peptide prediction, we reasoned that their low accuracy arose from the small training sets used (ApicoAP) and the use of cytosolic as well as endomembrane proteins in the negative training set (PATS). By using an expanded positive training set based on proteomic data and limiting our training sets to only signal peptide-containing proteins, we developed an algorithm with higher sensitivity than BioID and higher accuracy than previous apicoplast protein prediction models. Inevitably, some false positives from the BioID data set would have been used for neural network training and cross-validation. While this may slightly influence the PPV of the PlastNN list, we expect that the substantially larger fraction of true positives in the training set mitigated the effects of any false positives. Importantly, as more apicoplast and nonapicoplast proteins in P. falciparum parasites are experimentally validated, updated training sets can be used to retrain PlastNN. Moreover, PlastNN suggests testable hypotheses regarding the contribution of sequence-based and temporal regulation to protein trafficking in the ER. Overall, we have compiled a high-confidence apicoplast proteome of 346 proteins that are rich in novel and essential functions (Fig 5). This proteome likely represents a majority of soluble apicoplast proteins, since 1) our bait for proximity biotinylation targeted to the lumen and 2) most soluble proteins use canonical targeting sequences that can be predicted. An important next step will be to expand the coverage of apicoplast membrane proteins, which more often traffic via distinctive routes [53,54]. Performing proximity biotinylation with additional bait proteins may identify such atypical apicoplast proteins. In the current study, our bait was an inert fluorescent protein targeted to the apicoplast lumen to minimize potential toxicity of the construct. The success of this apicoplast GFP bait gives us confidence to attempt more challenging baits, including proteins localized to suborganellar membrane compartments or components of the protein import machinery. Performing apicoplast BioID in liver and mosquito stages may also define apicoplast functions in these stages. This compiled proteome represents a substantial improvement upon previous bioinformatics predictions of apicoplast proteins and provides a strong foundation for further refinement. In analogy to progress on the mammalian mitochondrial proteome, which over the course of decades has been expanded and refined by a combination of proteomic, computational, and candidate-based approaches [55,56], we expect that future proteomic, computational, and candidate-based approaches to identify apicoplast proteins will be critical for ultimately determining a comprehensive apicoplast proteome. Organellar proteomes are valuable hypothesis-generating tools. Already several candidates of biological interest based on their biochemical function annotations were validated. We demonstrated an unexpected role for the ATP-binding cassette protein PfABCF1 in apicoplast biogenesis. ABCF proteins are understudied compared to other ABC-containing proteins but tend to have roles in translation regulation [47]. An E. coli homolog, EttA, regulates translation initiation in response to cellular ATP levels [57,58], and mammalian and yeast ABCF1 homologs also interact with ribosomes and regulate translation [59–62]. By analogy, PfABCF1 may regulate the prokaryotic translation machinery in the apicoplast, although the mechanistic basis for the severe defect in parasite replication upon loss of PfABCF1 is unclear. We also validated PfROM7 as an apicoplast-localized rhomboid protease. Rhomboid proteases are a diverse family of intramembrane serine proteases found in all domains of life. In the Apicomplexa, rhomboids have been studied primarily for their roles in processing adhesins on the parasite cell surface [63], although the functions of most apicomplexan rhomboids are still unknown. Little is known about ROM7 other than that it appears to be absent from coccidians and was refractory to deletion in P. berghei [64,65]. However, a rhomboid protease was recently identified as a component of symbiont-derived ER-associated degradation (ERAD) -like machinery (SELMA) that transports proteins across a novel secondary plastid membrane in diatoms [66], indicating that ROM7 may similarly play a role in apicoplast protein import in Plasmodium parasites. Neither PfABCF1 nor PfROM7 had known roles in the apicoplast prior to their identification in this study, underscoring the utility of unbiased approaches to identify new organellar proteins. Moreover, the apicoplast is one of few models for complex plastids that permits functional analysis of identified proteins to investigate the molecular mechanisms underpinning serial endosymbiosis. A summary of all candidate proteins validated in this study is shown in S9 Table. A recent study aimed at identifying apicoplast membrane transporters highlights the difficulty in identifying novel apicoplast functions in the absence of a high-confidence proteome [43]. Taking advantage of the tractable genetics in murine Plasmodium species, Sayers and colleagues screened 27 candidates in P. berghei for essentiality and apicoplast localization. Following >50 transfections, 3 essential and 4 nonessential apicoplast membrane proteins were identified. One newly identified essential apicoplast membrane protein was then validated to be required for apicoplast biogenesis in P. falciparum. In contrast, even though our study was not optimized to identify membrane proteins, the combination of BioID and PlastNN identified 2 known apicoplast transporters, 4 of the new apicoplast membrane protein homologs, and 56 additional proteins predicted to contain at least 1 transmembrane domain. A focused screen of higher quality candidates in P. falciparum is likely to be more rapid and yield the most relevant biology. Our high-confidence apicoplast proteome will streamline these labor-intensive screens, focusing on strong candidates for downstream biological function elucidation. As methods for analyzing gene function in P. falciparum parasites continue to improve, this resource will become increasingly valuable for characterizing unknown organellar pathways. Human erythrocytes were purchased from the Stanford Blood Center (Stanford, California) to support in vitro P. falciparum cultures. Because erythrocytes were collected from anonymized donors with no access to identifying information, IRB approval was not required. All consent to participate in research was collected by the Stanford Blood Center. P. falciparum Dd2attB [31] (MRA-843) were obtained from MR4. NF54Cas9+T7 Polymerase parasites [67] were a gift from Jacquin Niles. Parasites were grown in human erythrocytes (2% hematocrit) obtained from the Stanford Blood Center in Roswell Park Memorial Institute (RPMI) 1640 media (Gibco) supplemented with 0. 25% AlbuMAX II (Gibco), 2 g/L sodium bicarbonate, 0. 1 mM hypoxanthine (Sigma), 25 mM HEPES, pH 7. 4 (Sigma), and 50 μg/L gentamicin (Gold Biotechnology) at 37 °C, 5% O2, and 5% CO2. Oligonucleotides were purchased from the Stanford Protein and Nucleic Acid facility or IDT. gBlocks were ordered from IDT. Molecular cloning was performed using In-Fusion cloning (Clontech) or Gibson Assembly (NEB). Primer and gBlock sequences are available in S10 Table. To generate the plasmid pRL2-ACPL-GFP for targeting transgenes to the apicoplast, the first 55 amino acids from ACP were PCR amplified with primers MB015 and MB016 and were inserted in front of the GFP in the pRL2 backbone [68] via the AvrII/BsiWI sites. To generate pRL2-ACPL-GFP-BirA* for targeting a GFP–BirA* fusion to the apicoplast, GFP was amplified from pLN-ENR-GFP using primers MB087 and MB088, and BirA* was amplified from pcDNA3. 1 mycBioID (Addgene 35700) [20] using primers MB089 and MB090. These inserts were simultaneously cloned into BsiWI/AflII-digested pRL2-ACPL-GFP to generate pRL2-ACPL-GFP-BirA*. To generate pRL2-SP-GFP-BirA*-SDEL for targeting GFP–BirA* to the ER, SP-GFP-BirA*-SDEL was PCR amplified from pRL2-ACPL-GFP-BirA* using primers MB093 and MB094 and was cloned into AvrII/AflII-digested pRL2-ACPL-GFP. For GFP-tagging to confirm localization of proteins identified by apicoplast BioID or PlastNN, full-length genes were amplified from parasite cDNA with primers as described in S10 Table and were cloned into the AvrII/BsiWI sites of pRL2-ACPL-GFP. For CRISPR-Cas9–based editing of endogenous ABCB7 and ABCF1 loci, sgRNAs were designed using the eukaryotic CRISPR guide RNA/DNA design tool (http: //grna. ctegd. uga. edu/). To generate a linear plasmid for CRISPR-Cas9–based editing, left homology regions were amplified with primers MB256 and MB257 (ABCB7) or MB260 and MB261 (ABCF1), and right homology regions were amplified with MB258 and MB259 (ABCB7) or MB262 and MB263 (ABCF1). For each gene, a gBlock containing the recoded coding sequence C-terminal of the CRISPR cut site and a triple HA tag was synthesized with appropriate overhangs for Gibson Assembly. This fragment and the appropriate left homology region were simultaneously cloned into the FseI/ApaI sites of the linear plasmid pSN054-V5. Next, the appropriate right homology region and a gBlock containing the sgRNA expression cassette were simultaneously cloned into the AscI/I-SceI sites of the resultant vectors to generate the plasmids pSN054-ABCB7-TetR-DOZI and pSN054-ABCF1-TetR-DOZI. Transfections were carried out using variations on the spontaneous uptake method [69,70]. In the first variation, 100 μg of each plasmid was ethanol precipitated and resuspended in 30 μL sterile TE buffer and was added to 150 μL packed RBCs resuspended to a final volume of 400 μL in cytomix. The mixture was transferred to a 0. 2 cm electroporation cuvette (Bio-Rad) and was electroporated at 310 V, 950 μF, infinity resistance in a Gene Pulser Xcell electroporation system (Bio-Rad) before allowing parasites to invade. Drug selection was initiated 3 days after transfection. Alternatively, 50 μg of each plasmid was ethanol precipitated and resuspended in 0. 2 cm electroporation cuvettes in 100 μL TE buffer, 100 μL RPMI 1640 containing 10 mM HEPES-NaOH, pH 7. 4, and 200 μL packed uninfected RBCs. RBCs were pulsed with 8 square wave pulses of 365 V x 1 ms separated by 0. 1 s. RBCs were allowed to reseal for 1 hour in a 37 °C water bath before allowing parasites to invade. Drug selection was initiated 4 days after transfection. All transfectants were selected with 2. 5 μg/mL Blasticidin S (Research Products International). Additionally, BioID–ER parasites were selected with 125 μg/mL G418 sulfate (Corning), and ABCB7 and ABCF1 TetR-DOZI parasites were grown in the presence of 500 nM ATc. Transfections for generating BioID constructs (Fig 1) and expression of GFP-tagged candidates (Figs 6 and 7) were performed in the Dd2attB background. Transfections for CRISPR editing were performed with the NF54Cas9+T7 Polymerase background. Clonal lines of ABCF1 and ABCB7 knockdown parasites were obtained by limiting dilution. Correct modification of transfectant genomes was confirmed by PCR. Briefly, 200 μL of 2% hematocrit culture was pelleted and resuspended in water, and 2 μL of the resulting lysate was used as template for PCR with Phusion polymerase (NEB). PCR targets and their corresponding primer pairs are as follows: integrated attL site, p1 + p2; integrated attR site, MW001 + MW003; unintegrated attB site, MW004 + MW003; ABCB7 unintegrated left homology region (LHR), MB269 + MB270; ABCB7 integrated LHR, MB269 + MB255; ABCB7 unintegrated right homology region (RHR), MB281 + MB278; ABCB7 integrated RHR, MB276 + MB278; ABCF1 unintegrated LHR, MB271 + MB272; ABCF1 integrated LHR, MB271 + MB255; ABCF1 unintegrated RHR, MB282 + MB283; and ABCF1 integrated RHR, MB276 + MB283. To label parasites for analysis by streptavidin blot, fixed imaging, or mass spectrometry, cultures of majority ring-stage parasites were treated with 50 μM biotin or with a DMSO vehicle-only control. Cultures were harvested for analysis 16 hours later as majority trophozoites and schizonts. To generate apicoplast-minus parasites, ring-stage cultures were treated with 10 μM actinonin (Sigma) and 200 μM IPP (Isoprenoids, LLC) and cultured for 3 days before analysis. Parasites were separated from RBCs by lysis in 0. 1% saponin and were washed in PBS. Parasite pellets were resuspended in PBS containing 1X NuPAGE LDS sample buffer with 50 mM DTT and were boiled at 95 °C for 10 min before separation on NuPAGE or Bolt Bis-Tris gels and transfer to nitrocellulose. Membranes were blocked in 0. 1% Hammarsten casein (Affymetrix) in 0. 2X PBS with 0. 01% sodium azide. Antibody incubations were performed in a 1: 1 mixture of blocking buffer and Tris-Buffered Saline with Tween 20 (TBST; 10 mM Tris, pH 8. 0,150 mM NaCl, 0. 25 mM EDTA, 0. 05% Tween 20). Blots were incubated with primary antibody for either 1 hour at room temperature or at 4 °C overnight at the following dilutions: 1: 20,000 mouse-α-GFP JL-8 (Clontech 632381); 1: 20,000 rabbit-α-Plasmodium aldolase (Abcam ab207494); 1: 1,000 rat-α-HA 3F10 (Sigma 11867423001); 1: 4,000 rabbit-α-PfClpP [71]. Blots were washed once in TBST and were incubated for 1 hour at room temperature in a 1: 10,000 dilution of the appropriate secondary antibody: IRDye 800CW donkey-α-rabbit; IRDye 680LT goat-α-mouse; IRDye 680LT goat-α-rat (LI-COR Biosciences). For detection of biotinylated proteins, blots were incubated with 1: 1,000 IRDye 680RD streptavidin for 1 hour at room temperature. Blots were washed 3 times in TBST and once in PBS before imaging on a LI-COR Odyssey imager. For live imaging, parasites were settled onto glass-bottomed microwell dishes (MatTek P35G-1. 5-14-C) or Lab-Tek II chambered coverglass (ThermoFisher 155409) in PBS containing 0. 4% glucose and 2 μg/mL Hoechst 33342 stain (ThermoFisher H3570). For fixed imaging of biotinylated proteins in cells, biotin-labeled parasites were processed as in Tonkin and colleagues [72], with modifications. Briefly, parasites were washed in PBS and were fixed in 4% paraformaldehyde (Electron Microscopy Science 15710) and 0. 015% glutaraldehyde (Electron Microscopy Sciences 16019) in PBS for 30 min. Cells were washed once in PBS, resuspended in PBS, and allowed to settle onto poly-L-lysine-coated coverslips (Corning) for 60 min. Coverslips were then washed once with PBS, permeabilized in 0. 1% Triton X-100 in PBS for 10 min and washed twice more in PBS. Cells were treated with 0. 1 mg/mL sodium borohydride in PBS for 10 min, washed once in PBS, and blocked in 3% BSA in PBS. To visualize biotin-labeled proteins, coverslips were incubated with 1: 1,000 AlexaFluor 546-conjugated streptavidin (ThermoFisher S11225) for 1 hour followed by 3 washes in PBS. No labeling of GFP was necessary, as these fixation conditions preserve intrinsic GFP fluorescence [72]. Coverslips were mounted onto slides with ProLong Gold antifade reagent with DAPI (ThermoFisher) and were sealed with nail polish prior to imaging. For immunofluorescence analysis, parasites were processed as above except that fixation was performed with 4% paraformaldehyde and 0. 0075% glutaraldehyde in PBS for 20 min, and blocking was performed with 5% BSA in PBS. Following blocking, primary antibodies were used in 5% BSA in PBS at the following concentrations: 1: 500 rabbit-α-PfACP [73]; 1: 1,000 rabbit-α-PfBip1: 1,000 (a gift from Sebastian Mikolajczak and Stefan Kappe); 1: 500 mouse-α-GFP JL-8 (Clontech 632381); 1: 100 rat-α-HA 3F10 (Sigma 11867423001). Coverslips were washed 3 times in PBS, incubated with goat-α-rat 488 (ThermoFisher A-11006), goat-α-mouse 488 (ThermoFisher A11029), or donkey-α-rabbit 568 (ThermoFisher A10042) secondary antibodies at 1: 3,000, and washed 3 times in PBS prior to mounting as above. Live and fixed cells were imaged with 100X, 1. 4 NA or 100X, 1. 35 NA objectives on an Olympus IX70 microscope with a DeltaVision system (Applied Precision) controlled with SoftWorx version 4. 1. 0 and equipped with a CoolSnap-HQ CCD camera (Photometrics). Images were taken in a single z-plane, with the exception of those presented in Figs 1D, 8A and S8A, which were captured as a series of z-stacks separated by 0. 2-μm intervals, deconvolved, and displayed as maximum intensity projections. Brightness and contrast were adjusted in Fiji (ImageJ) for display purposes. Image capture and processing conditions were identical for micrographs of the same cell line when multiple examples are displayed (S4 Fig) or when comparing untreated to actinonin-treated cells (S5 and S6 Figs). Biotin-labeled parasites were harvested by centrifugation and were released from the host RBC by treatment with 0. 1% saponin/PBS. Parasites were washed twice more with 0. 1% saponin/PBS, followed by PBS and were either used immediately for analysis or were stored at −80 °C. Parasite pellets were resuspended in RIPA buffer [50 mM Tris-HCl, pH 7. 4,150 mM NaCl, 0. 1% SDS, 0. 5% sodium deoxycholate, 1% Triton X-100,1 mM EDTA] containing a protease inhibitor cocktail (Pierce) and were lysed on ice for 30 min with occasional pipetting. Insoluble debris was removed by centrifugation at 16,000 xg for 15 min at 4 °C. Biotinylated proteins were captured using High Capacity Streptavidin Agarose beads (Pierce) for 2 hours at room temperature. Beads were then washed 3 times with RIPA buffer, 3 times with SDS wash buffer [50 mM Tris-HCl, pH 7. 4,150 mM NaCl, 2% SDS], 6 times with urea wash buffer [50 mM Tris-HCl, pH 7. 4,150 mM NaCl, 8 M urea], and 3 times with 100 mM ammonium bicarbonate. Proteins were reduced with 5 mM DTT for 60 min at 37 °C, followed by treatment with 14 mM iodoacetamide (Pierce) at room temperature for 45 min. Beads were washed once with 100 mM ammonium bicarbonate and were digested with 10 μg/mL trypsin (Promega) at 37 °C overnight. The following day, samples were digested with an additional 5 μg/mL trypsin for 3–4 hours. Digested peptides were separated from beads by addition of either 35% or 50% final concentration acetonitrile, and peptides were dried on a SpeedVac prior to desalting with C18 stage tips. Desalted peptides were resuspended in 0. 1% formic acid and analyzed by online capillary nanoLC-MS/MS. Samples were separated on an in-house–made 20-cm reversed phase column (100-μm inner diameter, packed with ReproSil-Pur C18-AQ 3. 0 μm resin [Dr. Maisch GmbH]) equipped with a laser-pulled nanoelectrospray emitter tip. Peptides were eluted at a flow rate of 400 nL/min using a 2-step linear gradient including 2%–25% buffer B in 70 min and 25%–40% B in 20 min (buffer A, 0. 2% formic acid and 5% DMSO in water; buffer B, 0. 2% formic acid and 5% DMSO in acetonitrile) in an Eksigent ekspert nanoLC-425 system (AB Sciex). Peptides were then analyzed using an LTQ Orbitrap Elite mass spectrometer (Thermo Scientific). Data acquisition was executed in data-dependent mode, with full MS scans acquired in the Orbitrap mass analyzer with a resolution of 60,000 and m/z scan range of 340–1,600. The top 20 most abundant ions with intensity threshold above 500 counts and charge states 2 and above were selected for fragmentation using collision-induced dissociation (CID) with isolation window of 2 m/z, normalized collision energy of 35%, activation Q of 0. 25, and activation time of 5 ms. The CID fragments were analyzed in the ion trap with rapid scan rate. In additional runs, the top 10 most abundant ions with intensity threshold above 500 counts and charge states 2 and above were selected for fragmentation using higher-energy collisional dissociation (HCD) with isolation window of 2 m/z, normalized collision energy of 35%, and activation time of 25 ms. The HCD fragments were analyzed in the Orbitrap with a resolution of 15,000. Dynamic exclusion was enabled with repeat count of 1 and exclusion duration of 30 s. The AGC target was set to 1,000,000; 50,000; and 5,000 for full FTMS scans, FTMSn scans, and ITMSn scans, respectively. The maximum injection time was set to 250 ms, 250 ms, and 100 ms for full FTMS scans, FTMSn scans and ITMSn scans, respectively. The resulting spectra were searched against a “target-decoy” sequence database [74] consisting of the PlasmoDB protein database (release 32, released April 19,2017), the Uniprot human database (released February 2,2015), and the corresponding reversed sequences using the SEQUEST algorithm (version 28, revision 12). The parent mass tolerance was set to 50 ppm and the fragment mass tolerance to 0. 6 Da for CID scans, 0. 02 Da for HCD scans. Enzyme specificity was set to trypsin. Oxidation of methionines was set as variable modification, and carbamidomethylation of cysteines was set as static modification. Peptide identifications were filtered to a 1% peptide false discovery rate using a linear discriminator analysis [75]. Precursor peak areas were calculated for protein quantification. To generate updated lists of PATS-predicted apicoplast proteins, nuclear-encoded P. falciparum 3D7 proteins (excluding pseudogenes) from PlasmoDB version 28 (released March 30,2016) were used to check for existence of a putative bipartite apicoplast targeting presequence using the artificial neural network predictor PATS [17]. Updated PlasmoAP-predicted apicoplast proteins were identified using the PlasmoDB version 32 proteome (released April 19,2017) by first checking for the presequence of a predicted signal peptide using the neural network version of SignalP version 3. 0 [76] and were considered positive if they had a D-score above the default cutoff. The SignalP C-score was used to predict the signal peptide cleavage position, and the remaining portion of the protein was inspected for presence of a putative apicoplast transit peptide using the rules described for PlasmoAP [18], implemented in a Perl script. P. falciparum proteins predicted to localize to the apicoplast by ApicoAP were accessed from the original paper [19]. Genes predicted to encode pseudogenes were excluded. A positive control list of 96 high-confidence apicoplast proteins (S2 Table) was generated based on either (1) published localization of that protein in Plasmodium parasites or T. gondii or (2) presence of that protein in either the isoprenoid biosynthesis or fatty acid biosynthesis/utilization pathways. To generate a negative control list of potential false positives, nuclear-encoded proteins (excluding pseudogenes) predicted to contain a signal peptide were identified as above, and 451 of these proteins were designated as negative controls based on GO terms, annotations, and the published literature. To generate the positive training set for PlastNN, we took the combined list of previously known apicoplast proteins (S2 Table) and apicoplast proteins identified by BioID (S1 Table) and removed proteins that (1) were likely false positives based on our negative control list (S2 Table) or published localization data, (2) were likely targeted to the apicoplast without the canonical bipartite N-terminal leader sequence, or (3) did not contain a predicted signal peptide based on the SignalP 3. 0 D-score. This yielded a final positive training set of 205 proteins (S4 Table). The negative training set was the previously generated list of known nonapicoplast proteins (S2 Table). The test set for PlastNN consisted of 450 proteins predicted to have a signal peptide by the SignalP 3. 0 D-score that were not in the positive or negative training sets. For each protein in our training and test sets, we took the 50 amino acids immediately after the end of the predicted signal peptide (according to the SignalP 3. 0 C-score) and calculated the frequency of each of the 20 amino acids in this sequence. The length of 50 amino acids was chosen empirically by trying lengths from 20–100; highest accuracy was obtained using 50. Scaled FPKM values at 8 time points during intraerythrocytic development were obtained from published RNA-Seq [39]. By combining the amino acid frequencies with the 8 transcriptome values, we represented each protein in our training and test sets by a feature vector of length 28. To train the model, the 205 positive and 451 negative training examples were combined and randomly shuffled. The training set was divided into 6 equal folds, each containing 109 or 110 examples. We trained models using 6-fold cross-validation; that is, we trained 6 separate models with the same architecture, each using 5 of the 6 folds for training and then using the 1 remaining fold as a cross-validation set to evaluate performance. Accuracy, sensitivity, specificity, NPV, and PPV are calculated on this cross-validation set. The final reported values of accuracy, sensitivity, specificity, NPV, and PPV are the average and standard deviation over all 6 models. When predicting on the test set, the final predictions are generated by a majority vote of all 6 models. Neural networks were trained using the RMSProp optimization algorithm with a learning rate of 0. 0001. Tensorflow version 1. 4. 1 was used to build and train the neural network. Logistic regression on the same data set was carried out using the caret package (version 6. 0–77) in R version 3. 3. 3. The BioID apicoplast proteome and the predicted proteomes from PATS, PlasmoAP, ApicoAP, and PlastNN were analyzed according to the following formulae: Accuracy= (TP+TN) / (TP+FP+TN+FN) Sensitivity=TP/ (TP+FN) Specificity=TN/ (TN+FP) NegativePredictiveValue (NPV) =TN/ (TN+FN) PositivePredictiveValue (PPV) =TP/ (TP+FP) Abbreviations are as follows: TP, true positive; TN, true negative; FP, false positive; FN, false negative. Because none of the 451 negative control proteins from the original list (S2 Table) were identified in our 187-protein BioID proteome, we manually inspected the BioID list, identified 5 likely false positives, and added these to the negative control list for the purposes of analyses presented in Fig 2 and S2 Fig. Proteins in the apicoplast proteome were manually categorized for having a potentially novel function based on PlasmoDB version 33 (released June 30,2017) gene product annotations. Gene products with annotations that could clearly assign a given protein to an established cellular pathway were labeled as “Known Pathway; ” gene products with a descriptive annotation that did not clearly suggest a cellular pathway were labeled as “Annotated Gene Product, Unknown Function; ” and gene products that explicitly contained the words “unknown function” were labeled as “Unknown Function”. To analyze the conservation of candidate apicoplast proteins identified by apicoplast BioID, OrthoMCL-DB ortholog group IDs were obtained from PlasmoDB. Based on OrthoMCL-DB version 5 (released July 23,2015), each ortholog group was then categorized as being present only in Plasmodium spp. , only in Apicomplexa, or present in at least 1 organism outside of the Apicomplexa. Genome-scale essentiality data for P. berghei or P. falciparum genes were accessed from the original manuscripts [41,42]. To facilitate molecular cloning, proteins identified by BioID or PlastNN were candidates for GFP tagging only if their corresponding gene sizes were less than 2 kb. With the exception of ROM7, which was selected based on the biological interest of rhomboid proteases, we focused on localizing conserved Plasmodium genes of unknown function due to interest in functional characterization of the Plasmodium genome. PF3D7_1472800, PF3D7_0521400 and PF3D7_0721100 (all from the BioID list) were chosen due to their diverse apicoplast: ER enrichment rankings in the BioID list (S1 Table) (PF3D7_1472800 ranked number 52/187 and was identified exclusively in BioID-Ap samples; PF3D7_0521400 ranked number 131/187 and was found in both samples but was enriched nearly 400-fold in BioID-Ap samples; and PF3D7_0721100 ranked 184/187 and was enriched in BioID-Ap samples only slightly above our 5-fold cutoff). From the PlastNN list, PF3D7_1349900 and PF3D7_1330100 were selected solely based on being proteins of unknown function with small gene sizes. Because of the small sample sizes of proteins selected for GFP-tagging and the nonrandom nature of selecting candidates, we note that the results of our experimental validation should not be extrapolated to be representative of the PPVs of the BioID and PlastNN data sets as a whole. Sorbitol-synchronized ABCB7 and ABCF1 TetR-DOZI parasites were washed multiple times to remove residual ATc and were returned to culture medium containing 500 nM ATc, 200 μM IPP (Isoprenoids, LLC), or no supplements. Samples for growth assays, DNA isolation, or western blotting were harvested every other day when the majority of parasites were trophozoites and schizonts. For growth assays, parasites were fixed in 1% paraformaldehyde in PBS and were stored at 4 °C until completion of the time course. Samples were then stained with 50 nM YOYO-1 and parasitemia was analyzed on a BD Accuri C6 flow cytometer. Samples for DNA isolation and western blotting were treated with 0. 1% saponin in PBS to release parasites from the erythrocyte, washed in PBS, and stored at −80 °C until analysis. Total parasite DNA was isolated from time course samples using the DNeasy Blood & Tissue Kit (Qiagen). Quantitative PCR was performed using Power SYBR Green PCR Master Mix (Thermo Fisher), with primers CHT1 F and CHT1 R targeting the nuclear gene chitinase or TufA F and TufA R targeting the apicoplast gene elongation factor Tu (0. 15 μM final concentration each primer) [46]. Quantitative PCR was performed on an Applied Biosystems 7900HT Real-Time PCR System with the following thermocycling conditions: initial denaturation 95 °C/10 min, 35 cycles of 95 °C/1 min, 56 °C/1 min, 65 °C/1 min, final extension 65 °C/10 min. Relative quantification of each target was performed using the ΔΔCt method. 95% confidence intervals were determined using the GraphPad QuickCalc for confidence interval of a proportion via the modified Wald method (https: //www. graphpad. com/quickcalcs/confInterval1/). Two-way ANOVAs were performed in GraphPad Prism version 7. 04.

Plasmodium parasites, which cause malaria, and related pathogens belonging to the phylum Apicomplexa contain a relict chloroplast called the apicoplast. During evolution, the apicoplast lost photosynthetic functions but retained critical metabolic pathways that are required for host cell infection. Because of its importance for parasite survival and its unusual evolutionary origin, the apicoplast is of major interest for its unique biology and its potential to yield new antimalarial drug targets. However, an accurate inventory of proteins that localize to the apicoplast is lacking, hindering our ability to study the most novel and biologically interesting aspects of apicoplast biology. To address this limitation, we combine proximity biotinylation-based proteomics and a new machine learning algorithm to identify apicoplast proteins in an accurate and unbiased manner. We identify 346 candidate apicoplast proteins with high confidence and find that many of these proteins are new and expected to be important for parasite survival. This proteomic resource will therefore be a valuable tool for elucidating the unique cell biology of the apicoplast and for identifying new antimalarial drug targets.

Abstract Introduction Results Discussion Materials and methods

b vitamins parasite groups methods and resources chemical compounds biotin plasmodium endoplasmic reticulum cell processes green fluorescent protein organic compounds parasitic protozoans parasitology membrane proteins apicomplexa luminescent proteins protozoans cellular structures and organelles malarial parasites proteins chemistry vitamins cell membranes biochemistry eukaryota organic chemistry post-translational modification proteomes cell biology secretory pathway integral membrane proteins biology and life sciences physical sciences signal peptides organisms

2018

Integrative proteomics and bioinformatic prediction enable a high-confidence apicoplast proteome in malaria parasites

Efforts are underway to scale-up the facial cleanliness and environmental improvement (F&E) components of the World Health Organization’s SAFE strategy for elimination of trachoma as a public health problem. Improving understanding of the F&E intervention landscape could inform advancements prior to scale-up, and lead to more effective and sustained behavior change. We systematically searched for relevant grey literature published from January 1965 through August 2016. Publications were eligible for review if they described interventions addressing F&E in the context of trachoma elimination programs. Subsequent to screening, we mapped attributes of F&E interventions. We then employed three behavior change frameworks to synthesize mapped data and identify potential intervention gaps. We identified 27 documents meeting inclusion criteria. With the exception of some recent programming, F&E interventions have largely focused on intermediate and distal antecedents of behavior change. Evidence from our analyses suggests many interventions are not designed to address documented determinants of improved F&E practices. No reviewed documents endorsed inclusion of intervention components related to behavioral maintenance or resilience–factors critical for sustaining improved behaviors. If left unaddressed, identified gaps in intervention content may continue to challenge uptake and sustainability of improved F&E behaviors. Stakeholders designing and implementing trachoma elimination programs should review their F&E intervention content and delivery approaches with an eye toward improvement, including better alignment with established behavior change theories and empirical evidence. Implementation should move beyond information dissemination, and appropriately employ a variety of behavior change techniques to address more proximal influencers of change. Behavior change reflects a process of modification and transformation. This process is complicated by an assortment of change antecedents, which are precursors that need to be addressed before behavior change and maintenance can occur. Antecedents include factors such as perceived risk of infection or disease, attitudes and normative beliefs regarding improved practices, perceived and actual abilities to perform improved practices, self-regulation, and intentions to initiate and maintain the adoption of improved practices [13,14]. Along with antecedents, other behavioral determinants (e. g. , barriers and facilitators) mediate the adoption and translation of improved behaviors into action. In other words, in addition to potentially impeding initial behavioral uptake, these factors may also prevent the actual manifestation of improved behaviors [15]. For instance, one can possess improved behaviors, but due to unfavorable environmental conditions (e. g. , contextual or technological barriers), be unable to translate those behaviors into the execution of improved practices. The complexity of behavior change is often reflected by the difficulty in achieving it; it requires more than simply raising awareness that change might be beneficial [16]. Although numerous behavior change intervention techniques exist, some are more successful at eliciting change than others. According to the Theory of Triadic Influence, various tiers of influence exist, and account for factors that have direct and indirect effects on behavior and practice [17]. See S1 Appendix. Supplemental material and S1 Fig for further information regarding this theory. Some intervention techniques operate at a more distal level of influence, and are therefore less likely to directly facilitate change when implemented in isolation. Others operate at a proximal level of influence, and are more likely to bring about change [17]. Critical evaluation of intervention attributes is necessary for understanding whether F&E content and delivery attend to key behavior change antecedents at proximal levels of influence. Throughout this paper, we distinguish between F&E content and delivery. Ensuring both intervention attributes are behavior-centered is critical to eliciting and maintaining improved behaviors and practices. Data from cross-sectional investigations suggest associations between improved F&E-related practices; more favorable water, sanitation, and hygiene (WASH) conditions; and lower prevalence of active trachoma [18–22]. Yet, few data from rigorous, experimental studies demonstrate the impact of specific F&E interventions on active trachoma and C. trachomatis infection [11,12,23]. It is crucial to consider the nature and fidelity of the interventions that generated effect estimates when interpreting evidence on the possible effectiveness of F&E in general. It is also important to understand the intervention landscape and potential gaps therein, to inform program improvement prior to scale-up. Syntheses of the F&E literature have yet to fully describe approaches the trachoma community has taken to design and implement F&E interventions, and whether and to what extent they address important behavioral change and maintenance factors. Previous reviews have only examined data from peer-reviewed literature [11,12,21–23]. No known review has examined information from the grey literature, where many implementing organizations publish information regarding F&E content, delivery, and related metrics. Existing reviews have not collated information on frameworks, theories, and metrics the trachoma community has used to assess behavioral change elicited by F&E interventions. There is utility in thoroughly documenting details related to F&E content and delivery so as to: Collating, synthesizing, and broadly disseminating information regarding F&E implementation from the grey literature can add to the existing understanding of the intervention landscape. More nuanced information regarding intervention content and delivery outlined in these materials is often not included in peer-review journal articles. These details can be used to identify programmatic gaps. The purpose of this review was to synthesize information regarding F&E intervention attributes to more fully describe the intervention landscape, so as to inform decision-making regarding policies, program implementation and evaluation, and possibly future research. We sought to determine which behavior change factors are addressed through F&E-related interventions endorsed in the grey literature. Our research questions were: We employed the PICOT framework [24], an extension of the traditional PICO framework [25], to construct our search strategy and guide decision-making regarding eligibility, data extraction, and analyses. PICOT is designed to improve precision of search results through articulation of details related to Participants, Interventions, Comparators, Outcome measures, and Types of eligible literature/study designs. We performed a mapping exercise [32] to broadly survey the nature of F&E content and delivery, and reported results regarding lessons learnt and promising practices. The purpose of this mapping phase was to extract data that could be synthesized during the subsequent analysis phase. One reviewer employed a semi-structured instrument to rapidly, yet systematically survey the content of all documents included in the review. A second reviewer employed the same instrument to complete a validity check of 10% of the reviewed literature. When inconsistencies arose, the two discussed the results, and came to a consensus. In addition to information related to document identification and publication, we mapped five categories of characteristics via our semi-structured data extraction instrument (S1 Appendix). Due to anticipated heterogeneity in methodology and reporting styles, we employed various, complementary qualitative methods for data synthesis. Based on the type of information provided in the F&E grey literature, we determined that thematic and narrative synthesis approaches were appropriate methods for the purposes of this review [33]. Our synthesis activities focused on categorizing intervention activities, cataloguing documented behavioral determinants of improved F&E adoption, and determining the types of behavioral antecedents and intervention techniques addressed through interventions endorsed in the grey literature. We used these synthesis activities, described in further detail below, to answer our review questions. Each activity resulted in a unique analytical output. Cumulatively, the syntheses generated a descriptive overview of F&E content and delivery endorsed in the grey literature, and pinpointed gaps in the intervention landscape. Table 1 lists the documents included, and provides a summary of publication information as well as the means by which we identified the document, and the type of content contained therein. The grey literature was produced by a range of trachoma and NTD networks/coalitions, non-governmental development organizations, multilateral and technical organizations, research institutions, and government partners. Several documents were produced by a combination of different stakeholder groups. The vast majority of included literature (96%, n = 26) was identified via keyword searches on target websites. Amongst those, nearly half (42%, n = 11) of the documents were also detected via another mode of identification. Only one document was not first identified through our web search. This was the Handbook on Community-Led Total Sanitation (CLTS) [37], which we confirmed was identified via our web search, but was initially screened out, as it met a pre-defined exclusion criterion (i. e. , it represented a general WASH-related resource that did not specifically touch on inclusion of CLTS in trachoma elimination programs). We later included this resource in the review because several practitioners identified it as a key resource for trachoma programming during the literature solicitation process. Over half of the documents included (55%, n = 15) represent technical guidance and resource documents. Six (22%) are implementer reports, and the remaining literature represented a combination of policy/advocacy pieces and other documents. Table 2 outlines types of interventions, content and delivery approaches, and level of implementation. Twenty-three (85%) of the documents contained F&E-specific content, while three (11%) more generally described content and delivery along the WASH-NTD nexus. The authors are aware that several funded trachoma elimination programs are using participatory hygiene and sanitation transformation (PHAST) and children’s hygiene and sanitation training (CHAST). However, no related resources were identified through our searches or literature solicitation process. As such, no other general WASH resources were deemed appropriate for inclusion in the review. Several of the documents included addressed the creation of enabling environments via cross-sectoral partnerships and collaboration, national and sub-national polices and guidelines, and community-level contracts and by-laws. Other documents focused largely on local, community, school, and individual-level intervention activities and implementation approaches. The literature covered a mix of intervention types, from F&E advocacy, design, planning, and implementation to monitoring and evaluation, collation, and synthesis of resulting information. Until recently, major limitations of F&E programming included a lack of behavior-centered F&E content and delivery, and more broadly, insufficient integration of behavior change theory and evidence into the design and implementation of related interventions. Our findings indicate that many F&E-related interventions do not align well with learning from the behavioral sciences [13,64–67]. This may explain, in part, why such interventions do not yield sustained improvements in F&E behaviors and practices. Our review demonstrated that a high proportion of documents endorse F&E interventions that focus on knowledge generation via health and hygiene information dissemination and other techniques that address more intermediate and distal behavior change factors. Evidence suggests that information dissemination does not motivate sustained behavior change in the absence of activities that address more proximal change factors, such as improved action capacity (i. e. , performance skills), and enhanced attitudes, perceptions, normative beliefs, and self-regulation [64,68–71]. Our findings reflect a belief from parts of the trachoma community that promotion of eye health via education will serve to achieve behavioral and environmental changes [58]. As indicated in Fig 2, F&E intervention activities endorsed in the literature also overlook factors that are more influential to uptake and sustained adoption of improved practices. The relative deficiency of F&E intervention activities that address more proximal antecedent factors of sustained behavioral change may limit the impact of these interventions. In addition to several other behavior change frameworks, the RANAS approach suggests that various intervention techniques are necessary for the habitual formation of new behaviors and practices [13,66,72]. Wood & Neal [67] suggest that a two-pronged approach of breaking unhealthy habits while simultaneously establishing healthy ones is optimal for facilitating enduring behavior change. Such frameworks include intervention components that move beyond information dissemination to address factors related to behavioral maintenance and resilience. The majority of identified barriers to improved F&E adoption documented in included literature reflect proximal psychosocial factors. However, F&E-related interventions have largely focused on information dissemination and water and sanitation infrastructure development: activities addressing more distal psychosocial, technological, and contextual factors. While the trachoma community has identified many barriers and facilitators of improved F&E adoption (see Figs 2 and 3), determinants are often either not addressed, or are confronted via inappropriate intervention techniques. A high proportion of documented F&E behavioral determinants reflected barriers, which suggests there is a need to further identify and leverage facilitators of improved behaviors and practices. Enhancements in formative assessments of behavioral determinants might include the performance of motive analyses [73], which aim to identify meaningful facilitators of change. Intervention content and delivery approaches that support behavioral maintenance and resilience are absent from the F&E programming we reviewed. None of the documents included in this review explicitly or implicitly touch on behavioral maintenance or resilience, key factors facilitating sustained behavior change (Fig 3 and S3 Table) [13,67]. Maintenance of new behaviors is an important step in habituation. Similarly, resilience is central to sustained adoption of improved F&E practices because it focuses on one’s perceived ability to recover from setbacks and continue the improved practices after disruptions. Recovery from relapse and setbacks is actually part of the behavioral change process, and when appropriately addressed, can serve to facilitate sustained change. Interventions that do not address these dimensions of ability factors often result in only temporary change that is susceptible to relapse [70,74]. This finding underscores the limited scope of F&E interventions as well as a lack of emphasis on behavior-centered content and delivery, the outputs of which can more readily facilitate sustained change. As such, it should not be surprising to observe behavioral slippage (i. e. , regression back to unimproved practices) in areas where initial F&E improvements were observed [75,76]. Current promising practices from the behavioral sciences and WASH sectors call for sanitation and hygiene programming that: Similar approaches can be employed by the trachoma community to design and implement F&E interventions. Implications of our findings are far-reaching, in that they touch on both the effectiveness and sustainability of F&E interventions. If F&E interventions do not address proximal behavior change factors in ways that bring about and sustain adoption, they may be ineffective. Therefore, it is important for implementers and researchers to assess the composition of intervention content and approaches to delivery. All relevant stakeholders should ensure that proximal behavior change factors are included in their intervention designs, and contextually adapted, as appropriate. Focus should be placed on an intervention mapping and design approach that leverages formative assessments, employs a combination of intervention techniques at different levels of influence, and incorporates intervention components that address behavioral maintenance and resilience. Our findings corroborate evidence from prior investigation that community-based F&E interventions have largely focused on hardware and resource provision plus information dissemination [77–80]. However, a few works highlight interventions that incorporate approaches or capitalize on non-health motives that address more proximal influencers of improved F&E practices [81,82]. With regard to the influence of health education interventions, evidence provided by Simms and colleagues [82] suggests that no health education is necessary for the uptake of E-related behavioral outcomes (i. e. , coverage, operation and maintenance, and utilization of household latrines), particularly when the intervention affects contextually important benefit beliefs and non-health motives, such as pride and embarrassment. Well-designed and executed interventions provide evidentiary support that participatory strategies can effectively change facewashing practices [81]. Such interventions reinforce improved behaviors and practices, facilitate communal commitment, and incorporate barrier identification and action capacity activities at the community, group, and individual levels. Other interventions incorporate more proximally influencing social constructs (such as social solidarity) in training guides to harness mutual responsibility and trust to improve utilization of personal and common pool resources (e. g. , water) [83]. While some evidence suggests well-designed interventions may be effective when implemented over a short time horizon [81], behavior change and maintenance of improved practices involves a protracted process of confronting underlying behavioral antecedents and determinants. Addressing underlying factors of behavior is not typically accomplished in a sustainable manner over a three to eight month period. This idea is supported within the trachoma community, as some note that intervention activities carried out in the months surrounding mass treatment with azithromycin do not address normative factors (e. g. , social disapproval) in such a timeframe [84]. Our work has some limitations. First, the scope of our review was relatively narrow, as we restricted our review to documents posted on key trachoma websites, published in a grey literature database, or submitted to our team by ICTC and WHO-CCT member organizations. While our electronic searches were systematic, they may not have captured all published F&E-related grey literature. We note that although our search dates back to 1965, no documents produced earlier than 1990 were identified. This is perhaps an artifact of website content and document archives. However, the entities engaged represent the majority of stakeholders working on F&E in trachoma elimination programming. Second, in contrast to a systematic review of the peer-reviewed literature, in which reviewers ideally identify and extract data from all eligible papers to reduce risk of biasing the results, the veracity of our review depends on the range of concepts presented, and whether they would be in agreement with those presented in non-reviewed grey literature [34]. As such, we aimed for ‘conceptual saturation’ as opposed to exhaustive discovery, and limited data extraction to intervention approaches endorsed in the literature. However, there is uncertainty regarding the validity of these results beyond the confines of the reviewed documents. Third, the frameworks that we used, particularly RANAS and Theory of Triadic Influence, are specific and reductionist in the sense that they focus on psychosocial behavioral factors, and do not fully incorporate contextual and environmental dimensions. A review grounded in other frameworks may yield slightly different results. Finally, we do not have a comprehensive synthesis of empirical evidence that the distal and proximal influencers of behavior that we postulate here are specifically relevant in the context of trachoma. While acknowledging that some progress has been made in the context of certain funded F&E-related programming, based on the results of this review, we recommend the following:

Trachoma is the world’s leading infectious cause of blindness. In light of a 2020 global target, the international trachoma community is intensifying its efforts to scale-up facial cleanliness and environmental improvement (F&E) components of the WHO-endorsed SAFE strategy for the elimination of trachoma as a public health problem. This is therefore an opportune moment to examine intervention content and delivery, and to consider their optimization. F&E interventions seek to change behaviors over the long term. Behavioral change and maintenance are predicated on behavioral antecedents, or precursors of change, and other determinants that may fall along different levels of influence. From the behavioral sciences literature, we know interventions are likely to be more effective if they target contextually-specific antecedents and causal determinants of behavioral change and maintenance. Evidence also suggests it is important to incorporate intervention techniques that are designed to address proximal change factors. However, F&E programming has largely focused on information dissemination via health and hygiene education, provision of water and sanitation facilities, and other interventions designed to address intermediate and distal behavioral factors. Intervention mapping and incorporation of techniques designed to address more proximal factors of behavioral change and maintenance have been deficient from F&E programming. Our review provides information regarding gaps in F&E intervention content and delivery that relevant stakeholders can use to understand why some interventions have not worked as intended, and strengthen F&E-related programming for trachoma elimination.

Abstract Introduction Methods Results Discussion

medicine and health sciences face tropical diseases health care preventive medicine bacterial diseases health services administration and management eye diseases global health neglected tropical diseases sanitation public and occupational health infectious diseases behavior hygiene head environmental health anatomy ophthalmology biology and life sciences trachoma

2018

Interventions to maximize facial cleanliness and achieve environmental improvement for trachoma elimination: A review of the grey literature

OVATE gene was first identified as a key regulator of fruit shape in tomato. OVATE family proteins (OFPs) are characterized as plant-specific transcription factors and conserved in Arabidopsis, tomato, and rice. Roles of OFPs involved in plant development and growth are largely unknown. Brassinosteroids (BRs) are a class of steroid hormones involved in diverse biological functions. OsGKS2 plays a critical role in BR signaling by phosphorylating downstream components such as OsBZR1 and DLT. Here we report in rice that OsOFP8 plays a positive role in BR signaling pathway. BL treatment induced the expression of OsOFP8 and led to enhanced accumulation of OsOFP8 protein. The gain-of-function mutant Osofp8 and OsOFP8 overexpression lines showed enhanced lamina joint inclination, whereas OsOFP8 RNAi transgenic lines showed more upright leaf phenotype, which suggest that OsOFP8 is involved in BR responses. Further analyses indicated that OsGSK2 interacts with and phosphorylates OsOFP8. BRZ treatment resulted in the cytoplasmic distribution of OsOFP8, and bikinin treatment reduced the cytoplasmic accumulation of OsOFP8. Phosphorylation of OsOFP8 by OsGSK2 is needed for its nuclear export. The phospphorylated OsOFP8 shuttles to the cytoplasm and is targeted for proteasomal degradation. These results indicate that OsOFP8 is a substrate of OsGSK2 and the function of OsOFP8 in plant growth and development is at least partly through the BR signaling pathway. OVATE gene was first cloned in tomato and demonstrated to encode a hydrophilic protein with putative bipartite nuclear localization signal, and a C-terminal domain of approximate 70 amino acids which is designated as the OVATE domain and conserved in tomato, Arabidopsis, and rice [1–3]. As a plant-specific transcription factor family, OVATE family proteins (OFPs) control multiple aspects of plant growth and development [2,4–6]. Sequence analysis showed that there are 18 OVATE genes in the Arabidopsis genome [2,5–7]. AtOFP1 was shown to function as an active transcriptional repressor to suppress cell elongation [2]. Arabidopsis plants overexpressing AtOFP1 exhibited abnormal morphological phenotypes because AtOFP1 suppresses the expression of AtGA20ox1, the key gibberellin (GA) biosynthesis enzyme gene [2]. AtOFP4 was reported to interact with KNAT7 (Knotted1-Like Homeodomain Protein 7) in planta, this interaction enhances KNAT7’s transcriptional repression activity and regulates the secondary cell wall formation [5]. AtOFP5 is required for normal embryo sac development in Arabidopsis by suppressing the activity of BELL-KNOX TALE complexes [7]. In rice, there are 31 putative OFPs identified in the genome [3]. Although increasing evidence in Arabidopsis demonstrates that AtOFPs participate in multiple aspects of plant growth and development by regulating the transcriptional levels of target genes, little is known about the function and action mode of OsOFPs in rice. Brassinosteroids (BRs) are a class of plant-specific steroidal hormones that are structurally related to animal and inset steroids. As a group of growth-promoting steroid hormones, BRs play pivotal roles in promoting cell expansion and division, regulating senescence, male fertility, fruit ripening, and modulating plant responses to various environmental signals [8]. Extensive studies in Arabidopsis have identified a nearly complete BR signaling pathway starting with BRI1 (Brassinosteroid insensitive 1) as the cell membrane receptor which perceives and binds to BR [9], then initiating a phosphorylation-mediated cascade involving BSK1 (BR-signaling kinase 1), BSU1 (BRI1 suppressor 1), BIN2 (BR-insensitive 2), and PP2A (Protein phosphatase 2A), and finally transducing the extracellular signal to the transcription factor BZR1 (Brassinazole resistant 1) [10–13]. In this signaling pathway, BIN2 acts as a negative regulator that interacts with and phosphorylates BZR1 to inhibit its function, thereby blocking BR signaling [14,15]. BIN2 can also phosphorylate Auxin Response Factor 2 (ARF2), resulting in the inhibition of the DNA binding activity of ARF2, thus promoting downstream auxin responses [16]. In addition, BR regulates stomatal development through BIN2-mediated phosphorylation of YDA, a mitogen-activated protein kinase kinasekinase (MAPKKK) [17,18]. These studies indicated that BIN2 acts as a multi-tasker in diverse cellular signal transduction pathways [19]. In rice, OsGSK2 is the counterpart of Arabidopsis BIN2, and acts as a negative regulator to mediate BR signaling [20]. The phosphorylated form of OsBZR1 was increased in OsGSK2 overexpression plants, and decreased in OsGSK2 RNAi plants, suggesting that OsGSK2 mediates BR signaling through OsBZR1 [20]. In addition to OsBZR1, two other proteins in rice have been found as substrates for OsGSK2. DLT (Dwarf and Low-Tillering), encoding a GRAS-family protein, is a direct target of OsGSK2 and functions similarly to OsBZR1 [20–22]. In contrast to DLT, OsLIC (LEAF and TILLER ANGLE INCREASED CONTROLLER), another substrate of OsGSK2, acts as an antagonistic transcription factor of OsBZR1 and plays a negative role in BR signaling [23]. These studies suggested the vital role of OsGKS2 in BR signaling. We report here the characteristics of OsOFP8, a member of OVATE family protein genes in rice. The gain-of-function mutant Osofp8 and OsOFP8 overexpression transgenic lines showed enhanced lamina joint bending, whereas OsOFP8 RNAi lines showed upright leaves and tight architecture. Further analysis revealed that OsGSK2 interacts with and phosphorylates OsOFP8, and phosphorylated OsOFP8 shuttles to the cytoplasm and is targeted for the proteasomal degradation. In the experiment of generation of T-DNA mutants, we identified a T-DNA insertion mutant showing lamina joint bending phenotype at the maturation stage, especially for the flag leaves (Fig 1A). To identify the gene in which T-DNA was inserted, we performed TAIL-PCR analysis [24], DNA sequence comparison showed that the T-DNA was inserted into the 3’ region of LOC_Os01g64430 at the site of 27 bp downstream of the stop codon (Fig 1B), and there is no other annotated genes in the 5. 5 kb region downstream of LOC_Os01g64430 (http: //rice. plantbiology. msu. edu). LOC_Os01g64430 encodes an OVATE family protein (hereafter designated as OsOFP8). To investigate the effect of T-DNA insertion on the expression of OsOFP8, we carried out quantitative real-time RT-PCR (qRT-PCR), which showed that T-DNA insertion causes an increase in the expression of OsOFP8 (Fig 1C), thus T-DNA insertion generates a gain-of-function mutant Osofp8. To further investigate the function of OsOFP8 gene, we generated both OsOFP8 overexpression and RNA-interference (RNAi) transgenic lines. OsOFP8 overexpression lines OE7 and OE10 phenocopied the Osofp8 mutant, showing increased lamina joint bending phenotype, by contrast, RNAi transgenic lines RNAi2 and RNAi4 showed upright leaves and tight architecture (Fig 1D). Furthermore, we examined the leaf inclination degrees of the three uppermost leaves. Compared to the wild-type (WT) plants, OsOFP8 overexpression plants showed largely increased leaf inclination for all three uppermost leaves, and the flag leaf showed the largest inclination angle, by contrast, RNAi transgenic plants showed reduced leaf angles (Fig 1E). Expression analysis by qRT-PCR showed that OsOFP8 expression was increased in overexpression lines and reduced in RNAi lines (Fig 1F). Tissue-specific expression of OsOFP8 was examined by qRT-PCR analysis, showing that OsOFP8 was expressed in various tissues (S1A Fig). Native promoter of OsOFP8 was fused to GUS gene to gain expression profile of OsOFP8, GUS activity was detected in different organs including roots, stem, leaf, lamina joint, inflorescence, and seeds (S1B–S1J Fig). Gain-of-function Osofp8 mutant and OsOFP8 overexpression lines showed obvious leaf lamina joint bending phenotype, which is a classic phenotype of BR response. We hypothesized that OsOFP8 may be involved in BR signaling pathway. To this end, we tested the sensitivity of wild-type plants, OsOFP8 overexpression and RNAi lines to 24-epibrassinolide (BL) in lamina joint bending experiments. BL treatment caused a dose-dependent lamina joint inclination in both WT and OE10 with the latter plants were more sensitive to BL treatment, whereas RNAi4 plants were insensitive to BL treatment (Fig 2A and 2B). To further confirm this phenomenon, lamina joint assay was performed in the dark-grown seedlings by the excised leaf segment method [22]. We observed more severe inclination of leaf angle in OE10 than in WT plants, whereas RNAi4 plants did not show too much change in leaf angle when treated with BL (Fig 2C). To investigate the effect of OsOFP8 on the expression of BR-related gene expression, we analyzed the expression levels of genes involved in BR biosynthesis and signaling. Rice D2/CYP90D2 (OsD2) gene is involved in the last step of brassinosteroid biosynthesis [25], and OsDWARF4/CYP90B2 functions in the rate-limiting step of brassinosteroid biosynthesis [26]. Overexpression of OsOFP8 suppressed the expression level of OsD2 but had little effect on the expression of OsDWARF4, whereas knockdown of OsOFP8 expression by RNAi led to significantly increased expression of OsD2 and OsDWARF4 (Fig 2D). OsOFP8 had little effect on the expression of OsGSK2, a negative regulator gene in BR signaling, by contrast, the expression levels of OsBZR1, a positive controller in BR signaling, were increased in OsOFP8 overexpression lines and significantly decreased in OsOFP8 RNAi lines (Fig 2D). We also measured the expression of OsOFP8 in BR signaling. BL treatment induced the mRNA transcript of OsOFP8 (Fig 2E). To further investigate the effect of BR signaling on OsOFP8 expression, we first treated the OsOFP8-YFP transfected protoplast cells with BRZ for 12 hr, and then BL was applied after the removal of BRZ. The protein level of OsOFP8 was induced by BL treatment (Fig 2F). The different responses of OsOFP8 overexpression and RNAi plants to BL treatment prompted us to further explore the possible functions of OsOFP8 in BR signaling pathway. We performed yeast two-hybrid (Y2H) analysis to test the interactions between OsOFP8 and the components of BR signaling. OsGSK2, OsBZR1, and DLT are three important components in BR signaling pathway, we investigated the possible interactions between OsOFP8 and these three components. Y2H analysis showed that OsOFP8 interacted with OsGSK2, but not with OsBZR1 and DLT (Fig 3A). The interaction between OsOFP8 and OsGSK2 required the full length of OsOFP8 because neither the N-terminal nor the OVATE domain-containing C-terminal of OsOFP8 interacts with OsGSK2 (S2A and S2B Fig). The interaction between OsOFP8 and OsGSK2 was also confirmed by the coimmunoprecipitation (Co-IP) assay. The HA-tagged OsGSK2 protein (3HA-OsGSK2) and YFP-tagged OsOFP8 protein (OsOFP8-YFP) were co-transfected into Arabidopsis protoplast cells, the fusion protein 3HA-OsGSK2 can be immunoprecipitated by OsOFP8-YFP fusion but not by YFP protein (Fig 3B). Furthermore, BiFC assay was used to confirm the interaction between OsOFP8 and OsGSK2 (Fig 3C). OsGSK2 protein phosphorylates proteins such as OsBZR1 and DLT with which it interacts [20]. In this scenario, we were interested to know whether OsGSK2 phosphorylates OsOFP8. To this end, we applied the biotin-pendant Zn2+-phos-tag and horseradish peroxidase-conjugated streptavidin method [27] to investigate the phosphorylation status of OsOFP8 when OsOFP8-YFP was expressed alone or co-expressed with 3HA-OsGSK2 in Arabidopsis protoplast cells. When OsOFP8-YFP was expressed alone in the protoplast cells, we only detected a faint band showing phosphorylated OsOFP8 (Fig 3D, asterisk), which is probably caused by the endogenous BIN2 of the protoplast cells. When OsOFP8-YFP and 3HA-OsGSK2 were co-expressed in protoplast cells, a stronger band was detected, indicating the increased phosphorylation status of OsOFP8 (Fig 3D, black arrow). The lower band showed the phosphorylated OsGSK2 (Fig 3D), which can be used as an internal reference for the system. This assay showed that OsGSK2 is able to phosphorylate OsOFP8. GSK3 kinases recognize a conserved sequence for phosphorylation (S/TXXXS/T, where S/T is serine or threonine and X is any amino acid), for example, BZR1 protein has 25 serine/threonine residues potentially phosphorylated by BIN2 [28]. Examination of OsOFP8 protein sequence revealed that there are 25 GSK3 kinase phosphorylation sites at the N-terminal region of OsOFP8 (S2C Fig), further supporting the notion that OsGSK2 phosphorylates OsOFP8. To investigate the subcellular localization of OsOFP8, we made various OsOFP8 fusions in which YFP protein was fused to either the N-terminus or the C-terminus of OsOFP8, the fused OsOFP8 protein was transiently expressed in Arabidopsis protoplast cells to monitor the localization of OsOFP8. This assay showed that OsOFP8 localizes to the nucleus (Figs 4A, S3A and S3B). NLS-mCherry gene is expressed in the nucleus, the merged image of NLS-mCherry and OsOFP8-YFP showed that majority of OsOFP8 exists in the nucleolus, which was indicated by the strong fluorescence intensity in the round structure of the nucleus and seen in the differential interference contrast (DIC) images (Fig 4B). In addition, the fluorescent signals of OsOFP8-YFP and NLS-mCherry were well overlapped, further supporting the nuclear localization of OsOFP8 (Fig 4B). To test whether the interaction with OsGSK2 alters subcellular localization of OsOFP8, we investigated the localization of OsOFP8 in the presence of OsGSK2. Co-expression with OsGSK2 clearly caused the cytoplasmic distribution of OsOFP8 (Fig 4C), and western blotting was carried out to show the presence of OsGSK2 (S3C Fig). A closer view of individual cells showed both the nuclear and cytoplasmic localization of OsOFP8 (Fig 4D, upper panel), and analysis of the fluorescent signal peaks showed that only one peak of OsOFP8-YFP was overlapped with that of NLS-mCherry, and three other peaks of OsOFP8-YFP were detected in the cytosol (Fig 4D, lower panel). These results indicate that interaction with OsGSK2 leads to the nuclear export of OsOFP8 to the cytoplasm. To analyze the effect of BR signaling on the subcellular distribution of OsOFP8, we treated the OsOFP8-YFP transfected protoplast cells with BL. After BL treatment for two hours, the protein level of OsOFP8 was increased in the nucleus, and OsOFP8 was not detected in the cytoplasm (Fig 4E). When treated with BRZ, a BR biosynthetic inhibitor brassinazole, the protein level of OsOFP8 was detected both in the nucleus and in the cytoplasm, indicating that BRZ treatment altered the subcellular localization of OsOFP8 (Fig 4E). OsGSK2 interacts with and phosphorylates OsOFP8, the phosphorylation status of OsOFP8 may be required for its nuclear export. To this end, we treated the protoplast cells co-transfected by OsOFP8-YFP and 3HA-OsGSK2 with bikinin which inhibits the activity of BIN2 by acting as an ATP competitor. Western blotting showed that the protein level of OsOFP8 in the cytoplasm was largely reduced when the cells were treated with bikinin for two hours, indicating that phosphorylation of OsOFP8 by OsGSK2 is needed for its cytoplasmic localization (Fig 4F). Because presence of OsGSK2 induced the nuclear export of OsOFP8 (Fig 4D and 4F), we further analyzed the state of OsOFP8 and the phosphorylated OsOFP8 in the nucleus and cytoplasm, respectively. Nuclear and cytoplasmic fractions were prepared both from the protoplasts co-transfected with OsOFP8 and NLS-mCherry and with OsOFP8 and OsGSK2. In the absence of OsGSK2, OsOFP8 was only detected in the nucleus, which is consistent with the previous findings (Fig 4A and 4E). In the presence of OsGSK2, OsOFP8 was detected both in the nucleus and in the cytoplasm, but the band in the cytoplasm was much weaker than that in the nucleus (Fig 4G). When the co-transfected protoplasts were treated with MG132, a proteasome inhibitor, the intensity of the band in the cytoplasm was increased (Fig 4G), which suggests that the phosphorylated OsOFP8 is cytoplasm-localized and targeted for proteasomal degradation. The OVATE gene was first identified as a major QTL controlling pear-shaped fruit in tomato [1,29], and later on, studies in Arabidopsis show that the OVATE family proteins control multiple aspects of plant growth and development [2,4–6]. The rice genome contains more number of OFPs than the Arabidopsis genome, but very few studies on OFP function in rice have been reported. We studied the function of OsOFP8 in rice, and demonstrated that OsOFP8 is involved in BR signaling. Elevated expression of OsOFP8 in rice leads to increased lamina joint bending phenotype and BR hypersensitivity (Fig 2). In Arabidopsis, Atofp1-1D is a dominant, gain-of-function mutant, which has a T-DNA inserted at the 4332 bp downstream of the stop codon of the AtOFP1 gene and shows increased expression of AtOFP1 [2]. Osofp8 is also a gain-of-function mutant with T-DNA inserted in the 3’ region of OsOFP8 gene and shows increased expression of OsOFP8 gene (Fig 1). This coincidence of gain-of-function mutants generated by T-DNA insertion may imply a common mechanism regarding the expression regulation of OFP genes. In Arabidopsis, Atofp1-1D mutant shows reduced lengths in all aerial organs including hypocotyls, rosette leaf, inflorescence stem and floral organs. By contrast, the rice Osofp8 mutant displays increased lamina joint inclination but does not show reduced length of aerial organs (Fig 1). Furthermore, AtOFP1 is involved in GA signaling by repressing the expression of GA20ox1, a gene encoding a key enzyme in GA biosynthesis, but OsOFP8 is involved in BR signaling pathway and shows normal response to GA treatment (S4 Fig), indicating the functional diversity of these two genes in Arabidopsis and rice. AtOFP1 is expressed in roots, shoots, vasculatures, trichomes, and in mature flowers. Similarly, OsOFP8 is expressed in roots, shoots, and inflorescences (S1 Fig). Sequence analysis of AtOFP1 and OsOFP8 proteins also showed different functional domains. OVATE domain is the common feature for all OFP proteins, besides this, AtOFP1 contains an LXLXL motif (where L is leucine and X for any amino acid) in its OVATE region, which is not present in OsOFP8. The LXLXL motif has been shown to play an important role in repression of gene expression [2], although it only contributes marginal repression function to the AtOFP1. OsGSK2 is an ortholog of Arabidopsis BIN2 gene and plays negative roles in BR signaling [20]. The expression levels of BR-biosynthesis related genes such as OsD2 and OsDWARF4 were increased in OsGSK2 overexpression line and decreased in OsGSK2 RNAi line [20]. By contrast, OsBZR1 and DLT plays positive roles in BR signaling, and OsD2 and OsDWARF4 expression levels were induced in OsBZR1 RNAi plants and dlt mutant [20,22]. On the contrary, the expression levels of OsD2 and OsDWARF4 were increased in OsOFP8 RNAi lines, and the expression of OsD2 was reduced in OsOFP8 overexpression lines although in these lines OsDWARF4 did not show much change in expression when compared to its expression in wild-type plants. BL treatment (1 μM) increases OsOFP8 mRNA transcript and the protein amount at the translational level (Fig 2E and 2F), at this concentration the mRNA level of OsBZR1 and DLT was decreased [22,23], suggesting OsOFP8 behaves differently to OsBZR1 and DLT. OsGSK2 interacts with and phosphorylates the nuclear protein DLT [20], but we do not know whether the interaction with OsGSK2 causes subcellular distribution of DLT. OsGSK2 interacts with and phosphorylates OsOFP8, phosphorylation of OsOFP8 by OsGSK2 is required for its cytoplasmic localization. The phosphorylated OsOFP8 is exported from the nucleus and targeted for the proteasomal degradation in the cytoplasm. This phenomenon resembles the interaction between BIN2 and BZR1 [23,30]. However, without BL treatment, BZR1 is located mainly in the cytoplasm [30], and BIN2 is distributed both in the nucleus and cytosol, as well as at the plasma membrane [31], but OsOFP8 is a nuclear protein. BR signaling converts phosphorylated BZR1 proteins to the dephosphorylated state [30], and BIN2 protein is rapidly depleted after 30 min treatment with 1 μM BL [32], whereas OsOFP8 protein level is increased under this treatment (Fig 2F). In rice, binding of 14-3-3 proteins to the phosthorylated OsBZR1 inhibits OsBZR1 function at least in part by reducing its nuclear localization [33]. Phosphorylated OsOFP8 may also adopt this mechanism to regulate its function. Indeed, there is a putative 14-3-3 motif in OsOFP8 protein (S3C Fig), providing a possibility that OsOFP8 may interact with 14-3-3 proteins to retain itself in the cytoplasm for degradation. Further studies are required to test the interaction between 14-3-3 and OsOFP8 and the possibility of cytoplasmic retention of phosphorylated OsOFP8 by 14-3-3 binding. Suppressing OsBZR1 expression by RNAi leads to dwarf phenotype and reduced lamina joint bending [33], however, reducing expression level of OsOFP8 by RNAi shows reduced lamina joint bending phenotype without dwarfism (Fig 1), suggesting OsOFP8 may have other biological functions in addition to participating in BR signaling. The wild-type rice (Oryza sativa L.) plants Zhonghua 11 (japonica cv. ZH11) and OsOFP8 transgenic plants were grown in the experimental field at South China Botanical Garden in Guangzhou during the rice growing season. The angles between the leaf blades and the culms were measured with a protractor. The Osofp8 mutant was identified from T-DNA transformation. Genomic DNA of the Osofp8 mutant was used as the template to amplify the flanking regions of the T-DNA insertion by high-efficiency thermal asymmetric interlaced PCR [24]. The primers are listed in S1 Table. For promoter analysis, a 1983 bp promoter sequence upstream of the translation start codon of OsOFP8 was amplified by PCR. The PCR product was digested with EcoRI and NcoI, and inserted into the pCAMBIA1391z vector to generate the PromoterOsOFP8: GUS construct. Ten independent transgenic lines were obtained and showed β-glucuronidase (GUS) activity. To overexpress OsOFP8, the full-length cDNA of OsOFP8 was PCR-amplified and inserted into the binary vector pCAMBIA1301-35S. The OsOFP8 RNAi lines were generated by RNA interference, using a 280 bp fragment of the OsOFP8 coding region. The fragment was inserted into an intermediate vector as positive and inverted directions, and then the whole cassette was cut out and inserted into the binary vector pCAMBIA1301-35S. The resulting constructs of overexpression and RNAi were introduced into Agrobacterium tumefaciens strain EHA105, respectively, and then transformed rice ZH11. The empty vectors were also transformed into ZH11 as controls. Full-length OsOFP8 and OsGSK2 cDNAs were inserted into pBI221-YFP and pBI221-3HA vectors, respectively, to generate OsOFP8-YFP and 3HA-OsGSK2 constructs. For BiFC assay, OsGSK2 and OsOFP8 coding sequence fragments were cloned into pSPYNE-35S and pSPYCE-35S vectors, respectively. The resulting constructs were co-transfected into protoplast cells, the transfected cells were incubated in dark for 12 h, and the fluorescence of YFP was observed. Total RNAs were isolated with Trizol reagent (Invitrogen) according to the manufacturer’s instructions. Total RNAs were pre-treated with DNase I, and first-strand cDNA was synthesized from 2 μg of total RNAs using oligo (dT) 18 as primers (Promega, http: //cn. promega. com/). The first-strand cDNA product was used as template in a 20 μL PCR reaction. For quantitative RT-PCR, SYBR Green I was added to the reaction system and run on a Roche real-time PCR detection system according to the manufacturer’s instructions. The melting curve was acquired at the end. The transcript data were calculated by Roche’s Software, and were normalized using OsActin 1 as an internal control; the relative expression level was calculated by 2-ΔΔCt. Each experiment was performed with three replicates. The primers are listed in S1 Table. GUS staining was performed according to the method as described [34]. Different tissues of the PromoterOsOFP8: GUS transgenic plants were incubated in a solution containing 50 mM NaPO4 buffer pH7. 0,5 mM K3Fe (CN) 6,5 mM K4Fe (CN) 6,0. 1% Triton X-100 and 1 mM X-Gluc at 37°C overnight. Images were taken under the stereomicroscope (Leica M165c). The lamina joint assay by the micro-drop method was performed as described previously [25]. A drop of ethanol (1 μL) containing 0,10,100 or 1000 ng of 24-epiBL, respectively, was spotted onto the top of lamina of the seedlings which were germinated for 2 days and grown for 3 days at 30°C. Images were taken after 3 days of incubation with 24-epiBL, and the angles of lamina joint bending were measured. The lamina joint assay using excised leaf segments was performed as described previously [35]. Synchronous seeds after 2 days germination were selected and grown in the dark for 8 days at 30°C. The entire segments comprising 1 cm of the second leaf blade, the lamina joint and 1 cm of the leaf sheath were floated on distilled water for 24 h and then incubated in 2. 5 mM maleic acid potassium solution containing 1 μM 24-epiBL for 48 h in the dark. Lamina joint angles were measured, this experiment was repeated three times with similar results. For transient expression assays, typically, 4×104 mesophyll protoplasts were isolated from 4-week-old Arabidopsis seedlings. Isolation of protoplasts and PEG-mediated transfection were as described [36]. For transient expression analysis of OsOFP8-YFP, 10 μg of the plasmid DNA were used to transfect the protoplast cells. The transfected cells were treated with 10 μM BRZ for 12 h, and then treated with 1 μM BL for 0,0. 5,1 h after the removal of BRZ by washing. To test the OsGSK2-mediated cytosolic translocation of OsOFP8, plasmid DNAs containing OsOFP8-YFP, NLS-mCherry, and 3HA-GSK2 were co-transfected into protoplasts. After 8 h incubation, the protoplasts were incubated with or without 10 μM MG132 for 1h. All transient transfection experiments were repeated at least three times with similar results. YFP and RFP fluorescence was observed with a confocal laser scanning microscope (ZEISS-510 Meta). The signal intensities of YFP and RFP were quantitatively determined using LSM Image Browser Rel. 4. 0 software. For yeast two-hybrid analysis, OsOFP8, OsBZR1, OsGAK2 and DLT were cloned into either pGBKT7 vector or pGADT7 vector, their combinations were tested for interaction. The reported gene assay was performed following the manufacturer’s instructions (Clontech). In addition, the truncated fragments of OsOFP8 were also ligated into pGBKT7 vector for the analysis of protein-protein interacting sites. For Co-immunoprecipitation analysis, OsOFP8-YFP, YFP, and 3HA-OsGSK2 were co-transfected into protoplasts in different combinations as indicated. After 8 h of incubation, total cell lysates from protoplasts were prepared in IP buffer (10 mM Tris-HCl pH 7. 4,150 mM NaCl, 0. 5 mM EDTA, 0. 2% Nonidet P-40,5% glycerol, 1 mM dithiobis and 1 x Complete Protease Inhibitor Cocktail) and were then incubated with GFP-Trap agarose beads (ChromoTek) for 4 h at 4°C in a top to end rotator. After incubation, the beads were washed four times with ice cold washing buffer (10 mM Tris-HCl, pH7. 4,150 mM NaCl, and 0. 5 mM EDTA) and then eluted by boiling in reducing SDS sample buffer. Samples were separated by SDS-PAGE and analyzed by immunoblot using appropriate antibodies. For phosphorylation analysis, Plasmids of OsOFP8-YFP and 3HA-OsGSK2 were co-transfected into protoplasts. After 8 h incubation, total cell lysates from protoplasts were prepared in IP buffer (10m MTris-HCl pH 7. 4,150 mM NaCl, 0. 5 mM EDTA, 0. 2% Nonidet P-40,5% glycerol, 1 mM dithiobis, 1 × Phosphatase Inhibitor Cocktail, and 1 x Complete Protease Inhibitor Cocktail) and were then incubated with GFP-Trap agarose beads (ChromoTek) for 4 h at 4°C in a top to end rotator. After incubation, the beads were washed four times with ice cold washing buffer (10 mM Tris-HCl, pH 7. 4,150 mM NaCl, and 0. 5 mM EDTA) and then eluted by boiling in reducing SDS sample buffer. Samples were separated by SDS-PAGE and followed by immunoblotting with biotin-pendant Zn2+-Phos-tag (BTL-111) according to the manufacturer’s instructions (Western Blot Analysis of Phosphorylated Proteins-Chemiluminescent Detection using BiotinylatedPhos-tag). Nuclear and cytoplasmic fractions in protoplasts were separated as described [37]. Protoplasts were lysed with a buffer (20 mM Tris-HCl, pH 7. 0,250 mM sucrose, 25% glycerol, 20 mM KCl, 2 mM EDTA, 2. 5 mM MgCl2,30 mM β-mercaptoethanol, 1 × protease inhibitor cocktail, and 0. 7% Triton X-100) and fractionated by centrifugation at 3000 g for 15 min at 4°C. The supernatant was taken as the cytosolic fraction. The pellet was further washed with a resuspension buffer (20 mM Tris-HCl, pH 7. 0,25% glycerol, 2. 5 mM MgCl2, and 30 mM β-mercaptoethanol) and reconstituted as the nuclear fraction. Each fraction was separated by SDS-PAGE and analyzed by Western blotting. For total protein extraction from protoplasts, transformed protoplasts were harvested by centrifugation at 200 g for 3 min, followed by resuspension in lysis buffer containing 25 mM Tris-HCl, pH 7. 5,150 mM NaCl, 1 mM EDTA, and 1 × protease inhibitor cocktail. The protoplasts were further lysed by vortexing for 2 min. The total cell extracts were then centrifuged at 15,000 g for 30 min at 4°C; the supermatant were total protein and analyzed by immunoblotting with appropriate antibodies [38].

OVATE family proteins (OFPs) are characterized as plant-specific transcription factors and mainly function in affecting fruit shape, but the molecular mechanisms by which they function are largely unknown. Rice genome contains 31 OFPs, the roles of these OsOFPs involved in plant development and growth are not understood. Brassinosteroids (BRs) are a class of steroid hormones involved in diverse biological functions. Here we report in rice that OsOFP8 plays a positive role in BR signaling pathway by interacting with OsGKS2, a negative regulator in BR signaling pathway. Our results shed light on studying the functions of OFPs and provide a chance to explore the new components of BR signaling pathway.

Abstract Introduction Results Discussion Materials and Methods

phosphorylation plant anatomy rna interference gene regulation regulatory proteins brassica dna-binding proteins cereal crops plant science model organisms rice transcription factors crops epigenetics plants cellular structures and organelles research and analysis methods arabidopsis thaliana grasses crop science genetic interference proteins gene expression leaves agriculture cytoplasm biochemistry rna plant biochemistry plant and algal models cell biology post-translational modification nucleic acids genetics biology and life sciences organisms

2016

OVATE Family Protein 8 Positively Mediates Brassinosteroid Signaling through Interacting with the GSK3-like Kinase in Rice

Lassa fever is caused by a viral haemorrhagic arenavirus that affects two to three million people in West Africa, causing a mortality of between 5,000 and 10,000 each year. The natural reservoir of Lassa virus is the multi-mammate rat Mastomys natalensis, which lives in houses and surrounding fields. With the aim of gaining more information to control this disease, we here carry out a spatial analysis of Lassa fever data from human cases and infected rodent hosts covering the period 1965–2007. Information on contemporary environmental conditions (temperature, rainfall, vegetation) was derived from NASA Terra MODIS satellite sensor data and other sources and for elevation from the GTOPO30 surface for the region from Senegal to the Congo. All multi-temporal data were analysed using temporal Fourier techniques to generate images of means, amplitudes and phases which were used as the predictor variables in the models. In addition, meteorological rainfall data collected between 1951 and 1989 were used to generate a synoptic rainfall surface for the same region. Three different analyses (models) are presented, one superimposing Lassa fever outbreaks on the mean rainfall surface (Model 1) and the other two using non-linear discriminant analytical techniques. Model 2 selected variables in a step-wise inclusive fashion, and Model 3 used an information-theoretic approach in which many different random combinations of 10 variables were fitted to the Lassa fever data. Three combinations of absence∶presence clusters were used in each of Models 2 and 3, the 2 absence∶1 presence cluster combination giving what appeared to be the best result. Model 1 showed that the recorded outbreaks of Lassa fever in human populations occurred in zones receiving between 1,500 and 3,000 mm rainfall annually. Rainfall, and to a much lesser extent temperature variables, were most strongly selected in both Models 2 and 3, and neither vegetation nor altitude seemed particularly important. Both Models 2 and 3 produced mean kappa values in excess of 0. 91 (Model 2) or 0. 86 (Model 3), making them ‘Excellent’. The Lassa fever areas predicted by the models cover approximately 80% of each of Sierra Leone and Liberia, 50% of Guinea, 40% of Nigeria, 30% of each of Côte d' Ivoire, Togo and Benin, and 10% of Ghana. Lassa fever (LF) is a viral haemorrhagic fever the pathogenic agent of which is an arenavirus Lassa virus (LASV) first discovered in 1969 in Nigeria, in a missionary nurse living in Lassa, a village close to the border with Cameroon [1]. Lassa fever is widespread in West Africa, affecting 2 million persons per annum with 5,000–10,000 fatalities annually [2]. Since its initial discovery, nosocomial outbreaks of Lassa fever have occurred repeatedly in Sierra Leone: Panguma, Kenema, 1971–83,1997, Liberia: Zorzor, 1972; Phebe 1972,1977,1982; Ganta 1977,1982 and Nigeria: Jos, 1970,1993; Onitsha, 1974; Zonkwa, 1975; Vom, 1975–77, Imo, 1989; Lafia, 1993; and Irrua, 2004 [3], [4], [5], [6], [7], [8], [9]. In Guinea, some acute but isolated cases were recorded in hospitals [10] and a single rural outbreak was recorded on the Sierra Leone border in 1982–83 [11]. Between these two areas, namely in Côte d' Ivoire, Ghana, Togo and Benin, no outbreak has ever been recorded, though isolated cases show evidence of viral circulation in that area [12], [13], [14]. Lassa fever therefore appears to have 2 geographically separate endemic areas: the Mano River region (Guinea, Sierra Leone, Liberia) in the West, and Nigeria in the East. The reservoir host of this virus is the multimammate rat, Mastomys natalensis, which was found infected for the first time in Sierra Leone and in Nigeria in 1972 [15], [16], and recently in Guinea [17]. In Upper Guinea, these commensal rodents aggregate in houses during the dry season, and disperse into the surrounding fields in the rainy season, foraging in cultivated areas before harvesting [18]. Villages where LASV-positive rodents have been trapped are all located in rain forest areas or in the transition zone between forest and savannah, within the 1500 mm rainfall isohyet. Rainfall seems to be an important ecological factor because a recent longitudinal study in rodents demonstrated that LASV infection was two to three times higher in the rainy season than in the dry season [18]. There are no studies to date indicating that the virus can survive better in humid than in dry soil, but evidence points in this direction. For example, the recent discovery of a new arenavirus in Mus minutoides (Kodoko virus [19]) and of hantavirus in Hylomyscus simus (Sangassou virus) in Guinea [20], were both made in rodents trapped in wet habitats, swamps or along river edges. In the USA, many new hantaviruses discovered within the last 15 years are found in damp or wet places such as arroyos or canyons, i. e. Black Creek canal virus, Blue river virus, El moro Canyon virus, Limestone Canyon virus. In the case of Sin Nombre virus, responsible for hemorrhagic fever with pulmonary syndrome, high risk areas are associated with higher elevation and mesic vegetation whereas low risk areas are associated with lower elevation and xeric vegetation. Soil moisture appears to be a key factor explaining the maintenance of this virus in high risk areas [21], [22]. In Europe, the transmission and persistence of Puumala virus, responsible for nephropathia epidemica, seems possible only if indirect transmission through a contaminated environment is included in a mathematical model. The combination of viral dynamics inside and outside the host, rodent demographic patterns and humid periods seems to explain the geographical distribution of this disease [23]. These advances all indicate the possible importance of rainfall patterns and humidity for Lassa Fever. We present our analysis of LF in West Africa in three steps: a first univariate analysis linking LF with high rainfall areas (Model 1) and the other two, multivariate analyses quantifying associations between LASV presence and a number of environmental parameters, derived from earth-observing satellites, that lead to the production of the first predictive risk maps for Lassa fever. One of these multivariate modelling approaches uses step-wise variable selection procedures (Model 2) whilst the other uses random combinations of predictor variables to identify the individual best predictors of LASV presence and absence (Model 3). Figure 1 shows the location of LF outbreaks (or areas of high human seroprevalence) from 1951 to 1989. The Jos plateau in Nigeria receives more rainfall than the surrounding areas and is disconnected from the wet coastal area by lowland areas of lower rainfall. Only the initial case in Lassa (800 mm/year) is located outside the high rainfall area. The map in Figure 1 suggests that areas with between 1200 mm and 1500 mm of rainfall per year are at relatively low risk of LF; areas with above 1500 mm have a much higher risk and, finally, areas with in excess of 3000 mm of rainfall annually appear to be at zero risk (i. e. had no outbreaks of LF in that period), although these very high rainfall areas are not widespread. The predictor variables chosen for the three different cluster versions of Model 2 are shown in Table 2 with their mean ranks across the 100 bootstrap models for each. The average accuracy of these models is shown in Table 3 and the mean values of the selected predictor variables for one of the top models from the 2 Absence: 1 Presence cluster combination is shown in Table 4. Figure 2 shows the mean predicted risk map of LF from the 100 bootstrap models using this same combination of absence and presence clusters. With only one cluster each, LF appeared to be over-predicted whilst with two clusters each LF appeared to be more strongly limited to the training set data points and their immediate surrounding areas (i. e. the disease was possibly under-predicted). The 2 Absence∶1 Presence cluster combination was therefore considered to give the best overall result. The rainfall variables were disproportionately selected by all cluster combinations in Model 1; each ‘top ten’ list in Table 2 contains four such variables, where the random expectation (5 satellite channels) is only two. At the same time, the vegetation index channels (NDVI and EVI) are under-represented, with only a single one of 20 such variables (10 Fourier variables per channel) chosen across all cluster combinations; the balance of the important predictor variables were thermal ones (either LST or MIR). The relatively high values for the average ranks of even the top variables in all cluster combinations in Table 2, however, reflects the fact that each of the 100 bootstrap samples gave rather different results in terms of the variables selected, and in their order of selection. This is a common feature of relatively small datasets. Despite the variability in the selected predictor variables, mean model accuracies were very high (Table 3) with, as expected, model accuracy increasing with increasing cluster numbers. The mean values of kappa put all models well within the ‘Excellent’ category of the Landis and Koch [41] scale (where kappa<0. 4 is ‘Poor’; 0. 4<kappa<0. 75 is ‘Good’; and kappa>0. 75 is ‘Excellent’). The mean values of the key predictor variables may differ considerably, or only by rather small amounts (Table 4). Table 4 shows that the mean values for the single clusters of presence points in the model are often intermediate between those of the two absence clusters. This applies to mean rainfall, night-time LST minimum, MIR phase 2 and daytime LST (mean and maximum). In other cases, mean values for the presence points are well outside those for either absence cluster. This applies to rainfall (amp1, amp3, phase1 and minimum) and NDVI phase 3. Concentrating on the important rainfall variables in Table 4 it is possible to suggest that LASV requires high (but not the highest) mean rainfall areas (rain mean), but with very high annual variation of this variable (rain amp1), and with peak rainfall occurring much later in the year (during August rather than during May or March, the months of peak rainfall of the absence clusters in Table 4, rain phase1). The significance of the higher amp3 rainfall value in Table 2 (the first selected variable) is unclear; often such higher harmonics act to modulate the lower frequency – annual or bi-annual – harmonics, and thus adjust the seasonal pattern of rainfall (extending or reducing high rainfall periods, depending on the timing of this tri-annual harmonic). The predicted risk map (Figure 2) captures most of the presence points in the database (the grey areas in Figure 2 in southern Nigeria and Cameroon are regions where cloud contamination is so continuous that it was not possible to obtain either sufficient cloud-free images or their temporal Fourier derivatives for modelling; these are therefore areas where it is not possible to make predictions of risk). The predicted risk areas in Figure 2 contract towards the coast in the ‘Dahomey gap’ between the western and central forests of Africa (see Introduction) but are still more extensive than the rainfall map and data in Figure 1 suggest. In fact the satellite rainfall image (CMORPH mean, not shown) also indicates a lower mean rainfall area in this region, so that the positive LASV predictions for this area must arise from the values of other key predictor variables. The differences between Figure 1 and Figure 2 in the basin of the River Zaire, towards Central Africa, arise because these areas (though high in rainfall) are environmentally quite distinct from those of the training set area and so the risk map models classify them as ‘No prediction’ areas (coloured grey in Figure 2). Tables 4 and 5 show results analogous to those of Tables 2 and 3 but for Model 3, where the important variables were identified using the combination method of Burnham and Anderson [40]. This method highlights even more the importance of rainfall variables (only 8 out of the 30 variables in Table 5 are not directly rainfall related), with slightly different combinations in each case for the different cluster combinations. Overall model accuracies are still excellent (Table 6) though not quite as good as those for Model 2. Figure 3 shows the mean predicted risk map obtained by using in Model 3 the selected combination of the top 10 variables for the same 100 bootstrap samples that were used in Model 2 to generate Figure 2. Figure 3 is less equivocal about risk areas than is Figure 2 (i. e. there are fewer regions of intermediate probability of LASV risk) for the simple reason that the same 10 variables were used throughout, whereas different combinations of variables were often selected in the Model 2 models, giving more variable results. Figure 3 again captures most of the presence sites within the training set, with rather different predictions for the Dahomey Gap region than those in Figure 2. The question that comes immediately to mind is: why does Lassa fever occur only in West Africa, whereas the range of its vertebrate host extends into East and Southern Africa? This is a recurrent question for other rodent-borne diseases (such as plague and hemorrhagic fevers with renal or pulmonary syndrome; see [42] for a review), which are also much more restricted in their distributions than are their hosts. Our analyses here show quite clearly that Lassa fever requires a particular combination of high (but not the highest) rainfall, and with a particular form of variability and seasonal timing, whereas its hosts can and do occur over regions experiencing a much wider range of rainfall conditions. Temperature appears to be less important in determining LASV distribution, although there are large differences between different areas; for example the annual mean and maxima in high risk areas are 27°C and 32°C respectively, whereas in low risk areas the mean temperature was approx. 38°C. Such high temperatures are known to increase LASV decay [43]. One curious feature of the present results is the seeming unimportance of vegetation variables in the predictor data sets. This lack of importance is not due to their strong correlation with rainfall variables (such a correlation might exclude them in step-wise inclusive variable selection), because Model 3 (using a method that avoids the problems of step-wise methods) independently and quite categorically failed to identify vegetation variables as important in determining LASV distribution. Taken together these results suggest that the survival of the virus outside of the vertebrate host might be a key to determining its distribution, and that this survival depends upon moisture or rainfall conditions above more or less all other environmental variables. This result differs from the conditions favouring other viral transmission; for example, low relative humidity and temperature favour avian influenza [44]. In the case of Lassa, the virus appears to survive better in humid conditions, during the rainy season. Rodents will be more often contaminated during their frequent movements at this season, for mating or dispersing into the surrounding fields [45]. Conversely, viral aerosol stability, seems to be higher when the humidity is lower [43], a condition that obviously occurs more frequently in the dry season. The experiments of Stephenson help to explain the numerous LF cases recorded in hospitals during the late dry season, between January and March in Sierra Leone and Nigeria ([46], Omilabu, pers. com.) but they do not necessarily throw much or any light on the persistence of Lassa fever in the general environment. We suggest that rainfall, within defined limits, is the single most important abiotic determinant of this persistence. M. natalensis, the most important host of LASV, does not occur in the western part of the region, in coastal Guinea and Sierra Leone and west to the 12th meridian. Only M. erythroleucus occurs in these regions, and our surveys have always found it to be negative for LASV infections [17]. The low human sero-prevalences recorded in these coastal areas are most likely due to the movement of people from highly endemic zones, or to human-to-human transmission. Towns and villages in these coastal areas, from Guinea to Gabon, have been invaded by the black rat Rattus rattus, and the domestic mouse, Mus musculus, probably taken there in historical times by Arab and European traders, explorers and colonisers. Absence of M. natalensis from coastal areas, for whatever reason (e. g. unsuitable habitats, or competition from other, non-Lassa-reservoir rodents), would explain the absence of Lassa fever in these areas, despite the apparently favourable (for LASV) climatic conditions (although the models suggest that some areas may be too wet for LASV). In Conakry for example, rodent sampling (330 specimens) showed that the most abundant species was M. musculus (70%), followed by R. rattus (25%) (unpublished data). In East and South Africa, the same reservoir species is present but the virus is replaced by other Lassa-like viruses such as Ippy, Morogoro and Mopeia, found in M. natalensis in CAR, Tanzania, Mozambique and Zimbabwe (CRORA database in Pasteur Institute website, http: //www. pasteur. fr/recherche/banques/CRORA/, [47], [48], [49]). These different Lassa –like viruses are not known to be pathogenic in humans and are considered ancestral by phylogenetic studies [50]. The scenario of multiple infection with both Lassa-like and Lassa virus is highly unlikely, and so we consider that central and eastern Africa are Lassa free. This is supported by many negative serological studies in Cameroon, in CAR, Congo, Equatorial Guinea and Gabon [24], [25], [26], [27]. However, the situation in south-west Cameroon bordering Nigeria remains problematic because this zone appears to be at high risk according to Figure 2. This is a volcanic area, which could provide a geographic barrier (Mt Cameroon, 4100 m, and the volcano chain up to the Adamaoua plateau). Furthermore, another species of Mastomys is suspected to be present in this area, M. kollmannspergeri, which is found in Niger, NE Nigeria, N Cameroon, S. Sudan and Chad [51]. In Zakouma National Park in Chad, some specimens were found in a village and in camps, indicating a potential synanthropy of this species [52]. The predictive risk map in Figure 2 identifies the central parts of Cameroon and CAR as risky areas, where it is possible that other Lassa-like viruses could occur, intermediate between Ippy/Mobala and Lassa (Mobala is another Lassa-like virus found in Praomys sp. , a closely related species to Mastomys spp, in CAR [53].). According to the risk maps shown here, with the reservations noted above, the LF risk area covers approximately 80% of the area of each of Sierra Leone and Liberia, 50% of Guinea, 40% of Nigeria, 30% of each of Côte d' Ivoire, Togo and Benin and 10% of Ghana. Such maps help public health policies and research, in targeting disease control and studies in potentially infected areas.

Previous studies on the eco-epidemiology of Lassa fever in Guinea, West Africa, have shown that the reservoir is two to three times more infected by Lassa virus in the rainy season than in the dry season. None of the intrinsic variables of the murine population, such as abundance or reproduction, was able to explain this seasonal variation in prevalence. We therefore here investigate the importance of extrinsic environmental variables, partly influenced by the idea that in the case of nephropathia epidemica in Europe contamination of the environment, and therefore survival of the pathogen outside the host, appears to be an important factor in this disease' s epidemiology. We therefore made an extensive review of the literature, gathering information about the geographical location of sites where Lassa fever has been certainly identified. Environmental data for these sites (rainfall, temperature, vegetation and altitude) were gathered from a variety of sources, both satellites and ground-based meteorological stations. Several statistical treatments were applied to produce Lassa ‘risk maps’. These maps all indicate a strong influence of rainfall, and a lesser influence of temperature in defining high risk areas. The area of greatest risk is located between Guinea and Cameroon.

Abstract Introduction Results Discussion

infectious diseases/neglected tropical diseases computational biology/ecosystem modeling infectious diseases/viral infections virology/emerging viral diseases

2009

Risk Maps of Lassa Fever in West Africa

Cells generate diverse microtubule populations by polymerization of a common α/β-tubulin building block. How microtubule associated proteins translate microtubule heterogeneity into specific cellular functions is not clear. We evaluated the ability of kinesin motors involved in vesicle transport to read microtubule heterogeneity by using single molecule imaging in live cells. We show that individual Kinesin-1 motors move preferentially on a subset of microtubules in COS cells, identified as the stable microtubules marked by post-translational modifications. In contrast, individual Kinesin-2 (KIF17) and Kinesin-3 (KIF1A) motors do not select subsets of microtubules. Surprisingly, KIF17 and KIF1A motors that overtake the plus ends of growing microtubules do not fall off but rather track with the growing tip. Selection of microtubule tracks restricts Kinesin-1 transport of VSVG vesicles to stable microtubules in COS cells whereas KIF17 transport of Kv1. 5 vesicles is not restricted to specific microtubules in HL-1 myocytes. These results indicate that kinesin families can be distinguished by their ability to recognize microtubule heterogeneity. Furthermore, this property enables kinesin motors to segregate membrane trafficking events between stable and dynamic microtubule populations. Understanding how cells generate intracellular structures and overall morphologies is one of the major goals of cell biology. For the cytoskeleton, strikingly different structures can be assembled from a set of highly conserved building blocks. For example, all microtubules are generated by the polymerization of a common α/β-tubulin subunit yet diverse microtubule populations can be generated (e. g. , axonemes, spindles, and radial arrays) that carry out distinct functions. One way microtubule diversity can be characterized is based on dynamic properties [1]. Some microtubules are dynamic and turn over rapidly by alternating between periods of microtubule growth (polymerization) and shrinkage (depolymerization). Other microtubules are stable (low turnover) and are defined by their resistance to drugs that result in depolymerization of microtubules, such as nocodazole. In vivo, microtubules frequently pause, undergoing neither polymerization nor depolymerization [2]. Microtubule diversity can also be characterized by structural differences, for example alterations in protofilament number, as well as by chemical differences between tubulin subunits due to differences in the expression of tubulin genes (isotypes) or the presence of post-translational modifications (PTMs) [1], [3]–[5]. What are the biological functions of microtubules diversity? Dynamic instability allows microtubules to explore three-dimensional space for rapid remodeling of the cytoskeleton during processes such as spindle assembly and cell migration [6]–[9]. Stable microtubules likely play important roles in cellular morphogenesis but how and why remain unclear [10]–[13]. The chemically diverse PTMs that mark stable microtubules may affect morphogenesis by stabilizing microtubules and/or by influencing distinct intracellular transport events [4]–[6]. The expression of tubulin isotypes can influence polymerization dynamics and plays a role in the formation of specific microtubule assemblies such as the flagellar axoneme [3], [14]. The challenge is to explain how the diversity of microtubule structures is translated into specific cellular functions. This likely requires the functions of a large number of microtubule associated proteins (MAPs). Of special interest are motor proteins of the kinesin and dynein families that use ATP hydrolysis to move cellular cargoes along microtubule tracks [15], [16]. The molecular and mechanistic properties of motor proteins have typically been studied in vitro using homogeneous microtubule assemblies. Thus, understanding how motor proteins could read microtubule diversity has been difficult to answer. We recently developed techniques for tracking kinesin motors at the single molecule level in the cytoplasm of live cells [17]. Here, we extend these techniques to two-color tracking and evaluate the motility of kinesin motors along heterogeneous populations of microtubules tracks in COS cells. We show that Kinesin-1 motors move preferentially along stable microtubules marked by PTMs whereas Kinesin-2 (KIF17) and Kinesin-3 (KIF1A) motors can utilize dynamic microtubule tracks. These results indicate that kinesin motors have evolved to recognize specific microtubule subpopulations and, thus, segregate membrane trafficking events within cells. Constitutively active kinesin motors can be generated by truncations that remove autoinhibitory and cargo-binding regions of the polypeptide. For this work, we generated KHC (1-560) (Figure 1A), a dimeric motor that has been well characterized in vitro and in vivo [16], [18], [19]. KHC (1-560) motors were tagged with three tandem copies of monomeric Citrine (mCit), a variant of enhanced yellow fluorescent protein (FP) (Figure 1A), and expressed in COS cells (Figure 1B). Single Kinesin-1 motors were tracked in live cells using a modified TIRF microscope (Figure 1C) in which the angle of illumination was varied to enable deeper imaging as described [17]. KHC (1-560) -3xmCit motors were observed to undergo both free diffusion and linear movement (Video S1). Linear motility occurred with an average speed of 0. 83±0. 08 µm/sec and average run length of 0. 91±0. 23 µm in live cells (Table 1, n = 372 events), consistent with previous work [16], [17]. To gain an understanding of the ensemble characteristics of Kinesin-1 motility events, we developed methods to sum all motility events of a time series into one image called the standard deviation map (SD Map) [17]. We compute this map by determining the standard deviation (SD) of the fluorescence intensity change for each pixel over the entire time series (Figure S1). Thus, pixels with little to no fluorescence variation over the time series have low SDs whereas pixels with large fluctuations in intensity (e. g. , where motility events occurred) are highlighted in the SD Map. As seen in the SD Map of a single molecule TIRF movie of KHC (1-560) -3xmCit in live COS cells (Figure 1D), multiple individual Kinesin-1 motors moved repeatedly on linear tracks. Only a few Kinesin-1 tracks were identified in each image series suggesting that the SD Map reveals “hot spots” of Kinesin-1 motility. That these hot spots are not due to TIRF imaging of microtubules at the bottom of the cell is indicated by the observation of many microtubules in a comparable area of COS cytoplasm (Figure 1E). We hypothesized that Kinesin-1 motors move preferentially on only a subset of microtubule tracks. To directly test the possibility that Kinesin-1 motors distinguish microtubule populations in cells, we used two approaches. We first performed two-color TIRF imaging (Figure 1F) of live cells expressing FP-tagged Kinesin-1 motors and microtubules. This approach allows us to simultaneously visualize motors and their tracks and can account for shifts in the position of individual microtubules during live cell imaging (see [20] and Figure S2). COS cells were first transfected with plasmids encoding mCherry-tubulin and 24 h later with plasmids encoding KHC (1-560) -3xmCit. After an additional 5 h of expression, the cells were imaged by TIRF microscopy. KHC (1-560) -3xmCit motility events were observed along mCherry-tubulin microtubules (Video S2). Merged images of the SD Map of KHC (1-560) -3xmCit motility events and the mCherry-tubulin fluorescence (Figure 1F, representative of n = 4 cells in two experiments) indicate that mCherry-tubulin microtubules overlapped with 96. 6%±5. 2% of the Kinesin-1 tracks. In contrast, Kinesin-1 motility events were observed on only 9. 1%±8. 1% of the mCherry-tubulin microtubules (Table 1). These data indicate that Kinesin-1 motors utilize only a subset of the available microtubule tracks. We then performed retrospective immunofluorescence (Figure 1G) after single molecule TIRF imaging. This approach avoids difficulties with double transfection and the possibility that mCherry-tubulin does not incorporate into all microtubules yet is hindered by the fact that cells shrink during fixation, often resulting in a shift in position of the entire microtubule population. For retrospective imaging, COS cells expressing KHC (1-560) -3xmCit motors were imaged by TIRF microscopy, fixed, and then stained with an antibody to total tubulin (Figure 1G). The previously imaged cell was again observed by TIRF microscopy. A comparison of the SD Map of Kinesin-1 motility events and the total tubulin staining confirmed that the motility events occurred on only a subset of microtubules present in the imaging field (Figure 1G, representative of n = 16 cells in six experiments). Taken together, these results confirm that Kinesin-1 motors preferentially utilize only a subset of the microtubules present in COS cells. We first tested whether Kinesin-1 motors move preferentially on dynamic microtubules. This population can be observed in live cells expressing FP-tagged plus-end tracking proteins (+TIPs) [21]. Two-color TIRF microscopy was used to analyze COS cells coexpressing KHC (1-560) -3xmCit with a mCherry-labeled version of the +TIP protein end binding (EB) 3 [22]. Very few Kinesin-1 motility events could be observed on microtubules extending back from the EB3-mCherry-labeled plus ends (Video S3). A comparison of the SD Map of KHC (1-560) -3xmCit motility events with the average EB3-mCherry fluorescence (Figure 2A, representative of n = 12 cells in four experiments) demonstrates that the microtubule tracks utilized by KHC (1-560) -3xmCit motors are distinct from the dynamic microtubules marked by EB3-mCherry. In some cases, multiple Kinesin-1 motility events occurred on a microtubule track that appeared to lie directly adjacent to an EB3-marked dynamic microtubule (boxed region in Figure 2A, kymographs in Figure 2B, 2C). Kinesin-1 motility events overlapped with only 2. 5%±1. 1% of the EB3-mCherry-marked microtubules (Table 1). These results indicate that the preferential motility of Kinesin-1 motors does not occur on dynamic microtubules. An avoidance of dynamic microtubules is likely to be advantageous to the motor since the speed of Kinesin-1 is greater than that of microtubule polymerization. In the representative cell shown in Figure 2, KHC (1-560) -3xmCit motors moved at an average speed of 0. 73±0. 16 µm/sec (Figure 2D, red traces) whereas EB3-labeled microtubules grew at an average rate of 0. 08±0. 03 µm/sec (Figure 2D, black traces). Thus, Kinesin-1 motors that move along dynamic microtubules could rapidly run off the end of the track. We then tested whether Kinesin-1 motors move preferentially on stable microtubule tracks. To do this, we performed retrospective immunofluorescence staining using antibodies that recognize the PTMs that mark stable microtubules. Cells expressing KHC (1-560) -3xmCit were imaged in the TIRF microscope, fixed, stained with antibodies to acetylated α-tubulin and total tubulin, and the previously imaged cells were again viewed on the TIRF microscope. The pattern of KHC (1-560) -3xmCit motility events in the resulting SD Map was similar to the pattern of acetylated microtubules (Figure 3A, representative of 11 cells in six experiments). Kinesin-1 motility events colocalized with 90. 3%±5. 5% of microtubules marked by acetylated tubulin (Table 1). This suggests that Kinesin-1 moves preferentially along stable microtubules marked by acetylation of α-tubulin. We then used retrospective immunofluorescence to test whether Kinesin-1 motility events correlate with the presence of other PTMs that mark stable microtubules. The pattern of KHC (1-560) -3xmCit motility events in the SD Map was similar to that of the microtubule tracks marked by detyrosination (Figure 3B, representative of six cells in five experiments), a modification that appears to mark the same microtubule tracks as acetylation (see [23] and Figure S3). Kinesin-1 motility events did not colocalize with microtubules marked by polyglutamylation (Figure 3C, representative of eight cells in three experiments), most likely due to the low levels of glutamylation on cytoplasmic microtubules in these (Figure S4) and other non-neuronal cells [24]. We conclude that Kinesin-1 motors move preferentially along microtubules marked by acetylation and detyrosination. Is preferential motility on stable microtubules a general feature of kinesin motors that drive vesicular transport events? To test this, we performed single molecule imaging of 3xmCit-tagged KIF17, a homodimeric member of the Kinesin-2 subfamily. KIF17 has been implicated in the transport of cargoes in dendrites of neuronal cells and in cilia of invertebrates and vertebrates [25]–[27]. Single molecule TIRF imaging of a constitutively active version of KIF17 [KIF17 (1-490) -3xmCit, Figure 4A] showed that these motors moved with an average speed of 1. 31±0. 05 µm/sec and average run length of 0. 56±0. 22 µm in live COS cells (Table 1, n = 233 events), consistent with the motile properties of the C. elegans homologue OSM-3 [28], [29]. To test whether KIF17 motors move preferentially on a subset of microtubules, we performed two-color TIRF imaging of COS cells co-expressing mCherry-tubulin and KIF17 (1-490) -3xmCit (Video S4). A comparison of the SD Map of KIF17 (1-490) -3xmCit motility events with the average mCherry-tubulin fluorescence (Figure 4B) demonstrates that KIF17 motility events occurred on nearly all available microtubule tracks (Table 1). In addition, retrospective immunofluorescence imaging indicated that KIF17 motility occurred on both acetylated and non-acetylated microtubules (Table 1 and Figure S5). Together these results indicate that KIF17 is a non-selective motor as it does not show preferential motility when presented with a heterogeneous population of microtubules in COS cells. Since motility on dynamic microtubules would appear to be disadvantageous to plus end-directed motors, we performed two-color TIRF imaging of COS cells co-expressing KIF17 (1-490) -3xmCit and EB3-mCherry to directly test whether KIF17 motors move on dynamic microtubules. A comparison of the SD Map of KIF17 motility events to the average EB3-mCherry fluorescence demonstrates that KIF17 moved on dynamic microtubules (Figure 4C and Table 1). Multiple individual KIF17 (1-490) -3xmCit motors could be observed moving on the same EB3-marked microtubule (Figure 4D, 4E). Surprisingly, KIF17 motors that reached the plus end of the microtubule did not dissociate immediately, but rather lingered at the growing plus end (Figure 4E). Whether this ability of KIF17 to track the plus ends of growing microtubules is associated with its cellular functions is presently unknown. We then tested whether a member of the Kinesin-3 family, KIF1A, moves preferentially on a subset of microtubules. KIF1A has been implicated in the axonal transport of synaptic vesicle precursors [15]. Constitutively active dimeric KIF1A (1-393) -3xmCit motors (Figure 5A and [30]) imaged by single molecule TIRF microscopy moved with an average speed of 1. 82±0. 04 µm/sec and average run length of 0. 55±0. 19 µm in live COS cells (Table 1, n = 305 events), consistent with the measured properties of this motor in vitro and its cargoes in vivo [15], [30]. To analyze KIF1A motility on microtubules in vivo, two-color TIRF imaging of COS cells co-expressing mCherry-tubulin and KIF1A (1-393) -3xmCit was performed (Video S5). A comparison of the SD Map of KIF1A (1-393) -3xmCit motility to the average mCherry-tubulin fluorescence indicates that KIF1A motility events were observed on all of the available microtubules (Figure 5B and Table 1) indicating that KIF1A is also a non-selective motor. Consistent with this, KIF1A motility events were observed on both acetylated and non-acetylated microtubules (Table 1 and Figure S5). That KIF1A motors utilize dynamic microtubules for motility was demonstrated by two-color TIRF imaging of COS cells co-expressing KIF1A (1-393) -3xmCit and EB3-mCherry (Figure 5C). Multiple KIF1A motility events were observed to occur on individual dynamic microtubules labeled at their plus ends with EB3-mCherry (Figure 5 D, 5E and Table 1). KIF1A (1-393) -3xmCit motors that overtook the plus end of the microtubules did not fall off but, surprisingly, remained at the plus ends of growing microtubules for significant periods of time (Figure 5E). Thus, the Kinesin-3 motor KIF1A does not select specific microtubule tracks for motility in COS cells and, when using dynamic microtubules, remains localized to the growing plus ends. How does the preferential motility of single Kinesin-1 motors on stable microtubules relate to Kinesin-1-driven transport events inside cells? To test whether stable microtubules provide preferred tracks for motility of vesicular cargoes transported by Kinesin-1, we tracked the Golgi-to-plasma membrane transport of a variant of the vesicular stomatitis virus G protein that can be restricted to Golgi-derived vesicles using a temperature shift protocol (VSVG-GFP [19], [31], [32]). After imaging (Video S6), we performed retrospective immunofluorescence with antibodies to total and acetylated tubulin. A comparison of the SD Map of the motility events to the acetylated microtubules in the same cell (Figure 6A, representative of 12 cells in four experiments) demonstrates that 68. 0%±15. 2% of VSVG-GFP-positive vesicles moved along microtubules marked by acetylated α-tubulin. Multiple VSVG-GFP-marked vesicles were observed to move independently along the same acetylated microtubule (Figure 6B, 6C). These results indicate that Golgi-derived vesicles moved by Kinesin-1 are transported preferentially along microtubules marked by acetylation. We then tested whether cargoes transported by “non-selective” motors, such as KIF1A or KIF17, reflect the properties of these motors. KIF1A transports presynaptic vesicles in neuronal cells, but a cargo for this motor has not been described in fibroblasts. For KIF17, we found that the steady-state cell surface levels of the voltage-gated potassium (Kv) channel Kv1. 5 in HL-1 atrial myocytes was decreased by expression of a dominant negative (DN) version of KIF17 but not Kinesin-1 (Figure S6). This was surprising as KIF17 has so far only been described as a dendritic or ciliary motor [15], [33]. Western blot analysis shows that the KIF17 protein is expressed in mouse brain and heart tissues as well as HL-1 myocytes (Figure S7). KIF17 participation in the motility of Kv1. 5-GFP labeled vesicles was examined in live cells by coexpressing Kv1. 5-GFP with DN versions of KIF17 or Kinesin-1 (KHC) in HL-1 cells (Figure 7A–7C). To synchronize the Kv1. 5-GFP vesicle population spatially and temporally, we used a temperature shift protocol to first restrict Kv1. 5-GFP to the trans-Golgi by a 19°C incubation and then initiate post-Golgi transport by incubation at 37°C [34]. Golgi-derived Kv1. 5-labeled vesicles were observed to move in a linear fashion interspersed with pauses. Expression of DN KIF17, but not DN Kinesin-1, resulted in a significant decrease in Kv1. 5 vesicle motility, both in the average distance traveled and average net velocity of the vesicles (Figure 7D, 7E). These results indicate that KIF17 contributes to the microtubule-based transport of Golgi-derived Kv1. 5 vesicles in HL-1 myocytes. To examine whether KIF17-driven motility of Kv1. 5 channels occurs along a subset of microtubules, retrospective immunofluorescence imaging was applied to HL-1 myocytes expressing Kv1. 5-GFP. After live cell imaging (Video S7), the cells were fixed and stained with antibodies for total or acetylated tubulin. A comparison of the SD Map of Kv1. 5-GFP vesicle motility to the image of the total microtubule population demonstrates that Kv1. 5 vesicles move along microtubule tracks (Figure 7F, representative of 20 cells in six experiments) but only 13. 9%±12. 7% of the motility events occurred on microtubules marked by acetylated α-tubulin (Figure 7G, n = 3 cells in three experiments). Thus, the motility of Kv1. 5 vesicles, like that of the KIF17 motor, does not occur preferentially on stable microtubules marked by acetylation. Our results demonstrate a new property that distinguishes kinesin families—the ability to respond to microtubule heterogeneity in cells. We show that individual Kinesin-1 motors undergo preferential motility along stable microtubules marked by PTMs whereas individual Kinesin-2 (KIF17) and Kinesin-3 (KIF1A) motors are not selective as they undergo motility on both dynamic and stable microtubules. Transport along stable microtubules would prevent the undesirable situation where a dynamic microtubule track “disappears” under Kinesin-1 and its associated cargoes. Why then do Kinesin-2 and Kinesin-3 motors not avoid dynamic microtubules? One possibility suggested by our live cell imaging of single KIF1A and KIF17 motors (Figures 4 and 5) is that, upon overtaking the plus end of the microtubule, these motors do not dissociate but rather remain at the plus ends of growing microtubules. This may be an important requirement for motors whose cargoes function at the interface of microtubule plus ends and the cell cortex. Interestingly, computer simulations and cell staining have suggested that Kinesin-6 motors may remain attached rather than fall off upon reaching the tips of dynamic microtubules [35], [36]. A second possibility is that these motors and/or their cargoes can prevent depolymerization of the microtubule track. For example, in yeast, transport of +TIP proteins can prevent microtubules from depolymerizing under minus end-directed kinesin motors [37]–[40]. Differences have been reported between kinesin motors in their transport direction in neuronal cells. Kinesin-1 motors accumulate at the tips of axons whereas Kinesin-2 and Kinesin-3 motors accumulate in both axonal and dendritic compartments [18], [19]. Recent work suggests that track selectivity is likely related to the ability to undergo polarized transport [41]. In this work, substitution of residues in the Kinesin-1 motor domain with the corresponding residues of the KIF1A motor domain resulted in a motor that could no longer distinguish between tyrosinated and detyrosinated microtubules in vitro and could not undergo polarized transport to axons in vivo [41]. Thus, a selective motor was converted into a non-selective motor by mutation of the microtubule-binding surface. Together with our work on imaging single motors, these data support the hypothesis that track selection is an inherent property of the motor-microtubule interaction. What biochemical cues enable Kinesin-1 motors to distinguish microtubule populations? Our results show that Kinesin-1 selects stable microtubules marked by detyrosination and acetylation for preferential motility. One possibility is that it is the PTMs themselves that influence Kinesin-1. This possibility has gained support from recent work in vitro and in vivo [42]–[46]. However, the recognition of PTMs by the Kinesin-1 motor is likely to be complex as mutation of motor surface abolished the ability of Kinesin-1 to recognize detyrosinated but not acetylated microtubules [41]. While glutamylation could play an important role in guiding motor-based transport in epithelial and neuronal cells [44], [47]–[49], it is not likely to play a critical role in fibroblasts. Recent work has indicated that the situation may be different for fungal motors as Kinesin-3 motors, but not Kinesin-1 motors, show track selectivity [50]. Thus, how different PTMs create a tubulin code that can guide motor protein transport events is an important area for future studies. The correlation of Kinesin-1 motility with specific PTMs does not rule out the possibility that other microtubule-based mechanisms influence this or other motors. Structural changes that occur in the microtubule lattice after polymerization and/or stabilization may influence kinesin motors. Also, MAPs that stabilize microtubules have been shown to negatively influence kinesin-based transport events in cells [51]. Thus, it may be that the modifications that occur along stable microtubules serve to decrease binding of MAPs to microtubules and thus clear the way for motor-based transport [48]. Our work provides the first demonstration that transport events can be segregated between stable and dynamic microtubules via kinesin motors that select these subpopulations of microtubule tracks. Stable microtubules are critical for morphogenesis during diverse biological events such as cytokinesis, cell motility, and neuronal polarity [10], [13], [52], [53]. The preferential motility of Kinesin-1 along stable microtubules may serve to direct Kinesin-1 transport during morphogenesis and maintenance of polarity in neuronal and epithelial cells [18], [19], [41], [44], [54], [55]. A transport module comprised of cargo/Kinesin-1/microtubule subsets is likely involved in other polarized trafficking events such as the delivery of mRNA complexes to the vegetal pole in Xenopus oocytes [56]. Dynamic microtubules are important for microtubule search and capture events in mitotic cells as well as in interphase cells during cell polarity and motility [1], [7], [57], [58]. Our results imply that dynamic microtubules can serve as tracks for kinesin-based transport of cargoes that likely function at the microtubule-cortex interface. Indeed, a transport module that comprises cargo/Kinesin-2,3/dynamic microtubule components may be important in trafficking of connexins, cadherins, +TIPs, and channels to cell-cell junctions [59]–[62]. This transport module may also influence retrograde trafficking events such as cytoplasmic dynein-driven movement of endoplasmic reticulum (ER) -derived vesicles to the central Golgi complex [63]. This work is the first to analyze the segregation of kinesin motors and their cargoes to distinct microtubules populations and subcellular destinations. Recent work using TIRF microscopy of detergent-extracted cells has indicated that unconventional myosins also differ in their ability to select actin filament tracks [64]. Thus, these types of experiments provide a starting point for exploring the ability of motor proteins to respond to structural and/or biochemical changes in cytoskeletal filaments inside cells as well as the relationship between track selection and cellular function. COS and HL-1 cells were cultured and transfected as described [17], [65]. The following antibodies were purchased: total β-tubulin (E7, Developmental Studies Hybridoma Bank, Univ. Iowa), acetylated α-tubulin (Sigma B-11-61), detyrosinated α-tubulin (Chemicon), GFP (Invitrogen), and fluorescently marked secondary antibodies (Jackson ImmunoResearch). A monoclonal antibody to polyglutamylated tubulin (GT335) was a gift from C Janke (CNRS, France). A polyclonal antibody that recognizes acetylated α-tubulin was generated against amino acids 29–52. Constitutively active versions of kinesin motors were generated by PCR cloning of the relevant sequences (aa 1-490 of human KIF17, aa 1-393 of rat KIF1A, and aa 1-560 of rat KIF5C) into the 3xmCit-N1 vector [17]. DN versions of KHC (aa 566-955) and KIF17 (aa 488-846) were cloned into mCherry-C1. All plasmids were verified by DNA sequencing. Plasmids encoding VSVG-GFP and EB3-mCherry were gifts from A Akhmanova and N Galjart (U Rotterdam). Kv1. 5-GFP has been described [65]. Objective-based TIRF microscopy was carried out as described [17]. Briefly, transfected COS cells on a glass-bottomed 35 mm dish (MatTek) were carefully rinsed with Ringers buffer (10 mM HEPES/KOH, 155 mM NaCl, 5 mM KCl, 2 mM CaCl2,1 mM MgCl2,2 mM NaH2PO4,10 mM glucose, pH 7. 2) and imaged on a custom-modified Zeiss Axiovert 135TV microscope equipped with a 1. 45 NA a-Plan Fluor objective, 2. 5× optovar, 505DCXR dichroic and HQ510LP emission filter (Chroma Technology), 488 nm line of a tunable single-mode fiber-coupled Argon Ion Laser (Melles Griot), and a back-illuminated EMCCD camera (Cascade 512B, Roper Scientific). The angle of illumination was adjusted for maximum penetration of the evanescent field into the cell (near-TIRF), enabling an imaging depth of ∼500 nm, which is sufficient to image nearly all the microtubules in the periphery of flat COS cells. For two-color TIRF, a yellow diode pumped solid-state laser (593 nm, CrystaLaser) was combined with the 488 nm laser using a dichroic mirror (Z488RDC). Fluorescence emissions were first passed though a FF495/605 dual-band dichroic mirror (Semrock) and then projected onto separate halves of the CCD camera by a Dualview beam-splitter (Optic Insights) equipped with a T585LP dichroic beam splitter and ET525/50M and HQ610LP emission filters (all Chroma Technology). In general, cells were imaged ∼6 h post-transfection to maintain low expression levels representative of endogenous protein behavior and optimal for single molecule TIRF imaging. Images were captured every 50–100 ms for 30–35 s. All experiments were carried out at room temperature (18–21°C). For measurement of EB3-mCherry microtubules, the lower average microtubule growth rate compared to those reported in other studies (e. g. [66] and references therein) is likely due to this lower temperature. Immediately after imaging, cells were fixed (3 min) in 3. 7% paraformaldehyde (Ted Pella), quenched (10 min) with 50 mM NH4Cl, and permeabilized (3 min) with 0. 2% Triton X-100. An equal volume of 1. 0% glutaraldehyde was then carefully added for an additional 7 min of fixation. After quenching with freshly prepared 1. 5 mg/ml NaBH4, primary antibodies in 0. 2% fish skin gelatin were incubated for 1–2 h at room temperature or overnight at 4°C. COS or HL-1 cells in glass-bottom dishes (MatTek) were imaged live on a Nikon TE2000 microscope with a Plan-APO 100×/NA 1. 4 objective and Photometrics CS ES2 camera. VSVG (tsO45) -GFP-expressing cells were incubated overnight at 39°C to accumulate protein in the ER. Two h after shifting cells to 33°C for synchronous protein transport through the Golgi complex and to the plasma membrane, images were obtained every 1 s. For Kv1. 5-GFP, cells were incubated at 19°C for 3 h to accumulate secretory proteins in the trans-Golgi and then shifted to 37°C and imaged every 5 s. Statistical analysis was done using one-way ANOVA analysis. Videos and images were prepared with ImageJ (NIH) and Photoshop and Illustrator (Adobe). Generation of the SD Maps is described in Figure S1 and [17]. Home-made plug-ins for ImageJ were used for measuring the speed and run length of motors and vesicles. For motors, only diffraction-limited fluorescence spots (5×5 pixels) were selected for analysis that were clearly separated from the neighboring fluorescence and moved in a linear fashion on microtubules tracks identified in the SD Maps. Motile events that did not appear in the SD Map were usually short and/or blurry events that could not be separated from diffusion. For colocalization of motor and microtubule tracks (Table 1), the SD Map of the motility events was overlaid with a static image of the mCherry-tubulin fluorescence. The relative overlap (expressed in %) was calculated as the ratio of microtubules with motility events to total number of microtubules. For colocalization of VSVG-GFP or Kv1. 5-GFP vesicles with acetylated tubulin, the SD Map of the vesicle motility was overlaid with the fixed acetylated tubulin image. The overlap (expressed in %) was calculated as the (number of vesicles on microtubules) / (number moving vesicles). Supplemental methods are described in Text S1.

Eukaryotic cells assemble a variety of cytoskeletal structures from a set of highly conserved building blocks. For example, all microtubules are generated by the polymerization of a common α/β-tubulin subunit, yet cells can contain diverse, discrete populations of microtubule structures such as axonemes, spindles, and radial arrays. This diversity must be read and translated by cellular components in order to carry out population-specific functions. We use single-molecule imaging to study how molecular motors navigate the heterogeneous microtubule populations present in interphase cells. We show that different kinesin motors select different subpopulations of microtubules for transport. This selectivity, based solely on the motor-microtubule interface, may enable kinesin motors to segregate transport events to distinct microtubule populations and thus to target cargoes to specific subcellular destinations.

Abstract Introduction Results Discussion Materials and Methods

biophysics/macromolecular assemblies and machines cell biology/morphogenesis and cell biology biophysics/experimental biophysical methods cell biology/membranes and sorting cell biology/cytoskeleton

2009

Single Molecule Imaging Reveals Differences in Microtubule Track Selection Between Kinesin Motors

Burkholderia pseudomallei is an environmental bacterium that causes melioidosis, a major community-acquired infection in tropical regions. Melioidosis presents with a range of clinical symptoms, is often characterized by a robust inflammatory response, may relapse after treatment, and results in high mortality rates. Lipopolysaccharide (LPS) of B. pseudomallei is a potent immunostimulatory molecule comprised of lipid A, core, and O-polysaccharide (OPS) components. Four B. pseudomallei LPS types have been described based on SDS-PAGE patterns that represent the difference of OPS–type A, type B, type B2 and rough LPS. The majority of B. pseudomallei isolates are type A. We used matrix-assisted laser desorption/ionization time-of-flight mass spectrometry (MALDI-TOF MS) followed by electrospray ionization quadrupole time-of-flight mass spectrometry (ESI-QqTOF MS) and gas chromatography to characterize the lipid A of B. pseudomallei within LPS type A isolates. We determined that B. pseudomallei lipid A is represented by penta- and tetra-acylated species modified with 4-amino-4-deoxy-arabinose (Ara4N). The MALDI-TOF profiles from 171 clinical B. pseudomallei isolates, including 68 paired primary and relapse isolates and 35 within-host isolates were similar. We did not observe lipid A structural changes when the bacteria were cultured in different growth conditions. Dose-dependent NF-κB activation in HEK cells expressing TLR4 was observed using multiple heat-killed B. pseudomallei isolates and corresponding purified LPS. We demonstrated that TLR4-dependent NF-κB activation induced by heat-killed bacteria or LPS prepared from OPS deficient mutant was significantly greater than those induced by wild type B. pseudomallei. These findings suggest that the structure of B. pseudomallei lipid A is highly conserved in a wide variety of clinical and environmental circumstances but that the presence of OPS may modulate LPS-driven innate immune responses in melioidosis. Burkholderia pseudomallei is a Gram-negative bacillus and the causative agent of melioidosis. This disease is endemic in Southeast Asia and northern Australia, with a mortality rate up to 40% in northeast Thailand and 20% in Australia [1,2]. The manifestations of disease vary from indolent chronic infection to acute and fulminant sepsis. The common clinical presentations in melioidosis include severe pneumonia, bacteremia, skin infection, and abscesses in several organs [2]. B. pseudomallei is encapsulated, highly adaptable and able to survive under several extreme conditions in the environment, and clearance of B. pseudomallei in patients is difficult to achieve. The underlying mechanisms that contribute to persistence and death in melioidosis are unclear. Acute melioidosis is characterized by the up-regulation of pattern recognition receptors (PRRs) and pro-inflammatory cytokine release [3]. Melioidosis patients have increased pro-inflammatory cytokines, such as IL-12, IL-15, IL-18 and IFN-γ and patients who die from melioidosis have higher levels of plasma IL-6 and IL-8 than those who survive [4]. These studies suggest the importance of the innate immune response in the control of infection and pathophysiology of sepsis and mortality in melioidosis. Host immune cells express several PRRs including membrane bound Toll-like receptors (TLRs) and cytoplasmic receptors that recognize distinct bacterial pathogen-associated molecular patterns (PAMPs) [5]. Recognition of bacterial produces leads to the activation of innate immune response and release of inflammatory cytokines and mediators. Our previous study conducted in 300 healthy Thai donors suggest that lipopolysaccharide (LPS) of B. pseudomallei is a potent immuno-stimulatory molecule in humans [6]. The innate immune response to heat-killed B. pseudomallei is highly correlated with the response to purified LPS. We have also showed that genetic variation in TLR4, which encodes the canonical LPS receptor, TLR4 is associated with susceptibility to melioidosis [7]. These data point to the likely importance of B. pseudomallei LPS in the host-pathogen interaction in melioidosis. Four B. pseudomallei LPS types have been described based on SDS-PAGE patterns which represent the difference of O-polysaccharides (OPS) –ladder type A, ladder type B, type B2 and rough LPS. Type A and B LPSs are structurally composed of three covalently linked domains of O-polysaccharide, a core-oligosaccharide and a lipid A moiety, while rough LPS lacks OPS. We previously reported that type A is the predominant LPS type expressed by 97% of B. pseudomallei isolates from Thailand and 80% of isolates from Australia [8]. Lipid A, the membrane anchor region of LPS is recognized by the host innate immune system. This process occurs by the presentation of LPS by LPS-binding protein (LBP) in concert with the accessory molecule CD14 to the TLR4/MD-2 complex [9]. This complex is present on a variety of cell types including macrophages and dendritic cells and upon binding of LPS, triggers the production of proinflammatory cytokines by host cells [10]. In several pathogenic Gram-negative bacteria, lipid A variation has been demonstrated both in vivo and in vitro to be associated with the induction of different innate immune responses. In Salmonella enterica serovar Typhimurium, the presence of Ara4N residues on phosphate groups occurs when the bacteria are grown under magnesium-deficient conditions. These modifications decrease the overall negative charge on the surface of bacteria resulting in the lower affinity for cationic antimicrobial peptides (CAMPs) [11]. Temperature-dependent lipid A modifications have been described in Porphyromonas gingivalis, Francisella novicida, and Yersinia pestis [12–14]. Pseudomonas aeruginosa isolates from patients with cystic fibrosis have different lipid A structures that are correlated with disease severity and progression [15,16]. The first published study of B. pseudomallei lipid A structure in a single B. pseudomallei isolate, strain KHW, proposed a penta-acylated bisphosphorylated disaccharide backbone modified with Ara4N [17]. A second report on B. pseudomallei strains K96243 and 1026b showed the lipid A structure to be a predominantly tetra-acylated with a minor proportion of penta-acylated lipid A [18]. Recently, Norris et al have reported that LPSs from a type B Thai strain 576a and a virulent rough central nervous system-tropic strain MSHR435 from Australia induced higher innate immune responses than the prototype Thai type A strain 1026b. Matrix-assisted laser desorption/ionization tandem time-of-flight mass spectrometry (MALDI-TOF/TOF MS) analysis demonstrated structural differences in the lipid A of these isolates. They also suggested that the different LPS types may play a role in variable host responses to LPS [19] though a limitation of this study is the minimal number of B. pseudomallei used. Due to the discrepancy in the findings of these reports, the first aim of this study was to characterize the lipid A structure in a large collection of clinical B. pseudomallei isolates using MALDI-TOF MS followed by electrospray ionization quadrupole- time-of-flight mass spectrometry (ESI-QqTOF MS) and gas chromatography. Because the vast majority of Thai and Australian patients infected with B. pseudomallei type A present with a wide range of clinical symptoms and incur different outcomes, we hypothesized that different innate immune responses to B. pseudomallei type A may be associated with variation in the lipid A structures of clinical isolates. The second aim was to examine the structural change of lipid A under varying laboratory conditions and the third aim was to evaluate the TLR4-dependent immuno-stimulatory effect of different B. pseudomallei type A isolates, as well as the role of OPS and capsule. The study was exempted for ethical review by the Ethics Committee of the Faculty of Tropical Medicine, Mahidol University (documentary proof of exemption, MUTM-EXMPT 2017–003). All bacterial isolates obtained from human were anonymous. Cultivation of B. pseudomallei was performed in a biosafety level 3 (BSL3) laboratory. Two existing collections of B. pseudomallei isolates were used as follows: (i) 136 isolates from 68 melioidosis patients presenting to Sunpasitthiprasong Hospital, Ubon Ratchathani Northeast Thailand who developed relapsed infection. The relapse was defined as the bacterial isolates from the same patient during distinct episodes, which showed a similar pulsed-field gel electrophoresis banding pattern or multilocus sequence type regarded as originating from the same clone [20]. The first episode of melioidosis was between 1986 and 2004 and relapse was 15 days to 3,757 days after recovery from the primary episode (median = 207. 5 days, IQR = 84. 5–594. 5) [20,21]; (ii) 35 isolates from 35 individual colonies of 7 clinical specimens (5 colonies for each specimen) from a Thai patient with acute melioidosis who was admitted in 2006 to the same hospital [22]. The clinical specimens were blood, tracheal suction, urine, pus from right leg, pus from left leg, pus from forehead, and wound swab from thigh. Reference or laboratory strains of B. pseudomallei (K96243,1026b, 153 and 164) were used for comparison. The B. pseudomallei wild type, OPS mutant and capsule mutant strains used during this study are described in S1 Table. Unless stated otherwise, B. pseudomallei was cultured on trypticase soy agar (TSA) and incubated at 37 oC for 16–18 h. All isolates were stored in trypticase soy broth (TSB) with 15% glycerol at -80 oC. The effect of a range of laboratory conditions on lipid A structure was tested for B. pseudomallei strain K96243. The bacteria were cultured on TSA (Oxoid, Hants, UK) for 16–18 h. Approximately 10 colonies were streaked onto agar plates and incubated for 16–18 h under one of the following conditions: (i) TSA (pH 7. 4) at 25 oC, (ii) TSA (pH 7. 4) at 37 oC, (iii) TSA (pH 7. 4) at 42 oC, (iv) TSA (pH 4. 5) at 37 oC, (v) TSA (pH 8. 5) at 37 oC, (vi) TSA (pH 7. 4) plus 2 mM H2O2 (Merck, Darmstadt, Germany) at 37 oC, (vii) TSA (pH 7. 4) plus 5 mM H2O2 at 37 oC, (viii) TSA (pH 7. 4) plus 50 mM NaNO2 (Sigma-Aldrich, MO, USA) at 37 oC, (ix) TSA (pH 7. 4) plus 100 mM NaNO2 at 37 oC, (x) TSA (pH 7. 4) plus 1 mM MgCl2 (Fisher Scientific, Leics, UK) at 37 oC, (xi) TSA (pH 7. 4) plus 8 μM MgCl2 at 37 oC, (xii) Luria-Bertani agar (LB) (BD, MD, USA) (pH 7. 4) at 37 oC, (xiii) Ashdown agar (pH 7. 4) at 37 oC, (xiv) blood agar (Oxoid, Hants, UK) (pH 7. 4) at 37 oC, (xv) MacConkey agar (Oxoid, Hants, UK) (pH 7. 4) at 37 oC, and (xvi) M9 minimal medium agar (Sigma-Aldrich, MO, USA) (pH 7. 4) at 37 oC. The bacteria were harvested using a 10 μl loop and further extracted for lipid A. Lipid A was extracted from bacteria using a microextraction method as previously described [23] with some modifications. Briefly, the bacteria were cultured on TSA for 16–18 h and harvested as described above. One loop of the bacteria was suspended in 400 μl of 70% isobutyric acid (Sigma-Aldrich, MO, USA) and 1 M ammonium hydroxide (Sigma-Aldrich, MO, USA) at a ratio of 5: 3 (v/v). The samples were heated at 100 oC for 1 h followed by incubation on ice for 5 min. After centrifugation at 2000 × g for 15 min, the supernatant was collected, diluted with 1 ml of distilled water and lyophilized. The dried material was washed twice with 1 ml of methanol (Merck, Darmstadt, Germany) by centrifugation at 2000 × g for 15 min. The pellet was extracted for lipid A using 100 μl of a mixture of chloroform: methanol: water in a ratio of 3: 1. 5: 0. 25 (v/v/v). The suspension was centrifuged at 2000 × g for 1 min. One microliter of the supernatant was spotted onto a MALDI target plate and allowed to dry in air. The lipid A spots were overlaid with 1 μl of 20 mg/ml norharmane matrix (Sigma-Aldrich, MO, USA). Norharmane was used as a matrix in this study because it has been shown to improve limit of detection and support the characterization of a wide range of lipid A analysis [24]. Each spot was measured in 500 shot steps for a total of 3,000 laser shots using a Bruker Autoflex Speed MALDI-TOF mass spectrometer (Bruker Daltonics, Germany). The instrument was operated in negative ion reflector mode, across the mass range 1500–2000 and calibrated with electrospray ionization (ESI) tuning mix (Agilent Technologies, CA, USA). The mass spectrum was processed for smoothing and baseline subtraction using FlexAnalysis version 3. 4 (Bruker Daltonics, Germany). Multiple spectra were displayed in a mass spectrum window for spectra comparison. To prepare heat-killed B. pseudomallei, bacteria were grown on media including TSA, LB agar, Ashdown agar, blood agar, MacConkey agar and M9 minimal agar at 37 oC in air for 16–18 h, harvested by obtaining two 10 μl loops and suspended in 1 ml of sterile phosphate-buffered saline (PBS) pH 7. 4. The bacterial suspension was centrifuged at 10,000 × g for 10 min and washed twice with 1 ml of PBS. The pellet was resuspended in 1 ml PBS. One hundred microliters of bacterial suspension was taken to make a ten-fold serial dilution in PBS. The colony count was performed by spreading 100 μl of each dilution onto TSA plates in triplicate and incubating at 37 oC for 16–18 h. The remaining bacterial suspension was heated at 80 oC for 1 h. One hundred microliters of heat-killed bacteria were plated on TSA in duplicate for a sterility test. Lipopolysaccharide (LPS) was extracted using a modified hot water-phenol method as previously described [25]. Briefly, each of four B. pseudomallei isolates (strains K96243,1026b, 153 and 164) and B. pseudomallei K96243 derivative with ΔwbiD (mutant defective in OPS synthesis) were cultured on TSA for 50 plates and incubated at 37 oC for 16–18 h. Bacteria were scraped off using a 10 μl loop and suspended in 75 ml of distilled water (DW). The bacterial suspension was mixed with 90% phenol solution at a ratio of 1: 1, heated at 80 oC for 5 min with constant mixing, and subsequently cooled down at room temperature. The extract was dialyzed in a dialysis tube with pore size 6–8 kDa (Spectrum, CA, USA) against DW. The dialysate was centrifuged at 8,000 × g for 20 min and the supernatant was lyophilized. The dried material were solubilized in 10 mM Tris-HCl (pH 7. 5) buffer containing 1 mM MgCl2,1 mM CaCl2 at ratio of 30 mg: 1 ml and sequentially treated with RNase and DNase at final concentration of 50 μg/ml at 37 oC with agitation for 3 h, and proteinase K (Invitrogen, CA, USA) at final concentration of 50 μg/ml at 60 oC for overnight. LPS was isolated from the supernatant by ultracentrifugation (Beckman Coulter) at 100,000 × g at 4 oC for 3 h. The gel-like pellet was suspended in 10 ml pyrogen-free water and lyophilized. The purified LPS was further analyzed using SDS-PAGE and silver staining as previously described [8]. Protein contamination was examined using Coomassie blue staining [18] and bicinchoninic acid (BCA) assay (Pierce, IL, USA). The bacteria were cultured on 10 plates of TSA at 37 oC for 16–18 h, harvested by 10 μl loop and suspended in 18 ml of PBS. Bacteria were inactivated by adding phenol to obtain 1% final concentration, and the bacterial suspension was incubated at 37 oC for overnight. The cell pellet was collected by centrifugation at 12,000 × g for 3 min and wash twice with 20 ml of PBS. The pellet was resuspended in 5 ml PBS. Five hundred microliters of bacteria suspension were plated on TSA for sterility test. The remaining bacterial suspension was lyophilized and collected as a dried bacterial cell. LPS fatty acids were derivatized to fatty acid methyl esters prior to gas chromatography (GC) analysis as previously described [26]. 20 mg of dried bacterial cells were suspended in 500 μl of DW. Five hundred microliters of 90% hot phenol (80 oC) was added and further incubated at 70 oC for 1 h with constant mixing. The cell suspension was incubated on ice 5 min and centrifuged at 10,000 × g for 10 min. The upper aqueous phase was obtained and washed with 2 ml of diethyl ether (Fisher Scientific, NJ, USA) by centrifugation at 3,000 × g for 5 min. Then, the lower aqueous phase containing LPS was collected and lyophilized. The 200 μg of dried LPS were treated with 200 μl of 2 M anhydrous methanolic HCL (Alltech, KY, USA) at 90 oC for 18 h. Then, the derivatized fatty acid methyl ester was cooled down at room temperature. Two hundred microliters saturated NaCl solution was added. The mixture was subsequently extracted with 400 μl of hexane (Sigma-Aldrich, MO, USA) by mixing using a vortex for 30 sec. The upper layer phase (~400 μl) was analyzed by a gas chromatography-flame ionization detector (GC-FID), HP 5890 series II with a 7673 autoinjector. Bacterial acid methyl ester CP mixture (Matreya, PA, USA) was used as standard control and pentadecanoic acid (Sigma Aldrich, MO, USA) was used as an internal control respectively. Lipid A was liberated from purified LPS by hydrolysis. Purified LPS was dissolved in 10 mM sodium acetate (pH 4. 5) with 1% SDS and then heated at 100 oC for 1 h. The sample was lyophilized and then washed with 170 μl of endotoxin free water and 850 μl of acidified ethanol mixture by centrifugation at 5,000 × g for 5 min. The sample was further washed with 1 ml of non-acidified 95% ethanol and followed by 1 ml of 100% ethanol. Finally the sample was lyophilized to yield lipid A. The structural characterization was analyzed on a Waters Synapt G2 QqTOF mass spectrometer (Waters Corporation, Milford, MA, USA) operated in sensitivity mode and negative ion mode. The lipid A was dissolved in a solvent mixture of chloroform and methanol (2: 1 v/v). The lipid solution was infused at 3 μl/min flow rate. The source block temperature was set to 150 oC. Tandem MS was carried out using trap collision induced dissection (CID). The instrument standard values (LM resolution 4. 7 and HM resolution 15. 0) were used for mass selection in the quadrupole. To mitigate the effect of instantaneous signal fluctuations, the data were averaged for 2–3 min. TLR4-transfected human embryonic kidney cells (HEK-BlueTM-hTLR4 cells) were purchased from InvivoGen (San Diego, CA, USA). The HEK-BlueTM-hTLR4 cells are HEK cells that are stably transfected with human TLR4 (hTLR4), myeloid differentiation factor-2, cluster of differentiation-14 (MD-2/CD14) and a secreted embryonic alkaline phosphatase (SEAP) reporter gene. The cells were cultured and maintained at 37 oC with 5% CO2 in complete Dulbecco’s modified Eagle medium (DMEM) (Gibco, NY, USA) which contained 10% heat inactivated fetal bovine serum (FBS) and 1× HEK-Blue selection medium (InvivoGen, San Diego, CA, U. S. A); an antibiotic mixture for maintenance of HEK-Blue hTLR4 cell lines. The cells were seeded at 2. 5 × 104 cells/well in a 96-well plate and stimulated for 24 h with heat-killed bacteria at 106 or 107 CFU/ml and purified LPS at 1,10,100,1000,10,000 ng/ml from B. pseudomallei isolates. Cells stimulated with ultrapure TLR4 ligand Escherichia coli O111: B4 LPS (Sigma-Aldrich, MO, USA) was a positive control. The activation of nuclear transcription factor (NF) -κB in HEK-Blue hTLR4 cells in response to TLR4 agonists was determined by a SEAP reporter assay at a wavelength 620 nm using a microplate reader (TECAN, Grodig, Austria). The results of NF-κB activation in HEK-Blue hTLR4 cells were calculated to account for the difference in weight of wild type LPS and ΔwbiD mutant LPS by a factor 2. 5. The mass individually of the lipid A (1948), core (1258), and repeating O-antigen unit (423+ 4250 = 4673) were used to calculated the weight of wild type LPS. The mass individually of the lipid A (1948) and core (1258) were used to calculated the weight of the mutant LPS. Susceptibility testing to polymyxin B was performed with B. pseudomallei clinical isolates using a broth microdilution method according to the Clinical and Laboratory Standards Institute (CLSI) guidelines [27]. The bacterial isolates used in this experiment were randomly selected from different groups of infections that the lipid A results showed slightly different mass intensity at m/z 1801. 6. These included 14 isolates from 7 Thai patients with primary infection (strains: 226a, 855e, 1234a, 1304a, 1620a, 1790a and 2944b) and relapse after recovery from primary infection (strains: 240a, 855h, 1234b, 1304b, 1620d, 1790b and 2944d) and 7 isolates from within-host infection [H3921b (C5), H3921c (C3), H3921d (C2), H3921e (C2), H3921f (C2), H3921g (C3), H3921h (C5) ]. The MIC interpretive criteria used for polymyxin B were established in CLSI standard M100-S24 as follows: susceptible ≤ 2 μg/ml, intermediate 4 μg/ml and resistant ≥ 8 μg/ml [28]. Pseudomonas aeruginosa ATCC 27853 was used as a control strain. B. pseudomallei isolates were recovered from the freezer vials at -80 oC by streaking onto Columbia agar and incubating aerobically at 37 oC for 24 h. Bacterial colonies were harvested and suspended in normal saline 0. 85%. Bacteria suspension was adjusted at OD600 nm to obtain a concentration of 1 × 108 CFU/ml. Bacteria at a final concentration of 5 × 105 CFU/ml as recommended in CLSI document M07-A9 [29] were used for susceptibility testing of polymyxin B (catalog number P4932; Sigma-Aldrich) at concentrations of 0,0. 25,0. 5,1, 2,4, 8,16,32,64,128,256, and 512 μg/ml in duplicate. The MIC was read as the lowest drug concentration at which no visible growth was observed following aerobic incubation at 35 ± 2 oC for 16–20 h. Four B. pseudomallei mutants including three OPS mutants and a capsule mutant were constructed from wild types K96243 and 4095A, respectively using a fragment mutagenesis method as described [30] (S1 Table). K96243ΔwbiD was defective in wbiD and did not synthesize the OPS, K96243ΔwbiA and K96243ΔoacA were both defective in acetylation of the OPS. The capsule mutant (4095a ΔwcbB) was defective in wcbB and did not have capsule [30]. Characterization of OPS was examined by SDS-PAGE of proteinase K extracts, silver staining and Western blot using a monoclonal antibody to OPS [8]. Characterization of capsule was examined by latex agglutination and Western blot using a monoclonal antibody to CPS [31]. Statistical analysis was performed using GraphPad Prism (version 5; GraphPad Software). One-way ANOVA was used to test the difference among the means of NF-κB activation from different B. pseudomallei isolates. Unpaired t tests were used to test the difference between the mean of results from B. pseudomallei wild type and mutant strains. The data were presented as mean ± standard deviations. Differences were considered statistically significant at a P value < 0. 05. We initially analyzed the lipid A isolated from the reference strain B. pseudomallei K96243. Negative ion mass spectra of K96243 lipid A showed a complex pattern of ions with major ions at m/z 1511,1575,1590,1670,1686,1698,1714,1721,1801,1817,1954,1970 (Fig 1A). The predicted structures of fatty acid, phosphate, and carbohydrate substituents on lipid A of K96243 backbone are shown in Fig 2 and Table 1. Based on overall chemical compositions, the data predicted that B. pseudomallei contained penta- and tetra-acylated lipid A species. The ion at m/z 1511 was a representative of tetra-acylated monophosphorylated GlcN disaccharide backbone possessing two C14: 0 (3-OH) residue in ester linkage and two C16: 0 (3-OH) residues in amide linkage with one phosphorylated and one 4-amino-4-deoxy-arabinose (Ara4N) attached to the phosphate group. The ion at m/z 1575 was a representative of tetra-acylated bisphosphorylated GlcN disaccharide backbone possessing one C14: 0 residue and one C14: 0 (3-OH) residue in ester linkage and two C16: 0 (3-OH) residues in amide linkage with two phosphorylated and one Ara4N attached to the phosphate group. The ion at m/z 1590 was a penta-acylated monophosphorylated GlcN disaccharide backbone possessing one C14: 0, two C14: 0 (3-OH) and two C16: 0 (3-OH) residues with one phosphorylated. The ion at m/z 1670 was a penta-acylated bisphosphorylated GlcN disaccharide backbone possessing one C14: 0 residue, two C14: 0 (3-OH) residues and two C16: 0 (3-OH) residues with two phosphorylated. The ion at m/z 1686 was a representative of ion at m/z 1670 with the substitution of fatty acid C14: 0 for C14: 0 (2-OH). The ion at m/z 1721 was a representative of ion at m/z 1590 modified with one Ara4N residue attached to the phosphate group (Δm/z + 131). The ion at m/z 1801 was a representative of ion at m/z 1670 modified with an Ara4N residue attached to the phosphate group. The ion at m/z 1817 was a representative of ion at m/z 1686 modified with an Ara4N residue attached to the phosphate group. The ion at m/z 1954 was a representative of ion at m/z 1670 with sodium adduct and modified with two Ara4N residues attached to the phosphate group. The ion at m/z 1970 was a representative of ion at m/z 1686 with sodium adduct and modified with two Ara4N residues attached to the phosphate group. The ions at m/z 1698 and 1714 were detected only by MALDI-TOF and could not be detected for structural identification by ESI-QqTOF. The total fatty acid of B. pseudomallei K96243 lipid A was validated by gas chromatography (GC). The result is presented in Fig 3. Fatty acid composition of B. pseudomallei strain K96243 LPS confirmed the MALDI-TOF MS results for the presence of tetradecanoic acid (C14: 0), 2-hydroxytetradecanoic [C14: 0 (2-OH) ], 3-hydroxytetradecanoic acid [C14: 0 (3-OH) ], hexadecanoic acid (C16: 0), and 3-hydroxyhexadecanoic acid [C16: 0 (3-OH) ]. Our data suggest that the lipid A species of B. pseudomallei K96243 was predominantly penta-acylated with a combination of C14: 0 (2-OH), C16: 0 (3-OH), C14: 0 (3-OH) or C14: 0. ESI-QqTOF MS analysis with negative ion mode of B. pseudomallei K96243 lipid A showed major ions at m/z 1574. 8,1590. 0,1606. 0,1670. 0,1685. 9,1721. 1,1801. 0,1817. 1,1932. 1 and 1948. 1 (Fig 1B). ESI results, where ESI-QqTOF gives higher mass accuracy, as compared to MALDI-TOF are shown at Fig 1B. The two MS results are very reproducible. These ions were further characterized by trap collision-induced dissociation (S1 Fig) and the tandem MS results confirm the predicted structures (Fig 2). The ion at m/z 1606. 0 was a representative of m/z 1590. 0 with the substitution of fatty acid C14: 0 for C14: 0 (2-OH). The ion at m/z 1932. 1 was a representative of ion at m/z 1670. 0 with two Ara4N residues. The m/z 1698 and 1714 were detected only by MALDI-TOF and could not be detected for structural identification by ESI-QqTOF. We next investigated the variation of lipid A structure in 136 clinical isolates of B. pseudomallei grown on TSA plates. The mass spectrum of lipid A was determined in two groups of existing Thai isolate collections. The first group represented primary isolates (S2 Fig) and relapse isolates (S3 Fig) from 68 melioidosis patients who had at least one episode of relapse. MALDI-TOF lipid A spectra of representatives of these isolates are shown in Fig 4. Using a negative reflector mode MALDI-TOF MS analysis, we observed similar lipid A spectra in a mass range of 1500–2000 Da in all isolates from both primary and relapse infections. All lipid A species of these isolates contained four main clusters of peaks around m/z of 1575,1670,1801 and 1954 corresponding to tetra-acylated modified with one phosphorylated and one Ara4N group, bisphosphorylated penta-acylated, bisphosphorylated penta-acylated modified with only one Ara4N group and bisphosphorylated penta-acylated modified with two Ara4N group, respectively. The second group consisted of 35 B. pseudomallei isolates from a variety of specimen type (blood, tracheal suction, urine, pus from right leg, pus from left leg, pus from forehead, and wound swab from thigh) of a Thai patient with acute melioidosis. The lipid A species of these within-host isolates were not different (S4 Fig), and represented the same four main clusters of peaks at approximately m/z of 1575,1670,1801 and 1954. Together, these comprehensive set of data indicate a striking lack of variation in lipid A structure in a large group of clinical isolates cultured from patients with primary melioidosis, as well as in isolates obtained during relapse infection, and from multiple isolates obtained from numerous sites within a single host. We considered the possibility that the lipid A structures of B. pseudomallei isolates may be modified during growth on TSA at 37 oC, the conditions used in the initial experiments. To address this, we tested whether the lipid A structure of B. pseudomallei K96243 differed following culture under a range of different laboratory conditions. All of lipid A spectra of B. pseudomallei cultured in different laboratory conditions had the same four main clusters of peaks at approximately m/z of 1575,1670,1801 and 1954 (S5 Fig). These data indicate that B. pseudomallei K96243 lipid A molecule was structurally conserved and suggested that the lack of variability in lipid A structure observed in our 136 clinical isolates was not a result of identical growth conditions. Cationic antimicrobial peptides (CAMPs), such as polymyxin B exerts their activity through electrostatic interactions with the lipid A moiety of LPS causing disruption of the outer membrane and cell death. Bacteria can develop resistance to colistin by modifying the structures of the lipid A moiety, hindering colistin binding. Aminoarabinose (AraN4) modification of pathogenic bacteria lipid A has been related to resistance to polymyxin B [32,33]. B. pseudomallei has been reported to be resistant to polymyxin B [34]. Our lipid A structural characterizations showed that B. pseudomallei lipid A species were modified with Ara4N residues at the terminal phosphate groups that was observed by the ion at m/z 1801, and other associated ions. We postulated that the varying ion intensity of lipid A modified with Ara4N may contribute to variable polymyxin B resistance. To test the susceptibility of B. pseudomallei to polymyxin B, 21 representative isolates were chosen with notable varying ion intensity at m/z 1801. The results demonstrated all 21 isolates were resistant to polymyxin B with a minimum inhibitory concentrations (MICs) for all isolates ≥ 512 μg/ml. Our previous studies demonstrated that LPS of B. pseudomallei activates the innate immune responses in human monocytes through TLR4 [6,35]. In this study, we considered whether different B. pseudomallei isolates cultured on the same medium differentially activate TLR4-dependent responses. We prepared heat-killed bacteria from four isolates (K96243,1026b, 153 and 164) grown on TSA and stimulated HEK-Blue hTLR4 cells, expressing hTLR4-MD2-CD14 (Fig 5). All heat-killed B. pseudomallei isolates at 106 CFU/ml were able to induce NF-κB activation after 24 h at OD levels between 0. 49–0. 73 (median = 0. 63, IQR = 0. 58–0. 68). At bacterial concentrations of 107 CFU/ml, the NF-κB activation was induced at OD level between 1. 02 and 2. 17 (median = 1. 48, IQR = 1. 24–1. 72), respectively. We demonstrated no significant difference in TLR4-dependent NF-κB activation in cells stimulated with the four heat-killed bacterial isolates at 106 CFU/ml (P = 0. 12) or 107 CFU/ml (P = 0. 38). To test whether B. pseudomallei isolates cultured on various media differentially activate TLR4-dependent responses, we prepared heat-killed B. pseudomallei K96243 grown on six different agar media (TSA, LB, Ashdown agar, blood agar, MacConkey agar, M9 minimal medium agar) and stimulated HEK-Blue hTLR4 cells with bacteria at 106 and 107 CFU/ml (Fig 6). We demonstrated no significant difference in NF-κB activation induced by heat-killed B. pseudomallei K96243 on different media at 106 CFU/ml (P = 0. 47). However, NF-κB activation induced by 107 CFU/ml of heat-killed B. pseudomallei K96243 cultured on M9 minimal medium agar was significantly lower than that induced by K96243 cultured on other media (P < 0. 0001). We previously demonstrated that LPS is a main driver of the innate immune response to B. pseudomallei [6]. Therefore, we further investigated whether LPS from different isolates potentially induce variable NF-κB activation. SDS-PAGE and silver stain of purified LPSs from all four isolates showed typical patterns of LPS type A (Fig 7). The Coomassie blue staining and BCA assay results showed no protein contamination. We determined NF-κB activation of HEK-Blue hTLR4 cells at 24 h after stimulation with LPSs from four B. pseudomallei isolates (K96243,1026b, 153 and 164) at concentrations of 1,10,100,1000, and 10,000 ng/ml. The LPSs of all isolates induced a significant increase of NF-κB activation in HEK-Blue hTLR4 cells above the baseline level in dose-dependent manner (Fig 8). Our results showed no significant difference among different B. pseudomallei isolates in TLR4-dependent NF-κB activation in cells stimulated with LPSs at 1 ng/ml (P = 0. 68), 10 ng/ml (P = 0. 14), 100 ng/ml (P = 0. 33), 1000 ng/ml (P = 0. 28) or 10,000 ng/ml (P = 0. 29). Although lipid A is the ligand for TLR4 [35], we hypothesized that there may be an effect of the OPS component of the LPS on this interaction. To investigate whether the presence of OPS on B. pseudomallei cells was associated with altered TLR4 signaling, we stimulated HEK-BlueTM-hTLR4 cells with heat-killed wild type K96243 and OPS mutant strains containing unmarked deletions of wbiD, wbiA, and oacA, which are involved in B. pseudomallei OPS biosynthesis [30]. We have previously shown that–like wild type B. pseudomallei K96243 –K96243 ΔwbiA and ΔoacA have type A ladder patterns, whereas K96243 ΔwbiD completely lacks OPS [30]. As shown in Fig 9, NF-κB activation was significantly higher in cells stimulated with 106 (P < 0. 001) and 107 (P < 0. 01) CFU/ml of B. pseudomallei K96243 ΔwbiD compared to those of wild type. We demonstrated no significant difference in TLR4-dependent NF-κB activation in cells stimulated with heat-killed bacteria obtained from B. pseudomallei K96243 ΔwbiA and ΔoacA compared to wild type at 106 CFU/ml or 107 CFU/ml. These data show that the absence of OPS does permit enhanced TLR4 signaling by heat killed B. pseudomallei. To further test whether the presence of capsule of B. pseudomallei cells interfered with TLR4 signaling, we stimulated HEK-BlueTM-hTLR4 cells with heat-killed wild type strain 4095a and a mutant containing unmarked deletions of wcbB (4095a ΔwcbB), which is involved in B. pseudomallei capsule biosynthesis [30]. Confirmatory testing with latex agglutination, SDS-PAGE, and Western blot of heat-killed bacteria indicated that the 4095a ΔwcbB did not express capsule. The results in Fig 10 demonstrated that NF-κB activation was not different in cells stimulated with 106 (P = 0. 90) and 107 (P = 0. 37) CFU/ml of B. pseudomallei 4095a ΔwcbB, as compared to those of wild type. Thus, the deletion of the capsule gene does not affect TLR4 signaling by heat killed B. pseudomallei. We then tested the impact of the OPS mutant on TLR4-dependent NF-κB activation by B. pseudomallei LPS. LPS was extracted and verified by SDS-PAGE and silver staining confirming the structural difference in LPS between wild type and isogenic mutant. The results in Fig 11A revealed ladder pattern with the presence of OPS for K96243 wild type but not for the K96243 ΔwbiD mutant, indicating the lack of OPS biosynthesis [30]. Stimulation of HEK-BlueTM-hTLR4 cells with different concentrations of LPSs from wild type and ΔwbiD mutant demonstrated that NF-κB activation was significantly higher in cells stimulated with ΔwbiD LPS at 0. 1 ng/ml (P < 0. 001) and 1 ng/ml, (P < 0. 0001) compared with NF-κB activation in cells stimulated with wild type LPS. We observed no difference in NF-kB activation between cells stimulated with 10 ng/ml LPS of wild type and cells stimulated with 10 ng/ml LPS of the ΔwbiD mutant. We noted that the lipid A structure of K96243 ΔwbiD mutant and the wild type were identical (S6 Fig). These data provide additional confirmation that the absence of OPS allows enhanced TLR4 signaling by B. pseudomallei LPS. Human melioidosis is characterized by hyperinflammatory responses leading to high mortality rates even in patients who have received appropriate antimicrobial therapy. The initial interaction between B. pseudomallei and innate immune receptors such as the LPS-TLR4 interaction is important in the immunopathogenesis and outcome of infection. In this study, we characterized lipid A structure of B. pseudomallei isolates by MALDI-TOF MS followed by ESI-QqTOF MS and GC and examined the innate immune responses to the different bacterial isolates. Our data revealed the same lipid A profile for B. pseudomallei from a large collection of diverse isolates of LPS type A and demonstrated minimal variability in TLR4 signaling. Despite being cultured in different laboratory-induced conditions, B. pseudomallei K96243 lipid A expressed similar lipid A profiles, as characterized by MALDI-TOF MS. Moreover, we demonstrated that B. pseudomallei OPS interfered with TLR4 activation. Our data show that the major lipid A species of B. pseudomallei contain a mixture of tetra- and penta-acylated lipid A species that are non-stoichiometrically substituted with Ara4N residues at both phosphate groups. We observed the presence of fatty acid C14: 0, C14: 0 (2-OH), C14: 0 (3-OH), C16: 0, and C16: 0 (3-OH). In comparison with other bacterial species, the lipid A structure of B. pseudomallei is unique. Previous published data showed that the fatty acid C14: 0 (2-OH) substituted into lipid A backbone of B. pseudomallei was not present in other closely related Burkholderia species, such as B. thailandensis, B. cepacia and B. mallei [17,36,37]. Although genomic analysis of within-host isolates showed substantial divergence from the founder genotype during a short period of acute infection [22], we found a remarkable lack of lipid A structural variation in 35 clinical B. pseudomallei isolates cultured from multiple body sites from a single patient with disseminated melioidosis. The data suggest that the lipid A is essential for B. pseudomallei and the modification during short period of acute infection is not required for bacterial fitness. Lipid A has an important role as it is the anchor for LPS on the outer membrane of B. pseudomallei. We previously showed that the synthesis of lipid A molecules is of vital importance among the various components that are responsible for outer membrane assembly [38]. Surprisingly, our study revealed similar lipid A spectra of 136 clinical isolates obtained from primary and relapse infections of 68 melioidosis patients, thus confirming that all B. pseudomallei isolates in our collection expressed a similar lipid A structure. In comparison to B. pseudomallei, B. mallei also produce both tetra- and penta-acylated lipid A species that are potent stimulators of hTLR-4-dependent cytokine production [36]. Other studies demonstrated that B. cenocepacia strain LMG 12614 expressed only penta-acylated lipid A species and it can induce stronger TNF-α and IL-6 responses than B. multivorans strain LMG 14273 that expressed both tetra- and penta-acylated lipid A [39]. Gram-negative organisms have evolved several LPS modification that benefit these organisms in their interactions with the host innate immune system and hostile environments during infections. The alteration of lipid A structure can promote resistance to antimicrobial peptides and interfere with host recognition by the innate immune system. Alterations are accomplished by many enzymes that modify the lipid A moieties by adding or removing acyl chains and phosphate groups. In some bacteria, lipid A are modified to evade immune recognition and survive within a host by reducing the charge of the bacterial surface via adding sugar onto lipid A molecules. The addition of phosphoethanolamine and aminoarabinose to lipid A molecules can protect bacteria against cationic antimicrobial peptides (CAMPs) [40,41]. In Salmonella serovar Typhimurium, aminoarabinose modification at both phosphate groups have increased resistance to cationic antimicrobial peptides [11]. Our characterization of lipid A revealed that the lipid A species of all B. pseudomallei isolates were modified with aminoarabinose residues, thus demonstrated the resistance to polymyxin B with MIC value >512 μg/ml. We observed no alteration of lipid A structure when B. pseudomallei strain K96243 was cultured under standard laboratory-induced conditions. However, we observed a slight decreased TLR4-dependent NF-κB activation after culture using M9 media. A previous study demonstrated that a variety of growth conditions alter the expression of Escherichia coli LPS polysaccharide formation especially in nutrient-depleted conditions [42]. The changes in LPS expression may extend to lipid A synthesis and result less TLR4-dependent NF-κB activation. In contrast, various other Gram-negative bacteria modify their lipid A structure when exposed to the external environment in vitro. For example, temperature-dependent structural changes in lipid A structure have been shown in Porphyromonas gingivalis and Yersinia species. P. gingivalis expresses mainly nonphosphorylated and monophosphorylated tetra-acylated lipid A structure at normal body temperature, whereas the major lipid A produced are monophosphorylated penta-acylated lipid A when the bacteria are grown at 41 oC. The temperature-dependent alteration in P. gingivalis lipid A structure is associated with more potent TLR4-dependent innate immune activation by LPS from bacteria grown at 41 oC than LPS from bacteria cultured at 37 oC [12]. Three pathogenic Yersinia; including Y. pestis, Y. enterocolitica, and Y. pseudotuberculosis synthesize predominantly hexa-acylated lipid A at 21 oC which induces higher levels of cytokine production compared to tetra-acylated lipid A at 37 oC [14]. Additionally, Salmonella enterica serovar Typhimurium undergoes hydroxylation of fatty acid C14: 0 to fatty acid C14: 0 (2-OH) when grown under magnesium-depleted conditions [11]. We also observed that the heat-killed bacteria and LPSs from different B. pseudomallei isolates activated human TLR4 and we demonstrated no significant difference among different B. pseudomallei isolates in TLR4-dependent NF-κB activation. Similarly, LPS from four clinical isolates resulted in similar TLR4-dependent NF-κB activation. These results are concordant with the invariable MALDI-TOF results characterizing B. pseudomallei lipid A. While increased natural diversity or heterogeneity of specific components of LPS, such as lipid A, can produce dramatic changes in host responses, this is not the case for B. pseudomallei. Our results differ from recent reports. Norris et al. described different lipid A profiles among each of different LPS types in B. pseudomallei and the effect on innate immune activation suggesting that clinical strains with different LPS types may be the cause of clinical outcomes. However, their study used only one strain of B. pseudomallei for each LPS type (type A, B, B2 and rough) [19]. Our study was conducted entirely using LPS type A, which is predominant in Thailand, yet despite analysis of LPS from over 170 clinical isolates, found no differences that could potentially explain clinical variations. Another reason for the discrepancy in results may be due to the different MALDI-TOF matrix used. We used norharmane as a matrix for MALDI-TOF in contrast to the 2,5-dihydroxybenzoic acid used by Norris and colleagues. Norharmane is an optimum matrix for overcoming limitations of lipid A detection [24]. Despite our previous identification of B. pseudomallei lipid A as a TLR4 agonist [35], we also show here that TLR4-dependent NF-κB activation induced by an OPS mutant (rough LPS) is significantly greater than those induced by LPS from wild type B. pseudomallei. However, the response induced by a B. pseudomallei capsule mutant was comparable to those induced by the wild type. This observation is concordant with findings from a previous study showing that rough B. pseudomallei LPS induces higher nitric oxide and TNF-α levels in RAW 264. 7 macrophages than smooth LPS [19], and suggests that there may be interference by the OPS on the LBP/CD14-mediated presentation or binding of LPS to the TLR4/MD-2 complex [9]. In conclusion, we have demonstrated that the structural features of B. pseudomallei lipid A are extremely conserved among different clinical B. pseudomallei isolates with LPS type A and result in no variation in TLR4-dependent innate immune activation. Further studies are required to evaluate the lipid A structure of environmental isolates and the role of OPS in the LPS-TLR4 interaction.

Melioidosis is a tropical infection caused by Burkholderia pseudomallei. Disease ranges from mild to lethal, and patients who die from melioidosis have greater inflammatory responses than survivors. Identifying variation in bacterial structures that alter the immune response is therefore important. B. pseudomallei expresses lipopolysaccharide, a major stimulator of the immune response. Studies of other bacteria suggest that differences in the lipid A component of lipopolysaccharide result in variable stimulation of the immune system. In this study, we analyzed the chemical structure of lipid A of 171 clinical B. pseudomallei isolates, structural changes caused by different growth conditions, and innate immune responses induced. Surprisingly, we did not observe any variation of lipid A structure or different innate immune activation from different B. pseudomallei isolates. Our data suggest that lipid A structure of B. pseudomallei is highly conserved and that different inflammatory responses in patients may be caused by other factors.

Abstract Introduction Materials and methods Results Discussion

medicine and health sciences chemical compounds phosphates melioidosis immunology bacterial diseases lipid structure lipid analysis research and analysis methods infectious diseases lipids chemistry mass spectrometry molecular biology immune response matrix-assisted laser desorption ionization time-of-flight mass spectrometry biochemistry analytical chemistry biology and life sciences physical sciences fatty acids spectrum analysis techniques macromolecular structure analysis

2018

Comprehensive analysis of clinical Burkholderia pseudomallei isolates demonstrates conservation of unique lipid A structure and TLR4-dependent innate immune activation

The cohesion of sister chromatids is mediated by cohesin, a protein complex containing members of the structural maintenance of chromosome (Smc) family. How cohesins tether sister chromatids is not yet understood. Here, we mutate SMC1, the gene encoding a cohesin subunit of budding yeast, by random insertion dominant negative mutagenesis to generate alleles that are highly informative for cohesin assembly and function. Cohesins mutated in the Hinge or Loop1 regions of Smc1 bind chromatin by a mechanism similar to wild-type cohesin, but fail to enrich at cohesin-associated regions (CARs) and pericentric regions. Hence, the Hinge and Loop1 regions of Smc1 are essential for the specific chromatin binding of cohesin. This specific binding and a subsequent Ctf7/Eco1-dependent step are both required for the establishment of cohesion. We propose that a cohesin or cohesin oligomer tethers the sister chromatids through two chromatin-binding events that are regulated spatially by CAR binding and temporally by Ctf7 activation, to ensure cohesins crosslink only sister chromatids. Proper transmission of eukaryotic chromosomes during cell division requires DNA replication and three other DNA-dependent processes: recombination-dependent DNA repair, sister chromatid cohesion, and chromosome condensation. Each of these diverse processes requires protein complexes containing two members of the highly conserved structural maintenance of chromosomes (Smc) family of proteins [1–3]. Smc complexes likely share a common core activity of chromosome crosslinking, either within a chromosome, as in chromosome condensation, or between chromosomes, for sister chromatid cohesion and recombination-dependent DNA repair. How Smc complexes mediate chromosome crosslinking is unknown. Smc molecules are composed of five structural domains (Figure 1A) [4,5]: a globular N-terminal domain containing a Walker A motif, a globular C-terminal domain with Walker B and Signature motifs, two long α-helical domains, and a globular Hinge domain. Smc monomers fold in half at the Hinge domain, allowing the two α-helices to form a long antiparallel coiled-coil domain [6]. This folding juxtaposes the N- and C-terminal globular domains and the Walker A and B motifs, creating an Smc head domain with ATPase activity. Folded Smc monomers resemble a flexible dumbbell, with the Hinge and head domains separated by ∼40 nm of coiled coil [6,7]. Smc complexes are composed of two Smc molecules, a kleisin subunit, and at least one accessory protein [6,8, 9]. Smc monomers dimerize primarily through interactions between their Hinge domains [6,10]. The head domains of Smc molecules are also tethered together through the shared binding of a single kleisin subunit and two ATP molecules [6,11,12]. These interactions at both the head and Hinge domains give Smc dimers the potential to form large rings that have been observed in preparations of purified Smc complexes [13]. One of the most studied Smc complexes is cohesin, which mediates sister chromatid cohesion. Cohesin is composed of Smc1, Smc3, Scc3, and the kleisin subunit Mcd1/Scc1 [14–18]. The association of cohesin molecules with chromatin requires the integrity of all cohesin subunits and the ability of the Smc molecules to bind and hydrolyze ATP [19–22]. Multiple cohesins bind proximal to each centromere, forming a large pericentric domain. Cohesins also bind along chromosome arms. In budding yeast, these cohesin-associated regions (CARs) extend over approximately 1 kb of DNA and are spaced at roughly 10-kb intervals [23–25]. Several auxiliary factors contribute to the establishment, maintenance, and eventual dissolution of sister chromatid cohesion. The binding of cohesin to chromosomes at any phase of the cell cycle requires the loading factors Scc2 and Scc4 [26]. However, only cohesin binding in S phase, coupled with the function of Ctf7, results in the establishment of cohesion [16,27]. Pds5 binds cohesin and helps maintain cohesion during S and G2 phases of the cell cycle [28–30]. Finally, separase promotes removal of cohesin at the onset of anaphase by cleavage of Mcd1 [31]. These observations have led to two distinct models to explain how cohesin molecules crosslink sister chromatids. The embrace model posits that cohesin rings encircle chromosomes prior to replication and make no specific contacts with chromatin [6,20]. A topological interaction between cohesin and chromatin is supported by the fact that cohesin can be released from chromatin either by a single cleavage of the DNA or a single proteolytic cleavage of a cohesin subunit [19,20,32]. Passage of the replication fork through cohesin rings leaves both sister chromatids trapped inside, establishing sister chromatid cohesion. In contrast, oligomerization models, on the basis of observations of other Smc complexes, posit that cohesins bind to both sister chromatids. Then, cohesins on one sister chromatid oligomerize with cohesins on the other sister chromatid to generate cohesion [33–36]. To resolve these and other models will require a better understanding of how cohesins bind chromatin and the relationship between this chromatin binding and establishment of cohesion. Mutagenesis of cohesin subunits provides one approach to gain insight into the chromatin binding of cohesin mutants. Indeed, site-directed mutagenesis of conserved Smc1 residues in the Hinge, Walker A, Walker B, and Signature motifs demonstrated that Smc1 must bind Mcd1 and Smc3, as well as bind and hydrolyze ATP in order for cohesin to bind chromatin [7,21,22,37]. While informative, these mutational analyses of cohesin–chromatin association leave key questions unanswered. Do these mutations define all the domains of cohesin subunits required for chromatin binding? Is chromatin binding of cohesin anywhere on chromatin prior to DNA replication sufficient to generate cohesion? Are additional constraints on cohesin necessary to generate cohesion? To address these types of questions, one needs a way to identify rare mutant forms of cohesin that modulate rather than eliminate its activity. In the past, knowledge of a protein' s structure (sites for modification or interaction with other subunits) has been used to make dominant negative mutants that alter its activity or incorporation into a fully functional complex. We rationalized that the reciprocal would also be true; surveying an entire polypeptide chain for rare insertions with a dominant negative phenotype should provide a highly efficient means to identify precise regions of a protein important for its activity and/or assembly with other subunits. Furthermore, the study of these alleles should be highly informative for elucidating the molecular mechanism of a protein/complex. With this in mind, we screened a library of random insertion mutations in SMC1 for those that cause a dominant negative phenotype in the budding yeast, Saccharomyces cerevisiae. This strategy is henceforth referred to as random insertion dominant negatives (RID). Here, we successfully use RID to identify rare SMC1 alleles that are highly informative in dissecting Smc1′s role in the assembly and function of cohesin. We also use these mutants along with a ctf7 mutant to provide important insights into cohesin and sister chromatid cohesion. As the basis of our RID mutagenesis of SMC1, we constructed a minichromosome containing the SMC1-TAP gene under control of the galactose-inducible GAL1 promoter (see Materials and Methods). The level of Smc1-TAP protein expressed from the GAL1 promoter relative to the endogenous SMC1 promoter decreases 0. 5-fold under uninducing conditions and increases 100-fold under inducing conditions (unpublished data). The Smc1-TAP protein is functional, as it restores viability and normal growth rate to cells carrying the temperature-sensitive smc1–2 allele strain deleted for the essential SMC1 gene under repressing, uninducing, and inducing conditions (Figure S1 and unpublished data). The SMC1-TAP gene was mutagenized by a Tn7-based in vitro system that resulted in five amino acid insertions at random positions within the Smc1 protein. For brevity, the initial SMC1-TAP gene is henceforth referred to as the SMC1 or wild-type allele, and the insertion derivatives are named based upon the position of the insertion in the amino acid sequence. Minichromosomes harboring the mutagenized library of SMC1 were transformed into budding yeast and assayed for their effects on cell viability and cohesion. A total of 13 transformants out of 2,500 candidates showed a marked decrease in cell viability under inducing conditions (Figure S1). The fact that these transformants are viable under uninducing conditions, where significant mutant smc1 expression occurs, indicates that these mutants are not dominant under low expression. To examine whether the mutant proteins are functional, minichromosomes containing inducible wild-type SMC1 or dominant negative smc1 alleles were introduced into cells carrying the temperature-sensitive smc1–2 allele at the endogenous locus [17]. Under uninducing or inducing conditions, the temperature sensitive growth of the smc1–2 strain is complemented by expression of wild-type SMC1, but not the insertion alleles (Figure S1B), indicating that the products of the insertion alleles are defective for Smc1 function under all conditions. The smc1–2 strains were then used to test the ability of the smc1 insertion alleles to generate cohesion. All smc1 insertion alleles are dramatically impaired for the establishment and maintenance of sister chromatid cohesion (unpublished data) (Figure 1B–1D). All 13 insertions mapped to the SMC1 ORF, and 11 of them were unique (Figure 2A and 2B). A total of five insertions map to regions of known functional importance: the Walker A motif (35W), Signature motif (1129S), globular Hinge region (657H), and C terminus of Smc1 (1192C and 1215C) [7,21,22]. The remaining six insertions are either in 209L1–1 and 209L1–2 or cluster around Loop1 (189L1,191L1–1,191L1–2, and 235L1). Loop1 was previously defined through bioinformatics as one of three small regions within the α-helical domains of Smc molecules that are predicted to disrupt coiled-coil formation [38]. Since the insertions within and around Loop1 exhibit similar phenotypes (see below), these insertions define a new functional region of Smc1, which we call the Loop1 region. We wanted to determine the efficacy of RID mutagenesis to define regions of Smc1 important for cohesin assembly. For this purpose we analyzed the ability of the 11 RID smc1 mutants (expressed at physiological levels) to coimmunoprecipitate with Mcd1 and Smc3. By comparison with wild-type Smc1, four RID mutants are dramatically reduced in their ability to coimmunoprecipitate with Mcd1, while all can coimmunoprecipitate Smc3 (Figure 3A and 3B). The failure to identify insertions that block Smc3 association is not surprising, since mutants defective in Smc1/Smc3 dimerization do not interact with any other cohesin subunit [21,22,39] and, therefore, are not likely be dominant negative. The four Smc1 insertions defective for Mcd1 binding lie within the Walker A motif (35W), the Signature motif (1129S), the HH helix (1192C), or the S15 β strand (1215C) (Figures 2B and 3C). These are four of the five motifs known to be required for Mcd1 binding. The Walker A and Signature motifs, through ATP binding, tether together the Smc1 and Smc3 head domains so that they can both associate with a single Mcd1 [21,22,39]. The helix HH and S15 β strand provide two of the three major contacts between Smc1 and Mcd1 (Figure 3C) [39]. These motifs are widely spaced within the polypeptide chain and constitute a target of only ∼2% of the total residues. The efficient identification of these small disperse motifs by RID validates it as an extremely efficient means to identify regions of proteins important for complex assembly. In addition to being a powerful tool to dissect cohesin assembly, we anticipated that RID would also efficiently identify informative alleles for cohesin function. The remaining seven RID alleles in the Hinge and Loop1 regions encode mutant Smc1 proteins that assemble with Smc3 and Mcd1 (Figure 3B) as well as Scc3 (unpublished data). Since they assembled with all known cohesin subunits, they were candidates for alleles that blocked cohesin function. All previously published alleles of cohesin subunits block chromatin binding. Therefore, we tested whether cohesins with these RID smc1 mutants were competent for chromosome binding. Minichromosomes were generated that express wild-type Smc1,657H, or 209L1–2 fused to a 3XHA epitope, again under control of the GAL1 promoter. These alleles will henceforth be referred to as Smc1-HA, 657H-HA, and 209L1–2-HA, respectively. Cultures of smc1–2 cells containing these minichromosomes were released from G1 under conditions that inactivate the smc1–2 protein and induce expression of the galactose-regulated allele. These cells were then arrested prior to anaphase. Nuclei from these cells were spread on slides and processed for immunofluorescence to detect the chromosome association of different Smc1 proteins (Figure 4A). Like wild-type Smc1-HA, the insertion alleles 657H-HA and 209L1–2-HA associate with chromosomes. We also monitored the chromosome localization of epitope-tagged Mcd1, a cohesin subunit, and Pds5, a cohesin accessory factor whose chromatin binding is mediated by cohesin (Figure 4B) [29,30]. Epitope-tagged Mcd1 and Pds5 localize to chromosomes in cells expressing the wild-type 657H or 209L1–2 allele, but not in cells with only inactivated smc1–2 (empty vector). The fact that the 657H and 209L1–2 alleles mediate Mcd1 and Pds5 localization to chromosomes suggests that these smc1 alleles bind chromosomes as part of the cohesin complex. While it is clear that upon induction, the Hinge and Loop1 mutant cohesins can associate with chromosomes, their levels of chromosomal staining are reduced compared to induced wild-type cohesin (Figure 4C). Thus, the first question we wanted to ask was whether this reduction was sufficient to explain the cohesion defect of the Hinge and Loop1 mutants. To assess the level of cohesin binding to chromosomes, cells expressing Smc1,657H, or 209L1–2 were arrested in mitosis. Nuclear spreads were prepared, and the levels of Smc1 and Pds5 immunostaining on chromosomes were quantified (Materials and Methods). The levels of chromosome binding for induced 657H-HA and its associated Pds5 are nearly identical to the binding of Smc1-HA and Pds5 under uninducing conditions (Figure 4C), a level of binding that is sufficient to generate cohesion and normal cell growth. Therefore, the level of chromosome association for 657H complexes should be sufficient to generate cohesion. The levels of 209L1–2-HA binding and its associated Pds5 are reduced by a maximum of 40%. This reduction may be below a threshold needed to generate cohesion. However, 50% of chromosome-bound cohesins in yeast meiosis and 90% in mammalian mitosis can be removed without eliminating cohesion [40,41]. Therefore, the level of chromosome association for 209L1–2-HA complexes is also likely to be sufficient to mediate at least partial sister chromatid cohesion, yet this seems not to be the case (Figure 1C). If the quantity of chromatin binding for the mutant complexes is sufficient to generate cohesion, then the quality of their chromatin binding must be defective. The mutant complexes may bind chromatin by a nonphysiological mechanism. Previous studies have shown that the binding of cohesin to chromatin requires Mcd1 and ATP. The ATP dependence is mediated through the Walker A, Walker B, and Signature motif. To test whether the mutant complexes also bound in an ATP-dependent manner, we constructed Hinge and Loop1 mutants that carried a second insertion allele in the Signature motif. While the protein products of these double mutants are stable (unpublished data), they failed to associate with the chromatin (Figure 4A), indicating that the mutant complexes bind by ATP-dependent mechanism. One method to address whether the mutant cohesin complexes require Mcd1 to bind chromosomes would be to inactivate the temperature sensitive mcd1–1 in strains harboring the smc1–1 and the smc1 loop or hinge mutants. Unfortunately, mcd1–1 and smc1–1 are synthetically lethal. As an alternative, we followed the chromosome association of the Hinge and Loop1 mutant complexes through a cell cycle. Cohesin binds chromosomes only after G1, when Mcd1 is expressed [14,15], and dissociates from chromosomes upon the onset of anaphase, when Mcd1 is cleaved [19]. Cohesin with 657H-HA or 209L1–2-HA is absent from chromosomes in G1, localizes to chromosomes in metaphase, and is lost by the following G1, a pattern identical to cohesin with Smc1-HA (Figure 4D), strongly suggesting that the mutant complexes, like wild-type complexes, require Mcd1. The fact that, like wild- type, the Hinge and Loop1 mutants bind chromatin in an ATP- and Mcd1-dependent manner suggests that they all bind chromatin by a similar mechanism. Despite these similarities to wild-type chromatin binding, the mutant complexes could still fail to generate cohesion if their binding is too late in the S phase to establish cohesion or too unstable to maintain cohesion until M. To test if the mutant cohesins bind chromatin in a timely manner, cells were released synchronously from a G1 arrest and analyzed at different intervals for DNA content and chromosome association of wild-type and mutant cohesins (Figure 5A). The timing of chromosome association during S phase for 657H-HA or 209L1–2-HA cohesins is very similar to that seen for uninduced levels of Smc1-HA cohesin, which is sufficient to generate cohesion. Also, the amount of cohesin bound per nuclei during S phase increased with similar kinetics for Smc1-HA, 657H-HA, and 209L1–2-HA cohesins (unpublished data). Hence, cohesin containing 657H or 209L1–2 exhibits a normal timing of chromosome association. To examine the stability of chromosome binding for cohesin, nuclei were prepared from mitotic cells expressing Smc1-HA, 657H-HA, or 209L1–2-HA and spread in the absence of fixative. Spread nuclei were incubated in ∼50 ml buffer with varying amounts of KCl for 30 min. After incubation, fixative was added, and the association of cohesin with chromosomes was examined by immunofluorescence (Figure 5B). No change in chromosome binding is observed for Smc1-HA in the presence of 150 mM KCl, corroborating previous biochemical analyses that infer stable association between cohesin and chromosomes [32]. Similarly, no significant change in chromosome binding is observed for 657H-HA. Both wild-type and mutant cohesins are extracted completely from spread chromosomes at 250 mM KCl (Figure 5C, unpublished data). Similar results were obtained with the 209L1–2-HA mutant (unpublished data). Therefore, cohesin complexes with wild-type, 657H-HA, or 209L1–2-HA appear to exhibit the same stability of chromosome binding under these conditions. Together, our analyses of the Hinge and Loop1 mutants reveal a new type of cohesin complex. Like wild-type, these mutant complexes bind stably to chromatin, bind chromatin with proper cell cycle timing, and require ATP and Mcd1. However unlike wild-type, these mutant complexes fail to generate cohesion. These mutants also suggest that binding, per se, prior to DNA replication is not sufficient to generate cohesion. Since we could not explain the cohesion defect of these mutants by changes in their general chromosome-binding properties, we tested the specificity of their chromatin binding by examining their enrichment at CARs using chromatin immunoprecipitation (ChIP) [23–25]. Cohesin enrichment was analyzed initially at the centromere of Chromosome III and CARC1, a CAR approximately 11 kb from the centromere (Figure 6A and 6B). Under uninduced levels of Smc1 expression, Mcd1-6HA is enriched at CARC1 and around CEN3 as reported previously [24,25]. Conversely, Mcd1-6HA enrichment at CARC1 and CEN3 is eliminated for cohesin complexes containing 657H or 209L1–2. Similar results were obtained for CARL1 on Chromosome XII (unpublished data). Therefore, cohesin complexes with 657H or 209L1–2 fail to be enriched at CARs. Because these mutant complexes do bind to chromatin as assayed by chromosome spreads, they apparently are bound to sites other than CARs. One possibility is that the mutant cohesins bind to chromatin through the Scc2/Scc4 loading factor, but are trapped in a nonproductive preloading complex. To test this, we examined cohesin binding at Scc2/Scc4 chromatin-binding sites [42], but we observed no enrichment of the mutant cohesins at these sites (Figure S2). The failure to observe enrichment of the mutant cohesins in our ChIP experiments does not exclude the possibility that they bind randomly within these regions. The enrichment of wild-type cohesin at CARs is only 10-fold above background, and CARs are spread at approximately every 10 kb, hence the dispersal of this signal to random sites would dilute the signal to background levels. While the position of the ectopic binding remains to be elucidated, the fact that the Hinge and Loop1 mutants cause ectopic binding indicates that Smc1 plays an active role in the specific localization of cohesins to CARs and pericentric regions. Furthermore, since this mislocalization is the only severe cohesin defect we have been able to observe in the Hinge and Loop1 mutants, it suggests that cohesin localization to CARs is critical for cohesion. Our studies show that the Hinge and Loop1 regions of Smc1 are required for the establishment of cohesion and the enrichment of cohesin at CARs and pericentric regions. The accessory protein Ctf7/Eco1 is also required for the establishment of cohesion [16,27]. In a ctf7 mutant, cohesins pellet with chromatin [16]. This technique does not distinguish between ectopic and specific chromatin binding. Indeed, if ctf7, like Hinge and Loop1 mutants, showed ectopic binding, then this would suggest that Ctf7 interacts with the Hinge and Loop1 regions of Smc1 to ensure specific binding to CARs. To test the function of Ctf7 in cohesin enrichment at CARs, we compared the localization of cohesin on chromatin in wild-type and ctf7 temperature-sensitive strains. Wild-type and ctf7 mutant cells expressing Mcd1-HA as the sole source of Mcd1 function were synchronized at the permissive temperature of 23 °C in G1. These cells were released to the the nonpermissive temperature of 37 °C in media containing hydroxyurea (HU) to inactivate mutant ctf7 protein and block DNA replication [27], respectively. Cohesin association to chromatin was monitored by ChIP (Figure 7A and 7B). In wild-type cells arrested in HU, cohesin is enriched at CARs and pericentric regions as described previously [24,25]. In ctf7 strains arrested in HU at the nonpermissive temperature, cohesin is again enriched at CARs and pericentric regions. Thus Ctf7 is not needed to localize cohesins to CARs or pericentric regions. The above results indicate that Ctf7 function is distinct from that of the Hinge and the Loop1 regions of Smc1. Previous experiments showed that Ctf7 function is required during S phase; however, this activity was not ordered relative to cohesin loading at CAR sites. To test whether Ctf7 is required coincident with or after cohesin binding to CARs, wild-type and ctf7 mutant cells were allowed to progress from G1 to early S (HU arrest) at the nonpermissive temperature and then shifted to the permissive temperature for the remainder of the cell cycle. The results show that cell viability was 93% ± 4% for wild-type cells and 88. 2% ± 5% for ctf7 mutant cells, indicating that Ctf7 function is required after cohesin loads to CARs. This result is strongly supported by cell cycle mapping studies that show inactivation of ctf7 from early S (HU arrest) and prior to the end of DNA replication leads to cell death [16,27,43], indicating that Ctf7 performs its essential function during this window of the cell cycle. Thus, Ctf7 appears to function in the establishment of cohesion after the Smc1-, Loop1-, and Hinge-dependent localization of cohesin to CARs and pericentric regions, and during S phase. Our results validate RID as a very efficient strategy to identify Smc1 alleles that provide important structural and functional information about Smc1 and cohesin. First, of the 11 smc1 RID alleles, ∼40% blocked Smc1 association with Mcd1 but not Smc3. Thus, RID efficiently identifies alleles that trap partially assembled cohesin. Furthermore, these Smc1 insertions lie within four of the five small structural elements (each element is approximately ten residues each spread throughout a total of 1,300), which are required for Mcd1 binding. The efficacy and remarkable precision of RID suggests that it can be used effectively on less well-characterized proteins for the de novo identification of candidate regions of that protein that mediate its binding to interacting partners. Second, the remaining seven RID alleles of Smc1 identify a new functional domain (the Loop1 region), and a new function for the Hinge beyond its established function in dimerization. These Hinge and Loop1 alleles also generated cohesin complexes that unlike any previous cohesin mutants retain the ability to bind chromatin. The isolation of these unusual alleles underscores the combined power of using random mutagenesis, which allows one to avoid the inherent biases of directed mutagenesis, and the imposition of the dominant negative phenotype, which allows rapid identification of rare partially functional alleles in a sea of common null mutations caused by misfolding and truncations. Given the success of the RID strategy for Smc1/cohesin, RID should be a useful tool to dissect the structure and function of many multi-subunit complexes. Our study of RID smc1 alleles has provided important new insights into cohesin binding to chromatin. First, our results show that in vivo the Hinge and Loop1 regions are needed for binding to CARs, two domains not implicated in chromatin binding by either the embrace or snap model. The joint requirement for these two domains is even more surprising given their apparent physical separation (∼40 nm) based upon electron micrographs of cohesin [13]. Interestingly, in vitro analyses of bacterial Smc complexes have implicated the hinge domain in single-stranded and double-stranded DNA binding [37,44] and DNA-dependent stimulation of the ATPase activity of the head [44]. Thus, our in vivo study combined with these in vitro studies support DNA-dependent functional interactions between the opposite ends of Smc complexes. The second interesting feature of cohesins containing the RID smc1 mutants is that they appear to bind chromatin ectopically. This conclusion is based upon the observations that cohesins exhibit general chromatin binding as assayed by chromatin spreads but are no longer enriched at CARs. It is possible that the cohesin binding is compromised at CARs in some way that they bind there but subsequently dissociate. Consistent with this, the level of binding of overexpressed RID mutant proteins is reduced compared to overexpressed wild-type protein. However, under our assay conditions using chromatin spreads, the Hinge, Loop1, and wild-type cohesins appear to be bound with similar stability. Furthermore, the level of chromatin binding of cohesin, when the Hinge mutant is overexpressed, is still comparable to the level of binding for uninduced wild-type, which is sufficient for cohesion function. Alternatively one could argue that the ectopic binding observed by chromatin spreads reflects some artifact of this assay. This is extremely unlikely, as chromatin binding observed by spreads for the RID mutants share many of the chromatin-binding features of wild-type cohesin, including Mcd1 dependence, ATP dependence, and proper cell cycle loading in S and unloading in M. Together these results suggest that cohesins are capable of general chromatin binding, mediated by Mcd1 binding and ATP functions, but are then targeted to CARs through an additional function (s) mediated by the Hinge and Loop1 regions. In one scenario the Loop1 and Hinge regions may target cohesins to CARs by binding targeting factors that recognize specific chromatin features; the RID mutants perturb the binding of the targeting factors. There are precedents for this idea, since the Hinge and Loop1 regions mediate interactions with non-SMC factors in other Smc-related proteins [45–47]. In addition, histone modifications have already been demonstrated to be critical to target cohesin binding to heterochromatin and double strand breaks [48–50]. In this model, cohesin binding to chromatin resembles RNA polymerase; both are targeted to specific sites through specificity factors that recognize changes in chromatin, and in their absence load with lower efficiency at abundant cryptic sites. As an alternative, cohesin itself may be capable of recognizing specific features of chromatin. Indeed the MutS mismatch repair complex, which shares similarities to Smc complexes, forms a ring that topologically traps generic DNA, while residues within the ring specifically recognize mismatch DNA [51–53]. In this scenario, the Smc1, Hinge, and Loop1 mutant subunits allow a topological interaction of cohesin with chromatin, but perturb its ability to make intimate chromatin/DNA interactions. Interestingly, in vitro studies have implicated residues on the interior of the bacterial Smc ring for DNA binding and DNA-dependent stimulation of the ATPase [44]. The chromatin-binding properties of cohesin in ctf7 and RID mutants also provide important insights into the mechanism of sister chromatid cohesion. First, the binding of cohesin to chromatin in early S at CARs (ctf7 mutants) or ectopic sites (Hinge and Loop1 mutants) is not sufficient to generate cohesion. Second, cohesin binding to CARs and centromeres appears necessary for cohesion. Third, an auxiliary factor, Ctf7 enables cohesins already bound at CARs to generate functional cohesion. These observations all contradict a one-step mechanism for cohesion like the simple embrace model, which requires only chromatin binding of cohesin anywhere on the chromosomes followed by DNA replication. Rather, our results support tethering of sister chromatids by two chromatin-binding events, which require specific binding of cohesins at CARs followed by a Ctf7-dependent step. Given these new constraints, we propose the following working model for the establishment of cohesion: cohesins bind to CAR sites as they emerge from the replication fork and are subsequently activated by Ctf7 to initiate the capture of the homologous CAR on the sister chromatid. This capture could occur by activating a second chromatid binding event by a single complex (for example, a second embrace) or by activating oligomerization of cohesins bound to each CAR (the oligomerization models) [33–36]. In either version of this two capture model, the specific binding to CARs and Ctf7 steps would be critical. Because of local proximity during replication, a cohesin molecule bound to a CAR on one sister chromatid will find the sister CAR and/or cohesin bound to that site before finding other sites/cohesins in the genome. We speculate further that the activation by Ctf7 is programmed to be local/transient. As a result, cohesin bound to a random site on a chromatid will be unlikely to remain active long enough to find a random CAR or another randomly bound cohesin. Thus, the targeting of cohesin by CAR binding and its local activation by Ctf7 would provide spatial and temporal regulation of cohesin to activate the second capture and ensure that cohesin generates a crosslink only between only sister chromatids and not random chromatin. Finally, the specificity of tethering by Smc complexes in other DNA processes may be achieved through specific chromatin binding coupled with Ctf7-like activators. Yeast strains were grown in YEP, SC-URA, or SC-URA-TRP media [54] supplemented with 2% dextrose (D), 2% raffinose 2% galactose (RG), 3% glycerol 2% lactic acid (LA), or 3% glycerol 2% lactic acid 2% galactose (LAG), as indicated. Glucose, raffinose, and galactose were purchased from Sigma-Aldrich (http: //www. sigmaaldrich. com), glycerol from EMD Biosciences (http: //www. emdbiosciences. com), and lactic acid (40% v/v stock, [pH 5. 7]) from Fisher Scientific (http: //www. fishersci. com). A PCR-based strategy was used to generate a complete genomic replacement of SMC1 with the Schizosaccharomyces pombe His5+ gene [55], creating the yeast strain Ymm1. Ymm1 is dependent on the plasmid pMM26 for viability. The genotypes of strains used in this work can be found in Table S1. Yeast transformation and genetic methods were as described previously [56]. The plasmids described were generated through use of the Echo Cloning System (Invitrogen, http: //www. invitrogen. com), which results in the expression of genes fused to COOH-terminal V5 epitopes under control of the GAL1 promoter. In all cases, the V5 epitope was replaced by a TAP tag using a PCR-based tagging strategy [57,58]. The plasmid pMM14 contains the SMC1 ORF. Restriction digestion by SmaI and PmeI, followed by religation, destroyed a unique PmeI site and generated AMH4. The plasmid pMM26 was derived from pMM14. GAL1 was replaced with a 412-basepair fragment immediately 5′ of the SMC1 ORF using AgeI and XhoI restriction sites, creating pMM24. Loss of the URA3 gene by BsgI digestion, followed by T4 blunting and religation, generated pMM26. Expression of Smc1-TAP from pMM24 and pMM26 is capable of rescuing growth of a temperature sensitive smc1–2 strain (1360–7C) and a strain deleted for endogenous SMC1. The plasmid pMM25-3HA expresses Smc1 fused to a COOH-terminal TAP-3HA tag under control of the GAL1 promoter. Through a PCR-based strategy, a BamHI site within the TAP coding sequence of pMM14 was destroyed by a silent mutation and replaced by a BamHI site immediately 5′ of the TAP stop codon [57]. TRP1 was lost by SmaI digestion, followed by religation, resulting in pMM14-PIBS. The XbaI/SmaI fragment of pMM14-PIBS was subcloned into pRS305, resulting in pMM27. A 3XHA coding sequence was inserted in frame at the BamHI site of pMM27, resulting in pMM27-3HA. Replacing the XbaI/SmaI fragment of pMM25 with the XbaI/SmaI fragment of pMM27-3HA resulted in pMM25-3HA. The plasmid pMM25-3HA is able to rescue viability of an SMC1 deleted strain under either uninduced or induced expression conditions. The plasmids 209L1–2-HA and 657H-HA were generated by replacing the XhoI/XbaI fragment of pMM25-3HA with those from 209L1–2 and 657H. Insertion mutagenesis of AMH4 was performed using the GPS-Linker Scanning system (New England Biolabs, http: //www. neb. com). In vitro mutagenized AMH4 was transformed into Escherichia coli. Approximately 3,100 colonies were scraped from plates, and plasmid DNA was isolated using a Plasmid Maxi kit (Qiagen, http: //www. qiagen. com). The majority of plasmids received only a single insertion (unpublished data). Plasmid DNA in the primary library was linearized by PmeI digestion, gel purified, religated, and transformed into E. coli. Plasmid DNA from this secondary library was isolated from approximately 4,000 colonies from plates using the Plasmid Maxi kit. The secondary library was transformed into YMM1/pMM26, with transformants selected for on SC-URA-TRP-D plates at 23 °C. Individual transformants were picked and patched to new SC-URA-TRP-D plates and grown at 23 °C. Patches were replica plated to an identical plate and a SC-URA-TRP-RG plate and grown for 3–5 d at 23 °C. From patches that were impaired for growth on SC-URA-TRP-RG plates, plasmids were isolated from the identical patch growing on the SC-URA-TRP-D plate and retested. For plasmids that retested, the insertion mutations were mapped by restriction endonuclease digestion and sequenced. To isolate functional alleles, the secondary library was transformed into an smc1–2 strain, 1360–7C. Transformants were selected for on SC-URA-D plates at 23 °C, patched to new SC-URA-D plates, and grown at 23 °C. Patches were replica plated to an SC-URA-RG plate and grown at 37 °C. Plasmids were isolated from patches that grew at 37 °C and retested. For plasmids that retested, the insertion mutations were mapped by restriction endonuclease digestion and sequenced. Exponentially dividing cell cultures were initially grown in SC-URA-LA at 23 °C. α-Factor (1. 5 × 10−8 M [Sigma]) was added to cultures in mid-log phase (approximately 0. 5 × 107 cell/ml). To simultaneously inactivate smc1–2 and release them from G1 arrest, cells were washed twice in either 37 °C YEP-LA or 37 °C YEP-LAG containing 0. 1 mg/ml Pronase (Sigma). Cells were resuspended in either 37 °C YEP-LA or YEP-LAG and grown at 37 °C. To arrest G1 released cells in metaphase, nocodazole was added to a final 15 μg/ml (1. 5 mg/ml in DMSO stock [Sigma]) and cultures were grown for 3 h. To arrest in early S phase, G1 released cells were grown in the presence of 0. 2 M hydroxyurea (HU) (Sigma) for 3 h. Cell cycle arrest was assessed by flow cytometry and cell morphology [27]. Cells were processed to visualize GFP by microscopy or to measure DNA content by flow cytometry as described previously [27,59]. Protein extracts for coimmunoprecipitation and immunoblotting were prepared as described previously [15,20] from cultures grown in YEP-D media. ChIP was performed as described [60]. Information about the primers used in this study is available upon request. PCR and data analysis for ChIP was performed as described [49]. All experiments were done at least twice and a representative dataset is shown. Chromosome spreads and indirect immunofluorescence of spread nuclei were performed as described previously except spheroplasting was done at 25 °C for 30 min [41]. To assess the level of cohesin association per chromosome, the average pixel intensity for each chromosome mass was determined. The background pixel intensity for each slide was determined by measuring the average pixel intensity for areas similar in size to spread nuclei. Subtracting the background intensity from each chromosome mass gave relative pixel intensity. At least 100 nuclei were analyzed per slide to generate an average relative pixel intensity per chromosome mass. To assay the stability of cohesin association with chromosomes, nuclei were spread on multiple slides in the absence of fixative with 0. 25% Triton X-100 in PHEM buffer (60 mM Pipes, 25 mM Hepes, [pH 6. 95], 10 mM EGTA, and 4 mM MgCl2) and incubated for 10 min. Each slide was then placed in a single coplin jar containing ∼50 ml 0. 25% Triton X-100 in PHEM buffer with varying amounts of KCl for 30 min, while gently shaking. Following this incubation, nuclei were fixed by 4% PFA with 0. 25% Triton X-100 in PHEM buffer, as before. Immunofluorescence was performed as above.

Complexes containing members of the structural maintenance of chromosomes (Smc) family regulate higher order chromosome architecture in diverse aspects of DNA metabolism including chromosome condensation, sister chromatid cohesion, DNA repair, and global control of transcription. Smc complexes are thought to regulate higher order chromosome folding by tethering together two strands of chromatin. However, the mechanism of tethering is poorly understood in part because of a poor understanding of the function of the core Smc subunits. To gain insight into the structure and function of Smc subunits, we developed a novel strategy of mutagenesis called random insertion dominant negative (RID), which generates informative alleles with high efficiency and should provide an effective tool to study any multi-subunit complex. Using RID we generated novel alleles of a Smc subunit from the cohesin complex. The cohesin complex tethers together newly replicated chromosomes (sister chromatids). The analyses of these RID mutants suggest that the tethering activity of cohesin (and possibly other Smc complexes) is generated by two sequential chromatin-binding events (first the capture of one piece of chromatin followed by the capture of the second piece of chromatin), which are regulated both spatially and temporally. We speculate that the spatial and temporal regulation of cohesin ensures that it tethers together only sister chromatids rather than randomly crosslinking the entire genome.

Abstract Introduction Results Discussion Materials and Methods

yeast and fungi genetics and genomics saccharomyces

2007

A Multi-Step Pathway for the Establishment of Sister Chromatid Cohesion

It has long been known that loss of the retinoblastoma protein (Rb) perturbs neural differentiation, but the underlying mechanism has never been solved. Rb absence impairs cell cycle exit and triggers death of some neurons, so differentiation defects may well be indirect. Indeed, we show that abnormalities in both differentiation and light-evoked electrophysiological responses in Rb-deficient retinal cells are rescued when ectopic division and apoptosis are blocked specifically by deleting E2f transcription factor (E2f) 1. However, comprehensive cell-type analysis of the rescued double-null retina exposed cell-cycle–independent differentiation defects specifically in starburst amacrine cells (SACs), cholinergic interneurons critical in direction selectivity and developmentally important rhythmic bursts. Typically, Rb is thought to block division by repressing E2fs, but to promote differentiation by potentiating tissue-specific factors. Remarkably, however, Rb promotes SAC differentiation by inhibiting E2f3 activity. Two E2f3 isoforms exist, and we find both in the developing retina, although intriguingly they show distinct subcellular distribution. E2f3b is thought to mediate Rb function in quiescent cells. However, in what is to our knowledge the first work to dissect E2f isoform function in vivo we show that Rb promotes SAC differentiation through E2f3a. These data reveal a mechanism through which Rb regulates neural differentiation directly, and, unexpectedly, it involves inhibition of E2f3a, not potentiation of tissue-specific factors. The simplicity of the retina makes it an ideal tissue to study neurogenesis. Its development proceeds through three overlapping steps starting with retinal progenitor cell (RPC) proliferation, followed by birth of post-mitotic retinal transition cells (RTCs, also referred to as precursors), and ending with terminal differentiation of seven major cell types (Figure 1A) [1]. RPCs are multipotent and exit the cell cycle to generate different RTCs at specific time periods in development [1]. This process of RTC “birth” requires coupling of differentiation and cell cycle exit. Once born, post-mitotic RTCs migrate and form different retinal layers. Rods and cones make up the outer nuclear layer (ONL); horizontal, bipolar, and amacrine cells, as well as Müller glia cell bodies, reside in the inner nuclear layer (INL); and ganglion and displaced amacrine cells form the ganglion cell layer (GCL) (Figure 1A). The outer plexiform layer (OPL) and inner plexiform layer (IPL) house synaptic connections separating the ONL/INL and INL/GCL, respectively. The retinoblastoma protein (Rb) is critical for cell cycle exit during retinal transition cell birth. Rb knockout (KO) RTCs continue to proliferate inappropriately and some (rod, ganglion, and bipolar cells) die by apoptosis [2,3]. Rb controls the cell cycle by binding and inhibiting E2f transcription factors (E2fs) (Figure 1B), first defined as transcription factors that bind adenoviral E2 regulatory elements and subsequently shown to be critical cell cycle regulators [4,5]. E2fs bind to DNA as heterodimers with proteins of the related Tfdp family. E2f1, E2f2, and E2f3a are “activating E2fs” that are required for fibroblast division. They are strong transcriptional activators that can drive G0 fibroblasts into cycle, and are inhibited when bound to Rb [4,5]. Ectopic division in Rb KO embryos can be rescued to various extents in different tissues by knocking out E2f1, E2f2, or E2f3 [6–9], but which member (s) drive division in Rb KO RTCs is unknown. Other members of the family, such as E2f4 and E2f5, are known as “repressive E2fs” because they are weak activators and appear to be primarily involved in gene silencing in quiescent or differentiated cells. Activating E2fs may also promote apoptosis in the Rb KO retina (Figure 1B). Originally, E2f1 was considered the primary pro-apoptotic member of the family [10]. However, this view was reevaluated when it was shown that either E2f1 or E2f3 deletion rescues apoptosis in the developing central nervous system (CNS) of Rb KO embryos [6,11]. Subsequently, CNS apoptosis was shown to be an indirect result of placental defects and probable hypoxia [12–14]. Indeed, E2f3-induced apoptosis in fibroblasts has recently been shown to require E2f1 [15]. Thus, it is controversial whether E2f3 is required for apoptosis of any Rb KO cell type. Determining which activating E2fs promote death in distinct Rb KO tissues requires conditional rather than germ line models of Rb deletion to avoid secondary indirect effects (such as hypoxia). E2f family diversity is expanded by E2f3 isoforms. Alternative promoters generate two forms (a and b) that are identical except for distinct first exons [16]. E2f3a is a strong activator, and, like other activating E2fs, its expression is induced when quiescent cells are stimulated to divide [16]. E2f3b, like repressive E2fs, is present in both quiescent and dividing cells, and in quiescent fibroblasts it associates primarily with Rb, suggesting that it mediates repression [16–18]. Indeed, silencing the Cdkn2d (p19Arf) locus in unstressed cells relies on E2f3b [19]. Other E2fs may also exist in isoforms since at least two mRNA species have been detected for E2f1 and E2f2 [16]. The roles of E2f isoforms in vivo are unknown. E2fs are also regulated by subcellular localization. Although this feature has been best characterized for repressive E2fs [20–22], it also affects activating E2fs [23–25]. The distribution of E2f isoforms has never been assessed. It has been known for many years that Rb loss perturbs neuronal differentiation [26–29]. However, prior work could not exclude the possibility that differentiation defects are simply an indirect consequence of abnormal division and death. If Rb does regulate differentiation directly it is unclear whether it does so in all or a subset of neurons. Moreover, the mechanism has never been solved. In other cell types where Rb may promote differentiation directly, such as muscle and bone, it seems to do so through E2f-independent means by potentiating tissue-specific transcription factors (Figure 1B) [30–33]. In the retina, others have noted abnormally shaped Rb KO rods and have suggested Rb may directly promote their morphogenesis by activating retina-specific factors [29]. However, differentiation defects in any Rb KO neuron could be an indirect effect of ectopic division and/or apoptosis (Figure 1B). Thus, it is critical to study differentiation of Rb KO cells in the absence of ectopic proliferation and death. Here, we establish that Rb suppresses RTC division and death by inhibiting E2f1, not E2f2 or E2f3. When these defects were rescued, most retinal neurons, including rods, survived, differentiated, and functioned normally. Thus, unexpectedly, retina-specific differentiation factors function independently of Rb. However, comprehensive assessment of the Rb/E2f1 double-null rescued retina revealed a differentiation defect in cholinergic starburst amacrine cells (SACs). Recent breakthroughs have revealed that these interneurons are critical for direction selectivity and developmentally important rhythmic bursts [34–36]. However, their differentiation is poorly understood. Contrary to the prevailing view that Rb promotes differentiation through E2f-independent tissue-specific transcription factors, we show that Rb facilitates SAC development through E2f3. Defects in Rb null SACs correlated with specific E2f3 expression in these cells, and E2f3 expression was absent in neurons that differentiated without Rb. E2f3 is also present in a specific subset of other CNS neurons, implying that this may be a general mechanism by which Rb facilitates neurogenesis. To define the mechanism in even more detail, we determined which E2f3 isoform Rb targets to control SAC differentiation. E2f3b mediates Rb function in quiescent fibroblasts [19], yet no prior studies to our knowledge have dissected E2f3a or E2f3b functions in vivo. Using an isoform-specific null mouse we show that Rb drives SAC differentiation through E2f3a. Thus, independent of E2f1-mediated effects on division and death, Rb does regulate neuronal differentiation, but only in specific neurons and, unexpectedly, through E2f3a, not tissue-specific differentiation factors. We used the α-Cre transgene to delete floxed Rb exon 19 at embryonic day (E) 10 in peripheral retina [2]. RbloxP/loxP; α-Cre mice were bred with strains lacking E2f1 or E2f2 in the germ line, or a strain carrying a floxed E2f3 allele [5]. RbloxP/loxP; E2f1+/− and RbloxP/loxP; E2f1+/−; α-Cre mice were bred to produce RbloxP/loxP; E2f1−/−; α-Cre mice at a frequency of 1/8 and littermate controls at the same or higher (1/4) frequency. For simplicity we will refer to the RbloxP/loxP; E2f1−/−; α-Cre peripheral retina as the Rb/E2f1 double knockout (DKO) retina. Similar strategies were employed in the case of E2f2 or E2f3. Cre-mediated excision of Rb and E2f3 alleles in the retina was confirmed by PCR as described previously [2,5]. To measure ectopic cell division, mice were pulse-labelled with bromodeoxyuridine (BrdU) 2 h before sacrifice and the peripheral retina analyzed for BrdU incorporation by immunofluorescence. As reported before [2,3], Rb KO retinas exhibited both spatial and temporal ectopic DNA synthesis (Figures 1C and S1A). This is easily detected at E14, E16, and postnatal day (P) 0 in the inner retina where abnormal BrdU+ ganglion and amacrine RTCs are located, or on the outermost region of the P0 retina, where BrdU+ photoreceptor RTCs reside (Figures S1A and S2, arrows) [2]. Ectopic RTC division in Rb KO retinas is even more obvious at P8 or P18, when division is completed in wild-type (WT) retina (Figures 1C and S1A). Strikingly, the ectopically positioned S-phase cells at E14, E16, and P0 and all the abnormal division at P8 and P18 were completely suppressed in the Rb/E2f1 DKO retina (Figures 1C, 1E, 1F, S1A, and S2). In contrast, deletion of E2f2 or E2f3 had no effect at any stage of development. Analysis of mitotic cells with anti–phosphohistone 3 (PH3) –specific antibodies confirmed that loss of E2f1, but not E2f2 or E2f3, suppressed ectopic division (Figure S3). Deleting one E2f1 allele partially suppressed ectopic S-phase and mitosis in Rb KO RTCs (Figures 1C, 1E, 1F, S1A, S2, and S3), suggesting that E2f1 drives ectopic division in Rb KO RTCs in a dose-dependent fashion. These data contrast with previous findings in the lens and CNS of Rb KO embryos, where deletion of any activating E2f suppresses ectopic division to some extent [6–9]. Loss of Rb in the retina results in considerable RTC apoptosis, eliminating most bipolar and ganglion cells as well as many rods (Figure 2A–2D) [2,3]. The loss of Rb KO rods is evident from the thinner ONL, and the death of these cells as well as bipolar and ganglion neurons can be detected directly by double labelling for apoptotic and cell-type-specific markers [2] (M. P. and R. B. , unpublished data). Loss of peripheral Rb KO ganglion cells is also evident from thinning of the optic nerve (D. C. and R. B. , unpublished data). Deleting E2f1, but not E2f2 or E2f3, blocked this ectopic cell death in a dose-dependent fashion (Figures 1D, 1G, and S1B). To investigate the molecular mechanism that underlies the unique role of E2f1, we assessed the expression of known E2f targets as well as other genes that regulate the cell cycle and apoptosis. Numerous positive and negative cell cycle and apoptotic regulators were up-regulated in the Rb KO retina (Figure 1H). Among the E2f family, E2f1, E2f2, E2f3a, and E2f7 were induced following Rb loss, but E2f3b, E2f4, and E2f5 were unaffected. Consistent with the BrdU and terminal dUTP nick-end labelling (TUNEL) analyses, E2f1 deletion specifically reversed all these molecular defects, but E2f3 deletion had no effect (Figure 1H). Because E2f1 deletion blocks abnormal division and death in the Rb KO retina, the Rb/E2f1 DKO retina provided a unique opportunity to evaluate whether Rb controls differentiation independent of cell cycle effects. The Rb/E2f1 DKO retina had many Sag+ (S-antigen/rod arrestin) photoreceptors, Pou4f2+ (Brn3b) ganglion cells, and numerous Prkca+/Cabp5+ bipolar neurons (Figure 2A–2D). In contrast, there was no such rescue of cell types in Rb/E2f2 or Rb/E2f3 DKO retinas (Figure S4). Analysis with general neuronal markers Mtap2 (MAP2) and Snap25, as well as other markers expressed in bipolar cells (Chx10, Rcvrn, Vsx1, Tacr3, and Atp2b1) and rod photoreceptors (Rho and Rcvrn) confirmed rescue of the Rb/E2f1 DKO retina (Table S1). Moreover, neurons exhibited the same complex morphology as in WT retina. Bipolar cell bodies were located in the INL and had ascending and descending processes ending in the OPL and IPL, respectively (Figure 2A). In addition, the Rb/E2f1 DKO retina had a healthy appearing ONL consisting of morphologically normal rods with descending processes ending in the OPL and ascending processes that terminated in inner and outer segments (Figure 2A). It was suggested that Rb might regulate photoreceptor differentiation, possibly through rod-specific transcription factors (Figure 1B) [29]. However, if Rb does regulate photoreceptor differentiation, it does so by inhibiting E2f1, not by potentiating rod differentiation factors, such as Otx2, Crx, or Nrl. It is impossible to tell whether E2f1 perturbs differentiation directly, by affecting the expression of genes that modulate maturation, and/or indirectly through its effects on proliferation and survival (Figure 1B). As with ectopic division and apoptosis (Figure 1C and 1D), the rescue of Rb KO retinal bipolar, ganglion, and rod cells was dependent on E2f1 dose (Figure 2A–2D). Separate from its role in driving ectopic division of Rb KO RTCs, E2f1 also promotes normal RPC division since in its absence RPC proliferation drops ~2-fold (D. C. and R. B. , unpublished data). This modest reduction of RPC numbers in the absence of E2f1 accounts for the slight reduction in the number of ganglion cells at P0, in the number of bipolar cells at P18 or P30, and in the thickness of the ONL at P18 or P30 in the E2f1 KO and Rb/E2f1 DKO retina (Figure 2B–2D). The morphology of E2f1 KO neurons was WT (Figure 2A). Despite a slight drop in absolute cell numbers, the proportion of Rb/E2f1 DKO and E2f1 KO bipolar cells was the same as WT (data not shown). Slightly reduced cell numbers were not due to residual RTC death since we have not observed ectopic apoptosis at any embryonic or postnatal stage in the developing Rb/E2f1 DKO retina (Figures 1D, 1G, and S2). Moreover, deleting Ccnd1, which acts upstream of Rb proteins, also reduces RPC number, but does not suppress any defect in the Rb KO retina (D. C. and R. B. , unpublished data). Thus, slightly reduced RPC division and dramatic rescue of severe defects in Rb KO RTCs are distinct effects stemming from the deletion of E2f1. The discovery that E2f1 loss rescues even the morphology of Rb KO neurons is surprising because Rb is thought to regulate differentiation primarily through E2f-independent pathways [30–33]. However, normal morphology may not equate to completely normal differentiation. Thus, we compared the electroretinograms (ERGs) of adult WT (α-Cre), E2f1−/−, α-Cre; RbloxP/loxP, and α-Cre; RbloxP/loxP; E2f1−/− mice. ERGs functionally assess visual signalling in the mammalian retina from photoreceptors to amacrine cells (but usually not gangion cells), and are dominated by rod and cone bipolar cells. Typically, an ERG signal begins with a negative deflection initiated by the photoreceptors (the a-wave), which is terminated by a large positive deflection due to the activation of ON bipolar cells (the b-wave). Responses to dim light in dark-adapted (scotopic) conditions specifically assess the rod system, and were defective in the Rb KO retina (Figure 2E). The substantial reduction of both a- and b-waves is consistent with rod and bipolar cell apoptosis [2]. The sensitivity of the residual response appeared unchanged, suggesting it arose from the Cre-negative portions of the retina. Responses were about the same in the WT and E2f1 KO retina, and, most importantly, also the Rb/E2f1 DKO response median lay at the lower end of the normal range for most intensities (Figure 2F). Thus, E2f1 deletion almost completely rescued the rod system in the Rb KO retina. Light-adapted (photopic) recordings to specifically assess the cone system yielded comparable results. Cones represent only 3% of photoreceptors and, unlike rods, develop without Rb, but they require rods for survival, and in the Rb KO retina, they have abnormal morphology and their synaptic targets, bipolar cells, are much depleted [2]. The photopic response, a product of cone and mainly bipolar activity, was much reduced by Rb loss, but was rescued considerably in the Rb/E2f1 DKO retina (Figure S5). Again, the median amplitude lay at the lower end of the E2f1 KO range. The photopic response in E2f1 KO mice was slightly reduced relative to WT (Figure S5B), possibly because E2f1 is required for maximal expansion of embryonic RPCs, and the E2f1 KO retina has, as noted earlier, slightly fewer cells than the WT retina, although cell type proportions are unaffected (D. C. and R. B. , unpublished data). Thus, marginally subnormal photopic responses in the Rb/E2f1 DKO retina can be attributed to both a reduction of cone numbers in E2f1 KO mice alone, and a “genuine” slight reduction in cone function attributable to Rb loss relative to WT. This slight effect may relate to a true differentiation defect in a subset of amacrine cells discussed below. This discussion should not obscure the major outcome that E2f1 deletion recovers most of the ERG response. Thus, E2f1 deletion not only rescues morphology but also both rod and cone system function in the Rb KO retina. ERGs primarily assess photoreceptor and bipolar cell function, but may miss differentiation defects in other cells. To test for subtle differences we stained the Rb/E2f1 DKO retina with 43 markers (Table S1). Thirty-two proteins displayed identical patterns in WT, E2f1 KO, and Rb/E2f1 DKO retina (Table S1). The other 11 markers revealed a cell-cycle– and apoptosis-independent differentiation defect in SACs. We first studied Calb2 (calretinin), which marks a subset of amacrine and ganglion cell bodies as well as three tracks corresponding to their processes in the IPL (Figure 3A). Normal Calb2 staining was seen in the E2f1 KO IPL (data not shown). However, only one Calb2+ track was evident in the Rb KO IPL, and this defect was not rescued in the Rb/E2f1 DKO retina (Figure 3A). We quantified Calb2+ cell bodies in the Rb KO INL (corresponding to amacrine cell staining only) and observed a reduction from P8 onwards (Figures 3C and S6). Of the three Calb2+ tracks in the IPL, the two outer tracks are from SACs, named after their extensive dendritic-tree-like morphology [37]. SACs are cholinergic, represent ~5. 2% of amacrine neurons [38], and are critical for both direction selectivity [34,35] and spontaneous rhythmic activity that occurs during normal retinal development [36]. SACs in the INL synapse in the OFF layer of the IPL that responds to decreasing light, while displaced SACs in the GCL have processes that synapse in the nearby ON layer of the IPL that responds to increasing light (reviewed in [39]). Mature SAC processes stain specifically for Slc18a3 (vesicular acetyl choline transporter, VAChT) [37], and, significantly, this marker was absent in the peripheral Rb KO or Rb/E2f1 DKO P18 retina (Figures 3A and S7B). Chat, expressed from the same locus, is also SAC specific, but marks both cell bodies and processes of mature SACs [37]. Chat was seen in fewer cells in the mature Rb KO retina, and was present in the soma but absent from processes (Figure 3B). We obtained similar results for Sv2c, a synaptic vesicle protein found in SACs [40]; Kcnc1b and Kcnc2, potassium channels expressed on SAC soma and dendrites as well as a very small number of ganglion cells [41]; gamma-aminobutyric acid (GABA), an inhibitory neurotransmitter present in about half of amacrine cells including SACs, as well as horizontal and some bipolar neurons [37]; and Calb1 (calbindin), which is expressed in many amacrine cells and labels SAC process faintly (Figure S7A and S7B; Table S1; and data not shown). Finally, we also examined the effect of Rb deletion on SAC differentiation using a Chx10-Cre transgene that is expressed in a mosaic pattern across the retina, generating patches of Cre-expressing cells [42]. Consistent with the mosaic deletion pattern, we observed markedly reduced Chat/Slc18a3 staining in the IPL of Chx10-Cre; RbloxP/loxP retina compared to WT (Figure S7C). Together, these results suggest a role for Rb in SAC differentiation. The above findings could indicate a defect in SAC specification, SAC survival, or the levels and/or transport of the markers described above. Camk2a marks both SACs and ganglion cells [37], but because ganglion cells are eliminated in the Rb KO retina, Camk2a is a specific SAC marker in this context. Importantly, Camk2a+ tracks and dendrites were present in both the WT and Rb KO retina (Figure 3B), and the number of Camk2a+ soma was similar in WT and Rb KO retina at P30 and beyond, although fewer cells stained in Rb KO retina at P18, suggesting a delay in its appearance (Figures 3C and S6B). Thus, Rb is not required for SAC survival or for process outgrowth, but rather it seems to regulate the expression and/or stability of Calb2, Calb1, Chat, Slc18a3, Sv2c, Kcnc1b, Kcnc2, and GABA in SACs, but leaves Camk2a expression virtually unaffected. The presence of Chat in some cell bodies but never in processes (Figure 3B) also suggests a transport defect. The developmental pattern of Slc18a3 expression also supported this notion. In mature WT SACs Slc18a3 is only in processes, but in early postnatal SACs, it is found in the cell body, and moves into emerging processes at approximately P4–P6. As noted above, Slc18a3 was absent at P18 in the Rb KO retina (Figure 3A); at P4 or P5 it was in cell bodies, yet was rarely present in Rb KO processes (Figures 4A and S6). Slc18a3 became virtually undetectable in Rb KO SACs by P8 (Figures 3C and S6C). These data suggest that Rb affects both the synthesis/stability and transport of SAC markers. Rb binds more than 100 proteins [43] and in some non-neuronal cells, such as skeletal muscle, adipocytes, and bone, Rb is thought to bind and potentiate tissue-specific transcription factors that promote differentiation [31–33]. Thus, we expected that Rb might interact with retina-specific factors to facilitate SAC differentiation. A direct role for E2f in mediating Rb-dependent differentiation defects (independent of cell cycle or death defects) has to our knowledge not been described, but because E2f can regulate some differentiation genes [44–48], we first tested whether E2f2 or E2f3 might perturb Rb KO SAC maturation. At multiple time points, E2f1 deletion suppressed ectopic mitosis (PH3+ cells), but did not reverse the SAC defect, and E2f2 deletion had no effect on either defect (Figure 4A). Remarkably, although E2f3 deletion did not reverse ectopic mitosis, it rescued Calb2, Slc18a3, Chat, GABA, Kcnc1b, Kcnc2, and Sv2c staining at multiple times (Figure 4A and data not shown). Rb/E2f3 DKO SAC tracks were slightly more disordered than in WT retina, most likely because of the absence of synaptic partner cells, which are killed by E2f1. Indeed, this minor defect was rescued in the Rb/E2f1/E2f3 triple knockout retina, where bipolar and ganglion cell death was rescued and SAC differentiation was restored (Figure 4A). E2f3 deletion alone did not affect SAC differentiation (Figure 4A); thus, it is unleashed E2f3 activity that is detrimental, and the critical role for Rb is to inhibit E2f3. We quantified the fraction of Camk2a+ SACs in different genotypes and found that 60% of WT P30 Camk2a+ cells expressed Chat and Slc18a3, which dropped to only 5. 6% in the Rb KO retina, and remained low at 3. 7% in the Rb/E2f1 DKO retina, but rose to 91% in the Rb/E2f3 DKO retina (Figure 4B). The latter fraction is higher than WT because ganglion cells, which normally make up ~40% of Camk2a+ cells, are killed by apoptosis. To quantify the effect of different E2fs on ectopic division specifically in SACs, we exploited Isl1 (Islet1). This marker is expressed in both SACs and ganglion cells, thus Isl1+ cells in the INL are exclusively SACs [49]. We found that 98. 2% ± 1. 8% of Isl1+ cells in the forming inner INL at P5 were also Slc18a3+, confirming that Isl1 is an excellent SAC marker (Figure 4C). Moreover, Isl1, unlike Slc18a3, is nuclear, which facilitates scoring of Isl1+/Mki67+ cells. It is also expressed earlier than Slc18a3, permitting analysis of SACs soon after their birth at ~E15; thus, we could study retina at P0, a time when ectopic division is high in the inner retina and prior to Rb-independent cell cycle exit associated with terminal differentiation [2]. At P0, no WT Isl1+ cells in the inner neuroblastic layer (NBL) (which is the future INL) were dividing, but 57 ± 14 Isl1+/Mki67+ cells were detected in the Rb KO inner NBL (Figure 4D). Indeed, about one-third of all Isl1+ cells in the entire inner NBL were dividing in the Rb KO retina, or ~50% in the periphery where Cre is expressed (Figure 4E and data not shown). This defect was suppressed in the Rb/E2f1 DKO retina, where we detected only 1 ± 1 dividing SAC, but not the Rb/E2f3 DKO retina, where there were 53 ± 8 dividing SACs (Figure 4D and 4E). We observed similar effects at P0 with Calb2, which marks newborn SACs and other amacrine cells (data not shown). Thus, in Rb KO SACs, E2f1 deletion suppresses ectopic division but not aberrant differentiation, whereas E2f3 deletion suppresses aberrant differentiation but not ectopic division. The unique effect of E2f3 in disrupting the differentiation of SACs but not other retinal neurons might be due to cell-type-specific expression or cell-type-specific activity of E2f3. Determining between these two possibilities is not easy, as E2f immunostaining in mouse tissues is problematic. We did not solve this issue for E2f1 or E2f2, but used a modified protocol [50] to successfully track E2f3 expression (Figure 5). At P0, E2f3 was detected in RPCs, consistent with a putative role in normal proliferation (Figure 5A). The signal was specific as it was absent in the E2f3 KO peripheral retina (Figure 5A). As the retina differentiated and RPC division diminished, the number of E2f3+ cells also dropped, and by P8, when division is virtually over, only a subset of post-mitotic cells in the inner retina expressed E2f3 (Figure 5A). By P18, E2f3 was also detected in two tracks in the IPL (Figure 5A and 5B), reminiscent of SAC markers such as Chat and Slc18a3 (c. f. Figures 3 and 4). This cytoplasmic E2f3 staining was also specific, as it was absent in the E2f3 KO peripheral retina of α-Cre; E2f3loxP/loxP mice (Figure 5A). Indeed, double labelling with E2f3 (red) and Chat plus Slc18a3 (green) confirmed that E2f3 is present in both SAC soma and dendrites (Figure 5B). Rb protein was also detected in the inner retina (Figure 5A), and showed a similar distribution as E2f3 in SACs (Figure 5B), and was also present in mature ganglion cells and Müller cells as reported [51]. Rb staining in SAC processes was specific as it was absent in the peripheral retina of αCre; RbloxP/loxP mice (Figure 5A). These data suggest that Rb and E2f3 colocalize in SACs and that E2f3 triggers defects in SAC differentiation because it is specifically expressed in these retinal neurons. We also found that E2f3 is present in a specific subset of mature neurons in various brain regions (data not shown). For example, in the P20 amygdala, E2f3 colocalized with the general neuronal markers Mtap2 and Mecp2 [52], but not with Calb2, which marks a subset of neurons, or with the glial marker Gfap (data not shown). Unlike in retinal SACs, E2f3 was not coexpressed in Chat+ or Slc18a3+ cholinergic neurons located in various regions of the brain and spinal cord (data not shown). In agreement, we could not detect defects in cholinergic Rb KO neurons in the developing forebrain, but other Rb KO neurons in this region showed differentiation defects that were rescued by deleting E2f3 [53]. Together, these results suggest that the common mechanism by which Rb promotes neural differentiation is through E2f3 inhibition. As noted above, E2f3 and Rb staining in SACs was both nuclear and cytoplasmic (Figure 5A and 5B). The antibody that worked in immunostaining recognizes a C-terminal region and thus, does not distinguish a/b isoforms. To our knowledge, the subcellular location of E2f3 isoforms has not been determined in any cell type. To verify the dual locations of E2f3 and to determine which isoforms were present in retina, we analyzed nuclear and cytoplasmic fractions by Western blot at different times during development. Analysis with the pan-E2f3 antibody (sc-878, Santa Cruz Biotechnology) detected a 55-kD E2f3a species and a 40-kD E2f3b polypeptide (Figure 6). To confirm that the upper species in our retinal lysates was E2f3a, we exploited novel mice that lack E2f3 exon 1a and thus express E2f3b exclusively (R. O. and G. L. , unpublished data). The genotyping strategy is discussed in detail later and is outlined in Figure 7A. Western analysis confirmed that the upper band was absent in E2f3a−/− mice (Figures 6 and S8). Consistent with the drop in E2f3-expressing cells during WT retinal maturation (Figure 5A), the total amount of E2f3a was less at P18 compared to P0 (Figure 6). E2f3b was present in similar amounts at both time points. At P0 and P18, E2f3a was present in both nuclear and cytoplasmic fractions, but in marked contrast, E2f3b was exclusively nuclear at both times (Figure 6). Two closely migrating E2f3a bands were detected, more clearly evident at P18, of which the faster migrating species was dominant in nuclear and the slower species was dominant in cytoplasm (Figure 6). The identity of both as E2f3a species was confirmed by their absence in the P18 E2f3a KO retina (Figure S8). Analysis of Pou4f2, a nuclear transcription factor expressed in ganglion cells, showed that nuclear proteins had not contaminated the cytoplasmic fraction, and analysis of Slc18a3, a cytoplasmic SAC marker, confirmed that the reverse had also not occurred (Figure 6). These data show, to our knowledge for the first time, that E2f3a and E2f3b exhibit distinct patterns of subcellular distribution, and raise the possibility that E2f3a localization may be regulated by as yet unknown post-translational modifications. We also examined the distribution of other cell cycle regulators during retinal development. Like E2f3a, Rb was present in both the WT cytoplasm and nucleus at P0, but at P18, when the levels of Rb had increased, it was primarily nuclear (Figure 6). A very faint cytoplasmic Rb signal was evident at P18, which is consistent with Rb staining of SAC processes (Figure 5B), and with the very small proportion of SACs in the retina [38]. E2f1 was also detected in both nuclear and cytoplasmic fractions, although unlike E2f3a it was predominantly nuclear both at P0 and P18 (Figure 6). The E2f dimerization partner, Tfdp1, which lacks a nuclear localization signal [54], was primarily cytoplasmic at both P0 and P18, and the Cdk inhibitors Cdkn1a and Cdkn1b showed a similar pattern of distribution (Figure 6). Thus, among the cell cycle regulators we examined, most showed bivalent distribution, and E2f3b was unusual in its solely nuclear compartmentalization. To test which E2f3 isoform is responsible for aberrant Rb KO SAC differentiation we exploited E2f3a−/− mice (Figure 7A). The genotyping strategy outlined in Figure 7A was used to distinguish the E2f3a, WT, and null alleles. Reverse transcriptase PCR (RT-PCR) confirmed the presence of both E2f3a and E2f3b RNA species in the developing WT retina, and the specific absence of E2f3a RNA in the E2f3a−/− retina (Figure 7B). E2f3a protein was absent in E2f3a−/− retinal lysate (Figures 6 and S8). Importantly, the levels of E2f3b message were similar in the Rb KO and Rb/E2f3a DKO retina, ruling out the possibility that any effects of E2f3a deletion we might observe were due to down-regulation of E2f3b (Figure 7C). Also, the levels of other E2fs were the same in the Rb KO, Rb/E2f3 DKO, and Rb/E2f3a DKO retina, ruling out any cross-regulatory effects (Figure 7C) [55]. E2f3a can trigger cell cycle induction, but because SAC defects are not linked to cell cycle perturbation (Figures 3A and 4), and in view of the predominant association between E2f3b and Rb in quiescent cells [16,19], we suspected that E2f3b may perturb differentiation in Rb KO SACs. Unexpectedly, however, E2f3a deletion suppressed the Rb KO SAC defect (Figure 7D). Thus, separate from its role in cell cycle control, Rb regulation of E2f3a is critical to ensure proper neuronal differentiation. Work in the early 1990s showed that Rb loss triggers defects in neuronal cell cycle exit, survival, and differentiation [26–28]. Much of the death is an indirect consequence of probable hypoxia linked to placental defects [12–14]. However, targeted KO and chimeric studies reveal that Rb autonomously promotes cell cycle exit in newborn neurons, and is required for survival of a subset of neurons, particularly in the retina [2,3, 13,14,56–59]. However, whether Rb also regulates differentiation is obscured by potentially indirect effects of ectopic division and death. Moreover, a mechanism though which Rb may regulate neuronal maturation has not been elucidated. Here, deleting E2f1 specifically rescued ectopic division and death in the Rb KO retina. Importantly, major Rb/E2f1 DKO neurons differentiated normally, and ERGs revealed the rescue of rod- and cone-mediated function, implicating a regular signal flow from photoreceptors to bipolar and amacrine cells. Division and death genes were induced in Rb KO cells, and deleting E2f1, but not E2f2 or E2f3, reversed these molecular events. E2f1 may also regulate differentiation targets, but whether this contributes to defects in retinal cell maturation is impossible to separate from potentially indirect consequences of deregulated division and death. In any case, it is clear that in most retinal cells, including photoreceptors [29], transcription factors that promote differentiation function independently of Rb. We have also found that E2f1 deletion rescues cell-autonomous ectopic division, death, and differentiation defects in sporadic Rb KO clones generated using a Cre retrovirus vector (M. P. and R. B. , unpublished data). These data are consistent with the observation that E2f1 overexpression in newborn photoreceptors drives ectopic division and apoptosis [60], and add to the growing evidence indicating that E2f1 is the major, and perhaps only, member of the three activating E2fs required to induce apoptosis in Rb KO cells [10,15]. Thus, deregulated E2f1 activity in the retina, whether resulting from the inactivation of Rb or from overexpression, promotes unscheduled cell division and triggers apoptosis in susceptible RTCs. E2f1, rather than other E2fs, may be a potential target for novel therapeutics to prevent retinoblastoma in RB1+/− humans. Our ERG studies revealed rescue of the Rb KO rod–bipolar system, and almost complete restoration of the cone–bipolar system following E2f1 deletion. There was a slightly lower response in the Rb/E2f1 DKO retina relative to the E2f1 KO control retina. This difference might reflect a role for Rb in the development of cones, bipolar cells, or other cells that may contribute to the photopic ERG, including potentially SACs, which do have a serious defect in the Rb/E2f1 DKO retina. Comprehensive marker analysis revealed that, in striking contrast to other retinal neurons, E2f1 deletion did not suppress defects in Rb KO cholinergic SACs. Instead, we observed E2f1-independent defects in the synthesis and transport of a large cohort of SAC proteins. These data expand insight into the development of these important interneurons, but more critically, provide to our knowledge the first unambiguous evidence that Rb regulates neurogenesis beyond terminal mitosis. Rb binds more than 100 factors [43], and in several non-neuronal cells, such as skeletal muscle, adipocytes, and bone, it binds and potentiates tissue-specific transcription factors that promote differentiation [31–33]. The idea that Rb promotes muscle differentiation by potentiating Myod1 activity was contested [61], and other mechanisms proposed [62,63], but not involving E2f repression. Strikingly, however, we discovered that Rb promotes SAC differentiation through E2f3 (Figure 8). Rb regulation of SAC differentiation through E2f3 was independent of its role in controlling division or death: E2f3 deletion rescued Rb KO SAC defects but did not suppress aberrant proliferation or death, whereas E2f1 deletion reversed abnormal proliferation and death but did not rescue SAC differentiation. Double labelling confirmed that E2f1 but not E2f3 deletion reversed Rb KO SAC division. Moreover, deleting E2f1, but not E2f3, reversed deregulated expression of cell cycle and apoptotic genes in the Rb KO retina. E2f3 is expressed in a subset of CNS neurons (this work) and drives specific cell-cycle–independent defects in Rb KO forebrain neurons [53]. Thus, E2f3 inhibition is the first, and may be the only, mechanism by which Rb participates directly in neuronal differentiation. To further dissect the mechanism of action of Rb in SACs we determined the E2f3 isoform it targets to promote differentiation. E2f3b was the primary candidate, since Rb and E2f3b collaborate to repress targets in quiescent cells in vitro [19]. However, in the first work to our knowledge to examine the function of any E2f protein isoform in vivo, we made the surprising observation that Rb regulates SAC differentiation through E2f3a (Figure 8). Formally, we cannot exclude the possibility that deleting E2f3b might also rescue SAC differentiation, but definitive proof will require analysis of E2f3b null mice. Nevertheless, our data prove that Rb definitely regulates SAC differentiation through the activating E2f3 isoform. The subcellular location of E2f isoforms has not to our knowledge been addressed before. E2f3a and E2f3b share 110 C-terminal amino acids that encode the NLS, DNA-binding, marked box, transactivation, and Rb-binding domains [16], yet they exhibit different subcellular distribution in developing retinal cells. E2f3a is both nuclear and cytoplasmic, but E2f3b is always nuclear. The unique 121- and six-residue N-termini of E2f3a and E2f3b, respectively, likely mediate this difference. This region in E2f1, E2f2, and E2f3a binds Ccna2, establishing a negative regulatory loop that deactivates E2fs in mid-late S-phase [64,65]. However, even E2f3b, which lacks this domain, binds and is regulated by Ccna2 [18], so the domain difference may not explain the unique distributions we observed. Rb family and Tfdp proteins can also determine E2f localization [20–22], and we found that a portion of both Rb and Tfdp1 proteins are cytoplasmic in retinal cells. Indeed, immunostaining revealed that Rb and E2f3 colocalize to SAC processes. The nuclear localization of E2f3b contrasts with that of other repressive E2fs in differentiating muscle, where E2f5 switches from the nucleus to cytoplasm, while E2f4 remains in both compartments [23]. The distinct compartmentalization of E2f3a and E2f3b in the retina suggests temporally and functionally distinct activities. Rb distribution matches that of E2f3a, consistent with its critical role in supporting SAC differentiation through E2f3a. Rb is critical to ensure that many types of terminally differentiating cells leave the cell cycle (e. g. , neurons, gut and skin epithelia, muscle, and lens fibres) (reviewed in [66]). Early overexpression studies in vitro suggested Rb might temper expansion of cycling cells, but KO studies in vivo indicate that its major role is to block division in terminally differentiating cells. In its absence, many (but clearly not all) aspects of differentiation go ahead relatively unperturbed. In the retina, differentiating transition cells are born in the absence of Rb, migrate to the correct layer, and express appropriate markers ([2] and this work). In brain, Rb KO neurons migrate away from the ventricular zone and switch on Tubb3 (βIII-tubulin), but continue to incorporate BrdU [13], and in gut epithelia, differentiated enterocytes migrate up the villi and activate expression of serotonin, yet continue to incorporate BrdU [67]. In the case of SACs, the differentiation defects we observed (e. g. , loss of Slc18a3 and Chat) were not due to aberrant division, but it is possible there are other problems with these cells that are caused by ectopic division. Nevertheless, it is clear that many aspects of differentiation in multiple cell types are compatible with ectopic division. However, division of terminally differentiating cells is dangerous, since it may facilitate transformation, as is the case in retinoblastoma (reviewed in [66]). E2f3a could disrupt SAC differentiation through its well known role as a transcriptional activator, or, in view of the discovery that it is partially cytoplasmic, E2f3a may affect processes other than gene regulation. Both scenarios are feasible since E2fs regulate differentiation genes [44–48], and cell cycle regulators, such as Cdkn1b, have cytoplasmic activities that influence differentiation [68,69]. Many transcription factors shuttle between nucleus and cytoplasm during neurogenesis (e. g. , [70] and references therein). It may be difficult to identify E2f3a-specific target genes or cytoplasmic proteins in SACs since these neurons are a small proportion (<1%) of the total retina and only ~5. 2% of amacrine neurons [38]. Others have suggested that Rb promotes differentiation in non-neuronal cells through E2f-independent means [31–33]. However, these studies did not assess whether these cell types differentiate normally if Rb is deleted along with one or more E2f family members. One study reported that Rb mutants that do not bind E2f still induce differentiation [30]. However, the binding assays were performed in solution, and we have found that several of these mutants do bind E2f, albeit weakly, on chromatin (T. Yu and R. B. , unpublished data). It is possible that Rb-mediated potentiation of tissue-specific transcription factors may, at least in some cases, be a redundant activity, and that the only critical Rb function is to inhibit E2f. Our study is the first to our knowledge to assess comprehensively whether Rb KO cells can differentiate in the absence of different E2fs. In light of our findings, it will be important to reassess differentiation defects in other Rb KO tissues in the absence of individual and combined activating E2f family members. Mice were treated according to institutional and national guidelines. α-Cre mice (P. Gruss), Chx10-Cre mice (C. Cepko), RbloxP/loxP mice (A. Berns), E2f1–/– mice, E2f2–/– mice, E2f3loxP/loxP mice, and E2f3a−/− mice were maintained on a mixed (NMRI × C57/Bl × FVB/N × 129sv) background. A detailed description of E2f3a−/− mice will be published elsewhere. Mice of different genotypes were compared within the same litter and across a minimum of three litters. We have not noted any phenotypic differences in separate litters. Genotyping was performed as before [2,5], and the primers used for genotyping E2f3a−/− mice were E2f3a KL (5′-CTCCAGACCCCCGATTATTT-3′), E2f3a KR1 (5′-TCCAGTGCACTACTCCCTCC-3′), and E2f3a KM (5′-GCTAGCAGTGCCCTTTTGTC-3′). Eyeballs were fixed in 4% paraformaldehyde for 1 h at 4 °C, embedded in OCT (TissueTek 4583, Sakura, http: //www. sakuraeu. com), frozen on dry ice, and cut into 12-μm sections on Superfrost plus slides (VWR, http: //www. vwr. com). For S-phase analysis, BrdU (100 μg/g of body weight) was injected intraperitoneally 2 h prior to sacrifice. BrdU+ cells were detected using a biotin-conjugated sheep polyclonal antibody (1: 500, Maine Biotechnology Services, http: //www. mainebiotechnology. com). All other antibodies are described in Table S1. For E2f3, Mki67, and Rb staining, antigen retrieval was performed by boiling sections in citric acid solution for 15 min according to Ino [50], except on frozen sections. TUNEL was performed as described [13]. Briefly, sections were incubated for 1 h at 37 °C with 75 μl of mixture solution consisting of 0. 5 μl of terminal deoxynucleotide transferase, 1 μl of biotin-16-dUTP, 7. 5 μl of CoCl2,15 μl of 5× terminal deoxynucleotide transferase buffer, and 51 μl of distilled water. After three washes in 4× SSC buffer, sections were incubated with Alexa 488– or Alexa 568−streptavidin (1: 1,000; Molecular Probes, http: //probes. invitrogen. com) for 1 h at room temperature. Primary antibodies or labelled cells were visualized using donkey anti-mouse Alexa 488 or Alexa 568, donkey anti-rabbit Alexa 488 or Alexa 568, donkey anti-goat Alexa 488 or Alexa 568, and streptavidin Alexa 488 or Alexa 568 (1: 1,000; Molecular Probes). Nuclei were counter-stained with 4,6-diamidino-2-phenyindole (DAPI; Sigma, http: //www. sigmaaldrich. com). Labelled cells were visualized using a Zeiss (http: //www. zeiss. com) Axioplan-2 microscope with Plan Neofluar objectives and images captured with a Zeiss AxionCam camera. For double-labelled samples, confocal images were obtained with a Zeiss LSM 5. 0 laser scanning microscope. The retina was separated into three bins by dividing the ventricular edge of the retina into equal parts and extending a line to the vitreal edge [2]. Bin 1 contains only cells that expressed Cre as progenitors; bin 3 is at central retina and contains cells derived from progenitors that did not express Cre. For cell counts or thickness measurement we used a region 0–100 μm peripheral to the boundary separating bins 1 and 2. Measurements were performed on an Axioplan-2 microscope using Axiovison software. Quantification of S-phase, M-phase, and apoptotic cells was performed on horizontal sections that included the optic nerve. Quantification of differentiated cell types was performed using horizontal sections at equal distances from the optic nerve. A minimum of three sections per eye and three eyes from different litters were counted. Total RNA was isolated from dissected peripheral retina using TRIzol reagent (Invitrogen, http: //www. invitrogen. com) followed by digestion with RNase-free DNase (DNA-free, Ambion, http: //www. ambion. com) to remove DNA contamination. First-strand cDNA was synthesized from 0. 2–0. 5 μg of total RNA using the SuperScript II first-strand synthesis system (Invitrogen). PCR primers are listed in Table S2. Real-time quantitative PCR was performed using an Applied Biosystems (http: //appliedbiosystems. com) PRISM 7900HT. Tests were run in duplicate on three separate biological samples with SYBR Green PCR Master Mix (Applied Biosystems) exactly as we described previously [71]. Briefly, master stocks were prepared such that each 10-μl reaction contained 5 μl of SYBR Green PCR Master Mix, 0. 1 μl of each forward and reverse primer (stock 50 μM), 0. 8 μl of blue H2O (0. 73% Blue Food Colour; McCormick, http: //www. mccormick. com), 2 μl of diluted cDNA template, and 2 μl of yellow H2O (0. 73% Yellow Food Colour). PCR consisted of 40 cycles of denaturation at 95 °C for 15 s and annealing and extension at 55 °C for 30 s. An additional cycle (95 °C, 15 s, 60 °C) generated a dissociation curve to confirm a single product. The cycle quantity required to reach a threshold in the linear range was determined and compared to a standard curve for each primer set generated by five 3-fold dilutions of genomic DNA or cDNA samples of known concentration. Values obtained for test RNAs were normalized to Hprt1 mRNA levels. Mouse retinas were homogenized by passing them through a 30-gauge BD 9 http: //www. bd. com) needle 5–10 times in 1× PBS solution. Nuclear and cytoplasmic proteins were extracted using the NE-PER Nuclear and Cytoplasmic Extraction Kit (Product# 78833, Pierce Biotechnology, http: //www. piercenet. com). Proteins were separated by 10% SDS-PAGE and transferred to nitrocellulose. After blocking overnight at 4 °C in 5% skim milk, membranes were incubated in the primary antibody for 2 h at room temperature. After three 10-min washes in TPBS (100 mM Na2HPO4,100 mM NaH2PO4,0. 5 N NaCl, 0. 1% Tween-20), membranes were incubated for 30 min at room temperature in the secondary horseradish peroxidase-conjugated antibody (Jackson ImmunoResearch Laboratories, http: //www. jacksonimmuno. com). Blots were developed using the ECL-Plus chemiluminescent detection system (Amersham Pharmacia Biotech, http: //www. pharmacia. ca), according to the manufacturer' s instructions. The following primary antibodies were used: E2f-1 (SC-193), E2f-3 (SC-878), Cdkn1a (p21, SC-471), Cdkn1b (p27, SC-528), Pou4f2 (Brn3b, SC-6062), and Tfdp1 (Dp1, SC-610) from Santa Cruz Biotechnology (http: //www. scbt. com), pRB (554136) from BD Science-Pharmingen (http: //www. bdbiosciences. com), and Slc18a3 (VAChT, G448A) from Promega (http: //www. promega. com). ERGs were recorded from dark-adapted mice as described [72]. Briefly, mice were dark-adapted overnight and anaesthetized by subcutaneous injection of ketamine (66. 7 mg/kg body weight) and xylazine (11. 7 mg/kg body weight). The pupils were dilated and single-flash ERG recordings were obtained under dark-adapted (scotopic) and light-adapted (photopic) conditions. Light adaptation was accomplished with a background illumination of 30 candela (cd) per square meter starting 10 min before recording. Single white-flash stimulation ranged from 10−4 to 25 cd·s/m2, divided into ten steps of 0. 5 and 1 log cd·s/m2. Ten responses were averaged with an inter-stimulus interval of either 5 s (for 10−4,10−3,10−2,3 × 10−2,10−1, and 3 × 10−1 cd·s/m2) or 17 s (for 1,3, 10, and 25 cd·s/m2). Band-pass filter cut-off frequencies were 0. 1 and 3,000 Hz. Different genotypes were evaluated using analysis of variance (ANOVA) followed by the Tukey honestly significant difference (HSD) test or Fisher test (XLSTAT program, http: //www. xlstat. com). The GenBank (http: //www. ncbi. nlm. nih. gov/genbank) accession numbers for the major genes and gene products discussed in this paper are Camk2a (NM_009792), Chat (NM_009891), E2f1 (NM_007891), E2f2 (NM_177733), E2f3 (NM_010093), Rb (NM_009029), and Slc18a3 (NM_021712).

The retinoblastoma protein (Rb), an important tumor suppressor, blocks division and death by inhibiting the E2f transcription factor family. In contrast, Rb is thought to promote differentiation by potentiating tissue-specific transcription factors, although differentiation defects in Rb null cells could be an indirect consequence of E2f-driven division and death. Here, we resolve different mechanisms by which Rb controls division, death, and differentiation in the retina. Removing E2f1 rescues aberrant division of differentiating Rb-deficient retinal neurons, as well as death in cells prone to apoptosis, and restores both normal differentiation and function of major cell types, such as photoreceptors. However, Rb-deficient starburst amacrine neurons differentiate abnormally even when E2f1 is removed, providing an unequivocal example of a direct role for Rb in neuronal differentiation. Rather than potentiating a cell-specific factor, Rb promotes starburst cell differentiation by inhibiting another E2f, E2f3a. This cell-cycle–independent activity broadens the importance of the Rb–E2f pathway, and suggests we should reassess its role in the differentiation of other cell types.

Abstract Introduction Results Discussion Materials and Methods Supporting Information

developmental biology cell biology mus (mouse) neuroscience

2007

Rb-Mediated Neuronal Differentiation through Cell-Cycle–Independent Regulation of E2f3a

Accumulating evidence supports the role of an aberrant transcriptome as a driver of metastatic potential. Deadenylation is a general regulatory node for post-transcriptional control by microRNAs and other determinants of RNA stability. Previously, we demonstrated that the CCR4-NOT scaffold component Cnot2 is an inherited metastasis susceptibility gene. In this study, using orthotopic metastasis assays and genetically engineered mouse models, we show that one of the enzymatic subunits of the CCR4-NOT complex, Cnot7, is also a metastasis modifying gene. We demonstrate that higher expression of Cnot7 drives tumor cell autonomous metastatic potential, which requires its deadenylase activity. Furthermore, metastasis promotion by CNOT7 is dependent on interaction with CNOT1 and TOB1. CNOT7 ribonucleoprotein-immunoprecipitation (RIP) and integrated transcriptome wide analyses reveal that CNOT7-regulated transcripts are enriched for a tripartite 3’UTR motif bound by RNA-binding proteins known to complex with CNOT7, TOB1, and CNOT1. Collectively, our data support a model of CNOT7, TOB1, CNOT1, and RNA-binding proteins collectively exerting post-transcriptional control on a metastasis suppressive transcriptional program to drive tumor cell metastasis. Metastasis is a complex process in which tumor cells disseminate from the primary tumor site to form life-threatening lesions at distant sites. To successfully complete the metastatic process tumor cells must activate a series of molecular functions. These include motility and invasion to escape the primary site and penetrate the parenchyma at the secondary organ, anti-apoptotic programs to survive anoikis during transit through the lymphatic or hematic vasculature, and proliferative programs to establish clinically relevant macroscopic lesions [1]. Each of these programs requires the action of multiple genes in a coordinated fashion. As a result, control of transcriptional programs in the metastatic cascade has been the focus of many studies over the past decade. For example, activation of embryonic programs through up-regulation of transcription factors is thought to be important in the migratory and invasive steps of the metastatic cascade [2,3]. Post-transcriptional control of metastasis-associated genes by microRNAs has also been the subject of a variety of studies [e. g. [4,5]]. Activation or suppression of these pleiotropic factors, through mutation, amplification, or deletion, therefore plays critical roles in tumor evolution and progression. In addition to activation or suppression of whole transcriptional programs, factors that significantly alter transcriptional units might also alter the ability of a tumor cell to complete one or more of the steps of the metastatic cascade. Studies in recent years have demonstrated an important role for inherited polymorphism in gene expression programs [6,7] suggesting that inherited factors can significantly influence tumor phenotypes. Our laboratory previously demonstrated that inherited polymorphisms significantly influence metastatic outcome [8] and that inheritance plays a role in the establishment of transcriptional profiles that discriminate patient outcome [9]. More recently we have integrated gene expression analysis and susceptibility genetics studies to identify co-expressed transcriptional networks associated with metastatic disease. One such network was centered on Cnot2, a scaffolding component of the CCR4-NOT RNA deadenylase complex [10]. In vivo validation studies demonstrated that varying CNOT2 levels significantly influenced tumor metastatic capacity and implicated the CCR4-NOT complex as a novel determinant of tumor cell metastatic potential [10]. The CCR4-NOT complex is a modular, multifunctional protein complex highly conserved in eukarya [11]. Components of CCR4-NOT are found both in the nucleus and cytoplasm, and mediate transcriptional and post-transcriptional regulatory functions [12,13]. In mammalian cells, CCR4-NOT has reported roles in epigenetically mediated transcriptional regulation [14], nuclear hormone receptor-mediated transcription [15], and initiation of transcript decay by deadenylation [16–18]. These observations suggest that the CCR4-NOT complex is a pleiotropic regulator of transcript abundance. Cnot2 depletion has been shown to disrupt CCR4-NOT deadenylase activity [19] which may be expected to alter metastasis-associated transcriptional programs. The absence of CNOT2 catalytic activity led us to hypothesize that other CCR4-NOT effector functions may drive metastasis. Moreover, the co-expression network analyses that identified Cnot2 also implicated the CCR4-NOT deadenylase Cnot8 and its binding partner Tob1 as candidate metastasis driving genes [10], suggesting that CCR4-NOT deadenylase function may be an important determinant for metastatic progression. Deadenylation, the progressive 3’-to-5’ shortening of the polyA tail, is a rate-limiting step of transcript destabilization [20]. Translational inhibition and transcript decay mediated by microRNAs [21], AU-rich elements [22], RNA binding proteins [23–25], and nonsense-mediated decay [26] occur through the recruitment of deadenylase complexes. The CCR4-NOT complex therefore plays an important role in maintaining mRNA equilibrium and coordinated control of transcriptional programs. We previously demonstrated that modulation of transcriptional elongation, mediated by Brd4 [27], significantly altered the metastatic capacity of mammary tumor cells. In this study we extend these results by demonstrating that transcriptional decay, mediated by the deadenylation activity of Cnot7 is also an important determinant in tumor progression. Our findings are consistent with the existence of post-transcriptional regulatory deadenylase complexes that promote metastasis by destabilizing metastasis suppressive transcriptional programs. Importantly, our work identifies CNOT7 deadenylase activity as a novel therapeutic target for anti-metastatic therapy. Metastasis assays by orthotopic implantation were performed to test if perturbation of Cnot7 or Cnot8 expression alters metastatic capacity. Cnot7 was knocked down in three independent murine mammary tumor cell lines [28,29] by stable transduction of short hairpin RNAs (shRNAs), resulting in reductions of protein and transcript abundance (Fig 1). Cnot8 was knocked down with the same method in 6DT1 and Mvt1 cells. After selection, transduced cells were implanted into the fourth mammary fat pad of syngeneic mice. At assay endpoint (t = 30 days), Cnot8 knockdown did not produce consistent results between the tested cell lines (S1 Fig) and therefore was not included in further studies. Cnot7 knockdown consistently reduced pulmonary metastasis without significant effect on primary tumor mass in vivo (Fig 1). To validate these results in an independent experimental system we next assessed the effect of diminished Cnot7 expression in an autochthonous transgene-driven metastasis model. The MMTV-PyMT transgenic mammary tumor model [30] was bred to Cnot7 hemizygous knockout [31] mice to produce PyMT+ Cnot7+/- and PyMT+ Cnot7 wild type mice (Fig 2A). Quantitative real time polymerase chain reaction analysis confirmed that Cnot7+/- mice expressed Cnot7 transcript approximately two-fold lower than wildtype mice in the spleen and tumor (Fig 2B & 2C). Consistent with observations in the orthotopic metastasis model, deletion of one copy of Cnot7 significantly reduced the metastatic incidence and burden with no effect on primary tumor mass (Fig 2D–2F). To test if reduction of Cnot7 expression in the stroma influenced metastasis, we then crossed C57BL/6J Cnot7+/- to FVB/NJ or Balb/cJ mice to generate Cnot7+/- and Cnot7+/+ mice that are immune-tolerant to the FVB- or BALB-derived tumor cells, respectively (Fig 3A). Wild type tumor cells were then orthotopically implanted into Cnot7+/- and Cnot7+/+ mice and animals were aged for 28 days prior to necropsy. No consistent difference in tumor mass or metastasis was observed (Fig 3B–3J) suggesting the primary role of Cnot7 in metastatic progression is tumor cell-autonomous. Tumor cell colonization of distant organs is thought to be the rate-limiting step of the metastatic cascade [32,33]. Lung colonization assays were therefore performed by intravenous injection of Cnot7-depleted tumor cells into the tail vein of mice (Fig 4A). Three of four Cnot7 knockdown conditions resulted in significant suppression of lung colonization (Fig 4B–4D). Cross-sectional area of metastatic lesions was subsequently measured in hematoxylin and eosin (H&E) stained lung sections to determine if differences in colonization occurred secondary to proliferative differences. No difference in metastatic size between control and Cnot7-depleted cells was observed for either 6DT1 or 4T1 cells (Fig 4E & 4F). These results are consistent with a role of Cnot7 promoting early stages of lung colonization in a proliferation-independent manner. To investigate the potential cellular mechanisms underlying the suppression of metastatic capacity upon Cnot7 depletion proliferation, motility, and colony formation in low attachment conditions were assessed. Cnot7 depletion resulted in a consistent reduction of cellular proliferation in all three cell lines tested (S2 Fig) although, as noted above, this proliferative suppression was not observed in the in vivo orthotopic implantation assays. Motility assays were performed on Cnot7 depleted 6DT1 and Mvt1 cells. Cnot7 knockdown in Mvt1 cells resulted in a reduced motility in a wound healing assay. In contrast, Cnot7 knockdown in 6DT1 using the same shRNA constructs showed no consistent difference in motility as measured by wound healing assay (S2 Fig). Similar discrepancies across cell lines were observed in soft agar assays, where Cnot7 depletion in 6DT1 resulted in significant reduction of colony formation while no difference was observed in Mvt1 cells (S2 Fig). Due to the lack of consistency among the in vitro assays, wound healing and soft agar assays were not performed for 4T1 Cnot7 depleted cell lines. Overall the ambiguous results of the in vitro assays suggested that further investigations into the mechanisms of the role of Cnot7 in metastatic disease would be best examined in vivo. Further efforts therefore focused on metastatic capacity based on orthotopic transplant assays. The CCR4-NOT complex has been implicated in multiple cellular functions, including RNA deadenylation and degradation as well as transcriptional control [12]. To determine whether CNOT7-mediated metastasis promotion was deadenylase-dependent, we stably expressed the wild type and a deadenylase-inactive point mutant (D40A) [34–36] in mammary tumor cell lines. Transduced cultures were selected for approximately equal expression of CNOT7 and D40A protein (Fig 5A–5C), and used for in vivo orthotopic transplant assays. Ectopic expression of CNOT7 or the D40A mutant showed no differences in primary tumor mass in Mvt1 or 6DT1 cells (Fig 5D & 5F). However, CNOT7 but not D40A significantly promoted metastatic burden for 6DT1 and Mvt1 (Fig 5G, 5I, 5J and 5L). In 4T1 cells, a statistically significant increase in tumor mass was observed for CNOT7 but not for D40A expressing cells (Fig 5E). Normalization of metastatic burden by tumor mass to account for difference in primary tumor growth still resulted in a significant difference in metastatic capacity for CNOT7 expressing 6DT1 or 4T1 cells and borderline significance in Mvt1 cells. Furthermore, although the 4T1 D40A expressing cells exhibited increased metastasis compared to control, a significant reduction of metastatic capacity was observed compared to CNOT7-wild type expressing cells (Fig 5H and 5K). Overall these results are consistent with a major role of the deadenylase function of CNOT7 in modulating metastatic capacity of mammary tumor cell lines. CNOT7 is a non-specific RNA-binding deadenylase protein [35]. Specificity for transcripts is mediated by sequence specific RNA binding proteins that interact with the CCR4-NOT complex. TOB1 is an adaptor protein that recruits CNOT7 to specific RNA-binding proteins, while CNOT1 is a scaffolding protein for the CCR4-NOT complex [37,38] (Fig 6A and 6B). Previous work from our laboratory found that Tob1 expression was correlated with metastasis in the [PyMT x AKXDn]F1 mice [10] (Fig 6C). In addition, human breast cancer datasets–available through the Gene expression-based Outcome for Breast cancer Online (GOBO) database, an expression array-based meta-analysis data set of 1,881 breast cancer patients [39]–also showed that high expression of either TOB1 or CNOT1 correlated with poor survival (Fig 6D & 6E). Furthermore TOB1 has previously been associated with poor distant metastasis free survival in breast cancer patients [40]. Orthotopic metastasis assays were conducted to test the role of the CCR4-NOT adaptor proteins in metastatic disease. Attempts to generate stable Cnot1 knockdown cells were unsuccessful and therefore orthotopic assays were not performed. Tob1 knockdown in 6DT1 cells (Fig 6F) showed diminished metastasis with no effect on primary tumor mass (Fig 6G–6J). Experimental metastasis assays showed that Tob1 knockdown suppressed lung colonization of 6DT1 cells (Fig 6K), consistent with Tob1 acting at the same stage of the invasion-metastasis cascade as Cnot7. In 4T1 cells, Tob1 knockdown suppressed tumor mass and pulmonary metastasis (S3 Fig). Significant suppression of metastasis was observed after normalizing by tumor mass, consistent with an effect other than just on tumor growth (S3 Fig). To test if CNOT7-mediated metastasis promotion was dependent on a complex with TOB1 or CNOT1, we constructed expression vectors of CNOT7 mutants that disrupted interaction with TOB1 or CNOT1. CNOT7 E247A/Y260A mutations have previously been shown to disrupt interaction with BTG/TOB family proteins but maintain interaction to CNOT1 [36]. Conversely, CNOT7 M141R substitution abolishes complex formation between CNOT7 and CNOT1 [34], but interaction with TOB1 has not previously been assessed. Co-immunoprecipitation (IP) experiments were therefore performed to address this question. As expected, IP of CNOT7 E247A/Y260A but CNOT7 M141R co-precipitated CNOT1. In contrast, the CNOT7 M141R mutant retained the ability to co-precipitate TOB1 but no longer interacted with CNOT1 (Fig 7A & 7B), indicating that the two mutant constructs specifically disrupted interaction with the TOB1 or CNOT1 adaptor proteins. CNOT7 and the mutant constructs were then expressed in 4T1 cells to achieve equal levels of CNOT7, CNOT7 E247A/Y260A, and CNOT7 M141R protein (Fig 7C) and the cells were then implanted orthotopically into mice. No significant changes in the rate of tumor growth or primary tumor mass at endpoint were observed (Fig 7D). Consistent with previous observations, overexpression of CNOT7 promoted tumor cell metastatic potential but expression of CNOT7 E247A/Y260A or CNOT7 M141R showed no change in metastasis compared to control (Fig 7E and 7F). These results indicate that CNOT7-mediated metastasis promotion depends on contact with both TOB1 and CNOT1. The above results suggest that Cnot7 mediates its effect on metastasis by modulating the RNA equilibrium. Therefore, to gain a better idea of the global gene expression program affected by Cnot7, we identified mRNAs that exhibited an inverse relationship with Cnot7 expression in 4T1 cells in which Cnot7 was knocked down or over-expressed. Array-based transcriptome analysis yielded 842 significantly dysregulated transcripts (p<0. 01, S1 Table). Of these, 514 transcripts were upregulated upon Cnot7 knockdown and down-regulated upon CNOT7 overexpression (t<0, S1 Table; Cnot7-anticorrelated transcripts). 3’ untranslated regions (3’UTRs) were then interrogated for known consensus RNA-binding protein (RBP) sequence motifs enriched in the Cnot7 inversely-correlated transcripts (S4 Fig). The inversely correlated transcripts showed enrichment for the cytoplasmic polyadenylation element (CPE) [41], Pumilio binding element (PUM) [42], Nanos response elements (NRE) [43], and cleavage and polyadenylation stimulation factor binding element (CPSF) [44]. In contrast neither permissive (AUUUA) nor stringent (UUAUUUAUU) AU-rich elements (ARE) [45] were found to be enriched indicating that, in 4T1 cells, Cnot7 preferentially mediates the degradation of a specific subset of mRNAs (S4 Fig). To identify transcripts directly regulated by CNOT7, RNA-immunoprecipitation was performed using the anti-FLAG (M2) antibody in 4T1 cells overexpressing FLAG-CNOT7. Co-precipitated RNA was subjected to high throughput sequencing (RIP-seq). 149 transcripts showed enrichment relative to input and control (M2 RIP in 4T1 empty vector cells not expressing FLAG-CNOT7) and showed an inverse correlation with Cnot7 expression (Fig 8A). These transcripts showed 3’UTR enrichment for CPE, CPSF, and NRE sites (Fig 8B). PUM and ARE sites were not considered since fewer than 10 of the 149 genes contained these binding motifs. Seventy-one of 149 transcripts (48%) possessed CPE, CPSF, and NRE sites (Fig 8C, S2 Table) suggesting that CPEB, CPSF, and Nanos family proteins may collectively constitute specificity factors that cooperatively drive metastasis by targeting CNOT7 to metastasis-associated transcripts. We next tested if the 71 transcripts that shared the tripartite motif were prognostic in human breast cancer data sets. Forty-six (65%) human orthologs of the CPE/CPSF/NRE containing genes were present in GOBO [39] (S3 Table). These 46 genes were applied as a gene signature, weighted by their inverse correlation with CNOT7 expression, to determine whether they could discriminate patient outcome. Consistent with the possibility that CNOT7 drives progression by degrading metastasis suppressing mRNAs, high expression of the signature was correlated with favorable distant metastasis free survival (DMFS, Fig 8D). Since CNOT7-bound and inversely regulated transcripts were reproducibly enriched for the CPE/CPSF/NRE motifs, we speculated that this tripartite motif may specify CNOT7 target transcripts, which are enriched for metastasis-associated genes. We thus interrogated the entire genome for transcripts that possessed CPE, CPSF, and NRE 3’UTR elements, and filtered this set for those transcripts expressed in 4T1 cells. This analysis identified 3424 genes (12. 5% of the mouse genome, Fig 8E). Filtering this set for those inversely-correlated with Cnot7 expression yielded 293 genes (S4 Table). Human orthologs for 217 (74%) of the 293 genes were available in GOBO, and were able to significantly discriminate DMFS (Fig 8F). High expression of this signature predicted favorable survival, suggesting that this set of tripartite motif containing, Cnot7-anticorrelated transcripts constituted a metastasis suppressive post-transcriptional program. Subjecting this 293-gene set to Ingenuity Pathway Analysis identified cancer as the most highly represented disease annotation, with 247 (84%) transcripts previously annotated as cancer-associated. The top five represented canonical pathways associated with extravasation signaling, and cancer-associated HER2 signaling, HGF signaling, and MAPK-signaling. The top five upstream regulators in this network included Kras, Erbb2, and Tgfb1 signaling (S5 Table). Gene sets regulated by each of these upstream regulators predicted distant metastasis free survival (S5 Fig). Consistent with our model, Erbb2 and Kras are previously reported upstream signaling components that regulate the activity of CNOT7 adaptor and deadenylation cofactor TOB1, transducer of ERBB2 [46,47]. CCR4-NOT is a highly conserved protein complex that has been implicated in diverse functions associated with gene regulation [12,13]. It exists in both cytoplasmic and nuclear forms and is thought to play different roles depending on its subcellular localization and subunit composition [12]. In the nucleus the CCR4-NOT complex has been implicated in numerous activities, including chromatin modification, transcriptional elongation, RNA export, nuclear RNA surveillance and transcription-coupled DNA repair [14,15]. In the cytoplasm the CCR4-NOT complex is thought to be the main RNA deadenylase, initiating both mRNA decay and translational repression by polyA tail shortening [12,13]. The CCR4-NOT complex also participates in miRNA-mediated gene silencing through interactions with the GW182/Argonaute complex [21]. This large, multi-functional protein complex therefore has the potential to play a variety of important roles in establishing and maintaining cellular function and response to extracellular cues. Previously our laboratory implicated the CCR4-NOT complex as an important determinant for metastatic mammary cancer. Generation of co-expressed gene maps from mouse strains with differing inherited sensitivity for pulmonary metastasis identified a network module centered on the CCR4-NOT component Cnot2 that was capable of discriminating breast cancer patient outcome. Furthermore, in vivo modeling demonstrated that suppression or over-expression of CNOT2 within tumor cells resulted in enhanced or reduced pulmonary metastases, respectively, indicating that Cnot2 has metastasis suppressing activities [10]. CNOT2 however does not have known enzymatic activities [48]. CNOT2 coordinates the interaction of CNOT3 with the core CCR4-NOT complex, as well as additional regulatory molecules including HDAC3 [14]. Despite lacking catalytic function, CNOT2 has been shown to be an important positive regulator of CCR4-NOT deadenylase activity [19], cellular apoptosis, and mouse embryonic stem cell pluripotency [49]. The contribution of Cnot2 to metastatic capacity through the CCR4-NOT complex could therefore occur through a variety of CCR4-NOT molecular functions. In this study we have begun to dissect the role of CCR4-NOT in mammary tumor metastasis by investigating the role of RNA deadenylation in tumor progression. The integrated genetics and gene expression analysis that initially identified Cnot2 also implicated other genes associated with RNA deadenylation (Cnot8, Angel2, Tob1) as potential modulators of metastatic disease, suggesting that this function of CCR4-NOT might be a critical determinant [10]. Bioinformatics analysis indicated that CNOT8 and TOB1 were associated with distant metastasis free survival in human patients. We therefore selected Cnot8 and its highly conserved paralog Cnot7 to determine whether the deadenylation function of CCR4-NOT plays a critical role in tumor progression. Cnot7 and Cnot8 are members of the DEDD superfamily of deadenylases [35]. Both genes are expressed ubiquitously in tissues of adult animals [50] and are thought to have overlapping functions [36]. Biochemical studies suggest that only one of the two proteins exist in the CCR4-NOT complex at a time, suggesting unique functions for the mutually exclusive CNOT7- or CNOT8-containing complexes, in addition to redundant functions [34,51]. This interpretation is consistent with the differing results observed for the shRNA knockdowns of the two genes in our studies. Cnot7 knockdown had little or no effect on primary tumor growth, indicating that its role in tumor progression is related to the metastatic process. In contrast Cnot8 suppressed both primary tumor growth and metastatic disease, suggesting a more general role in regulating tumor cell proliferation. Further elucidation of the commonalities and differences in molecular pathways controlled by these two deadenylases would likely provide interesting insights into tumor growth and progression. Importantly, due to the multifunctional nature of the CCR4-NOT complex, the ability of Cnot7 to promote metastatic disease was dependent on deadenylase activity. Point mutations eliminating enzymatic activity or that disrupted interactions mediating recruitment of RNA binding proteins to the CCR4-NOT complex suppressed the pro-metastatic activity of CNOT7. Knockdown of the adaptor protein TOB1, which acts as a bridge between CNOT7 and RNA binding proteins CPEB3 and CPEB4, which recruit RNAs to the complex for deadenylation, had similar effects. Attempts to knockdown CNOT1, which is responsible for recruitment of the NANOS1 and PUM2 RNA binding proteins, was unsuccessful. However, due to the central role of CNOT1 in the CCR4-NOT complex and increased apoptosis in CNOT1-depleted cells [52–54], this result was not unexpected. Taken together however, the point mutant and Tob1 depletion results suggest that the majority of the effect on metastasis was likely due to the influence of CCR4-NOT on RNA equilibrium or translational efficiency, rather than on the other many functions ascribed to the complex. The CCR4-NOT complex is thought to be one of two general deadenylase complexes in mammalian cells [18]. However, the difference in phenotypes for Cnot7 and Cnot8 knockdowns suggest that the different complexes likely target overlapping subsets of RNAs within the cell [36]. Alternatively, the two paralogs may be differentially expressed and regulate different post-transcriptional programs in different cell types. To gain a better understanding of what subset Cnot7-containing CCR4-NOT complexes target global gene expression analysis was performed in two independent experiments. We focused specifically on genes that were inversely correlated with Cnot7 levels to enrich for those that were likely direct targets rather than those dysregulated due to secondary effects on the transcriptome. Since transcript recruitment to the CCR4-NOT complex is specified by RNA binding proteins a screen for RNA binding proteins was performed [23–25,38,55][56]. This screen revealed an enrichment of some but not all of the known RNA binding protein motifs, suggesting that the Cnot7 metastatic suppressive program is mediated by specific RNA binding partners. Further investigation of these RNA binding proteins and their RNA targets will likely provide additional insights into the molecular pathways important for metastatic progression. Encouragingly enrichment of three of the four RNA binding protein motifs (CPE, CPSF and NRE) was replicated in the Cnot7 RIP-seq experiment, providing more direct evidence of the interaction of the CNOT7-containing CCR4-NOT complex with the Cpsf, Nanos and Cpeb families of RNA binding proteins. Furthermore, almost half (71/149) of the RNAs identified by the RNA-immunoprecipitation contained all three motifs. When used as a weighted gene signature, this set of transcripts was capable of discriminating metastasis outcome in human breast cancer. Global transcriptome analysis in 4T1 mammary tumor cells revealed only 293 expressed genes bearing all three RNA binding protein motifs, consistent with previous findings of only limited numbers of genes changing after Cnot7 knockdown [36]. Like the 71 genes identified by RIP-Seq, the 293 triple motif containing genes effectively discriminate outcome in breast cancer patients, suggesting enrichment of genes and molecular functions associated with breast cancer patients. This interpretation was further supported by pathway analysis, which identified previously known metastasis-associated functions such as extravasation and Tgfb1 signaling as enriched in the 293-gene set. Overall this study supports the hypothesis that in addition to initiation of specific transcriptional programs, such as epithelial-to-mesenchymal transition, degradation of specific RNAs may play an important role in the establishment of metastatic capacity. Further investigations of the RNA binding proteins that recruit transcripts for deadenylation and studies into possible roles of the other CCR4-NOT deadenylase subunits Cnot6 and Cnot6l in metastatic progression may reveal additional important insights into tumor autonomous metastatic mechanisms. Moreover, these results also suggest that targeting Cnot7 deadenylase activity may be useful for anti-metastatic therapy. Cnot7 knockout animals are viable, with limited known phenotypes, indicating that pharmaceutical suppression of Cnot7 deadenylase activity may not be unacceptably toxic in the clinic. If true, this would provide a novel class of therapeutic agents to suppress colonization or the emergence of disseminated but dormant tumor cells, ultimately leading to a reduction in the morbidity and mortality associated with metastatic disease. The research described in this study was performed under the Animal Study Protocol LCBG-004, approved by the NCI Bethesda Animal Use and Care Committee. Animal euthanasia was performed by cervical dislocation after anesthesia by Avertin. The mouse mammary carcinoma cell lines 4T1,6DT1, Mvt-1 [28] (provided by Dr. Lalage Wakefield) and human embryonic kidney HEK293 cells were cultured in Dulbecco’s Modified Eagle Medium (DMEM) (Gibco) supplemented with L-Glutamate (Gibco), 9% fetal bovine serum (FBS) (Gemini BioProducts), and 1% Penicillin and Streptomycin (P/S) (Gemini BioProducts). Two milliliter suspensions of 105 cells were incubated at 37°C in 5% CO2 overnight. Cells were then infected with lentivirus suspension, and selected 30 hours post-infection with 5mg/mL blasticidin for over-expression (Invitrogen) constructs or 10ug/mL (shRNA) puromycin for shRNA constructs. Co-immunoprecipitation was conducted as described in [57] using mouse origin anti-FLAG and Protein G Dynabeads magnetic beads (Invitrogen). Proliferation and wound healing assays were performed on the Incucyte ZOOM (Essen BioScience) system following the previously described protocols [58]. For soft agar assays, 5,000 trypsinized cells were seeded in triplicate in 0. 4% low-melting-point agarose (Sigma) on top of a 1% agarose layer and colonies enumerated 21 days later. Array-based transcriptome profiling of Cnot7-knockdown and CNOT7-overexpressing 4T1 tumor cells was performed on Affymetrix GeneChip Mouse Gene 1. 0 ST arrays by the Microarray Core in the NCI Laboratory of Molecular Technology. Library preparation was performed using the NEBNext Ultra Directional RNA Library Prep Kit for Illumina with NEBNext multiplexing oligos using manufacturer’s protocol. RNA sequencing was conducted on the Illumina HiSeq 2500. RNA was isolated from tumors and cell lines using RNeasy kit (Qiagen) or TriPure (Roche) and reverse transcribed using iScript (Bio-Rad). Real-Time PCR was conducted using VeriQuest SYBR Green qPCR Master Mix (Affymetrix). cDNA sequences of human FLAG-CNOT7, FLAG-CNOT7-D40A, and FLAG-CNOT7-E247A/Y260A were describe previously [36]. Lentiviral expression vectors were produced with Multisite Gateway recombination. An entry clone using the murine Pol2 promoter was recombined with the cDNA entry clone and N-terminal entry clone encoding the MYC (EQKLISEEDL) or FLAG (DYKDDDDK) epitope tag into a Gateway destination vector pDest-658. pDest-658 is a modified version of the pFUGW lentiviral vector which contains the enhanced polypurine tract (PPT) and woodchuck regulatory element (WRE) to provide higher titer virus. It also contains an antibiotic resistance gene for blasticidin resistance. Entry clones were subcloned by Gateway Multisite LR recombination using the manufacturer’s protocols (Invitrogen). Expression clones were transformed into E. coli STBL3 cells to minimize unwanted LTR repeat recombination, and verified by agarose gel electrophoresis and restriction digest. Transfection-ready DNA for the final clones was prepared using the GenElute XP Maxiprep kit (Sigma). A control vector (8166-M24-658) was generated by standard Gateway LR recombination of a stuffer fragment made up of a non-coding DNA into the pLenti6-V5-DEST vector (Invitrogen). CNOT7 and CNOT7-D40A lentivector constructs were generated by the Protein Expression Laboratory and the Viral Technology Group in NCI, Frederick, MD. Site directed mutagenesis was employed to generate CNOT7-M141R mutant using the primers listed in S6 Table. The cDNA segment containing these mutations was subcloned into the wild type CNOT7 lentivector by restriction digest and ligation. 4T1 cells expressing 8166-M24-658 control vector or FLAG-CNOT7 constructs were grown to ~85% confluence in ten 15cm culture plates. Cells were then rinsed with 15mL ice cold sterile PBS, scraped off, pelleted (~1mL pellet), and snap frozen in liquid nitrogen. Cells were thawed on ice and lysed in 3. 5mL lysis buffer (100mM NaCl, 5mM MgCl2,10mM HEPES (pH 7. 3), 0. 5% NP40,200 units RNasin (Promega), Protease inhibitor cocktail (Roche) ). The resulting 4mL of cell lysate was spun down twice at 21,000*g for 20 minutes at 4°C. 40uL of lysate was saved 1% input to confirm immunoprecipitation. 200uL of lysate was saved for 5% RNA-seq input and was purified by TriPure (Roche) RNA extraction. 25uL beads per 1mL lysate of Protein G Dynabeads were blocked with 1mL 0. 5% BSA in PBS at room temperature for 20 minutes then washed twice with 1mL NT2 wash buffer (50mM Tris-HCl (pH 7. 4), 150mM NaCl, 1mM MgCl2,0. 05% NP40). 24ug antibody was added to each sample and incubated at 4°C overnight then beads were added to each sample and rotated for 30 minutes at room temperature. Beads were washed four times in 1mL NT2 buffer. Eighty five percent of beads were subjected to RNA extraction with TriPure. Protein from the remaining 15% of beads was eluted with Laemli buffer to confirm immunoprecipitation. Protein was extracted with Pierce lysis buffer, vigorously homogenized, and incubated on ice for twenty minutes. 20ug lysate per sample in NuPage LDS Sample Buffer and NuPage Reducing Agent (Invitrogen) was used for western blotting. PVDF membrane (Millipore) containing transferred proteins was incubated overnight in solution of 5% milk protein, tris-buffered saline supplemented with 0. 05% Tween-20, and primary antibody. The membrane was then incubated with horse-radish peroxidase linked anti-mouse (GE Healthcare), anti-rat, or anti-rabbit (Santa Cruz Biotechnology) IgG secondary antibodies. Immunoblot was visualized using Amersham ECL Prime Western Blotting Detection System and Amersham Hyperfilm ECL (GE Healthcare). Rabbit origin anti-CNOT7 was a generous provided by G. Sebastiaan Winkler. Commercial antibodies used in this study include rabbit origin anti-TOB1 (GeneTex), rabbit origin anti-CNOT1 (Protein Tech), rat origin anti-HA (Roche), mouse origin monoclonal anti-FLAG (M2, Sigma). Female FVB/NJ or Balb/cJ mice from Jackson Laboratories were injected at 6–8 weeks of age. Two days prior to orthotopic injections, cells were placed in non-selective media. On the day of injection, 1x105 cells were injected orthotopically into the fourth mammary fat pad of age-matched virgin females. After 30 days the mice were euthanized by intraperitoneal injection of 1mL Tribromoethanol with subsequent cervical dislocation. Primary tumors were resected, weighed, and snap frozen in liquid nitrogen. Lungs were resected, surface metastases were counted; lungs were inflated with 10% nitrate-buffered formalin and sent for sectioning and staining. For tail vein injection, 105 were injected into the lateral tail vein, mice were euthanized 22 days post-injection. All procedures were performed under the Animal Safety Proposal (LCBG-004) and approved by the NCI-Bethesda Animal Care and Use Committee. Statistical analysis comparing two samples were conducted using the Mann-Whitney test on Prism Version 5. 03 (GraphPad Software, La Jolla, CA). Multiple-comparison data was analyzed by Kruskal-Wallis test with post-hoc Conover-Inman correction for multiple analyses by R-script. Survival data was conducted with the Mantel-Cox test on Prism. RBP motif sites were mapped on the 3’UTR regions of genome genes (27305 mRNA FASTA format) by using a Perl script. RBP motif site enrichment analysis was performed by the random sampling the genes from genome in the same number (the final overlapping gene number) and calculate the numbers of the gene with the RBP motif site and of the RBP motif site and repeat the sampling 1000 times. The p-value was estimated by one tail t-test. Differential gene expression analyses were done by t-test using the software package R. Array-based gene expression and RIP-seq studies from this study have been submitted to the NCBI Gene Expression Omnibus (GEO; http: //www. ncbi. nlm. nih. gov/geo/ under the accession numbers GSE73296 (array) and GSE73366 (RIP-seq).

The majority of human cancer related death is due to the effects of metastasis, the process of cancer dissemination to and growth in distant organs. Primarily due to its complexity, the metastatic process remains incompletely understood. This complexity stems from the tumor cell’s dependence on multiple cellular and molecular systems for the successful colonization of distant sites. In this study, we demonstrate that one of the factors that contributes to metastatic progression is the control of tumor cell RNA stability. Previously, we demonstrated that a structural component of the CCR4-NOT transcription regulatory complex was an inherited metastasis susceptibility gene. Here we demonstrate that one of the enzymatic components of the CCR-NOT complex, Cnot7, is also a metastasis-associated gene, and that enzymatic activity of Cnot7 is required for its promotion of metastatic disease. These results suggest that large-scale control of RNA abundance may be modulating specific metastasis-related transcriptional programs, and that inhibition of specific RNA deadenylases may be a viable avenue in the development of anti-metastatic therapeutics.

Abstract Introduction Results Discussion Materials and Methods

rna-binding proteins medicine and health sciences breast tumors protein interactions cancers and neoplasms basic cancer research oncology immunoprecipitation secondary lung tumors research and analysis methods proteins lung and intrathoracic tumors gene expression breast cancer precipitation techniques biochemistry metastasis genetics biology and life sciences

2016

Post-transcriptional Control of Tumor Cell Autonomous Metastatic Potential by CCR4-NOT Deadenylase CNOT7

Human metapneumovirus (hMPV) is a paramyxovirus that is a common cause of bronchiolitis and pneumonia in children less than five years of age. The hMPV fusion (F) glycoprotein is the primary target of neutralizing antibodies and is thus a critical vaccine antigen. To facilitate structure-based vaccine design, we stabilized the ectodomain of the hMPV F protein in the postfusion conformation and determined its structure to a resolution of 3. 3 Å by X-ray crystallography. The structure resembles an elongated cone and is very similar to the postfusion F protein from the related human respiratory syncytial virus (hRSV). In contrast, significant differences were apparent with the postfusion F proteins from other paramyxoviruses, such as human parainfluenza type 3 (hPIV3) and Newcastle disease virus (NDV). The high similarity of hMPV and hRSV postfusion F in two antigenic sites targeted by neutralizing antibodies prompted us to test for antibody cross-reactivity. The widely used monoclonal antibody 101F, which binds to antigenic site IV of hRSV F, was found to cross-react with hMPV postfusion F and neutralize both hRSV and hMPV. Despite the cross-reactivity of 101F and the reported cross-reactivity of two other antibodies, 54G10 and MPE8, we found no detectable cross-reactivity in the polyclonal antibody responses raised in mice against the postfusion forms of either hMPV or hRSV F. The postfusion-stabilized hMPV F protein did, however, elicit high titers of hMPV-neutralizing activity, suggesting that it could serve as an effective subunit vaccine. Structural insights from these studies should be useful for designing novel immunogens able to induce wider cross-reactive antibody responses. Human metapneumovirus (hMPV) was first isolated in 2001 from respiratory specimens collected from children with respiratory tract infections [1]. Sequence analysis was used to classify hMPV in the Metapneumovirus genus of the Pneumovirinae subfamily of paramyxoviruses. This subfamily also includes the Pneumovirus genus in which human respiratory syncytial virus (hRSV) is the best known prototype. Like all members of the Paramyxovirus family, hMPV and hRSV are enveloped, single-stranded, negative-sense RNA viruses that share many characteristics of their respective life cycles with other paramyxoviruses [2]. Sequence analysis of hMPV samples indicate that there are two main genetic lineages (A and B), each divided into at least two sub-lineages (A1, A2, B1 and B2) [3]. Clinical manifestations of hMPV infections are similar to those of hRSV, ranging from mild respiratory illness to bronchiolitis and pneumonia in children less than five years of age [2]. Although the frequency of severe lower respiratory tract infections is highest for hRSV, hMPV contributes to a significant fraction of the worldwide burden of bronchiolitis and pneumonia in young children [4]. As for hRSV, hMPV infections are also a frequent cause of morbidity and mortality in the elderly [5,6] and immunocompromised adults [7,8]. Despite their clinical significance, vaccines are not yet available for hMPV and hRSV. hMPV encodes three glycoproteins (SH, G and F) that are inserted into the viral membrane. The SH protein is a small hydrophobic protein whose function is unknown, although it has been claimed to inhibit NF-kappaB transcriptional activity [9]. The G glycoprotein is heavily glycosylated with multiple O- and N-linked sugar chains, resembling mucins [10], and serves as the putative viral attachment protein via interactions with cell-surface factors such as proteoglycans [11]. Finally, the fusion (F) glycoprotein mediates fusion of the viral and cellular membranes to allow entry of the viral ribonucleoprotein into the cell cytoplasm and thus initiate a new infectious cycle [12,13]. Recombinant hMPV with deletion of the G gene, the SH gene or both, retains the ability to replicate in epithelial cell lines, although these viruses have an attenuated phenotype in vivo [14]. Hence, at least in the G-deletion mutants, the F glycoprotein has to perform both the attachment and fusion steps. Indeed, it has been shown that F can bind to cell-surface molecules, such as proteoglycans [15] and certain integrins [16]. The interaction of the F glycoprotein with integrins requires a RGD motif conserved in all hMPV strains [12,17], and the interaction likely occurs after the initial binding of hMPV F to proteoglycans [15]. Paramyxovirus membrane fusion is thought to occur at the plasma membrane in a pH-independent manner. However, it was recently shown that hMPV particles are internalized via clathrin-mediated endocytosis in a dynamin-dependent manner [18] before pH-independent fusion of the viral and endosomal membranes takes place, except for a minority of strains that require acidic pH for efficient membrane fusion [19,20]. Even in certain cells, such as monocyte-derived dendritic cells, hMPV uptake occurs preferentially by macropinocytosis, a process that is partially inhibited by SH and G glycoproteins [21]. In all cases, however, hMPV F is the main player in the membrane fusion process. hMPV F is a class I fusion glycoprotein, synthesized as an inactive precursor (F0) that needs to be cleaved to become fusion competent. Proteolytic cleavage generates two disulfide-linked subunits (F2 N-terminal to F1) that assemble into a homotrimer. Cleavage occurs at a monobasic cleavage site immediately upstream of the hydrophobic fusion peptide. Cleavage can be achieved in tissue culture by addition of trypsin to the medium [19,22] but in vivo other serine proteases, such as TMPRSS2, are likely to be more relevant for cleavage [23]. The F trimer is incorporated into the virus particle in a metastable, “prefusion” conformation. To initiate membrane fusion, hMPV F is activated by still ill-defined mechanisms leading to a series of stepwise conformational changes in the F protein that drive membrane fusion and result in hMPV F adopting a highly stable “postfusion” conformation. Much of our current knowledge about the F protein conformational transition comes from the atomic structures of related paramyxovirus F proteins in either the prefusion [24–27] or postfusion conformation [28–31]. Among other changes, the prefusion-to-postfusion transition includes formation of a pre-hairpin intermediate in which heptad repeat A (HRA) sequences of the F1 subunit refold into a long continuous α-helix. This allows insertion of the hydrophobic fusion peptide, located at the N-terminus of HRA, into the target membrane. Refolding of this intermediate leads to merging of the viral and target membranes concurrently with the assembly of HRA and HRB sequences of the F1 subunit into a highly stable six-helix bundle (6HB), which is characteristic of the postfusion conformation [32]. Protection against hMPV infection is mediated mainly by neutralizing antibodies that presumably block refolding of the F glycoprotein and hence membrane fusion [33]. In contrast to other paramyxoviruses, the F glycoprotein is the only viral antigen of hMPV capable of inducing neutralizing and protective antibodies [34]. In addition, antibodies to the G glycoprotein are not protective [35]. Escape mutants selected with hMPV F-specific monoclonal antibodies (mAbs) have identified residues located in the hMPV F protein primary structure at sites equivalent to those of the antigenic sites identified in hRSV F [36]. Recently, mAb 54G10, isolated against hMPV F, was shown to cross-neutralize hRSV in vitro and to protect BALB/c mice against hRSV infection [37]. Another mAb, MPE8, was also recently described that cross-neutralized not only hMPV and hRSV but additionally two other viruses of the Pneumovirinae subfamily: bovine RSV and pneumonia virus of mice [38]. However, a global picture of the cross-reactivity potential of hMPV and hRSV F-specific antibodies is still missing. In order to advance our understanding of hMPV F structure and antigenicity (which lags behind other paramyxoviruses), we engineered a homogeneous preparation of soluble hMPV F folded in its postfusion conformation. This process required genetic manipulations that were not necessary for other paramyxovirus F proteins. We also crystallized the stabilized hMPV F postfusion trimer and determined its structure by X-ray diffraction analysis. Comparison with the postfusion hRSV F structure revealed a high degree of similarity between the two proteins in multiple regions, including a previously characterized antigenic site of hRSV recognized by mAb 101F [39,40]. Indeed, we found that 101F binds to hMPV F and cross-neutralizes hMPV. Although immunization of BALB/c mice with purified postfusion hMPV F induced very limited cross-binding and cross-neutralization with hRSV, the elicited sera had robust neutralizing activity against hMPV, indicating that the hMPV postfusion F protein may be an effective subunit vaccine antigen. Animal studies were performed under the regulations of the Spanish and European legislation concerning vivisection and the use of genetically modified organisms. Protocols were approved by the “Comité de Ética de la Investigación y del Bienestar Animal” of “Instituto de Salud Carlos III” (CBA PA 19_2012). The F protein ectodomain (amino acids 1–489, see Fig 1) was amplified from a pCAGG plasmid carrying the F gene from either the NL/1/00 strain (A1 sublineage) or the NL/1/99 strain (B1 sublineage) of hMPV [41]. Amplification was carried out with forward and reverse primers containing sequences from the beginning and the end of the F ectodomain and a C-terminal 6xHis-tag, incorporating EcoRI and NcoI sites. After digestion with these enzymes, the amplified DNA was inserted in the pRB21 plasmid digested with the same enzymes (Fig 1A, F Mon, protein 1). Subsequently: i) the amino acid change G294E was introduced in the F protein of the A1 sublineage since B1 sublineage already has Glu at position 294 and ii) the foldon trimerization domain [42] was added at the C-terminus of the F protein ectodomain, flanked upstream by a TEV protease site and downstream by a Xa protease site, followed by the 6xHis-tag (Fig 1A, F Foldon, protein 2). Hence, the complete amino acid sequence of the C-terminal appendage was: SGRENLYFQGGGGGSGYIPEAPRDQAYVRKDGEWVLLSTFLGGTEGRHHHHHH. TEV and Xa sequences are in italics and underlined and foldon sequences are in boldface. Other sequences correspond to linkers and histidines. pRB21 plasmid encoding the F protein of Fig 1A was further mutagenized by PCR, using the Phusion Site-Directed Mutagenesis kit (ThermoFisher Scientific) to: i) insert sequences of cleavage site II from hRSV F into the hMPV F cleavage site and ii) delete the first nine residues of the fusion peptide (amino acids 103–111, F Furin ΔFP Foldon, protein 3). The different recombinant plasmids were used to generate the matching recombinant vaccinia viruses by the method of Blasco and Moss [43]. CV-1 cells were infected with vaccinia viruses expressing the hMPV F protein ectodomains (moi 0. 05 pfu/cell) of the previous paragraphs (Fig 1A). In the case of constructs 3 and 4 shown in Fig 1A, cells were additionally co-infected with a vaccinia virus recombinant that expresses furin (moi 0. 03 pfu/cell) (a kind gift of Manuel Ramos, Centro Nacional de Microbiología, Madrid) [44]. In all cases, culture supernatants were collected 48 hours after infection, and these were concentrated and buffer-exchanged using Vivaflow membranes (Sartorius). Then, they were loaded onto Ni2+ columns in 50 mM Na2HPO4 pH 8. 0,300 mM NaCl, 10 mM imidazole buffer and, after washing, proteins were eluted with the same buffer containing 250 mM imidazole. Finally, proteins were concentrated with Amicon (Millipore) and exchanged to buffer without imidazole before being loaded onto a HiLoad 16/600 Superdex 200 pg gel filtration column (GE Healthcare) equilibrated and eluted with the same buffer (Fig 1B). Protein purity and integrity were checked by SDS-PAGE and Coomassie-blue staining under reducing conditions. Thermostability of purified proteins was assessed by heating samples in different tubes in a thermoblock. Starting at 30°C, samples were incubated for 10 minutes at this temperature before raising it to the next step and incubation continued for another 10 minutes. This stepwise increase in temperature and incubation was repeated until reaching 100°C. After incubation at each temperature, one of the sample tubes of each protein was withdrawn and kept at 4°C until the end of the incubation period. Then proteins were tested for antibody binding in ELISA as indicated below. Purified proteins were applied to glow-discharged carbon-coated grids and negatively stained with 1% aqueous uranyl formate. Images were recorded with a Gatan ERLANGHEN 1000 W CCD camera in a JEOL JEM-1011 electron microscope operated at 100 kV at a detector magnification of 20,400X or a FEI Eagle CCD camera in a Tecnai G2 electron microscope operated at 200 kV at a detector magnification of 69,444X. Xmipp software package [45] was used to select 112 and 100 images of 101F Fab in complex with hRSV F and hMPV F, respectively, and to obtain 2D averages with the CL2D routine. Crystallization conditions for Foldon-removed hMPV A1 postfusion F protein at 5 mg/ml in 2 mM Tris-HCl pH 8. 0,200 mM NaCl were screened by the sitting-drop vapor-diffusion method with an NT8 nanoliter-volume liquid handler (Formulatrix). Initial crystallization hits were obtained in 16. 8% PEG 3,350,10% 2-methyl-2,4-pentanediol (MPD), 0. 2 M lithium sulfate and 0. 1 M imidazole pH 6. 5 [46]. The crystal used for structure determination was grown in 18. 5% PEG 3,350,11% MPD, 0. 2 M lithium sulfate, 0. 01 M nickel chloride and 0. 1 M imidazole pH 7. 0 at a 2: 1 protein: reservoir ratio. Crystals were directly frozen in liquid nitrogen and data were collected to 3. 3 Å resolution at the Structural Biology Center beamline 19-ID (Advanced Photon Source, Argonne National Laboratory). Diffraction data were processed using the CCP4 program suite; data were indexed and integrated in iMOSFLM [47] and scaled and merged with AIMLESS [48]. A molecular replacement solution was found by PHASER [49] using a search model generated by replacing domains in the hRSV F postfusion structure (PDB ID: 3RRR) [29] with those from the antibody-bound monomeric fragment of hMPV F (PDB ID: 4DAG) [50]. The structure was built manually in COOT [51] and refined with PHENIX [52] using non-crystallographic symmetry restraints. PHENIX-generated feature-enhanced maps were particularly helpful during this process [53]. Data collection and refinement statistics are presented in Table 1. All experiments were carried out on a Biacore X100 using single-cycle format. Anti-mouse IgG (GE Healthcare) was covalently coupled to both the sample and reference cells of a CM5 chip at 10,000 response units (RU). Approximately 200 RU of mAb 101F were bound to the anti-mouse IgG. Then, purified proteins were injected at five different concentrations, as noted in the figure legends, at a flow rate of 50 μl/min. Association and dissociation phases were 108 seconds and 300 seconds, respectively. The chip was regenerated using 30 mM HCl, and the binding data were fit to a 1: 1 Langmuir binding model for the calculation of the kinetic parameters kon and koff. The KD was then calculated as the ratio of these two rate constants (koff/kon). Four hundred nanograms of the mAbs indicated in the figure legends was used to coat each well of 96-well microtiter plates for 16 hours at 4°C. Non-specific binding was blocked with 0. 5% bovine serum albumin (BSA) in PBS. Then, serial dilutions of soluble proteins were added and incubated for 1 hour at room temperature, followed by an excess of a biotinylated anti-His mAb, streptavidin-peroxidase and OPD (Sigma) substrate. Extensive washing with water was done after each step. Optical density was measured at 490 nm. In the case of mouse sera, the 96-well plates were coated with 40 ng/well of purified protein in PBS. Non-specific binding was blocked as described above. Then, serial dilutions of sera in blocking solution were added, and bound antibodies were detected with peroxidase-labelled goat anti-mouse Igs and OPD as substrate (GE Healthcare). Predetermined amounts of GFP-expressing hMPV recombinant viruses (NL/1/00 A1 sublineage or NL/1/99 B1 sublineage, a kind gift of Bernadette van den Hoogen and Ron Fouchier, Rotterdam, the Netherlands) or GFP-hRSV (A2 strain, a kind gift of Mark Peeples, Columbus, Ohio, USA) were mixed with serial dilutions of mouse serum before being added to cultures of Vero-118 cells. Twenty-four to forty-eight hours later, the medium was removed, PBS was added and the amount of GFP per well was measured with a Tecan microplate reader M200. Fluorescence values were represented as percent of a virus control without antibody. Vero-118 cells growing in either 8-well chamber slides or 24-well plates were infected with hMPVA1-GFP virus for 36 hours. At this time, medium was removed and after washing with PBS, cells were incubated sequentially with the mAbs of interest (30 ng/μl) followed by biotinylated sheep anti-mouse or anti-human IgG (GE Healthcare) and then streptavidin-RPE (red phycoerythrin, Southern Biotech). After washing with PBS, cells in the chamber slides were fixed with 2% paraformaldehyde in PBS and observed with a Leica TCS SP5 AOBS confocal microscope whereas cells in the 24-well plates were detached with 5 mM EDTA in PBS, fixed with 1% paraformaldehyde and analyzed with a Becton Dickinson FACSCanto analyzer. Groups of five BALB/c female mice (8 weeks old, from Envigo Rms Spain) were inoculated once intramuscularly in the hind legs with 10 μg of the soluble postfusion hMPV F A1 strain that was used for crystallization or the equivalent protein of the B1 strain. In addition, mice were inoculated with soluble postfusion hRSV F expressed similarly to the hMPV proteins [54]. Samples in 50 μl of PBS were mixed with an equal volume of CpG. Four weeks later, mice were euthanized, blood was collected and sera were obtained after coagulation. Following strategies previously used for the expression of postfusion F protein ectodomains from hRSV [29,54,55], parainfluenza virus type 3 (hPIV3) [31] and Newcastle disease virus (NDV) [30], sequences encoding amino acids 1–489 of the hMPV F protein ectodomain with a C-terminal 6xHis-tag were inserted into a vaccinia virus recombinant by the method of Blasco and Moss [43] (Fig 1A, construct 1). CV-1 cells infected with this vaccinia virus produced a soluble F protein that was secreted into the culture medium since it lacked the transmembrane (TM) region and the cytoplasmic tail. Whereas the soluble ectodomains of hRSV, hPIV3 and NDV F proteins could be purified as postfusion trimers, hMPV FTM- eluted from the gel-filtration column at a retention volume corresponding to a monomer (Fig 1B, black line) and migrated as an uncleaved band of the expected size in SDS-PAGE (inset). Negatively stained samples of this hMPV FTM- protein did not show discernable macromolecular assemblies when observed by electron microscopy (EM) (Fig 1C, panel 1). Addition of the fibritin trimerization domain (Foldon) from T4 bacteriophage [42] to the C-terminus of the hMPV F ectodomain (Fig 1A, construct 2) shifted elution of the new FTM- protein towards the size of a trimer (Fig 1B, green line). The amino acid change G294E, which is found in other hMPV strains [41] was introduced in this construct after observing that this change increased protein expression. Although the protein was heterogeneous when observed by EM, some cone-shaped molecules that resembled the previously described postfusion hRSV F trimer became visible [55] (Fig 1C, panel 2). The source of this heterogeneity has not been investigated further. hMPV growth in cell culture requires addition of trypsin to the culture medium to cleave the F protein at the monobasic site preceding the fusion peptide. Since no trypsin was present during production and purification of the hMPV FTM- monomer or trimer, these proteins remained uncleaved after purification, as seen by SDS-PAGE (Fig 1B, inset). Because circumstantial evidence suggests that cleavage enhances stability of paramyxovirus postfusion F proteins, the uncleaved hMPV F postfusion trimer of Fig 1A (construct 2) was treated with limited amounts of trypsin. Essentially all trypsin-treated molecules were seen by EM as cone-shaped molecules aggregated in rosettes (S1 Fig), presumably by intermolecular interactions of their respective fusion peptides, as previously reported for hRSV F [56]. To promote cleavage of the soluble hMPV F trimer without added trypsin, the hMPV F cleavage site was replaced with the second furin-cleavage site of hRSV F. In addition, to prevent protein aggregation, the first nine amino acids of the fusion peptide (residues 103–111) were deleted. The modified hMPV F protein eluted from the gel-filtration column as a partially cleaved trimer. To increase cleavage efficiency, cells were co-infected with a vaccinia virus expressing recombinant furin [44]. The soluble hMPV F protein now eluted as a trimer of fully cleaved protomers, as seen by SDS-PAGE (Fig 1B, inset), and existed as a homogeneous population of cone-shaped molecules, as observed by EM (Fig 1C, panel 3). Thermostability of the uncleaved and cleaved F molecules (Fig 1A, constructs 2 and 3) was assessed by testing their reactivity with four different mAbs (MF1, MF14, MF17 and 101F) after heating stepwise up to 100°C (Fig 1D). mAb MF1 recognizes the 6HB domain of postfusion F [54], whereas mAbs MF14 and MF17 recognize neutralizing, non-overlapping epitopes that have not yet been mapped on hMPV F. Lastly, 101F is a site IV-specific mAb raised against hRSV F which cross-reacts with hMPV F (described in more detail below). Binding of mAbs MF1 and MF14 was essentially unchanged after heating the cleaved hMPV F protein up to 100°C, whereas binding was lost to great extent with uncleaved F at that temperature. Although reactivity of mAbs MF17 and 101F with the two proteins was lost after heating at 100°C, that loss occurred at temperatures 5–10°C lower with the uncleaved protein. Therefore, these results lend support to the idea that cleavage increases the stability of postfusion hMPV F and differ from those reported with hRSV F, in which changes in reactivity with three different mAbs after heating were essentially the same for uncleaved and cleaved postfusion F [57]. Finally, the F protein described in the previous paragraph was cleaved with TEV protease to release the Foldon domain and affinity tag (Fig 1A, construct 4), which were separated from the authentic hMPV F ectodomain by gel filtration (Fig 1B, orange line). Removal of these residues made the F1 subunit migrate faster in SDS-PAGE (Fig 1B, inset) and the cone-shaped molecules looked slightly shorter by EM (Fig 1C, panel 4), demonstrating that the proteolysis was complete. Crystals of the hMPV F A1 subtype were obtained in space group P41212 and after optimization diffracted X-rays to 3. 3 Å resolution. A molecular replacement solution was obtained using a composite search model containing regions from the postfusion hRSV F trimer structure [29] and the antibody-bound hMPV F monomer structure [50]. The asymmetric unit contained two postfusion hMPV F trimers, which allowed non-crystallographic symmetry restraints to be used during refinement. After manual building, the structure was refined to an Rwork and Rfree of 22. 1% and 27. 0%, respectively, with no Ramachandran outliers as determined by MolProbity [58,59]. Data collection and refinement statistics are presented in Table 1. The structures of the two trimers in the asymmetric unit are very similar, with a root-mean-square deviation (rmsd) of 0. 26 Å for 1,325 Cα atoms. The structures are nearly complete, with no missing loops, and only a few disordered residues at the C-terminus of the F2 subunit and at the N- and C-termini of the F1 subunit. In addition, electron density for one or more core glycans is visible at some of the three N-linked glycosylation sites (N57, N172 and N353) on the different polypeptide chains in the asymmetric unit. Since the protein used for crystallization and structure determination was expressed from CV-1 cells without the addition of any glycosylation inhibitors or treatment of the protein with endoglycosidases, complex-type glycans are likely present at each N-linked site, with the electron density for most of the glycans being disordered due to heterogeneity and flexibility. The overall shape of the trimeric postfusion hMPV F protein resembled that of an elongated cone (Fig 2), consistent with the images observed by negative-stain EM (Fig 1C) and with previously determined structures of F proteins in the postfusion conformation from PIV3 [31], NDV [60] and hRSV [28,29]. The mature ectodomains of hMPV F and hRSV F have approximately 38% sequence identity, and overall their postfusion structures are similar, with an rmsd of 1. 48 Å for 419 Cα atoms in an F2–F1 protomer (Fig 3A). Secondary structures are well conserved, as are the conformations of the two major neutralizing epitopes retained on the postfusion conformation (antigenic sites II and IV of hRSV F). In contrast, the hMPV postfusion F structure has a much greater divergence from the hPIV3 (Fig 3B) and NDV (Fig 3C) postfusion F structures. Although the overall folds are similar, the secondary structures do not align well, consistent with the lower sequence conservation between hMPV and these two Paramyxovirinae subfamily members. The poor conservation of antigenic sites II and IV suggest that it is unlikely that a single antibody against either of these sites could neutralize viruses in both paramyxovirus subfamilies. 101F is a widely used mAb, originally produced by a hybridoma obtained from mice inoculated with a recombinant vaccinia virus expressing the hRSV F glycoprotein from the Long strain. The 101F epitope has been mapped to antigenic site IV of hRSV F [39,61], and a crystal structure of 101F Fab bound to a peptide corresponding to a linear portion of the epitope has been determined [40]. This structure was used in modeling studies to identify additional contact residues in prefusion [25] and postfusion [29] hRSV F. Given the structural similarities of soluble postfusion trimers of hMPV and hRSV F, particularly in antigenic site IV (Fig 3A), binding of mAb 101F to both proteins was tested by surface plasmon resonance (SPR) (Fig 4). Two preparations of fully cleaved postfusion hMPV F were included in the assay: one derived from A1 sublineage (strain NL/1/00), whose structure is shown in Fig 2, and the other derived from B1 sublineage (strain NL/1/99), which was also observed as a population of homogeneous cones by EM (S2 Fig). In parallel, a previously described soluble form of postfusion hRSV F was also tested for binding to mAb 101F [55,62]. Association rate constants (kon) for the binding of 101F to both hMPV F proteins were about 2–5 times slower than that to hRSV F, and dissociation rate constants (koff) for 101F binding to the hMPV F proteins were almost three times faster than that to hRSV F. Consequently, the affinities (KD) of 101F for the soluble hMPV F proteins were 5–10 times weaker than for hRSV F (Fig 4A), but remained in the 15–40 nM range. Binding of mAb 101F to soluble hMPV and hRSV postfusion F proteins was also tested by ELISA (Fig 5). The binding curves of 101F to hMPV F proteins from A1 or B1 lineages (Fig 5A) were similar to the binding curves of mAb MF14 (Fig 5B), which is a murine mAb specific for hMPV F (Fig 5B). In contrast, mAb 47F, specific for hRSV F [63], bound to the postfusion hRSV F protein as efficiently as 101F, but failed to bind significantly to the hMPV F proteins (Fig 5C). The antibodies used in the ELISA tests were also used in microneutralization assays with the same viruses from which the F proteins originated (Fig 5). mAb MF14 neutralized the two hMPV strains with similar efficiency, and mAb 101F also neutralized those two viruses, although it was slightly less efficient than MF14 in neutralization of the NL/1/99 strain (B1 sublineage, Fig 5D and 5E). In agreement with the ELISA results, 47F failed to neutralize hMPV, but neutralized hRSV almost as efficiently as 101F (Fig 5F). MF14 did not neutralize hRSV infectivity, consistent with its lack of binding as determined by ELISA. Since neutralization of hRSV by 101F is likely due to its binding to prefusion F [64], the binding of 101F to prefusion hMPV F was tested by an indirect method, since a prefusion-stabilized form of hMPV F is still unavailable. Thus, Vero-118 cells were infected with hMPVA1-GFP for 36 hours and then stained with 101F and control mAbs (S3 Fig). Antibody binding was revealed by confocal microscopy (left panels) or flow cytometry (right panels). The staining intensities of mAbs MF14 (a neutralizing antibody specific for hMPV F, Fig 5) and 101F were equivalent to that of mAb MPE8, which preferentially binds to the prefusion conformation [38]. Staining with mAb MF1, specific for the 6HB motif of postfusion F, was 5–10 times lower (see mean fluorescence intensity values in the flow cytometry panels). Although these data suggest that 101F is able to bind prefusion hMPV F and that this is the predominant conformation at the time point of S3 Fig, the relative affinities of the antibodies for each conformation are unknown. Therefore, a similar flow cytometry experiment was performed with or without heating of the cells at 50°C for 10 minutes, which should be sufficient to convert the majority of hMPV F from the pre- to postfusion conformation. As expected, the binding of MF1 was enhanced with the heat treatment (S4 Fig), whereas the binding of antibodies MF14,101F and MPE8 were largely unchanged or modestly decreased. Collectively, these data, along with the SPR (Fig 4) and neutralization (Fig 5) results, indicate that 101F binds to both the pre- and postfusion forms of hMPV F. To provide a structural basis for 101F cross-reactivity, we compared its epitope in both the hRSV and hMPV postfusion F structures (Fig 6). Based on our previous crystal structures of 101F in complex with hRSV F-derived peptides, we defined the minimal epitope as hRSV F residues 427–437, which corresponds to hMPV F residues 395–405 [40]. In the postfusion F structures, these 11 residues are in a similar conformation, with an rmsd of 2. 38 Å for the Cα atoms (Fig 6A). Importantly, all four residues in the center of the epitope are the same in both viruses, including the critical Lys433 residue (hRSV numbering; 401 in hMPV) that is altered in a 101F-escape variant (K433T) [39]. Substitutions at other residues within the epitope likely account for the decreased affinity of 101F for hMPV F. As expected based on the structural similarity, visualization of the complex by negative-stain EM revealed a binding mode essentially identical to that observed for the complex of 101F bound to postfusion hRSV F (Fig 6B). The images agreed well with models of 101F bound to hRSV and hMPV postfusion F proteins, which were based on superpositions of the peptide-bound 101F complex and the postfusion F proteins (Fig 6C). Postfusion forms of hRSV F, either as soluble ectodomain [28] or full-length protein aggregated in rosettes [65], have shown their potential to induce neutralizing antibodies in cotton rats and to protect them against a virus challenge. Similarly, soluble forms of hMPV F have been used to immunize cotton rats [66], Syrian golden hamsters [67], BALB/c mice [68] and macaques [69], demonstrating the capacity of hMPV F to induce neutralizing antibodies and protection. However, the soluble forms of hMPV F used in previous immunizations were not stabilized and purified as described above and probably represented a heterogeneous mixture of different conformers. Hence, it was pertinent to test the immunogenic potential of the well-characterized and crystallized hMPV postfusion F protein from the NL/1/00 strain (A1 sublineage) in BALB/c mice. Mice were also immunized with the equivalent F protein from the NL/1/99 strain (B1 sublineage) as well as an equivalent soluble form of postfusion hRSV F. Mice were inoculated after mixing the proteins with CpG as adjuvant. A single dose rather than multiple doses was used to discriminate better the specificity of the antibodies induced by each protein. The sera of mice inoculated with the hMPV F proteins showed high levels of antibodies binding to homologous and heterologous hMPV postfusion F proteins, but had no binding activity to the soluble postfusion hRSV F (Fig 7A). In addition, the binding titers were significantly higher for the homologous hMPV F protein versus the heterologous hMPV F protein, particularly in sera of mice inoculated with NL/1/99 (B1 sublineage). Mouse sera were also tested in a microneutralization assay (Fig 7B). As with ELISA, the sera of mice inoculated with the hMPV F protein from the NL/1/00 strain (A1 sublineage) neutralized both the homologous and the heterologous hMPV strains with similar efficiencies. In contrast, the sera of mice inoculated with the F protein from the NL/1/99 strain (B1 sublineage) had significantly higher neutralizing titers against the homologous strain than against the heterologous strain. Of note, the sera of mice inoculated with either of the two hMPV F proteins failed to neutralize hRSV (A2 strain). These data indicate that the postfusion conformation of the hMPV F protein is highly immunogenic, but fails to elicit hRSV-neutralizing activity despite a high degree of structural similarity with the postfusion hRSV F protein (Fig 3) and cross-reactivity of antibodies 101F (this article), 54G10 [37] and MPE8 [38]. Reciprocally, the sera of mice inoculated with a soluble form of postfusion hRSV F neutralized hRSV but failed to neutralize either of the two hMPV strains (Fig 7B). Data accumulated during the last 10–15 years have demonstrated extensive similarities between the clinical manifestations and epidemiology of hMPV and hRSV [2], and the two viruses share many steps of their respective life cycles. However, there are also important differences in gene order and number of genes encoded in their genomes [10] as well as in individual gene products. One of these differences resides in the protease maturation of their respective F glycoproteins. Whereas the hMPV F precursor is cleaved only once by trypsin-like proteases, the hRSV F precursor is cleaved at two distinct furin sites separated by 27 amino acids [70], which is unique among paramyxoviruses. It is still unresolved whether or not this difference has an impact on structural and functional properties of their respective F glycoproteins, but it has been shown that insertion of the two cleavage sites of hRSV F into Sendai virus F protein changed the hemagglutinin-neuraminidase (HN) dependence of Sendai F for activation [71] and altered Sendai virus thermostability [72]. It is conceivable therefore that hMPV F may be less stable than other fusion proteins and hence requires the genetic manipulations shown in Fig 1A to fold into a stable postfusion trimer. Indeed, whereas the ectodomain of hMPV F without an additional trimerization domain (Fig 1A, construct 1) is found mainly as a monomer in the supernatant of cultured cells, the ectodomains of uncleaved hPIV3 F [31], NDV F [30] and hRSV F [56] or cleaved hRSV F [28,29] fold spontaneously as stable trimers. Stabilization of the hMPV F postfusion trimer was dependent on cleavage of the protein precursor. As seen in Fig 1D, cleavage increased thermostability of several epitopes in postfusion hMPV F, an effect that had not been previously observed for the homologous hRSV F [57]. Given the high similarity between postfusion hMPV and hRSV F structures (Fig 3A), the source of their apparent differences in stability remains to be determined. As mentioned in the Introduction, cross-neutralization of hMPV and hRSV has been observed with certain anti-F mAbs. Particularly, 54G10, a human-derived mAb selected for binding to hMPV F, was shown to neutralize hRSV in vitro and to confer passive protection of mice against a hRSV challenge [37]. 54G10 selected hMPV escape mutants with an alteration (V397G) at a residue located in the corresponding region of the hRSV F antigenic site IV (Fig 6A). In a reciprocal manner, we have shown here that 101F, a murine mAb selected for binding to hRSV F, is capable of cross-binding to hMPV F and cross-neutralizing hMPV infectivity. Fig 6 provides an explanation for 101F cross-reactivity with the two F proteins since its binding site is fairly conserved in both hMPV F and hRSV F structures. The small differences in the structure of site IV between the two proteins may account for the noted affinity differences (Fig 4), which nevertheless are not linearly translated into neutralization potency (Fig 5). It is worth mentioning that 101F also binds to hRSV F in the prefusion conformation [64], and this is likely responsible for its neutralizing activity. In a similar manner, the neutralization of hMPV by 101F probably requires binding to its prefusion F. This hypothesis is supported by the results shown in S3 and S4 Figs, which indicate binding of 101F to prefusion F expressed at the surface of hMPV-infected cells. Despite the noted cross-reactivity with certain mAbs, neither significant cross-reactivity nor cross-neutralization were evident with the polyclonal antibodies induced in mice after inoculation with purified postfusion hMPV F or hRSV F. It should be noted that the cross-reactivity of mAb 101F with hMPV and hRSV F was revealed only after testing more than twelve different mAbs raised against hRSV F, including seven that competed with 101F for antigen binding. Although the structures of postfusion hMPV F and hRSV F are similar (Fig 3A), there is limited overall sequence identity (33–35%) [10]. Since antigen binding is dictated mainly by interactions of specific amino acid side-chains in the antibody with those in the antigen, it is likely that sequence changes in F, not reflected in the overall structure, account for the lack of polyclonal serum cross-reactivity and the scarce cross-reactivity of mAbs between postfusion hMPV F and hRSV F. An interim conclusion might be that antibodies that cross-neutralize hMPV and hRSV probably represent a minority of the global repertoire of specificities present in a polyclonal response, which is not just a pool of mAbs. Nevertheless, the sporadic isolation of cross-neutralizing mAbs that inhibit hMPV and hRSV infectivity opens the possibility of designing modified forms of the Pneumovirinae F capable of inducing highly cross-reactive and cross-protective antibody responses, as recently shown for the influenza virus hemagglutinin [73].

Human metapneumovirus (hMPV) is a frequent cause of severe lower respiratory tract infections in very young children and a vaccine is not yet available. Protection against hMPV infection is afforded mainly by neutralizing antibodies directed against the fusion (F) glycoprotein. After iterative rounds of protein engineering, we generated a soluble form of the hMPV F protein in the postfusion conformation and determined its crystal structure. The structure is similar to that of the related human respiratory syncytial virus (hRSV) F glycoprotein, and two neutralizing epitopes are particularly well conserved, thus providing a structural basis for the cross-neutralizing activity of several monoclonal antibodies. Immunization of mice with the engineered hMPV F postfusion protein elicited high hMPV-neutralizing antibody titers, suggesting that this protein could be an attractive subunit vaccine antigen. These results also open the possibility of designing novel cross-protective immunogens.

Abstract Introduction Materials and Methods Results Discussion

flow cytometry medicine and health sciences immune physiology crystal structure enzyme-linked immunoassays immunology condensed matter physics membrane fusion protein structure crystallography antibodies immunologic techniques cross reactivity cellular structures and organelles research and analysis methods immune system proteins solid state physics protein structure determination proteins antigens immunoassays molecular biology cell membranes spectrophotometry physics biochemistry cytophotometry cell biology physiology biology and life sciences physical sciences spectrum analysis techniques macromolecular structure analysis

2016

Engineering, Structure and Immunogenicity of the Human Metapneumovirus F Protein in the Postfusion Conformation

Gene expression analysis has become a ubiquitous tool for studying a wide range of human diseases. In a typical analysis we compare distinct phenotypic groups and attempt to identify genes that are, on average, significantly different between them. Here we describe an innovative approach to the analysis of gene expression data, one that identifies differences in expression variance between groups as an informative metric of the group phenotype. We find that genes with different expression variance profiles are not randomly distributed across cell signaling networks. Genes with low-expression variance, or higher constraint, are significantly more connected to other network members and tend to function as core members of signal transduction pathways. Genes with higher expression variance have fewer network connections and also tend to sit on the periphery of the cell. Using neural stem cells derived from patients suffering from Schizophrenia (SZ), Parkinson' s disease (PD), and a healthy control group, we find marked differences in expression variance in cell signaling pathways that shed new light on potential mechanisms associated with these diverse neurological disorders. In particular, we find that expression variance of core networks in the SZ patient group was considerably constrained, while in contrast the PD patient group demonstrated much greater variance than expected. One hypothesis is that diminished variance in SZ patients corresponds to an increased degree of constraint in these pathways and a corresponding reduction in robustness of the stem cell networks. These results underscore the role that variation plays in biological systems and suggest that analysis of expression variance is far more important in disease than previously recognized. Furthermore, modeling patterns of variability in gene expression could fundamentally alter the way in which we think about how cellular networks are affected by disease processes. In studying biological systems, we tend to think of groups as being defined by specific, measurable parameters, and of the important differences between those groups as being defined by a significant average difference in those parameters. Much of the language we use in describing biological systems is based on this bias and we talk about genes being expressed in a tissue at a particular level, or about differences in gene expression between groups reflecting the mechanism driving their phenotypic differences. This view of biological systems has been extremely useful in that nearly all of our understanding of biological systems is based on interpreting average behavior. Variance in this context is only used as a measure of the significance of those mean differences (through the use of statistical measures such as a t-test or ANOVA). Rarely has the variability across a population been considered in the analysis of transcriptional differences between populations. Arguably, variance has been largely ignored because it has been considered solely in the context of experimental reproducibility, and therefore something that must be reduced. This was a reasonable bias in the early days of microarrays, but the robustness and reproducibility of the current generation of array platforms [1] allows us to look at additional drivers of variance in gene expression studies. Increasingly there is evidence that biological sources of variation may play an important role in determining cellular and organismal phenotypes [2]-[8], as well as in helping to explain a wide range of biological phenomena ranging from reduced penetrance [9], [10] to evolutionary fitness [11]. The direct link between genetics and reduced penetrance, expression variability and phenotype was elegantly demonstrated in C. elegans by Raj and colleagues, who showed that variation in the number of transcripts expressed in any individual cell, as well as the number of cells expressing the transcript of interest, directly influenced the development of the worm' s intestine [6]. Mutations in key developmental transcription factors affected not just the mean expression of target genes, but also the variance of their expression levels. Their model proposed a threshold effect of absolute gene expression on phenotype, where the availability of the transcripts dictated cell fate. Penetrance of the mutant trait was therefore determined by quantifiable variance in gene expression levels. Variation, in genetic and phenotypic terms, has long been considered an important component of population fitness and adaptability. Similarly, one way to interpret the association between expression variance and phenotype is to consider how this might play a role in determining phenotype. Consider a pathway that plays a role in a developmental process or in a cell' s response to a particular environmental stimulus. If the genes in that pathway have very low variance, a natural interpretation is that those genes are themselves highly constrained, and that the spectrum of potential responses from activation of that pathway is itself limited. This interpretation was explored by examining the difference in gene expression variance in neural stem cells and fibroblasts derived from patients suffering from Schizophrenia (SZ), patients suffering from Parkinson' s disease (PD), and healthy donors (controls). We demonstrate how gene expression variance can be used as a way to distinguish between phenotypic groups, the way in which constraint provides information about network topology, and to provide insight into the mechanisms associated with disease and normal states. Of course it should be noted that any analysis of variance must be carefully considered in light of the methods used to collect and analyze biological datasets. Variation in biological systems has long been considered “noise” to be minimized either through careful experimental design or through the use of data normalization methods designed to improve comparisons between individual samples. However variation includes both biological and experimental (or random) effects and it is the former, rather than the latter which is important in the current study. In our analysis, we considered highly-constrained and lowly-constrained genes, descriptors synonymous with low variance and high-variance states, respectively. These definitions, in part, helped to define our hypothesis: that the degree of variation in the expression of the genes associated with a particular cellular network is indicative of the plasticity [12] of that network. In this sense, high variance is associated with increased plasticity and low variance with diminished plasticity. Our approach provides important evidence supporting the hypothesis that variation is an essential feature of biological systems and one that influences disease phenotype. In particular we illustrated that patterns of variance in certain key pathways are not random, but provide a potential mechanistic understanding of the phenotypic differences that arise in the development and progression of particular neurological diseases. The olfactory neuroepithelium is a continually regenerating tissue, and stem cells isolated from biopsied material give rise to neurons and glia in culture [13], and in transplant models in the rat [14], [15]. Patient-derived human olfactory neurosphere-derived (hONS) cells have been shown to be an informative tissue-specific system for studying the etiology of human brain disorders like PD and SZ [16], [17]. In a previous study, gene expression data from these hONS stem cell lines were used to identify disease-specific cellular alterations by comparing absolute expression profiles of hONS between donor groups [16]. Here, we use this same data set to focus instead on patterns of variability as a means to assess how hONS deviate from the normal population in PD and SZ donors, and explore the implications that variability may have on these disease processes. Tissue biopsies from skin (fibroblasts) or the neuroepithelium of the nose were obtained from nine Schizophrenia (SZ) and eleven healthy control donors. Olfactory biopsies were taken from an additional thirteen donors with Parkinson' s disease (PD) [16]. Adult stem cell lines are grown from olfactory biopsies for several passages as primary cultures, then moved through a neurosphere process to enrich for neural stem cells [13]. hONS cells are monolayer cultures expanded from disaggregated neurospheres [13], [16]. Patient-specific lines were grown from primary olfactory mucosa biopsies (primary) and hONS from all donor groups, additionally skin fibroblast cell lines were grown from the SZ patients and control donors. Genome-wide transcriptional profiling was performed on individual donor lines with replication (see Material and Methods). Principal component analysis (Figs. 4–6 in Text S1) shows that the samples clustered by the disease status of the donors. All donors used in this analysis were male. Figure 1 presents an outline of our analysis pipeline. We first examined the genome-wide expression variance distributions between skin fibroblasts and hONS derived from the same donors in the control group (Figure 2). Of the 22,184 probes represented on the Illumina microarray; 14,986 probes were detected in at least one cell type. To minimize experimental effects in our analysis, great care was taken to standardize our laboratory protocols and the assays that were performed. To explore the potential contributions of experimental noise, we consider a number of normalization approaches for microarray data and show that these effects do not contribute to the differences that we observe. As a measure of variance, we used the coefficient of variation (CV) which is computed for each gene by dividing the standard deviation of its expression measures across a sample population by its average expression. We designate highly-constrained genes as those falling below the lower 25th percentile of the genome-wide CV distribution based on all donors and lowly-constrained genes as those above the upper 25th percentile; those genes in the range between the 25th and 75th percentile we refer to as the “Mid Variability” gene set. Basing our analysis on CV values protects against detecting patterns in variability influenced by trends in absolute expression alone (Text S1). Normalization procedures are assumed to reduce or stabilize variance. To assess the impact of normalization on expression variance, CV distributions were examined using five normalization regimes and summarized in Figure 2: i) log2 transformed; ii) log2 transformed and median normalised in the presence or; iii) absence of non-detected probes; iv) log2 transformed and quantile normalized; v) log2 transformed and robust spline normalization. The distribution of expression values was consistent across all of the normalization strategies, which most likely reflects the high level of reproducibility of the raw data. CV was robust to the normalization strategy used but was most impacted by background correction or detection thresholds when using median normalization: this is discussed further in the Text S1, however all subsequent analyses were run on log2 transformed, quantile normalized data. Background subtraction was not performed, but data was thresholded using the Illumina detection scores. While the absolute numbers of genes in the ‘low’, ‘medium’ and ‘high’ variance categories varied slightly between the normalization methods, concordance was high (Figure 2B >75% overlap low-variance genes; >85% overlap high-variance genes) and the patterns in the underlying data types were highly reproducible (Figs. 1–3 in Text S1). Using gene expression data from hONS and skin fibroblasts isolated from control donors, we next isolated the core attractor pathways whose differential expression distinguish a normal hONS stem cell expression phenotype, using the attract method that we recently developed [18]. Rather than testing individual genes, attract begins by using an ANOVA based method to test gene sets defined by KEGG pathways for their ability to distinguish between phenotypic states. Pathways ranked as significant are then each decomposed into “synexpression groups”—subsets with expression profiles that are both highly correlated and informative for distinguishing between phenotypes. These synexpression groups are then expanded to include genes with highly correlated profiles from within the original dataset, producing a collection of “core pathway modules. ” The top five attract modules were the MAPK signaling pathway, the focal adhesion pathway, the purine metabolism pathway, the neurotrophin signaling pathway and the cell cycle pathway (Table 1), each of which has previously been linked to important aspects of stem cell biology [19]–[21]. These modules were significantly different between the cell types tested at the 0. 05 level (after adjusting for multiple testing using the Benjamini-Hochberg method). They were also the most “representative pathways, ” in the sense that they contained the largest numbers of genes in the list of array probes significant between classes. Overall we found 21 significant pathways (adjusted P-values <0. 05). However, many of these pathways were overlapping in their gene content (see Table S1) and together represent three key common biological themes—immune response, growth factor signaling, and DNA replication—that are consistent with the phenotypic differences between the cell types. In overlaying variance onto these networks we noted that the numbers of high-variance or highly constrained genes did not follow the expected population patterns, with a trend towards more high-variance genes across most of the pathways (Figure 3; Fig. 7 in Text S1). The purine metabolism and cell cycle pathways in particular had fewer low-variance and more high-variance genes than expected (Chi-square goodness of fit P-value <0. 01 and P-value <1. 2×10−6 respectively), an observation consistent in the hONS and skin fibroblast datasets. In contrast, the neurotrophin signaling pathway contained far fewer high-variance genes than expected (Chi-square goodness of fit P-value <0. 01) suggesting that this network was under greater regulatory constraint. Any analysis of variance must consider the contribution of technical variation to the data. To address this issue we compared the magnitude of the intra-individual expression variance with the size of the inter-individual expression variance associated with each donor group. Four additional donors had each contributed two independently derived biopsies, resulting in four replicate samples per donor (see Material and Methods and Text S3). The intra-individual mean of the CV distribution was smaller than the inter-individual CV mean, confirming that the technical variation in this data set was less than the biological variation observed. We next examined whether this observation held up when controlling for differences in sample size, by constructing variance distributions based on the same number of samples (n = 4), and found this trend persisted. Amongst the four individuals, differences in the CV distributions, albeit slight were observed for the five core pathway modules. Collectively these results suggest that repeated sampling of the same individual is associated with less expression variance, pointing to a strong genetic component in biological variability. The distributions of expression variance for each of the patient groups highlighted an unexpected observation. Using a two sample t-test on log2 transformed CV values, we saw significant deviations from the control pattern in the SZ group (P-value <2. 2×10−16), and also the PD group (P-value <0. 001). This suggested that the SZ hONS lines demonstrated much less variation in their genome-wide expression patterns than was expected and in contrast, the PD hONS lines showed greater variation in their genome-wide expression patterns (Figure 4A; Text S2). Applying the same tests to identify disease-specific differences in genome-wide average expression showed no significant differences in either donor group. We then investigated whether the differences in expression variance observed at a whole genome level were also apparent in our five core stem cell networks (Table 1). For each of these, we found that the SZ and PD groups sit at opposite ends of the variability spectrum (Figure 4C). The SZ group had a marked reduction in variance signaling pathway, as evidenced by more highly-constrained genes whereas the PD group had greater variance than the control group. The deviation in frequency distributions between the disease group and the control group was statistically significant for both SZ and PD groups in hONS stem cells, as assessed by a Chi-squared test with two degrees of freedom. One might anticipate that increased expression variance, as seen in the PD group, is evidence of poor network integrity, a model that has been suggested as important in complex diseases. What was surprising, however, is the observation that the SZ group has significantly reduced expression variability in these core pathways, suggesting that both extremes of the variability spectrum may be implicated in disease processes. These disease-specific patterns were independent on the particular cut-off used to define the regions of high and low constraint in the expression variance distribution (Figure 4). When more stringent cut-offs were applied (e. g. 5% and 10%) the same general trend of reduced variance for the SZ group and increased variance for the PD group was still observed. Disease status was associated with different proportions of lowly-constrained and highly-constrained genes; this observation raised the possibility that these two classes may play distinct roles in maintaining or driving cellular phenotype. Using Gene Ontology terms (cellular component, molecular function and biological process categories) [22], we performed a representation analysis for each set of highly-constrained and lowly-constrained genes in the pathway modules for the three donor groups. We then mapped these genes to networks based on highly-annotated protein-protein interaction data and compared the patterns of transcriptional constraint between phenotypic states. We found that genes with high-variance/low constraints were both functionally and physically involved at the periphery of signal transduction pathways. In hONS cells, they functioned largely as cell surface receptors and tended to be localized in the membrane, transmembrane or extracellular matrix regions (Table S2). This might suggest that the hONS were heterogeneous for expression of growth factor receptors and as a population had dynamic interactions with the extracellular environment. In contrast, highly-constrained genes tend to function in signaling roles, such as protein kinases and phosphatases (see Text S4). This might imply that all of the cells in the hONS population were competent to transduce signals through the MAPK pathway, and were only restricted by the expression of receptors or the availability of ligands. The hONS stem cells derived from SZ donors demonstrated both loss of variability at the cell surface, and increased constraint in the intracellular signaling molecules. For MAPK signaling, we saw significant functional enrichment of signaling GO categories (Table S2) for low-variance genes in the SZ and control groups, while there was enrichment for high-variance genes in the PD group. This suggests that MAPK signaling is particularly important for distinguishing between these three groups. Just as cellular distribution of highly-constrained genes was not random, we observed a nonrandom pattern in the degree-distribution (connectivity) of genes based on their expression variance (Figure 5). The lowly-constrained genes had on average, a low degree whereas the highly-constrained genes were more highly connected to other genes in the network; this shift from a random distribution was marginally statistically significant for the control group (Chi-squared test; P-value = 0. 05958 for control group, P-value = 0. 1219 for SZ group, P-value = 0. 09595 for PD group, see Figs. 19 and 20 in Text S1), suggesting that not only are constraints imposed on genes linked to specific functional roles but they also have significantly distinct network topologies. Each disease group was associated with distinct deviations in degree distributions from those observed for the control group; the degree distributions between high and low-variance gene sets became more identical for the SZ group whereas in the PD group, the high and low-variance distributions appeared reversed from those observed in the control group. These observations suggest two different vehicles in which normal regulatory control through a signaling network may be disrupted or perturbed. In comparing our three patient groups, we observed a consistent correlation between the degree of a particular gene and its expression variance, suggesting that transcriptional variability is an inherent aspect of all cellular systems (Fig. 20 in Text S1). This is consistent with our intuition that genes that are highly connected, and therefore play a central role in signaling or other networks, must be more tightly regulated than those that play more peripheral roles, including as cell surface receptors and downstream effectors. It may be that the increased expressivity of those genes peripheral to the network provides a wider range of potential phenotypic response to external stimuli. In this way, disease processes that alter the variability in expression of particular genes may influence phenotype so that reduced expressivity limits the spectrum of response while increased expressivity tends toward loss of regulation over the cell' s end-stage response. In genetics, the concepts of expressivity and penetrance describe phenotype variability between individuals with shared genotypes and across populations, respectively. The implicit explanations for phenotypic variability are differences in genetic, epistatic, or epigenetic interactions. In genomic analysis, we often see this variability in terms of sequence or structural polymorphisms within the genome [23] and focus our efforts on understanding the link between genetic variation and phenotypic diversity. In this context, low genetic variation leads to poor evolutionary fitness, whereas the convergence of multiple variants on disease networks is increasingly thought to contribute to disease states [24], [25]. A key element in the central dogma of molecular biology is the role played by RNA as an intermediary between gene and protein (and ultimately with phenotype). In this light it is surprising that variability of expression levels has received so little attention. This may reflect the fact that in the analysis of gene expression data, variance is often associated with technical artifacts rather than being seen as an intrinsic property that reflects the normal range of phenotypic heterogeneity. If technical noise intrinsic to the platform was the major driver in variance, then one would expect these random effects to affect experimental samples equally. Indeed, as our goal was to do comparisons between phenotypic groups, we made every effort to minimize experimental noise, beginning with standardized protocols for sample collection and laboratory handling, quality control during RNA extraction, and labeling and hybridization together so as to avoid potential batch effects. Given that we see specific biases in the variance profiles that correlate in a meaningful way with phenotype argues that this is not the case. Our results suggest that this variability is, in fact, a much more important measure of phenotype than previously recognized and that changes in the spectrum of expressivity may be fundamentally linked to the development of distinct phenotypic states. The expressivity of individual genes such as the pluripotency factors Oct4 and Nanog has been highlighted by others as an important contributor to phenotypic robustness of a population of embryonic stem cells. Variance of gene expression in these cells is both predictable and essential – providing a dynamic range of pluripotency factors which is directly linked to the differentiation potential of individual cells within that population [26], [27]. If expression variance of one or two key regulators is an important modifier of phenotype, then it holds that this can be measured at a pathway level as well. Even across cells derived from different donors, the distribution of expression variance for a pathway is predictable, and deviations from this correlate with disease phenotypes. What are the likely sources of this ‘regulated noise’? The reduced variance observed between hONS lines derived from different biopsies from the same donor indicates a role for genetics. One possible scenario is that genetic polymorphism might impose a combinatorial impact on the expressivity of genes within a network, and the consequent alteration of dynamic range of that network outcome. Indeed, the degree of expression variance of a reporter construct was found to be a heritable trait in S. cerevisiae [28]. Epigenetic factors are also likely to play a key role in expressivity at individual loci, and others have shown stochastic, population-wide variance in epigenetic modification of key developmental loci. With increasing evidence that expression variance is an important phenotypic attribute of cell populations, it holds that variance profiles may also reflect abnormal genetic or epigenetic events contributing to disease phenotypes. In this context, it is perhaps not surprising that the most profound shifts in expression variance were found in hONS cells isolated from SZ donors. SZ is a life-long psychotic disorder, with age of onset in males in early adulthood and later in females. SZ is considered a disease of neurodevelopment, based on epidemiological, histological and genetic evidence [29]–[32]. It is clear that SZ is a complex genetic disease with a strong environmental component. There are several well replicated genome-wide studies that have implicated common polymorphisms in the etiology of the disease, although these account for a small component of the heritability and an emerging theory is that polygenic risks explain more of the genetic component of this disease. Reduced expression variance of hONS networks provides fresh clues into potential mechanisms underlying diseases like SZ. For example, if the patterns of gene expressivity found in the hONS reflect those found in the developing brain, then the orderly cascade of brain development may be altered. In addition, it is feasible that an overly constrained biological pathway would be less adept at buffering environmental stressors. Neuronal stem cells are the progenitors for neural cell states and the development of brain is rooted in the cell fate decisions that occur in these stem cells. Olfactory stem cells have been shown to serve as a good surrogate for neuronal stem cells and stem cell differentiation[14]. Our analysis of gene expression in olfactory stem cells cultured from patients with SZ, PD, and matched controls identified a number of key pathways that distinguished these three groups, including key signaling and developmental pathways. What was surprising about these pathways was that not only was there a notable difference in expression, but that expressivity, or variability in gene expression levels, was also significantly different between these different diseases. Further, when the expressivity was mapped to protein-protein interaction networks, there were distinctly different patterns of transcriptional constraint that depended on the connectivity of the proteins in the network. Further, these patterns depended not only on disease state, but also on the degree of connectivity within the corresponding protein-protein interaction pathway. In examining the network topology, we found that in the SZ patients, there was significantly greater regulatory control over genes at the highly-connected core of the corresponding pathways. For example, in the MAPK pathway, SZ patients exhibited far less expressivity in the kinases that represent the core of the pathway. On the other hand, the PD patients had significantly fewer constrained genes mapping to the same core pathway regions than did controls. The increased variance associated with some proteins such as receptors, may indicate that these proteins have fluctuating turnover rates in the cell populations, which in turn will influence the capacity of cells to interact with their environment. These data, together, paint a very interesting and compelling picture of the mechanism associated with disease. One of the defining characteristics of SZ is the disruption of normal cognitive processes and this is reflected in fewer neural synapses in parts of the brain. While it is no doubt a leap to interpret our results as having a direct link to thought processes in the disease, one could imagine that a high degree of constraint in key transcription factor networks that play a role in cellular differentiation and developmental processes. One could imagine that highly constrained transcriptional networks in neuronal stem cells could reduce the spectrum of cellular phenotypes that could be derived from those stem cells and reduce the plasticity resulting in the brain and altering its potential responses and thought patterns. It is worth noting that among the pathways that demonstrate highly-constrained gene expression in our SZ patients are those involved in signaling in cancer. One consequence of increased constraints, and reduced plasticity, is defects in the network are more likely to be disastrous. In this kind of model, defects in cancer-related networks would lead to loss of those cells, preventing the types of adaption observed in cancer development. Although there are conflicting reports regarding the risk of cancer in schizophrenics, the majority of reports suggest that patients with SZ are protected against cancer in general, and from lung and colorectal, despite increased smoking [33], [34] and drinking habits [35] in this population. Laboratory studies have also reported reduced tumor growth in animal models of schizophrenia [36]. Although speculative, the high degree of transcriptional constraint we found in the MAPK pathway in SZ patients may explain, in part, these observations. In contrast, disease states may arise where increased variance changes the predictability of network outcomes, resulting in dysregulation of the desired cell state. In PD, we observed an increase in the expression variance of core signaling pathways, which we predict will diminish the robustness of the network to external events. This may be an essential element that is shared amongst diseases of aging. Although SZ and PD represent very different conditions, our results suggest that changes in expressivity relative to the normal spectrum of variability, may play an important role in the development of disease phenotypes. One possible way to interpret this is in the context of models first proposed by Conrad Waddington [37] and later refined by Stuart Kauffman [38]. Waddington and Kauffman envisioned what we can interpret as a gene expression state-space landscape defined by possible gene expression states. In this landscape, stable cell states represent fixed points (Kauffman referred to these as “attractors”) connected by evolutionarily-defined “canals” representing the differentiation pathways connecting distinct cell phenotypes. In this model, the SZ patients would be characterized by more highly-constrained, “deeper, ” canals, limiting the potential end states that one might achieve during differentiation. PD patients, on the other hand, could be characterized by a flattening of the same canals (what has been referred to as “decanalization”), increasing dysregulation of the associated pathways and potentially allowing for a degradation of the well-defined cellular end states that might be available. While the potential importance of decanalization in disease has been discussed [25], we believe that our results are the first to suggest that over-canalization may be an equally important process in developing disease phenotypes. This work uses public expression datasets from patient-derived cells which were collected under the ethics approval of the Griffith University ethics committee. Illumina Human-Refseq8 v2 BeadChips (Illumina, Inc.) arrays were used to capture genome-wide gene expression profiles; the raw data was summarized using BeadStudio Version 3. 1. 7 (Illumina, Inc). Background correction and normalization methods were performed using the R/Bioconductor lumi package. All downstream analyses were performed using Quantile normalized data, without background correction, and only probes passing the Illumina detection threshold were included in variance analysis. A probes was considered to pass the Illumina detection score if it had a detection p-value ≤0. 01 in at least 75% of individuals in the same donor group, resulting in 14,986 probes. This expression data is available from ArrayExpress under the experiment accession number E-TABM-724. Illumina Human-Refseq8 v3 BeadChips (Illumina, Inc.) arrays were used to collect genome-wide gene expression profiles of the four replicate samples for each of the four donors. The donors were made up of two healthy controls and two PD patients, and each donor underwent two independently derived biopsies that were each replicated twice, giving rise to four samples per donor. The raw gene expression data set was summarized as above. The data set was filtered using the same detection filter, and the subset of 6,809 detected probes common to the v3 and v2 arrays was used for the comparison of intra-individual and inter-individual variance analysis. To gauge the effect of sample size on variance, the inter-individual variance was calculated by computing CV distributions based on a random subset of 4 individuals from each donor group, and comparing these to the intra-individual CV distributions that were calculated from the four samples for each individual. From the 100 random subsets generated, we observed a reduction in the difference between the inter-individual variance and the intra-individual variance when the sample size was reduced and fixed at four replicates. The attract package can be obtained from Bioconductor [http: //www. bioconductor. org/packages/devel/bioc/html/attract. html] and is available as a module in the MeV microarray analysis software [39] (http: //www. tm4. org/mev). To identify activated core pathways whose expression defines a control hONS phenotype, we ran attract on an expression data set consisting of the skin fibroblasts and two types of hONS lines from the control patients and a set of mesenchymal stem cell lines from a group of unrelated individuals. Attract was run using pathway modules defined by KEGG pathways represented in Bioconductor (version 2. 4. 1, Biobase version 2. 8. 0 and illuminaHumanv2BeadID. db version 1. 6. 0). A CV value was calculated for each detected probe by dividing the standard deviation of its expression in a donor group by its average group expression. Low and high-expression variance genes were identified as those genes below and above the 25th percentile of the genome-wide CV distribution based on values from all donors. P-values were obtained by comparing the counts of high, medium and low constraint genes in each of the control, SZ and PD groups and using a Chi-squared goodness of fit test where counts from the control group were designated as the expected counts. For each of the top five pathway modules, we applied a representational analysis to the set of highly-constrained and lowly-constrained genes for each of the donor groups. We used tools from the Bioconductor GOstats package (version 2. 14. 0, run on R version 2. 11. 1). P-values were adjusted for using the Benjamini-Hochberg method within each ontology class, donor group and pathway module (Cellular Component, Molecular Function, or Biological Process) and significant results obtained at the 0. 05 level. To focus on what functional terms were unique to genes of altered constraint, we excluded significant GO terms that appeared in both lists of highly-constrained and lowly-constrained genes and retained only those GO terms that were unique to each list. The significant GO terms for highly-constrained genes appear in the Supplementary File. A Kolmogorov-Smirnov test was used to test the degree distributions of the lowly and highly-constrained gene sets. A Gaussian kernel density estimator (the density function from R, using the default method to select bandwidth size) was used to produce the degree distribution density plots shown in Figure 4. Protein-protein interactions were defined using two knowledge-based annotation systems, Ingenuity Pathway Analysis (IPA) software and the GeneGo metacore tool. Both tools permitted identification of highly curated protein-protein interactions; using IPA we extracted the degree of connectivity for each gene in the attract-networks and an image of the interaction network was obtained from GeneGo.

Genes are a repository of information that provides the framework for cellular processes, with the flow of information from gene (DNA) to phenotype via an intermediate molecule—the messenger RNA. We understand that sequence variations in a gene may lead to phenotypic variations, but less well understood is how variation in the information flow itself might also impact on phenotype. In this study we demonstrated that disease phenotypes were correlated with expression variance. A change in expression variance might infer that the genetic networks representing information flow were less robust—surprisingly, we found that too little and too much variance were equally detrimental in the context of neurological disease.

Abstract Introduction Results Discussion Materials and Methods

genome expression analysis mathematics statistics genetics biology genomics biostatistics gene networks genetics and genomics

2011

Variance of Gene Expression Identifies Altered Network Constraints in Neurological Disease

Melioidosis is caused by the facultative intracellular bacterium Burkholderia pseudomallei and is potentially fatal. Despite a growing global burden and high fatality rate, little is known about the disease. Recent studies demonstrate that cyclooxygenase-2 (COX-2) inhibition is an effective post-exposure therapeutic for pulmonary melioidosis, which works by inhibiting the production of prostaglandin E2 (PGE2). This treatment, while effective, was conducted using an experimental COX-2 inhibitor that is not approved for human or animal use. Therefore, an alternative COX-2 inhibitor needs to be identified for further studies. Tolfenamic acid (TA) is a non-steroidal anti-inflammatory drug (NSAID) COX-2 inhibitor marketed outside of the United States for the treatment of migraines. While this drug was developed for COX-2 inhibition, it has been found to modulate other aspects of inflammation as well. In this study, we used RAW 264. 7 cells infected with B pseudomallei to analyze the effect of TA on cell survival, PGE2 production and regulation of COX-2 and nuclear factor- kappaB (NF-ĸB) protein expression. To evaluate the effectiveness of post-exposure treatment with TA, results were compared to Ceftazidime (CZ) treatments alone and the co-treatment of TA with a sub-therapeutic treatment of CZ determined in a study of BALB/c mice. Results revealed an increase in cell viability in vitro with TA and were able to reduce both COX-2 expression and PGE2 production while also decreasing NF-ĸB activation during infection. Co-treatment of orally administered TA and a sub-therapeutic treatment of CZ significantly increased survival outcome and cleared the bacterial load within organ tissue. Additionally, we demonstrated that post-exposure TA treatment with sub-therapeutic CZ is effective to treat melioidosis in BALB/c mice. Melioidosis is a tropical and often fatal disease caused by the aerobic, Gram-negative facultative intracellular bacterium Burkholderia pseudomallei [1]. Traditional B. pseudomallei infection is associated with environmental exposures during the monsoon season in the tropics. It has also been identified as a Tier 1 select agent due to its high mortality rate, ability to cause respiratory infection and its drug resistance. B. pseudomallei is most common in Southeast Asia and Northern Australia were it is found naturally as a soil-dwelling bacteria [2] [3], Evidence suggests the incidence of melioidosis is underreported and the global burden is increasing with an estimated 169,000 cases per year and 89,000 deaths across 34 countries annually [4]. Therefore, the need to establish multiple therapeutic strategies is paramount. B. pseudomallei is naturally resistant to many antibiotics, such as penicillin, many cephalosporins and aminoglycosides, but can be susceptible to such antibiotics as doxycycline, ceftazidime and chloramphenicol. The typical treatment regime for melioidosis lasts 20 weeks with both intravenous and oral phases of antibiotic administration. Due to these intense treatment requirements and antibacterial resistant isolates, relapse of the disease is common [2]. Recent evidence suggests that modulating the immune modulation by inhibiting cyclooxygenase-2 (COX-2) to reduce prostaglandin E2 (PGE2) expression is an effective post-exposure therapeutic. Another consideration of the current work is use of the COX-2 inhibition standard, NS-398, since is not approved for human use [5]. A COX-2 inhibitor with similar effects as NS-398 in a form administered easily to patients would be one step closer to developing a successful immune modulation regimen to treat melioidosis. This has profound implications as effective immune modulation treatment can reduce the selective pressure for bacteria to evolve to become drug resistant. Immune modulation can augment treatment of disease by either enhancing the effectiveness of a given antibiotic or by reducing the antibacterial dose required for treatment. Such interventions are promising developments toward the ultimate goal of eliminating an infectious disease by optimizing the host innate immune response [6]. Tolfenamic acid (TA) belongs to the fenamate class of NSAIDS and can be administered orally and intravenously to various animal species. Although not approved for human use in the United States, oral administration is used elsewhere in the world for treatment of migraines [7] [8]. While TA is primary known for its ability to inhibit COX-2, TA has also shown to be effective at modulating other key players in inflammation such as NF-ĸB. Evidence suggests a reduction in cytoplasmic NF-ĸB p65 activation in colon cancer cells and LPS stimulated RAW 264. 7 cells as well [9] [10]. In veterinary medicine, TA has been shown to be a potent inhibitor of NF-ĸB p65 canine-derived tumor cells [11]. With its wide range of inflammatory modulation implications, TA could prove to be a valuable augmentation to current treatment therapeutics for melioidosis. Additionally, because of the broad impact of TA on many inflammatory mediators, its use may further elucidate how B. pseudomallei causes mortality. In the present work, COX-2 inhibition was used to reduce the inflammatory response caused by B. pseudomallei. We examined the role of NF-ĸB, COX-2 and PGE2 during acute pulmonary infection with B. pseudomallei. For the first specific aim, we investigated the characteristics of the immune response and the potential of TA treatment to modulate the immune response and survival outcome. This work was done in-vitro by infecting RAW 264. 7 cells with the B. pseudomallei 1026b ΔpurM strain Bp82. These results served as the basis of a second specific aim to expand our BALB/c in-vivo study protocol and monitor the bacterial dissemination and organ system burden of B. pseudomallei 1026b in order to confirm the relationship between bacterial burden and dissemination. From there, the study focused on how treatment with TA and known effective antibiotics, alone or in combination, affect survival outcome over time. Dimethyl sulfoxide (DMSO) was purchased from ATCC (Manassas, VA). Dulbecco’s Modified Eagles Medium (DMEM) and trypsin used for cell culture were purchased from GE Healthcare Sciences (Hyclone) (Logan, UT) and fetal bovine serum plus (FBS +) was purchased from Atlas Biologicals (Fort Collins, CO). Tolfenamic acid was purchased from Cayman Chemical (Ann Arbor, MI). NF-ĸB monoclonal antibodies where purchased from Santa Cruz Biotechnology (Paso Robles, CA) and the secondary alexa fluor antibodies, 647 and HRP secondary antibodies used for all applications were purchased from Cell Signaling Technologies (Danves, MA). Luria-Bertani (LB) agar, cation adjusted Mueller-Hinton broth (ca-MHB) and COX-2 monoclonal antibodies were purchased from BD Sciences (Franklin Lakes, NJ). Dibutylhydroxytoluene (BHT), bovine serum albumin (BSA), Ceftazidime (CZ), crystal violet (CV) and sodium dodecyl sulfate (SDS) were purchased from Sigma-Aldrich (St. Louis, MO). Formalin and triton X-100 was purchased from ThermoFisher-Scientific (Waltham, MA). Ketamine for animal studies was purchased from Aurora Veterinary Supply (Aurora, CO). RAW 264. 7 cells were purchased from ATCC and maintenance was performed to the company’s specifications. Cells were grown in DMEM with 4 mM L-glutamine, 4500 mg/L glucose, 5 mM sodium pyruvate, 1500 mg/L sodium bicarbonate, and 10% FBS. Cells were grown in an incubator at 37°C with 5% CO2. Bp82, a ΔpurM B. pseudomallei 1026b mutant incapable of adenine and thymine biosynthesis, [12] was used as a 1026b BSL-2 surrogate organism for in-vitro experiments. 1026b [13] was used for all BSL-3 animal studies. Both strains were prepared by growing 1 colony of the respective bacterial cultures in 50 ml of LB broth for 48 hours at 37°C. Bacterial stock was then frozen back in ~1. 0 ml in LB broth and 10% glycol. RAW 264. 7 cells were plated in a 96-well plate with 100,000 cells per well and incubated for 24 hours. Wells were washed with sterile PBS, pretreated with either 100μM TA or 0. 01% DMSO for 30 minutes, and infected with Bp82 at an MOI of 5. Each time point also included an untreated/infected control and an uninfected/untreated control. At each time point, 100μL of 10% formalin in methanol was added to each treatment well and placed on a rocker for 15 minutes with 20 tilts per minute. The formalin solution was removed and 100μL of a 0. 5% solution of CV in 25% methanol/75% ddH2O was added to each well and placed on the rocker for 15 minutes. After staining, the 96 well plates were gently washed under tap water until the water ran clear from each well. The plate was air dried overnight until there was no visible liquid in any well. Finally, the CV was re-suspended from the cells by adding 100μL of 1% SDS solution to each well and placed on the rocker for 30 minutes at room temperature. The plate was then read on a plate reader at 570nm and 600nm. This procedure was adapted and optimized from Kursheed et al [14] and Castro-Garza et al [15]. Each treatment had a technical replicate of 10 and the experiment was run in biological triplicate for statistical significance. To determine if TA possessed any natural antibiotic properties, a MIC was established. The procedure outlined in [16] was used. Briefly, Bp82 stock was incubated in LB broth overnight (18 hours) at 37°C, passed at a 1: 100 dilution, and incubated for an additional 6 hours. This stock was then diluted to reach an optical density of 0. 1 at 600 nm in caMHB. This diluted culture was further diluted 1: 100 to achieve approximately 1x 106 CFU/ml inoculum. In 50μl of caMHB, 1: 2 serial dilutions of TA and CZ were prepared in a 96-well plate with the highest concentration being 128μg/ml to the lowest concentration of 0. 0625μg/ml. Each treatment was run in triplicate and inspected by different personnel for verification. 50μl of inoculum was added to each well of the 96-well plate and plate incubated overnight at 37°C for 18 hours. The MIC concentration was determined as the lowest concentration of treatment that resulted in no bacterial growth. RAW 264. 7 cells were cultured in six-well plates at a density of 2x106 cells per well for 24 hours. Cells were then pre-treated with 100μM TA or 0. 01% DMSO for 30 minutes. The untreated and uninfected control were pretreated with fresh media. After pre-treatment, cells were infected at an MOI of 5, and centrifuged for 5 minutes at 500 rpm to allow the infection to reach the monolayer. At 90 minutes or 6 hours post infection (NF-ĸB/COX-2) cells were washed twice with sterile PBS (pH7. 4) and removed utilizing 0. 025% trypsin and neutralized with complete DMEM. Cells were fixed in 3. 7% paraformaldehyde in PBS for 10 minutes and washed with PBS. Cell membranes were permeabilized with 0. 01% triton X-100 in PBS for 10 minutes and the primary antibody stain was added at a dilution of 1: 100. Samples were incubated at 42°C for 20 minutes, washed with PBS and re-suspended in 0. 01% triton X-100 in PBS and the secondary antibody at a dilution of 1: 100. Samples were washed and prepared for analysis in PBS. Flow cytometry was conducted on a Beckman Coulter CyAn ADP Flow Cytometer operating Summit v4. 3 software for data collection. All further data analysis was done with FlowJo software. Samples were run in biological triplicate with two technical duplicates. RAW 264. 7 cells were cultured as described in 2. 5. At each time point after infection, the supernatant was removed and 1% BHT solution was added to avoid the free-radical peroxidation as explained in [17]. Samples were then analyzed or stored at -20°C for later analysis. PGE2 analysis and quantification was conducted via ELISA using a PGE2 analysis kit (R&D Systems) in accordance with the manufacturer' s instructions, except samples were not diluted as indicated. This experiment was conducted three independent experiments run in duplicate. Ethics Statement: Animal experiments were performed in accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. The protocol was approved by the Colorado State University Institutional Animal Care and Use Committee, protocol number 15-6138A. Six to eight week old, female BALB/c mice (Jackson Laboratories) were maintained under pathogen-free conditions and allowed free access to sterile food and water with 12 hour light/dark cycles. 45 mice were separated into 9 groups (n = 5/group) depending on treatment and survival regime. For bacterial challenge, mice were anesthetized with ketamine/xylazine (100/10 mg/kg). The 1026b bacterial inoculum contained ~2×103 CFU suspended in 20 μl sterile saline and was delivered dropwise via pipet. Bacterial CFU were confirmed by plating the inoculum on LB agar. Treatments were initiated 3 hours post infection and repeated every 24 hours for two consecutive days as done by Asakrah et al [5]. Mice were euthanized when morbidity characteristics of hunched posture, loss of response to stimuli and loss of >20% body weight were reached. After euthanasia, the lung and spleen were removed and homogenized in 1 ml 0. 9% sterile saline. Serial dilutions of tissue homogenates were plated on LB agar and bacterial CFU were counted after 2 days of incubation at 37°C. For the effects of treatment on organ bacterial burden and dissemination, 30 mice were divided into 6 treatment groups (n = 5/group): TA (50 mg/kg) suspended in corn oil given via oral gavage, CZ (200 mg/kg and 25 mg/kg) dissolved in sterile PBS (pH 7. 4) given subcutaneously, or co-treated with TA (50 mg/kg) at a sub-therapeutic dose of CZ (25 mg/kg), untreated, or vehicle control (corn oil) via oral gavage. Once the untreated infected control group showed signs of morbidity (typically around 60 hours post infection), mice were euthanized, and the lung and spleen from 3 mice (selected randomly) from each group were homogenized and plated for bacterial load determination. For the survival study, mice were divided into 3 treatment groups (n = 5/group): 25 mg/kg CZ administered subcutaneously, 50 mg/kg TA administered via oral gavage in corn oil, and a co-treatment of 25 mg/kg CZ and 50 mg/kg TA. After the treatment, mice were monitored daily for signs of mortality for up the 37 days post-infection. Statistical analyses was performed using Prism 6. 0 software (Graphpad). Log-rank Mantel-Cox analysis was conducted for survival curves. All other data were analyzed using a one-way or two-way ANOVA followed by the Bonferroni post-test to determine statistical differences between groups or a two-tailed t-test for experiments with less than three groups. p<0. 05 was considered statistically significant. To determine if treatment with TA resulted in increased cell viability in-vitro, pretreated RAW 264. 7 cells were infected with Bp82 at an MOI of 5 and cell viability was monitored over 4,6, and 8 hours. Fig 1 reveals that Bp82 infection resulted in 44% reduction in cell viability after 4 hours, 66% after 6 hours, and 75% after 8 hours. RAW 264. 7 cell cultures pretreated with 100μM TA revealed a reduction in cytotoxicity induced by Bp82 by 8% at 4 hours, by 42% at 6 hours and 30% at 8 hours when compared to the vehicle control group. Experimentation after 8 hours was not conducted as greater that 75% cytotoxicity was shown in later time points. 100 μM TA is a significant treatment, as treatments of this magnitude have shown to limit cell growth and cause cell death in colon cancer cells via activation of apoptosis pathways [9]. Therefore, we needed to confirm TA did not possess an inherent chemical nature that affected the bacterial growth and proliferation of Bp82. The MIC of TA exceeds 489 μM (Table 1), which exceeds the treatment doses used in this study. This suggests that the ability of TA in enhancing cell viability is a result of the immune modulatory properties of TA. Additionally, the MIC of DMSO was confirmed to be greater than 10%, confirming that the 0. 01% vehicle concentration used in this study does not affect bacterial growth and proliferation. The MIC of CZ was used as an experimental control. The MIC of CZ fell within previously published results [16]. A combination of flow cytometry and western blotting were used to confirm Bp82 activation of the NF-ĸB pathway. Infection results in a 2. 5 fold increase in p65 mean channel fluorescence (MCF) in RAW 264. 7 cells (Fig 2) and pretreatment with TA reduced the MCF by 20% when compared to the DMSO control. This suggests that TA is able to significantly reduce p65 levels in Bp82 infected RAW 264. 7 cells. S3 Fig confirms the effects of TA on p65 via western blot. This is consistent with previously published work as TA was shown to reduced p65 levels in LPS activated RAW 264. 7 cells [10]. Infection with Bp82 resulted in a ~1. 5-fold increase in COX-2 MCF from uninfected groups as presented in Fig 3. We also determined that pretreatment of RAW 264. 7 cells pre-treated with 100 μM TA and infected with Bp82 resulted in a ~30% reduction in COX-2 MCF versus the vehicle control group. S5 Fig offers visual representation of this reduction in COX-2 via immunofluorescence. It has been shown that PGE2 plays an important role during infection and its suppression is a possible therapeutic strategy [18]. Additionally, it has been shown that B. thailandensis infection results in increased production of PGE2, suggesting that PGE2 is necessary for the bacteria’s intracellular survival [5]. Our next goal was to confirm that Bp82 upregulates the production of PGE2 in RAW 264. 7 cells and pretreatment with 100 μM TA reduces PGE2 production. Indeed, Bp82 infection resulted in 3-fold increase in supernatant PGE2 concentration after 4 hours of infection, and over a 4 fold increase in supernatant concentration at 6 and 8 hours. TA effectively reduced PGE2 levels by over 4-fold at all time points when compared to the vehicle control (Fig 4). Additionally, it would appear that pretreatment with TA reduced the PGE2 concentration to levels lower than the uninfected control, however, this was not significant. The only treatment successful at significantly reducing the bacterial burden in both the lung and spleen was 200 mg/kg CZ (Fig 5). This confirms 200 mg/kg CZ administered subcutaneously is effective at reducing bacterial burden as shown in previous studies [19]. To assess if immune-modulation can potentiate the efficacy of CZ against an acute B. pseudomallei infection in the murine model, TA was co-administered with a sub-therapeutic dose of CZ. This data also indicates that 25 mg/kg CZ was ineffective at reducing the bacterial burden in the lung and spleen. In TA-treated mice there was a 40% increase in survival time for 40% of the group (Fig 6). Co-treated TA/CZ group had 100% survival until 37 days post infection (pre-determined study endpoint). After 37 days, bacterial burden was assessed in the lung and spleen and compared to burden assessed at day 2. 5 (Fig 7). Notably, the lung showed a reduction in bacterial burden of ~5x107 CFU/ml and the spleen a slight reduction of ~1 x 103 CFU/ml between 2. 5 and 37 days post infection. Combating the problem of antibacterial resistance is a global responsibility. This is particularly relevant for B. pseudomallei due to its ability to efflux many antibiotics [20]. Novel conjunctive therapies should be investigated to reduce the current antibacterial treatment regimes. Immunomodulation utilizing COX-2 inhibition during B. pseudomallei infection may prove an effective strategy to increase antibacterial effectiveness and treat the disease. We corroborated previous studies by Asakrah et al [5] on cytotoxicity and the important role of COX-2 and PGE2 during melioidosis. These findings were expanded upon by investigating the effects of TA treatment on the NF-ĸB pathway during infection. We were able to mimic similar treatment outcomes using TA, which is approved for human use in certain areas around the globe [7]. Most importantly, we showed that the conjunctive treatment of immunomodulation with sub-therapeutic antibiotics significantly increased survival outcome and decreased organ bacterial burdens. PGE2 has been linked to the regulation of a number homeostatic biological functions. Of greatest relevance here is that its involvement in initiating the classic signs of inflammation including redness and swelling due to PGE2-mediated arterial dilatation and increased vascular permeability and pain from PGE2 acting on sensory neurons [21] Additionally, all of the aforementioned mechanisms can lead to tissue damage. This compromised tissue provides an opportune environment for bacterial proliferation. Limiting inflammatory damage during bacterial infections has been shown to be an effective therapeutic strategy [6]. Our in-vitro studies revealed that limiting COX-2 induction and PGE2 production using TA translated to a significant increase is cell viability in mouse macrophage-like RAW 264. 7 cells. This further confirms the results published by Asakrah et al [5]. Furthermore, pretreatment with TA was not only able to inhibit COX-2, as shown by the reduction of PGE2 production, but treatment also limited COX-2 induction. This may be linked to effects of TA on NF-ĸB and the interaction between the two inflammatory pathways. NF-ĸB may prove an important inflammatory target during B. pseudomallei infection as it is believed to activate during infection and lead to the translation of many pro-inflammatory cytokines [3]. These cytokines amplify the systemic inflammatory response, often times by inducing COX-2 and resulting in continued production of PGE2 and other prostaglandins. PGE2 can have a positive feedback loop with NF-ĸB by increasing its transactivation and enhancing the production of pro-inflammatory cytokines [22]. Previous studies indicate TA is effective at reducing NF-ĸB activation in stimulated RAW 264. 7 cells [10], but it is unclear whether this is linked to the PGE2/NF-ĸB relationship. Here, we show that treatment with TA reduces Nf-ĸB p65 during B. pseudomallei infection. Immunomodulation may be necessary for comprehensive treatment of bacterial infections, particularly to combat bacterial resistance [6]. Our in-vivo studies reveal that immune modulation with orally administered TA is effective at increasing survival outcome in BALB/c mice. This is consistent with the findings of Asakrah et al [5]; although those studies were conducted using NS-398, an experimental COX-2 inhibitor given intraperitoneally. TA treatment alone did not show the same efficacy as NS-398 by Asakrah et al [5] but this may be due to the present dose being too low. More research is needed to determine if a larger dose of TA affords greater protection from a lethal B. pseudomallei pulmonary challenge. Co-treatment of TA and CZ significantly increased survival outcome, and contributed to a nearly complete bacterial clearance after 37 days post infection. This suggests that the immunomodulation activity of TA allows for the exploitation and enhancement of the therapeutic benefit of CZ. To the best of our knowledge, this is the first study indicating that immune modulation with orally-administered TA enhances the therapeutic benefit of antibiotic treatment during pulmonary melioidosis. The increased survival outcome resulting from the reduction of PGE2 production during bacterial infections through COX-2 inhibition and reduction in NF-ĸB activation has profound implications. PGE2 is produced during a number of lethal bacterial infections (i. e Francisella tularesis [23]) and opportunistic bacterial infections (Pseudomonas aeruginosa and Staphylococcus aereus) for which antibacterial treatment is complex due to antibacterial resistance [24] [25]. Proving the efficacy of immunotherapy using commercially available, orally administered TA in combination with sub-therapeutic antibiotic treatment during melioidosis in BALB/c mice warrants further investigation in other animal models of melioidosis. Such experiments may further reveal the effectiveness of TA and other NSAIDS across a wide range of bacterial infections. While this may have a profound impact on treatment of human bacterial infections, additional animal studies are needed to ensure efficacy of this treatment strategy before human clinical trials.

Burkholderia pseudomallei is the causative agent of melioidosis, a fatal tropical disease endemic in parts of Southwest Asia and Northern Australia. While it was once believed to be isolated to these areas, recent research indicates the global burden on melioidosis is growing. Furthermore, treatment of melioidosis is difficult because of the high occurrence of disease relapse and increasing antibacterial resistance. Recent research suggests that immunomodulation via COX-2 inhibition to subsequently reduce with PGE2 production is an effective therapeutic strategy for melioidosis. The current study was built on this immunomodulatory principle by using an orally administered COX-2 inhibitor and evaluating its effects on the COX-2 and NF-ĸB pathways. We also investigated whether the conjunctive therapies of immunomodulation and antibiotics increased efficacy of the treatment. We confirmed immunomodulation is effective as a post-exposure therapeutic in BALB/c mice. More importantly, we found that conjunctive post-exposure treatment via immunomodulation increased antibacterial treatment efficacy. Conjunctive therapy may prove efficacious for other infectious diseases resembling melioidosis. Hence, further research is needed to identify the long-term effects of the described treatment (s) across multiple animal models.

Abstract Introduction Materials and Methods Results Discussion

antimicrobials medicine and health sciences immune physiology pathology and laboratory medicine spleen melioidosis drugs immunology microbiology animal models bacterial diseases model organisms signs and symptoms antibacterials antibiotics analgesics pharmacology immunomodulation research and analysis methods infectious diseases inflammation cox-2 inhibitors mouse models immune response pain management diagnostic medicine physiology microbial control biology and life sciences

2016

Immune Modulation as an Effective Adjunct Post-exposure Therapeutic for B. pseudomallei

Growing evidence supports other regulatory roles for protein ubiquitination in addition to serving as a tag for proteasomal degradation. In contrast to other common post-translational modifications, such as phosphorylation, little is known about how non-degradative ubiquitination modulates protein structure, dynamics, and function. Due to the wealth of knowledge concerning protein kinase structure and regulation, we examined kinase ubiquitination using ubiquitin remnant immunoaffinity enrichment and quantitative mass spectrometry to identify ubiquitinated kinases and the sites of ubiquitination in Jurkat and HEK293 cells. We find that, unlike phosphorylation, ubiquitination most commonly occurs in structured domains, and on the kinase domain, ubiquitination is concentrated in regions known to be important for regulating activity. We hypothesized that ubiquitination, like other post-translational modifications, may alter the conformational equilibrium of the modified protein. We chose one human kinase, ZAP-70, to simulate using molecular dynamics with and without a monoubiquitin modification. In Jurkat cells, ZAP-70 is ubiquitinated at several sites that are not sensitive to proteasome inhibition and thus may have other regulatory roles. Our simulations show that ubiquitination influences the conformational ensemble of ZAP-70 in a site-dependent manner. When monoubiquitinated at K377, near the C-helix, the active conformation of the ZAP-70 C-helix is disrupted. In contrast, when monoubiquitinated at K476, near the kinase hinge region, an active-like ZAP-70 C-helix conformation is stabilized. These results lead to testable hypotheses that ubiquitination directly modulates kinase activity, and that ubiquitination is likely to alter structure, dynamics, and function in other protein classes as well. The best-characterized function of ubiquitin is as a marker for protein degradation. The 76-residue ubiquitin protein can be covalently attached at its C-terminus to a lysine side chain amine group or N-terminus of another protein. Typically, additional ubiquitin molecules are also attached to the first ubiquitin, forming a polyubiquitin chain. The polyubiquitinated protein is then targeted to the proteasome for degradation, although ubiquitination can also result in lysosomal degradation [1–3]. Ubiquitin tagging for selective degradation is crucial to the regulation of many cellular processes [4]. Recently, alternate roles for ubiquitination that are not associated with degradation have gained attention. In 1992, Varshavsky et al. postulated that ubiquitin might have additional functions beyond proteasome targeting and that it might perturb the conformation of the proteins to which it is linked [5]. The residue linking two ubiquitin molecules along the polyubiquitin chain can influence whether the ubiquitinated protein is targeted for proteasomal degradation. The most common linkage, on K48, is closely associated with degradation by the proteasome [6]. K6, K11, K27, and K23 linkages have not been studied as extensively, but may also be associated with proteasomal degradation [7,8]. Levels of K63-linked ubiquitination, however, are not affected by proteasomal inhibition, suggesting alternate functions [6–8]. K63-linked polyubiquitination is also associated with protein trafficking, DNA repair, and inflammation [7]. Monoubiquitination also is not typically associated with degradation, and in some cases may regulate protein function by impeding or facilitating protein-protein association [8–10]. Finally, several ubiquitin-like proteins, such as NEDD8 and SUMO[3], are known to modify proteins in a similar manner to ubiquitin, and regulate function independent of degradation. Several pathways have been identified in which ubiquitination influences function. One example is calmodulin, which is site-specifically and reversibly monoubiquitinated, reducing its activity toward phosphorylase kinase [11]. Non-degradative ubiquitination both positively and negatively regulates the TGF-β signaling pathway through mono and polyubiquitinated Smad proteins [12–14]. TAK1 polyubiquitination with a K63 linkage also regulates TGF-β signaling by activating TAK1 [15]. These ubiquitin modifications may modulate protein-protein interactions within the pathway, or directly induce conformational changes in the modified protein. A second example of non-degradative ubiquitination controlling function is the T cell receptor-zeta (TCR-zeta), which is polyubiquitinated by Cbl-b and Itch with a K33 linkage. This modification regulates TCR-zeta phosphorylation and its association with protein kinase ZAP-70 in a non-proteolytic manner [16–18]. Rather than inducing degradation, the ubiquitin chain promotes phosphorylation, possibly by interacting with the kinase Lck, or by altering the conformation of the TCR-zeta. There are several possible mechanisms for how ubiquitin might regulate protein function independent of degradation. Computational structural modeling suggests that ubiquitination can directly affect protein function by blocking a functional domain or active site, sterically preventing oligomerization, preventing target protein binding, or restricting protein flexibility [19]. A perspective by Chernorudskiy and Gainullin details examples of ubiquitination that regulates protein structure and function without involvement of a ubiquitin binding domain (UBD) [20]. Monoubiquitination of histone H2B has a steric effect, preventing chromatin compaction which allows increased methylation of histone H3, and does not depend on recognition by a UBD [21,22]. Ras is another example of activation by site specific monoubiquitination [23]. Different Ras GTPase isoforms are monoubiquitinated at distinct sites which result in distinct mechanisms of Ras activation, and chemical attachment of monoubiquitin at these sites recapitulates this result [24,25]. This suggests a conformational change that is different depending on the ubiquitination site. Alpha-synuclein also shows a conformational dependence on ubiquitination site in terms of propensity to form fibrils and oligomers [26]. Certain ubiquitination sites on alpha-synuclein result in proteasomal degradation, while other sites do not [27]. Crystal structures of these ubiquitinated proteins would be very useful for determining what conformational changes occur upon ubiquitination and how these affect function. Two crystal structures of monoubiquitinated constructs have been solved for proliferating cell nuclear antigen (PCNA) [28,29]. These structures show that the ubiquitin is flexible, interacting with PCNA through the canonical hydrophobic patch at I44 or extended away from PCNA due to electrostatic repulsion, which is consistent with SAXS data [29,30]. The structures show changes in PCNA conformation upon ubiquitination, and mutational studies indicate that the ubiquitin hydrophobic patch can directly interact with the polymerase, facilitating polymerase exchange. The ubiquitin moiety also provides a steric restriction on proteins interacting with PCNA. A recent combined SAXS and computational study shows that SUMOylation has a different effect on PCNA conformation than ubiquitination [31]. A crystal structure of SUMOylated thymine DNA glycosylase (TDG) also reveals a conformational change in the modified protein caused by attachment of the SUMO moiety [32]. SUMOylation of the TDG C-terminal domain results in formation of a helix at the TDG-SUMO interface that disrupts TDG function [32], a mechanism which could also occur in ubiquitination. Crystal structures of ubiquitin conjugated to an E2 ubiquitin conjugating enzyme [33] and a HECT E3 ubiquitin ligase [34,35], reveal that a ubiquitinated protein can form a compact interaction with ubiquitin with the canonical hydrophobic patch either buried or solvent exposed. Although no crystal structures of a polyubiquitinated target protein are available, many structural studies have been performed on polyubiquiitn chains on their own, which have implications for the biophysical consequences of polyubiquitination. K48-linked diubiquitin and tetraubiquitin both adopt compact, globular conformations with the ubiquitin moieties bound together, burying the I44 hydrophobic patch at the interdomain interface [36–38]. However, other linkages form different conformations as a polyubiquitin chain. K63-linked diubiquitin adopts an open, extended conformation, and as a tetrameric chain it fluctuates between extended and compact conformations [39–44]. The rarer linkages have not been studied as thourghly, but K33-linked, K29-linked and linear diubiquitin are observed to adopt extended structures without significant contacts between the ubiquitin monomers, similarly to K63 [39,45]. In crystal structures, K11 and K6-linked diubiquitin chains form compact conformations that are globular, but distinct from K48-linked diubiquitin [46,47]. From these polyubiquitin structures, it seems likely that K48-linked polyubiquitin would have a different physical effect on its target than K63-linked polyubiquitin or one of the other linkages that adopts an extended conformation. In an extended structure, it is possible that only the proximal ubiquitin domain in the polyubiquitin chain would directly interact with the target protein. In this work, we focus on ubiquitin modifications of protein kinases, because many mechanisms for kinase regulation have been characterized previously. The role of post-translational modifications (PTM), especially phosphorylation, in regulating kinase activity has been extensively investigated. A common site of kinase phosphorylation is the activation loop, where the introduction of the negatively charged phosphate is usually critical to complete activation [48]. Other examples of kinase regulation induced by phosphorylation include activation by releasing an autoinhibitory element (PKA, PKC, EphB2 and TGF-β receptor tyrosine kinases) [48–50], inactivation by sterically hindering access to the active site (CDKs) [51], inactivation by triggering association with an autoinhibitory domain (Src kinases) [48,52], and inactivation by inhibiting dimerization (DAPK2) [53]. Phosphorylation at one site can also affect phosphorylation at another site, as in the case of AMPKα, where phosphorylation at S485 or S173 reduces phosphorylation of the activation loop (T172) [54]. Beyond phosphorylation, lysine acetylation can also play a role in kinase regulation. For example, AMPK is inhibited by acetylation of a regulatory subunit [54,55]. Studies of kinase regulation by ubiquitination have demonstrated decreased kinase activity by triggering proteasomal or lysosomal degradation, as in the case of tyrosine kinase cell surface receptors [56] and AMPK [54]. Ubiquitination of Sic1 activates the cyclin-CDk protein complex, when Sic1 dissociates from the complex before being degraded by the proteasome [57]. However, there are also indications that ubiquitin and ubiquitin-like proteins can regulate kinases in a degradation-independent manner. In T cells, Cbl-b ubiquitinates and negatively regulates PI3-kinase, and this is not associated with degradation [56]. SUMOylation was found to increase activity of AMPK-β by increasing activation loop phosphorylation [54]. Recent developments in mass spectrometry and antibody enrichment of samples for ubiquitinated proteins have enabled several cell-wide proteomics studies of ubiquitinated proteins [8,58–61]. These experiments identify the precise site of ubiquitination on the protein and can indicate whether that ubiquitination site is associated with proteasomal degradation; however, they lack the ability to distinguish between mono and polyubiquitination. Kim et al. found that 42% of the ubiquitination sites in their study did not increase by more than 2-fold when the proteasome was inhibited, indicating that ubiquitination at these sites may have a function other than proteasomal degradation. In particular, they observed that when ubiquitin itself was ubiquitinated at K63 (indicating K63-linked polyubiquitination of an unknown target), this was not associated with proteasomal degradation [8]. In this study, we report proteomics data on ubiquitination of 132 kinases in Jurkat and cells, both in the presence and absence of the proteasome inhibitor MG132. We identify many ubiquitination sites on protein kinases that do not increase substantially upon treatment with MG132. Combining our data with previously published data from the PTMfunc database [60], we find 107 kinase ubiquitin modifications that are proteasome sensitive, and 103 that are proteasome insensitive. This suggests that ubiquitination of some kinases likely has a downstream effect other than proteasomal degradation. The sites of ubiquitination are most commonly found within folded protein domains, as has been observed of ubiquitination generally [60,62–64]. In the kinase domains, these sites are also concentrated in regions known to be important for regulating activity. We hypothesized that ubiquitination, like other PTMs such as phosphorylation and acetylation, modulates the conformation and dynamics of the protein to which they are attached. Molecular dynamics (MD) simulations have been used to study the ways in which phosphorylation changes protein energy landscapes, and to thereby obtain new insights into the mechanisms by which phosphorylation affects protein function [65–69]. The primary effect of phosphorylation on the energy landscape is often electrostatic because the phosphate group introduces a -2 charge, which can disrupt hydrogen-bonding networks or introduce new stabilizing interactions. Using MD, lysine acetylation has also been shown to have allosteric effects on protein conformation and influence conformational stability [70,71]. Ubiquitin differs from these well-characterized modifications in that it is much larger and attached via a short but flexible linker, and it can be extended as a polyubiquitin chain with different linkage types. To begin exploring the ways in which ubiquitin modulates protein structure and dynamics, we performed a series of MD simulations for one kinase from our proteomics data, ZAP-70 tyrosine kinase, because it is ubiquitinated at several sites that are proteasome insensitive. ZAP-70 is a member of the Syk gene family of protein tyrosine kinases involved in T cell development and activation [72]. It consists of a kinase domain and two SH2 domains, which bind to the hinge region of the kinase, trapping it in an inactive conformation [73–77]. This autoinhibition is released when the SH2 domains bind to the phosphorylated form of the TCR complex ITAMs [75,76,78–80]. The SH2 domains detach from the kinase domain, exposing two conserved tyrosine residues on the linker between the SH2 and kinase domains, which are subsequently phosphorylated [72,73,75–77]. The crystal structure of the unphosphorylated ZAP-70 kinase domain without the SH2 domains bound adopts an active-like conformation [76,81]; however, phosphorylation on Y493 on the activation loop is required for full activation [72,75,76]. The downstream effect of ZAP-70 activation is Cbl phosphorylation and ultimately T cell activation [72]. Although a few studies have noted that ZAP-70 can be ubiquitinated in vivo [18,82,83], no one has examined how ubiquitination affects ZAP-70 conformation or activity. We performed MD simulations of two different monoubiquitinated constructs of ZAP-70 and characterized the sampled conformational ensembles. The proteomics data does not determine the number of ubiquitin molecules, so we attached only a single ubiquitin molecule, as this is the minimal possible modification and may have an effect that is relavent to all types of ubiquitination. These were compared to simulations of the unmodified ZAP-70 as well as other (non-ubiquitin) modifications to the same sites. The results demonstrate that ZAP-70 ubiquitination modulates structure and dynamics in a site-specific manner, and in ways that suggest possible regulation of kinase activity. For ubiquitination analysis in Jurkat cells, cells were first SILAC-labeled for at least 10 doublings in either “light” media containing standard amino acids or “heavy” media containing 13C6-lysine and 13C6,15N4-arginine [84]. Cells were then infected with VSV-pseudotype HIV (strain NL4-3) virus at a multiplicity of infection of 5. Both “light” and “heavy” cell populations were infected or mock infected in duplicate. Infected cells were either treated with MG-132 (Millipore) at 10 μM for 4 hours prior to harvest or were left untreated. Cells were harvested 48 hours after infection by lysis in a buffer containing 8 M urea, 0. 1 M Tris pH 8. 0,100 mM NaCl, and protease inhibitors (Complete EDTA-free, Roche). 10 mg of protein as measured by a Bradford assay (Quick Start Bradford reagent, Bio-Rad) was subjected to trypsin digestion and ubiquitin remnant immunoaffinity purification with 250 μg of ubiquitin remnant antibody (Cell Signaling) as previously described [85]. The diglycine remnant against which the ubiquitin remnant antibody was raised is also exposed for NEDD8 and ISG15-modified proteins. 100 μg of each trypsin-digested Jurkat lysate was reserved for protein abundance analysis. Peptides were fractionated by hydrophilic interaction chromatography (HILIC). The peptides were injected onto a TSKgel amide-80 column (Tosoh Biosciences, 2. 0 mm x 15 cm packed with 5 μm particles) equilibrated with 10% HILIC Buffer A (2% ACN, 0. 1% TFA) and 90% HILIC Buffer B (98% ACN, 0. 1% TFA) using an AKTA P10 purifier system. The samples were then separated by a one-hour gradient from 90% HILIC Buffer B to 55% HILIC Buffer B at a flow rate of 0. 3 ml/min. Fractions were collected every 1. 5 min and combined in 12 fractions based on the 280 nm absorbance chromatogram. Fraction were evaporated to dryness and reconstituted in 20 μl of 0. 1% formic acid for mass spectrometry analysis. For ubiquitination analysis in HEK293 cells, a HEK293 cell line was engineered to express HIV (strain NL4-3) with a tetracycline-responsive promoter element using the pTRE-Tight TetOn system (Clontech) [86] and a self-inactivating deletion in the 3’ UTR to prevent second round infection (publication forthcoming). All cells were treated with interferon alpha 1 (Cell Signaling) at 30 ng/ml and for +Dox samples HIV was induced by the addition of doxycycline (Sigma) at 1 μg/ml for 48 hours. Cells were either treated with MG-132 (Millipore) at 10 μM for 5 hours prior to harvest or were left untreated. All HEK293 experiments were performed in duplicate. 1 mg of protein as measured by a Bradford assay (Quick Start Bradford reagent, Bio-Rad) was subjected to trypsin digestion and ubiquitin remnant immunoaffinity purification with 31 μg of ubiquitin remnant antibody (Cell Signaling) as previously described [85]. All samples were analyzed on a Thermo Scientific LTQ Orbitrap Elite MS system equipped with an Easy nLC-1000 HPLC and autosampler system that is capable of maintaining backpressures of up to 10,000 psi for high-resolution chromatographic separations. The HPLC interfaces with the MS system via a nanoelectrospray source. Samples were injected onto a C18 reverse phase capillary column (75 um inner diameter x 25 cm length, packed with 1. 9 um ReproSil Pur C18-AQ particles). Peptides were then separated by an organic gradient from 5% to 30% ACN in 0. 1% formic acid over 112 minutes at a flow rate of 300 nL/min. The mass spectrometer collected data in a data-dependent fashion, collecting one full scan in the Orbitrap at 120,000 resolution followed by 20 collision-induced dissociation MS/MS scans in the dual linear ion trap for the 20 most intense peaks from the full scan. Dynamic exclusion was enabled for 30 seconds with a repeat count of 1. Charge state screening was employed to reject analysis of singly charged species or species for which a charge could not be assigned. Raw mass spectrometry data were analyzed using the MaxQuant software package (version 1. 3. 0. 5) [87]. Data were matched to the SwissProt human reference protein database (downloaded on May 15,2013). MaxQuant was configured to generate and search against a reverse sequence database for false discovery rate calculations. Standard variable modifications were allowed for methionine oxidation, and protein N-terminus acetylation. When searching ubiquitination samples, a variable modification was also allowed for lysine ubiquitin remnant addition. A fixed modification was indicated for cysteine carbamidomethylation. Full trypsin specificity was required. The first search was performed with a mass accuracy of +/- 20 parts per million and the main search was performed with a mass accuracy of +/- 6 parts per million. A maximum of 5 modifications were allowed per peptide. A maximum of 2 missed cleavages were allowed. The maximum charge allowed was 7+. Individual peptide mass tolerances were allowed. For MS/MS matching, a mass tolerance of 0. 5 Da was allowed and the top 6 peaks per 100 Da were analyzed. MS/MS matching was allowed for higher charge states, water and ammonia loss events. Data were searched against a concatenated database containing all sequences in both forward and reverse directions with reverse hits indicating the false discovery rate of identifications. The data were filtered to obtain a peptide, protein, and site-level false discovery rate of 0. 01. The minimum peptide length was 7 amino acids. Results were matched between runs with a time window of 2 minutes for technical duplicates. The HIV ubiquitination proteomics results from the HEK293 and Jurkat cells were analyzed using an in-house computational pipeline built for the analysis of post-translationally modified peptides with mixed effect models, implemented in the MSstats (v2. 3. 4) Bioconductor package [88]. For both datasets, protein identifiers were converted into modification site identifiers, contaminant and false positive MaxQuant search results were removed, and all samples were normalized per cell line by median-centering the log2-transformed MS1-intensity distributions. Then, the MSstats groupComparison function was run on the Jurkat dataset with the following options: no interaction terms for missing values, no feature interference, unequal intensity feature variance, restricted technical and biological scope of replication. The HEK293 dataset was processed in a similar fashion with the following options: imputation of missing values with the mean minimal observed MS1-intensity across samples, no feature interference, unequal intensity feature variance, restricted technical and biological scope of replication. We opted not to impute missing values for the SILAC Jurkat dataset since they are less prevalent due to the nature of SILAC assays. In addition to the ubiquitination proteomics data that we collected in Jurkat and HEK293 cells, we used the PTMfunc database of protein PTMs [60] to identify other PTMs present on human protein kinases. Functional annotations for some of these PTMs are included in the PTMfunc database [60]. In-house scripts were used to map PTMs from all human kinases onto the PKA and ZAP-70 structures based on an alignment using the ClustalW2 multiple sequence alignment software from EMBL-EBI [89–91]. When mapping PTMs onto the PKA or ZAP-70 structures, any PTM in a location with no alignment to the mapping protein was omitted. To determine whether a ubiquitination site is associated with proteasomal degradation, we used data collected in the presence of a proteasome inhibitor (MG-132). Sites whose abundance was at least two-fold greater in the presence of a proteasome inhibitor compared to control were categorized as proteasome sensitive, whereas those with a difference of less than two-fold were categorized as proteasome insensitive. If the same site was proteasome sensitive in one cell line or condition and proteasome insensitive in another, we categorized it as neither. For sites that were significantly more abundant in the presence of HIV, we determined the proteasome sensitivity based on MG-132 sensitivity in the presence of HIV only, since the level of ubiquitination without HIV present might be very low. We determined whether differences between groups (such as proteasome-sensitive and insensitive ubiquitin modifications) are significant by using the function prop. test in R [92] to find a p-value, which tests whether the proportions in different group are the same. We performed all-atom MD simulations starting from 9 different ZAP-70 kinase domain constructs. We had five constructs built from the active-state crystal structure: no modification (control), monoubiquitinated at K377, monoubiquitinated at K476, acetylated at K377, and with an immunoglobulin (Ig) domain from Cardiac Myosin Binding Protein C attached at K377. We also built four constructs from the inactive-state crystal structure: control, monoubiquitinated at K377, monoubiquitinated at K476, and acetylated at K476. As part of a larger study of HIV-host interactions, data were generated to investigate changes in ubiquitination in response to HIV infection and proteasome inhibition (manuscript in preparation). Ubiquitination sites (indistinguishable from NEDD8 or ISG15 sites) were identified using ubiquitin remnant immunoaffinity enrichment and quantitative mass spectrometry in Jurkat and HEK293 cells, both in the presence and absence of HIV infection [107]. The HEK293 cells were treated with interferon, to activate the expression of known antiviral substrates of ubiquitination and proteasomal degradation, while the Jurkat cells were not. Indeed, ubiquitination of APOBEC3C and the cellular receptor CD4 (stably transfected in these cells) were detected in HEK293 experiments responding to the expression of HIV and only in the presence of a proteasome inhibitor (S2 Fig). As interferon treatment also induces expression of the ubiquitin-like protein ISG15, which generates an indistinguishable ubiquitin remnant epitope, we used the data from the Jurkat cells, to examine ubiquitination on kinases, including sensitivity to proteasome inhibition. The Jurkat cells were not interferon treated and thus the sites identified reflect ubiquitination sites that are not confounded by the possibility of being ISG15 modifications. This is true even in the HIV infected Jurkat cells since HIV prevents the activation of interferon stimulated genes by suppressing IRF3 activity [108]. This was verified by checking the abundance of ISG15 and proteins known to be induced by interferon, none of which increased significantly with HIV infection (S1 Dataset). A total of 432 sites were identified on 132 protein kinases in the Jurkat cell studies (Fig 1). Combining this proteomics data with those in the PTMfunc database [60], we found that 201 human protein kinases are ubiquitinated at one or more sites, with a total of 845 ubiquitination sites; 107 of these ubiquitin modifications are proteasome sensitive, and 103 are proteasome insensitive (the proteasome sensitivity of the remaining sites is unknown). All 103 of the proteasome insensitive ubiquitination sites were ubiquitinated even in the absence of HIV infection, indicating that they do not result from nonspecific ubiquitination caused by infection. The frequency of HIV dependence for ubiquitination sites is reported in supporting information (S2 Table). The 103 proteasome insensitive ubiquitination sites are unlikely to be associated with degradation by the proteasome, and may serve another function such as regulation of kinase activity. A full list of all ubiquitination sites is provided in the supporting information (S2 Dataset). We found that most ubiquitin modifications occur on a folded domain (Table 1), in contrast to phosphorylation sites, which are much more likely to be located in unstructured regions (Table 1). This difference is statistically significant (p-value < 2. 2x10-16 with prop. test) and has been observed in studies of global ubiquitination [60,62–64], but the underlying reason is not clear. One possible explanation is that the E3 ligase requires a structured domain for ubiquitin attachment, whereas kinases require the segment around the phosphorylation site to be in an extended conformation. Many of the phosphorylation sites located in unstructured regions are not believed to be functional [60,109,110]. Since most ubiquitin modifications are attached to a structured region of the kinase, we speculate that those that are proteasome insensitive may regulate the kinase by affecting the stability, flexibility, or conformation of that structured region. Fig 2 maps all of the proteasome-sensitive and proteasome-insensitive ubiquitin modifications observed in kinase domains onto the protein kinase A (PKA) structure. Ubiquitin modifications occur on essentially all solvent exposed regions of the kinase domain, but are enriched on the region of the N-lobe opposite the C-helix and near the glycine loop, at the N-terminal end of the activation loop, and on the region of the C-lobe distal to the substrate pocket (Fig 2). These data are consistent with another study that found the glycine-rich loop and the region N-terminal to the activation loop of kinases to be enriched in ubiquitination sites [61]. Both proteasome-sensitive and proteasome-insensitive modifications are concentrated in these functionally important regions, but certain residues are more commonly associated with proteasome-insensitive ubiquitination. In particular there are residues associated with proteasome-insensitive ubiquitination on the N-terminal side of the glycine-rich loop, on the N-terminal side of the activation loop (near the DFG motif), and one on the C-helix itself. Ubiquitination at these sites may cause specific conformational changes that affect kinase activity or interaction with other proteins. We also examined whether kinase ubiquitination sites are conserved across different human kinases. We found that about half of modification sites are common across more than 5 of the 206 ubiquitinated kinases, and that the ubiquitinated residue is usually conserved as a lysine across more than 40 other human kinases (Table 1). As a point of comparison (null hypothesis), if the ubiquitination sites were chosen randomly from all the lysine residues on all the kinases, the percentage occurring at a conserved residue would be 29%. Instead, we observe that 80% of ubiquitin modifications occur at a conserved residue, indicating that these residues are preferred. Wagner et al. also reported that ubiquitinated residues are more conserved than other lysines, possibly because they tend to be in structured regions [62]. We hypothesized that proteasome-insensitive sites would be more conserved than those affected by proteasome degradation; however, we did not find any significant differences between their conservation. This implies that, whether or not ubiquitination is associated with proteasomal degradation, the ubiquitin modification tends preferentially to occur at conserved sites, many of which are critical to kinase structure or function. We also tested whether other PTMs, such as phosphorylation, co-localize with ubiquitination sites, and determined that 11% of ubiquitination sites occurred within 4 residues of a phosphorylation site on the same protein (Table 1). This fraction is higher than the 3% of collocated sites that would be found if ubiquitination sites were chosen randomly from among all lysine residues. Similarly, 3% of ubiquitination sites are located near a residue identified as a functional phosphorylation hot spot (sites where phosphorylation is believed to play a role in function, as determined by the PTMfunc database) [60], rather than the 0. 3% expected from a random distribution (Table 1). No difference was observed between proteasome-sensitive and proteasome-insensitive ubiquitination sites with regard to whether they are located near a phosphorylation site or hot spot. The kinase with the largest number of proteasome-insensitive ubiquitination sites is ZAP-70 (Table 2). For this reason, and because its structure, function and regulation are well studied, we chose ZAP-70 for an initial study of how proteasome-insensitive ubiquitination might affect kinase structure and dynamics. Fig 3 shows the locations of both proteasome-sensitive and proteasome-insensitive ubiquitination sites on ZAP-70. The first site we chose for modeling ZAP-70 ubiquitination was K377 (Fig 3), a site close to the C-helix, because the C-helix position is crucial for kinase activity. We hypothesized that a ubiquitin moiety attached near the C-helix might disrupt its position and shift the kinase conformation away from the active structure. To choose a second ZAP-70 ubiquitination site for modeling, we examined the proteasome-insensitive ubiquitination frequency of the various sites across human kinases (Fig 3). The most common site for proteasome-insensitive ubiquitination is K476 (ZAP-70 numbering), which is ubiquitinated on 6 kinases: ZAP-70, Lck, MAPK9, MAPK8, MAPK10, and CDK5. The K476 ubiquitination site is not as close to the C-helix and active site as K377, and therefore we hypothesized that simulations of K476-ubiquitinated ZAP-70 would have a smaller effect on active site conformation and dynamics, but might still disrupt the active state allosterically. Both K377 and K476 were ubiquitinated on ZAP-70 in cells not infected with HIV, so these modifications are likely relevant to normal ZAP-70 kinase regulation. We conducted simulations of ZAP-70 starting from the active and inactive-state crystal structures with a ubiquitin attached at K377 or K476. K377 is near the N-terminal end of the C-helix (Fig 4 (a) ), and K476 lies on the opposite side of the N-lobe from the C-helix, near the hinge region (Fig 4 (d) ). In all simulations of ubiquitinated ZAP-70, we attached only a single ubiquitin molecule, as this is the minimal possible modification. The proteomics data does not determine the number of ubiquitin molecules or, in the case of polyubiquitination, the attachment residue. Since these modifications were proteasome insensitive, they are most likely to be K63-linked or monoubiquitinated. We acknowledge that K63-linked polyubiquitination may have different effects from monoubiquitination, but we anticipate that there may be commonalities, especially since K63-linked polyubiquitin chains are extended, leaving only the proximal ubiquitin domain to interact with the target protein. As a practical matter, simulating a polyubiquitin attachment would also greatly increase the computational expense. For all ZAP-70 simulations we used a construct that was not phosphorylated at Y493 on the activation loop. The available crystal structures did not include phosphorylation at Y493, and we did not want our simulations to be complicated by the effects of two modifications away from the crystal structure. Also, although ZAP-70 must be phosphorylated at Y493 to be fully active, there is no data to indicate whether ZAP-70 ubiquitination occurs prior to or following phosphorylation, so a ubiquitinated but unphosphorylated ZAP-70 may be relevant in vivo. In our simulations, the ubiquitin domain interacted only transiently with the ZAP-70 kinase domain and sampled many different relative orientations (Fig 4), demonstrating that the starting position of the ubiquitin molecule was not important. The range of orientations sampled by ubiquitin attached at K476 is more restricted than those sampled by ubiquitin attached at K377, probably because K476 is surrounded by other parts of the kinase structure that interfere sterically with the ubiquitin motion. On the kinase domain, we observe that the glycine-rich loop area and the region N-terminal to the activation loop are preferentially ubiquitinated, as did Swaney et al. [61]. Phosphorylation hot spots also cluster to the activation loop and the glycine-rich loop, which are key regions for kinase regulation [60]. Ubiquitination at regions of the kinase structure associated with conformational regulation indicates that ubiquitin attachment could directly modulate kinase activity (whether or not this is the primary role of the ubiquitination). We expect only proteasome insensitive ubiquitination to be responsible for conformational regulation, however both proteasome sensitive and insensitive ubiquitination are enriched in these regions, as well as at conserved lysine residues and near phosphorylation hot spots, as reported previously for other proteins [58,61]. Although proteasome sensitive ubiquitination leads to degradation, these ubiquitin modifications may still be site-specific because proteases degrade substrates by unfolding them starting at the site of ubiquitination [113]. Adding a layer of complexity, the same site can be associated with proteasomal degradation on one kinase, and not on a different kinase or under different cellular conditions. There are also specific sites that are particularly associated with proteasome-insensitive ubiquitination. The conservation of these ubiquitination sites, like ZAP-70 K476, across kinases indicates that ubiquitination plays an important functional role at these sites, which is likely not related to proteasomal degradation. We also found that most ubiquitinated kinases had multiple sites of ubiquitination, as was previously observed in global ubiquitination studies [8]. For some of these kinases, such as ZAP-70, several of these sites are not associated with proteasomal degradation. In fact, these different ubiquitination sites may differentially affect conformation of the same protein. Our MD studies of two different ubiquitinated constructs of the ZAP-70 kinase domain indicate that attachment of a single ubiquitin molecule to a surface lysine modifies the kinase structure and dynamics in a site specific manner. Ubiquitination at K377, near the C-helix, causes ZAP-70 to adopt conformations closer to the inactive state, while ubiquitination at K476, near the hinge region, results in a conformational ensemble more similar to the active state. On the basis of these simulations, we hypothesize that ubiquitination at different sites may differentially modulate kinase activity and/or interactions. Since K476 is a conserved ubiquitination site on kinases, it is also possible that other kinases will undergo similar conformational changes to ZAP-70 when ubiquitinated at this site. Importantly, we recognize that incomplete sampling is a limitation of our MD results. Our simulations starting from the active and inactive ZAP-70 crystal structures are fundamentally the same since the autoinhibitory domains from the inactive crystal structure are removed in our simulations. The differences in ZAP-70 conformation that we observe are a result of different starting structures. The K369-E386 salt bridge that holds the C-helix in the active state is not broken in our active structure simulations and not formed in our inactive structure simulations. If the sampling were increased, we would observe these rare events and the ensembles sampled starting from the active and inactive states would converge. Because we have run many independent short simulations, we also do not sample motions that occur on long timescales. We ran a single long simulation for each of four ZAP-70 constructs (active control, active K377-ubiquitinated, inactive control, and inactive K476-ubiquitinated) to compare to our data combined from many short simulations (S5 and S6 Figs). In these long simulations we also do not observe breaking or forming of the K369-E386 salt bridge, but we do observe more opening of the substrate binding groove in the K476-ubiquitinated simulations starting from the inactive structure, indicating that this may be a motion that occurs on a longer timescale (S5 Fig). Overall we see that the data from a single long simulation has a narrower distribution than the data from many short simulations (S5 and S6 Figs), since the different time points from a single simulation are correlated. We therefore prefer many short simulations to a single long simulation for assessing the effects of ubiquitination on ZAP-70. Given total simulation times on the order of 100 ns, these short simulations provide better sampling of the true conformational distribution for ubiquitinated ZAP-70. Having simulated only the kinase domain, we also cannot directly infer what the effects of ubiquitination are on the full-length ZAP-70. However, our simulations may be a reasonable approximation of the ZAP-70 kinase domain following TCR complex binding and the release of the SH2 domains from the kinase domain. After this release, the SH2 domains only interact with the kinase domain through a flexible linker, and the kinase domain is catalytically active when the SH2 domains are cleaved [81]. Another limitation of this study is the presence of only one ubiquitination molecule in the ZAP-70 simulations. In vivo, a polyubiquitin chain may be attached; however, we cannot distinguish the number of ubiquitin molecules or the linkage residue for polyubiquitination. Given this ambiguity, we chose to simulate monoubiquitinated ZAP-70, which is the simplest possible construct. Our simulations indicate that ubiquitination has the potential to regulate kinase structure and dynamics, leading us to compare its effects to other PTMs like phosphorylation. We found that kinase ubiquitination occurs more frequently on a folded domain than on unstructured linkers or tails, unlike phosphorylation, which occurs more in unstructured regions. Similar trends were reported by Wagner et al. and Hagai et al. in studies of ubiquitination across all protein families [62,63], and Duttler et al. found that ubiquitination of nascent protein chains also is not enriched in regions of intrinsic disorder [64]. Several studies have noted that within these structured domains, ubiquitination tends to occur on loops or flexible regions rather than helices, or, in particular, β-sheets [114–116]. Beltrao et al. also found that ~75% of phosphorylation sites, 40% of acetylation sites, and 45% of ubiquitination sites occur outside PFAM globular domains, after looking at all proteins in the PTMfunc database [60]. Focusing only on kinases, we found that ubiquitination occurs outside folded domains even more rarely, only ~30%. While phosphorylation on linkers and tails allows these regions of kinases to become ordered or disordered, affecting kinase conformation, ubiquitination likely regulates kinases by a different mechanism. Acetylation, like ubiquitination, occurs in structured regions, and in fact one of its roles may be to protect against ubiquitination. Wagner et al. found that 30% of acetylated lysines are also ubiquitinated, and these residues show a lower dependence on proteasome inhibitor [62]. However, while acetylated lysines show an enrichment in basic residues in the flanking sequence, flanking regions of ubiquitination sites are enriched in acidic aspartate residues [115,116]. Both acetylation and ubiquitination sites tend to be near hydrophobic residues, which may indicate that they are more buried and modifications at these sites could result in large conformational changes [116]. Based on the structural preferences of ubiquitination for folded domains of kinases, we expect it to play a different role in kinase regulation than phosphorylation. Both phosphorylation and acetylation primarily affect protein conformation by perturbing the energy landscape electrostatically. For example, simulations of phosphorylated kinases reveal how the addition of a negatively charged phosphate group on the activation loop can stabilize a protein conformation that is aligned for kinase activity, or in some cases directly stabilize the correct orientation of the substrate side chain [65]. The few molecular dynamics studies of lysine acetylation suggest that, analogous to phosphorylation, its primary effect may be driven by electrostatics, i. e. , neutralization of the positive charge on the Lys side chain [70,71]. Like acetylation, ubiquitination neutralizes the positive charge on the lysine residue. However, in our simulations acetylation does not have the same affect as ubiquitination on conformation at either ZAP-70 site. Unlike acetylation and phosphorylation, the effect of ubiquitination on the kinase energy landscape is not primarily electrostatic. Ubiquitin is much larger than the acetyl group and this bulky group may have significant entropic consequences including solvent effects. We observe that ubiquitination at residues that are not critical to kinase activity may allosterically regulate the kinase by changing its overall conformation. In the case of ZAP-70 ubiquitination, this appears to be mediated by the position and structure of the C-helix and possibly the peptide-binding groove. Previous work by McClendon et al. used MD simulations of PKA and the mutual information between residues to reveal a network of correlated motions that could mediate long-distance allosteric coupling in kinases [117]. There are several ways that ubiquitin may be perturbing these correlated motions to alter kinase conformation. Our simulations of ubiquitinated ZAP-70 show that the ubiquitin moiety is attached flexibly to the kinase and samples many orientations (Fig 4). This supports earlier observations of multiple ubiquitin orientations for the ubiquitinated Ras protein from NMR and computational models [24], and indicates that the effects of ubiquitin may be entropic. Molecular dynamics simulations of ubiquitinated proteins by Hagai and Levy have characterized the effect of ubiquitination on thermodynamic stability of a protein [118]. They observe a decrease in protein stability upon ubiquitination that they attribute to entropic stabilization of the unfolded state of the protein. They conclude that destabilization, along with local unfolding at the ubiquitination site, may facilitate protein unfolding prior to degradation. In our K377-ubiquitinated ZAP-70 simulations, we observe partial unfolding of the C-helix, which also moves away from the rest of the N-lobe (Figs 5 & 6). Entropic stabilization of the locally unfolded state may explain this partial C-helix unfolding. We were surprised to observe that ubiquitin attachment at K377 leads to C-helix disruption, while the attachment of the similarly sized Ig domain at K377 (Fig 9) does not significantly disrupt the C-helix (Fig 8). This indicates that the specific sequence or shape of the ubiquitin molecule plays a role in disrupting the ZAP-70 C-helix structure. Fig 10 shows ZAP-70 contacts with both the ubiquitin and Ig-domain molecules when attached at K377. The contacts with ubiquitin are present in much more of the simulated ensemble than the contacts with the Ig domain. Some of the ubiquitin contacts likely result from forming hydrogen bonds, but ion pairs also transiently form between ubiquitin and the kinase. For K377-ubiquitinated ZAP-70, there are contacts between E376 and both R72 and R42 on ubiquitin, especially in the inactive state simulations. Interestingly, R72 and R42 are near the ubiquitin “hydrophobic patch” that is the site of recognition by ubiquitin binding domains [119], although the key hydrophobic patch residue, I44, does not contact the kinase. In the simulations with the Ig domain attached at residue 377, we do not observe any ion pair interactions between the two proteins, nor are there significant hydrophobic contacts formed. This may explain why no conformational change occurs in simulations with the Ig domain attached to ZAP-70. Electrostatic interactions, like what we observe between ZAP-70 and ubiquitin, are also present in the ubiquitinated structure of PCNA. K164-ubiquitinated PCNA adopts a different conformation from SUMOylated PCNA due to specific electrostatic interactions with ubiquitin R42, K48, E51, D58, and D39 [31]. In the SUMOylated structure, however there is electrostatic repulsion between two negatively charged surfaces, resulting in an extended conformation with few contacts [31], as we observe with Ig domain. Like our simulations, the studies of PCNA ubiquitination and SUMOylation only included one ubiquitin or SUMO molecule. Since ubiquitin moieties that are part of a polyubiquitin chain are known to interact with each other, there could be competition for the interactions with ZAP-70 if it were polyubiquitinated. However, the conformational effects might also be exaggerated with the presence of multiple ubiquitin moieties. Specific interactions between ubiquitin and ZAP-70 in the K476-ubiquitinated simulations are shown in Fig 11. In the K476-ubiquitinated simulations, both E415 and D365 on ZAP-70 interact with R74 on ubiquitin, and ZAP-70 R594 contacts ubiquitin E51. The ZAP-70 C-terminal F614 also interacts with hydrophobic ubiquitin residues. However, while K377-ubiquitinated ZAP-70 ionic interactions could directly perturb the C-helix linker, a more complex mechanism is likely at play for K476 ubiquitination, since ubiquitin does not directly contact the C-helix in this construct. Hagai and Levy note that the thermodynamic effects of ubiquitination vary depending on the attachment site [118]. Atomistic MD simulations also show that ubiquitination of Ubc7, an E2 enzyme, decreases the flexibility of certain regions of Ubc7 although there were no specific direct interactions between ubiquitin and the E2 [120]. We observe similar stabilization of the C-helix with ubiquitination at K476, also in the absence of any stable interactions between ubiquitin and ZAP-70. The Ubc7 simulations also show that the strongest effect occurred with K48-linked tetraubiquitination at known sites of degradative ubiquitination [120], indicating that polyubiquitination could have an even larger effect on ZAP-70 conformation than monoubiquitination. The K476 ubiquitination site is farther from the C-helix than K377, yet somehow the presence of ubiquitin shifts the conformational equilibrium toward an active kinase structure with a stable folded C-helix that remains close to the rest of the N-lobe. Coupling between the activation loop and the C-helix can allow communication between the active site and the back of the kinase, where K476 is located [48]; however, we could not find any specific residues between K476 and the active site that were perturbed by the presence of ubiquitin. Another possible mechanism involves the bulky ubiquitin modifying the interaction between the protein and solvent. In the simulations of unmodified ZAP-70 as well as all K377-ubiquitinated simulations, the solvent accessible surface area of the kinase domain remains similar (Fig 12). In contrast, K476-ubiquitination modifies the exposed surface area, decreasing it slightly when starting from the active state, and increasing it significantly when starting from the inactive state (Fig 12). These results imply a net decrease in solvent accessible surface area from the inactive to active state when K476 is ubiquitinated. Thus, ubiquitin attachment may indirectly impact protein structure and dynamics by modulating solvation, which could involve both enthalpic interactions between water and the protein, and increases in water entropy accompanying decreased solvent exposure. In the case of polyubiquitination, the attached moiety would be even larger, possibly increasing or modifying the solvent-mediated effects of ubiquitination on conformation. Our study suggests that ubiquitination may regulate kinases in a manner unrelated to degradation, and that this regulation could result from site-specific changes in conformational dynamics. Simulations of ZAP-70 predict that ubiquitination at K377 or K476 could respectively reduce or increase kinase activity. Ubiquitination may also affect conformation in other protein families, and future work should explore the role of ubiquitin in non-degradative signaling, including further investigation of the biophysical effects of ubiquitination.

Attachment of ubiquitin to another protein is typically used to mark the protein for degradation by the proteasome. However, recent studies show that many proteins are tagged with ubiquitin and not degraded. We hypothesized that ubiquitin can regulate the protein it is attached to by changing its structure and dynamics. We performed proteomics experiments to identify all of the kinase proteins tagged by ubiquitin in a human cell line as well as the site of ubiquitination. We found that kinases are often ubiquitinated in structured regions important for regulation and activity. We then performed molecular dynamics simulations of one kinase, ZAP-70, to see if a ubiquitin tag could affect the kinase structure. We found that ubiquitin does affect the structure of ZAP-70, and the effect depends on where the ubiquitin is attached. At K377, ubiquitin changes the ZAP-70 structure to resemble the inactive state, while ubiquitin attached at K476, on the other side of the protein, has the opposite effect. These simulations indicate that ubiquitin, like other post-translational modifications, may alter the structure and dynamics of proteins in ways that impact activity and function.

Abstract Introduction Methods Results Discussion

phosphorylation crystal structure chemical compounds enzymes condensed matter physics enzymology organic compounds basic amino acids protein structure amino acids crystallography solid state physics proteins protein kinases ubiquitination chemistry molecular biology proteasomes physics biochemistry biochemical simulations protein complexes organic chemistry post-translational modification biology and life sciences lysine physical sciences computational biology macromolecular structure analysis

2016

Non-degradative Ubiquitination of Protein Kinases

West Nile virus (WNV), a mosquito-borne single-stranded RNA flavivirus, can cause significant human morbidity and mortality. Our data show that interleukin-10 (IL-10) is dramatically elevated both in vitro and in vivo following WNV infection. Consistent with an etiologic role of IL-10 in WNV pathogenesis, we find that WNV infection is markedly diminished in IL-10 deficient (IL-10−/−) mice, and pharmacologic blockade of IL-10 signaling by IL-10 neutralizing antibody increases survival of WNV-infected mice. Increased production of antiviral cytokines in IL-10−/− mice is associated with more efficient control of WNV infection. Moreover, CD4+ T cells produce copious amounts of IL-10, and may be an important cellular source of IL-10 during WNV infection in vivo. In conclusion, IL-10 signaling plays a negative role in immunity against WNV infection, and blockade of IL-10 signaling by genetic or pharmacologic means helps to control viral infection, suggesting a novel anti-WNV therapeutic strategy. WNV has caused severe morbidity and mortality in both animals and humans since it first appeared in North America in 1999. WNV is maintained in an enzootic cycle between mosquitoes and birds; but humans, horses and other mammals can be inadvertently infected by mosquitoes that carry the virus. Although most human infections are asymptomatic, the elderly and individuals with a compromised immune system are particularly susceptible to life-threatening neurologic disease [1]. Viral pathogenesis is not completely understood, and a vaccine or specific therapy has not yet been approved for use in humans [2]. Experimental mice infected with WNV develop a systemic infection that results in encephalitis and death, thereby mimicking human neuroinvasive disease and providing a valuable model system within which to study viral pathogenesis and immunity. Mammalian T helper 1 (Th1) cytokine responses are essential for eradicating invading intracellular pathogens. However, if these responses are too strong, bystander effects may damage the host. Interleukin-10 (IL-10), a pleiotropic cytokine, plays a crucial immunosuppressive role during excessive Th1 responses, and can thereby protect the host from potentially damaging immunopathology [3]–[5]. However, excessive IL-10 production suppresses host immune responses and can inadvertently facilitate the ability of intracellular pathogens to escape host innate immune defenses. Epstein-Barr virus (EBV) takes advantage of the immunomodulatory effects of IL-10 by expressing a viral homolog of IL-10 during infections [6]. In addition, recent studies showed that lymphocytic choriomeningitis virus (LCMV) clone 13 elicits high levels of IL-10 production from the infected host, thereby resulting in exhaustion of virus-specific T cells and viral persistence [7]–[9]. Interestingly, neutralization of IL-10 during persistent viral infection leads to recovery of virus-specific T cell responses and reduction of viral load, suggesting a potential therapy to restore T cell function and prevent viral persistence [7], [8]. Type I Interferon (IFN) and Th1 cytokines, including IFN-γ, provide immediate defense against WNV replication and dissemination [10]–[14]. Cellular immunity, including CD4+ [15], [16], CD8+ T cell [17]–[21] and γδ T cell responses [22]–[24], also participate in host recovery from WNV infection. IL-10 [3] and other regulatory cytokines, such as IL-4 [25]–[28] and IL-13 [29], [30] have been documented to suppress Th1 responses. The role of these regulatory cytokines in modulating Th1 responses during WNV pathogenesis is, however, not completely understood. In the current study, we investigated the role of IL-10 in WNV infection, and our results suggest a promising therapeutic strategy to combat WNV infection through blockade of IL-10 signaling. To study host immune responses to WNV infection, we performed cytokine expression polymerase chain reaction (PCR) arrays in which cytokine expression profiles were analyzed in thioglycollate-elicited peritoneal macrophages from C57BL/6 mice after WNV infection (multiplicity of infection [MOI] = 1). We observed that IL-10 expression was up-regulated 4- fold at 5 h and 90-fold at 24 h post-infection (p. i.), while expression of IL-4, IL-13 and another immunosuppressive cytokine, TGF-β, were not significantly altered at both time points (data not shown). To confirm this finding, we measured IL-10 mRNA by reverse transcription quantitative real-time PCR (Q-PCR) and quantified secreted IL-10 protein in cell culture media from primary cultures of WNV-infected macrophages by ELISA. Both Q-PCR and ELISA results confirmed that IL-10 mRNA and IL-10 protein were markedly up-regulated after WNV infection in vitro (Figure 1 A and B). To further evaluate expression kinetics of IL-10 in vivo, we intraperitoneally (i. p.) infected a group of C57BL/6 mice with 2,000 plaque forming units (pfu) of WNV corresponding to a dose at which approximately 50% of wild-type (C57BL/6J) animals survive (LD50) [31]–[33] and measured time-dependent IL-10 secretion in plasma by ELISA (Figure 1 C). Results showed that IL-10 production began to increase in plasma as early as 6 h p. i. and reached approximately 18-fold at day 5 (120 h) compared to baseline (before infection, 0 h). These data raise the possibility that IL-10 may play a role in modulating immune responses during WNV infection. To evaluate a putative role for IL-10 in WNV pathogenesis, we challenged IL-10−/− mice with an LD50 dose of WNV via an i. p. route of administration [33], [34]. Mice were monitored twice daily for morbidity and mortality for three weeks. Q-PCR designed to measure WNV envelope gene (WNVE) revealed 2–3 fold reduced WNV RNA in blood samples from IL-10−/− mice compared to control mice at days 1 and 3 p. i. (p<0. 05, Figure 2 A). Plaque formation assays were also performed to measure infectious viral particles in plasma samples. Consistent with Q-PCR data, plaque formation assay results showed that IL-10−/− mice had a similar magnitude of lower viral burden compared with control mice (data not shown). In selected experiments, we randomly sacrificed half the number of mice to collect spleen samples at day 3 and to collect brains at day 7, while the remaining mice were observed for survival analysis. Total RNA extracted from spleen and brain hemispheres was used for Q-PCR analysis of WNV burden. Contralateral brain hemispheres were fixed in 4% paraformaldehyde (PFA) for histological analysis. Markedly less viral RNA was present in the spleens and brains of IL-10−/− mice vs. wild-type mice (p<0. 05, Figure 2 B). Confocal microscopy results were consistent with brain Q-PCR data in that IL-10−/− mice had little evidence of virus in the olfactory bulb (OB), cerebral cortex, striatum, cerebellum and brainstem, whereas WNV was readily detected in similar brain regions of control mice on day 7 p. i. (Figure 2 C and Figure S1). Co-immunostaining with the neuronal marker microtubule-associated protein 2 (MAP2) revealed that most of the infected cells were neurons (Figure 2 C and Figure S1). In addition, a higher number of CD45+ leukocytes infiltrated into brains of wild-type mice as compared to IL-10−/− mice (Figure 2 C). Similar results were observed when immunostaining for CD11b, a marker for macrophages/microglia (Figure S1). The survival ratio of IL-10−/− mice (70. 0%) was significantly higher than that of wild-type mice (33. 3%) (p<0. 01, Figure 2 D). As virus inoculation routes may alter host immune responses, we also challenged IL-10−/− mice with WNV via footpad inoculation (100 pfu) route and performed survival analysis. Similar to intraperitoneal inoculation, IL-10−/− mice also had increased survival after footpad infection (Figure 2 E). Collectively, viral burden and survival analyses demonstrate that mice deficient in IL-10 have increased resistance to WNV infection, suggesting that IL-10 signaling facilitates WNV pathogenesis. We further tested this hypothesis by taking a pharmacological approach to interrupt IL-10 signaling. Specifically, we systemically administered anti-IL-10 receptor monoclonal antibody (aIL-10r mAb) to block IL-10 signaling. We injected aIL-10r mAb or isotype-matched control IgG (i. p.) on day 1 prior to WNV infection. Mice were challenged with 2,000 pfu (i. p.) or 100 pfu (footpad inoculation) of WNV. Blood samples were collected from the mice infected intraperitoneally on day 3 (p. i.) for WNV burden Q-PCR measurement. Q-PCR results showed a trend toward reduced viral burden in blood of mice treated with aIL-10r mAb prior to intraperitoneal WNV infection (Figure 3 A, p = 0. 07). Survival analysis indicated that mice treated with aIL-10r mAb had increased survival compared to mice injected with isotype-matched IgG (p<0. 05, Figure 3 B and C). These data further confirm that IL-10 signaling facilitates WNV infection and promotion of lethal encephalitis. We have previously shown that macrophages can be infected with WNV, and are likely an important target cell in WNV pathogenesis [33], [34]. To dissect the mechanisms underlying how IL-10 signaling facilitates WNV infection, we measured cytokine expression in macrophages from IL-10−/− and wild-type mice upon virus infection in vitro. We infected thioglycollate-elicited macrophages obtained from peritoneal lavage with WNV, and quantified cytokines by ELISA or Q-PCR at selected time points. Consistent with a stronger anti-viral response, results showed that WNV-infected macrophages from IL-10−/− mice produced more TNF-α, IFN-α, IFN-β and IL-12/23 p40 than wild-type cells (Figure 4 A–D). The spleen is a WNV target organ, and splenic WNV infection is thought to be involved in host immune defense against infection [15], [35]. We therefore assessed cytokine production in splenic cells isolated from naïve and WNV-infected mice. We infected naïve splenocytes with WNV (MOI = 0. 5) and measured production of IL-12/23 p40 and IFN-γ by ELISA at selected time points. Splenocytes from IL-10−/− mice produced elevated IL-12/23 p40 compared to splenocytes from wild-type mice (Figure 4 E), while IFN-γ production was not detected by ELISA in either group. We also quantified cytokine production from splenocytes isolated on day 3 p. i. . Specifically, we stimulated splenocytes with WNV MHC class I epitope peptide NS4b [36], [37] ex vivo for 24 or 48 h. ELISA results showed that TNF-α and IL-12/23 p40 were significantly elevated in splenocytes from IL-10−/− mice (Figure 4 F and G), while IFN-γ production was undetectable. In addition, we evaluated cytokine secretion in peripheral blood of IL-10−/− compared to wild-type mice during WNV infection (LD50, i. p.) by ELISA or Q-PCR. ELISA results showed that IL-12/23 p40 and TNF-α production in plasma from IL-10−/− mice was higher than wild-type mice plasma samples on day 2 after infection (Figure 4 H and I). Because IFN protein production in plasma was below the detection level of ELISA, we measured IFN-γ mRNA levels in blood by Q-PCR. The Q-PCR data show that expression of IFN-γ in blood of IL-10−/− mice trended toward significantly lower than wild-type mice in all the selected time points, and reached statistical significance on day 3 p. i. (Figure 4 J). Collectively, these data suggest that genetic blockade of IL-10 signaling results in a significant elevation of type I IFN and proinflammatory cytokines in macrophages, splenic cells and peripheral blood during WNV infection. Increased type I IFN and innate immune proinflammatory cytokines might be expected to contribute to efficient control of WNV infection after IL-10 signaling blockade. To assess this, we evaluated WNV replication in macrophages from IL-10−/− and wild-type mice in vitro by Q-PCR and plaque formation assay. Both assays indicated lower viral burden in macrophages from IL-10−/− vs. wild-type mice, showing increased resistance of IL-10−/− macrophages to direct WNV infection (Figure 5 A and B). To further assess whether type I IFN plays an antiviral role in macrophages from IL-10−/− mice, we incubated macrophages with anti-IFN-α/β receptor neutralizing mAb and measured viral replication by Q-PCR after infection with WNV (MOI = 1) (Figure 5 C). Q-PCR results showed that neutralizing antibody treatment partially restored WNV replication in macrophages from IL-10−/− mice. We next measured type I IFN activity in the media of naïve splenic cells that were infected with WNV (MOI = 0. 5) using a virus protection assay (murine encephalomyocarditis virus [EMCV] bioassay). These results suggested that splenic cells from IL-10−/− mice secrete more type I IFN compared to wild-type mice after WNV infection in vitro (Figure 5 D). The results of the antibody treatment and bioassay experiments indicated that type I IFN production limits viral replication in vitro in WNV-infected IL-10−/− immune cells in vitro. To determine if something similar may occur in vivo, we measured type I IFN activity in plasma from wild-type and IL-10−/− mice by EMCV bioassay. Results showed similar type I IFN activity at days 1 and 5 (p>0. 1), while at day 3, plasma from IL-10−/− mice had significantly reduced activity compared with plasma from WT mice (p<0. 01, Figure 5 E), which may due to a lower virus burden in the periphery of IL-10−/− mice. CD4+ and CD8+ T cells play an important role in recovery of WNV infection in the mouse model [15], [17]–[21]. In addition, host IL-10 expression may result in CD4+ and CD8+ T cell exhaustion in chronic viral infections [7]–[9]. We investigated whether CD4+ and CD8+ T cells contribute to increased resistance to WNV infection noted in IL-10−/− mice. On day 7 p. i. , splenocytes were isolated from IL-10−/− and wild-type control mice and stimulated with PMA or peptide NS4b for 6 h. Total splenic cells were stained for CD4 and CD8 surface markers and for intracellular IFN-γ. Flow cytometry analysis showed that the frequencies of CD8+IFN-γ+ and CD4+IFN-γ+ cells were significantly reduced in IL-10−/− mice (Figure 6 A–C). These results are consistent with IFN-γ in blood of IL-10−/− mice, which trended toward reduced compared with wild-type mice (Figure 4 J). These data suggest that IFN-γ does not play a dominant role in mitigation of WNV infection in IL-10−/− mice. To exclude the possibility that T cells from IL-10−/− mice may not be as responsive as those that from wild-type mice, we stimulated spenocytes from naïve mice with PMA for 6 h and stained for CD4 and CD8 surface markers and for intracellular IFN-γ. The results show that CD4+ and CD8+ T cells from naïve IL-10−/− mice produce more INF-γ than those from wild-type mice (Figure 6 D), suggesting that response to PMA stimulation is intact in T cells from IL-10−/− mice. Collectively, these data suggest that elevated WNV-induced antiviral cytokines in IL-10−/− immune cells but not IFN-γ production in CD4+ and CD8+ T cells likely serve to limit viral replication in IL-10−/− mice. As blockade of IL-10 signaling via genetic means leads to enhanced anti-viral immunity mitigating WNV infection, we wanted to assess the effect of pharmacological IL-10 blockade on lethal WNV encephalitis. We used a monoclonal antibody directed against IL-10 (anti-IL-10 mAb) to neutralize IL-10 in vivo following WNV infection. Because a treatment course for WNV disease would begin after exposure to the virus, we first infected mice with WNV (LD50, i. p.) and followed by treatment with two daily i. p. doses of anti-IL-10 mAb or isotype-matched IgG control antibody. Antibodies were injected starting on day 2 or day 4 p. i. Survival curves indicate that anti-IL-10 antibody significantly increases survival of WNV-infected mice when administered as late as day 2 p. i. (p<0. 05, Figure 7). Yet, this effect is abolished when the neutralizing antibody is administrated on day 4 p. i. , as survival rates are similar between the two groups (data not shown). IL-10 is provided by various cellular sources in different disease models [38]–[40]. A better understanding of the cellular sources of IL-10 and mechanisms of induction is required to define an effective immunotherapeutic strategy aimed at IL-10. To identify the cellular source of IL-10 in vivo during WNV infection, we used IL-10-GFP reporter knock-in tiger mice, in which an internal ribosome entry site GFP element was inserted into the 3′ region of the IL-10 gene. During characterization of these mice, GFP and IL-10 expression were found to be closely correlated in all cell populations tested [40], thus allowing the identification of IL-10-producing cells ex vivo without subsequent manipulation. We examined splenocytes isolated from tiger mice for expression of GFP and various cell-surface markers on selected days after WNV infection. No significant GFP production from macrophages (CD11b+), dendritic cells (CD11c+), B cells (CD19+), or non-CD4+ T cells (CD3+CD4−) was observed at any selected time points (Figure 8 A). However, a significant amount of GFP was produced by CD4+ T cells (CD3+CD4+), which peaked on day 5 post infection (Figure 8 A). This time course of IL-10 induction in CD3+CD4+ cells is consistent with IL-10 production in plasma of wild-type mice during WNV infection (Figure 1 C). These data suggest that CD4+ T cells secrete copious amounts of IL-10 during WNV infection in mice. To further confirm this conclusion, we reconstituted RAG1−/− mice that lack T and B cells with naïve CD4+ T cells from wild-type or IL-10−/− mice or PBS (cell-free control) on day 1 prior to WNV infection (LD50, i. p.). We subsequently measured IL-10 levels in plasma by ELISA at serial time points after infection. Results showed that RAG1−/− mice reconstituted with CD4+ T cells from wild-type mice produced significantly higher amounts of IL-10 than mice reconstituted with PBS or CD4+ T cells from IL-10−/− mice, providing definitive evidence that CD4+ T cells produce a significant amount IL-10 during WNV infection in mice (Figure 8 B). It is worth noting that IL-10 was also detected in plasma of the RAG1−/− mice reconstituted with PBS or CD4+ T cells from IL-10−/− mice, albeit at lower levels (Figure 8 B), suggesting that cells other than CD4+ T cells (such as macrophages) play a certain but subdominant role in secreting IL-10 during WNV infection in vivo. Therefore, these data confirm that CD4+ T cells produce a significant amount, but not all, of the IL-10 in the murine WNV infection paradigm. Studies on host immune responses to WNV infection will not only result in a better understanding of viral pathogenesis, but may also lead to novel immunotherapeutic strategies against viral infection. IL-10 plays an immunosuppressive role through interaction between antigen presenting cells and T cells [3], [41], [42]. IL-10 was also found to induce T cell exhaustion during chronic viral infections [7], [8]. Systemic IL-10 production is increased in various human chronic viral infections, such as hepatitis C virus [43], [44], human immunodeficiency virus [45], [46] and hepatitis B virus [47]. Studies demonstrate that the IL-10/IL-10R pathway plays a key role in the establishment of chronic LCMV persistent infection [7], [8]. Blockade of IL-10 signaling converts a chronic LCMV infection into a rapidly controlled acute viral infection and prevents functional exhaustion of memory T cells [7], [8]. Further, a recent study suggests that IL-10 blockade facilitates DNA-vaccine induced T cell response in an LCMV mouse model [48]. Interestingly, IL-10 production correlates directly with plasma HIV viral load and inversely with CD4 cell counts [49]. In this study, we observed that IL-10 expression was dramatically up-regulated in vitro and in vivo in murine plasma during WNV infection, raising the possibility that IL-10 plays a significant role in modulating host immune responses to WNV infection. Through evaluation of WNV infection in IL-10−/− mice and in wild-type mice that were administered anti-IL-10r mAb or IL-10 neutralizing antibody to temporally block IL-10 signaling, we found that IL-10 signaling facilitates WNV infectivity and lethal encephalitis. WNV is a neuroinvasive virus that develops low viremia in mice by replicating in peripheral organs including spleen, and then infecting neurons causing meningitis and lethal encephalitis in mice [31], [34], [50]. WNV is most often cleared from blood and peripheral organs between one and two weeks p. i. . In IL-10−/− mice, WNV load was significantly reduced in the blood and spleen, suggesting that viral replication was efficiently inhibited in the periphery. Consequently, the likelihood that the virus would infect neurons was greatly reduced. We also performed direct intracranial injection of WNV into wild-type vs. IL-10−/− mice and evaluated survival. We did not detect a difference using this route of infection (data not shown), suggesting that IL-10 deficiency promotes WNV clearance at the level of the periphery as opposed to the brain. We next dissected the mechanisms by which IL-10 signaling facilitates WNV infectivity. IL-10 suppresses Th1 responses that are essential to restrict intracellular pathogens [3], [41], [42], and previous studies suggested that IFN and proinflammatory cytokine responses are important in anti-WNV immunity [11], [14], [51]–[53]. Consistent with this notion, we found that these antiviral cytokines were increased in IL-10−/− macrophages, splenocytes and mouse blood. Macrophages not only produce type I IFN, proinflammatory cytokines and IL-10 upon WNV infection, but they can also be directly infected by WNV. Interestingly, we noted MOI-dependent increases in IL-10 production from wild-type macrophages (MOIs from 0. 01 to 1. 0), supporting the notion that macrophage IL-10 production is directly correlated with viral abundance. We confirmed that increased abundance of type I IFN and proinflammatory cytokines were associated with inhibition of WNV replication in IL-10−/− macrophages, and this inhibitory effect was significantly reversed by antagonistic IFN-α/β receptor antibodies. In addition, EMCV bioassay results demonstrated that splenic cells isolated from IL-10−/− mice produced more type I IFN upon WNV infection. These data are in line with the notion that type I IFN plays an antiviral role in IL-10−/− immune cells in vitro. However, a lower type I IFN activity level in plasma from IL-10−/− mice compared with wild-type mice at day 3 p. i. was detected by EMCV bioassay (Figure 5 E). This may be resolved by the following reasoning. It is possible that higher viral burden in wild-type mice (Figure 2 A) stimulates wild-type compared with IL-10−/− immune cells to produce more type I IFN. This response may be especially evident in the early phase of infection, because at this stage production of regulatory cytokines (such as IL-10) occurs at relatively low levels (Figure 1 C). Studies on infections of bacteria, protozoa and some viruses suggest that expression of IFN-γ is up-regulated and participates in controlling infection in the absence of IL-10 signaling [42], [54], [55]. The ability of IL-10 to down-regulate IFN-γ production is a consequence of its ability to inhibit accessory cell function, including production of cytokines (such as TNF-α and IL-12) and expression of costimulatory molecules that are necessary for optimal stimulation of T cells [56]–[59]. In our WNV model, we found that IFN-γ was not inducible in IL-10−/− immune cells or mice during WNV infection. Studies in chronic infection models have shown that host IL-10 production results in exhaustion of virus-specific T cells and reduction of IFN-γ expression [7]–[9]. We next examined IFN-γ expression on T cells by flow cytometry analysis during WNV infection in IL-10−/− mice on day 7 p. i. . Results showed that the frequency of IFN-γ-producing CD4+ and CD8+ T cells was remarkably lower in IL-10−/− than in wild-type mice (Figure 6 A–C). Consistently, IFN-γ expression in blood from IL-10−/− mice trended toward reduced from day 1 to day 3 p. i. (Figure 4 J). In IL-10−/− mice, viral replication was significantly inhibited following infection (Figure 2 A and B), indicating that IFN-γ is dispensable for controlling WNV infection in the model of IL-10−/− mice, particularly during the early stage of infection. In a previous study, TNF-α was shown to participate WNV control in mouse models [60], and we observed that TNF-α expression was up-regulated in IL-10−/− mice after WNV infection. It is worth noting that TNF-α has also been shown to increase permeability of the blood-brain-barrier (BBB) during WNV infection [34]. There are a number of variables, including the roles of TNF-α in anti-viral immunity independent of a role in WNV BBB permeabilization, the relative importance of WNV BBB permeability at early vs. later stages of infection, and whether WNV entry into the brain via the BBB (as opposed to indirect brain entry via brain penetration of infected leukocytes [34]) is a critical pathway for brain infection and subsequent encephalitis, that are dynamically at play during WNV infection [32], [34], [61]–[63]. These variables make direct determination of the contribution of WNV-induced TNF-α expression in BBB permeability complex and difficult to determine in our IL-10−/− WNV infection system. We did not examine BBB permeability during WNV infection in the present study because viral replication is significantly inhibited in the periphery of IL-10−/− mice and the likelihood that WNV would infect neurons is therefore correspondingly low. Thus, it does not appear to be necessary to invoke a BBB permeability hypothesis to explain our current results. Upon viral infection, host innate immune sensors, including Toll-like receptors, double-stranded RNA-activated protein kinase (PKR), retinoic-acid-inducible protein I (RIG-I) and melanoma-differentiation-associated gene 5 (MDA-5), can be activated to produce type I IFNs and cytokines that limit viral replication and dissemination. In the majority of human infections, WNV is controlled and cleared by anti-viral immune responses before the virus invades the central nervous system. However, for the elderly and immunocompromised patients, for whom innate immune responses are too weak or impaired to efficiently control WNV infection, life-threatening neuroinvasive disease may result. In the present study, we found that IL-10 inhibits host innate immune responses during WNV infection. Importantly, blockade of IL-10 signaling through administration of neutralizing IL-10 antibody was able to partially protect WNV-infected mice from death, suggesting a promising immunotherapeutic strategy. The therapeutic effect of neutralizing IL-10 antibody was significant when administered as late as day 2 p. i. , and it is possible that its efficacy may be extended by using higher doses of IL-10 signaling blocking antibodies or small molecule compounds that inhibit the IL-10 signaling pathway. Blocking IL-10 signaling to stimulate host immune responses rather than directly targeting viral genes has the added advantage that it is less likely to lead to the emergence of mutant virus. This anti-WNV strategy may also be combined with other currently available supportive measures, such as IFN therapy, in order to reap better clinical outcomes. Identification of a major cellular source of IL-10 during WNV infection in mice would, in principle, provide a more accurate target for modulation of IL-10 signaling as a therapeutic strategy. We found that CD4+ T cells produce a significant amount of IL-10 and may be a major cellular source of IL-10 in vivo during WNV infection. Our data also indicate that macrophages produce IL-10 when exposed to WNV ex vivo (Figure 1 A and B), which is consistent with in vivo results showing that CD11b+ macrophages produce low levels of IL-10. It is becoming evident that naturally occurring or adaptive CD4+ regulatory T cells produce copious amounts of IL-10 in different disease models [38], [39], [42]. Further studies are warranted to identify which CD4+ T cell subset is a major IL-10 producing cell during WNV infection in vivo. However, it should be noted that such a therapeutic strategy would need to be carefully titrated, as IL-10 signaling can also be beneficial by limiting extensive immunopathological damage resulting from uncontrolled Th1 or lymphokine-responsive T effector cell responses. Thus, while modulation of IL-10 signaling may be a means to enhance anti-pathogen immunity or limit immunopathology, this must be approached with caution to avoid tipping the balance too far in either direction. It is worth noting that we did not observe any overt immunopathology in IL-10−/− mice during WNV infection, and surviving IL-10−/− mice appeared active and healthy at the end of our infection experiments. It may be feasible to temporally block or reduce IL-10 signaling in the early phase of acute WNV infection to stimulate the patient' s anti-viral immune responses without causing severe immunopathological side-effects. Of course, further pre-clinical studies focused on blocking IL-10 signaling as a therapeutic strategy against WNV infection would need to be explored both in terms of safety and efficacy before moving forward into clinical trials. C57BL/6, IL-10−/−and RAG1−/− mice were purchased from the Jackson Laboratory (Bar Harbor, ME). IL-10-GFP knock-in tiger mice were generated in the laboratory of R. A. Flavell as previously described [40], and were bred as heterozygotes. We performed all experiments on 7–8 week-old mice, and mouse groups were rigorously age- and sex-matched for each infection experiment. West Nile viral isolate 2471 [64] used in these studies was propagated one time in Vero cells and titered in a Vero cell plaque-formation assay. We inoculated mice i. p. with LD50 (2,000 pfu) of WNV isolate 2741 in 100 µl of PBS with 5% gelatin. We performed footpad inoculation by injecting 100 pfu of WNV diluted in 50 µl PBS with 1% heat-inactivated fetal bovine serum (FBS) and conducted intracranial inoculation by injecting 20 pfu of WNV in 20 µl PBS with 1% FBS according to a previous report [32]. We observed mice for up to 21 days after infection and checked them twice daily for morbidity (including lethargy, anorexia, and difficulty ambulating) and mortality. All animal experimental protocols were approved by the Yale University Institutional Animal Care & Use Committee. All animal infection experiments were performed in a Bio-safety Level 3 animal facility, according to the regulations of Yale University. One million peritoneal thioglycollate-elicited macrophages were prepared from wild-type or IL-10−/− mice and challenged with WNV isolate 2741 (MOI = 1) in 6-well culture plates in 2 ml of medium and incubated in a 37°C, 5% CO2 incubator. Macrophages were collected at 24 h after washes with PBS for total RNA extraction by Qiagen RNeasy Mini kit (Valencia, CA). The RT2Profiler™ PCR Array kits were purchased from SuperArray Bioscience Corporation (Frederick, MD), and all assay procedures were conducted in accordance with the kit user manual. Briefly, one microgram of total RNA from IL-10−/− or wild-type macrophages was transcribed into first strand cDNA and loaded into 96-well PCR array plates with 25 µL Q-PCR master mix per well. After performing Q-PCR, resulting threshold cycle values (Ct) for all genes were exported into the company-provided data analysis template Excel files for comparison of gene expression between IL-10−/− and wild-type macrophages. The Q-PCR array experiments were repeated three times with similar results. Total RNA was extracted from macrophages, blood, spleen and brain tissue using the RNeasy kit (Qiagen, Valencia, CA). RNA was used to synthesize cDNA by the AffinityScript Multi Temperature cDNA synthesis kit (Stratagene, La Jolla, CA). WNV-specific RNA was quantified using a Q-PCR technique as we previously reported [65]. Cytokine gene and viral gene copy number were expressed as a ratio to cellular β-actin cDNA copies measured by Q-PCR. The ratio of the amount of amplified gene compared with the amount of β-actin cDNA represented the relative levels in each sample. Plaque-formation assay procedures were carried out according to our previous report [66]. Mice were sacrificed under isofluorane anesthesia and transcardially perfused with 60 ml of ice-cold PBS. Brains were rapidly isolated and immersion fixed in 4% PFA in PBS overnight at 4°C, and cryoprotected in a graded series of sucrose (10%, 20%, and 30%, each overnight at 4°C). Para-median sagittal sections were cut at 25 µm using a cryostat. Tissue sections were PAP pen (Zymed Laboratories, South San Francisco, CA) applied and pre-blocked in serum-free protein block (Dakocytomation, Denmark) for 30 min at ambient temperature. Sections were then labeled overnight at 4°C with various combinations of primary antibodies against CD11b (1∶200, Serotec, Raleigh, NC), CD45 (1∶200, Serotec), MAP2 (1∶500, Invitrogen) and/or WNV antigen (from J. F. Anderson; 1∶250). After three rinses in PBS, sections were labeled with appropriate secondary antibodies conjugated with AlexaFluor488, -594, and/or -647 for 1 h at ambient temperature. Following three additional rinses in PBS sections were mounted in fluorescence mounting medium (ProLong Gold, Invitrogen). Images were acquired in independent channels using a Nikon C1 Eclipse laser scanning confocal microscope. Type I IFN activity in medium or plasma was detected and quantified using an EMCV bioassay of L929 cells as described [67]. Briefly, serially diluted samples were applied to monolayers of L929 cells (2×106 cells/well) in DMEM containing 10% FBS in 96-well plates. Following incubation for 14 h at 37°C, cells were infected with EMCV. Cells were inspected for cytopathic effects, and 7 h p. i. , IFN-mediated protection was assayed using a CellTiter 96 aqueous cell proliferation assay (Promega, Madison, WI). The percentage of protected cells was calculated as described [68], according to the following formula: (optical density at 492 nm [OD492] of plasma or supernatant-treated EMCV-infected cells/ (OD492 of non-EMCV-infected cells − OD492 of EMCV-infected cells) /OD492 of non-EMCV-infected cells) ×100%). Mouse peritoneal macrophages were elicited by i. p. injection with 1 ml of thioglycollate medium (BD BBL™, Franklin Lakes, NJ) and macrophages were collected 3 days later and cultured in DMEM containing 10% FBS and antibiotics. Splenocytes were prepared from naïve or WNV-infected mice at selected time points and cultured in RPMI 1640 medium and infected by WNV (MOI = 0. 5) or stimulated with 50 ng/ml of PMA (Sigma-Aldrich, St. Louis, MO) or WNV NS4b peptide (SSVWNTATTAI, 1. 0 µg/ml). Cell supernatants were collected for cytokine ELISA analysis using kits purchased from R&D Systems (Minneapolis, MN). Seven-week-old female C57BL/6 mice were i. p. injected with 250 µg of mouse anti-IL10r (clone 1B1. 3a) or isotype-matched IgG1 (clone R3-34) antibodies in 100 µl of PBS one day prior to WNV infection. Anti-IL10r mAb and isotype control antibodies were purchased from BD Bioscience Pharmingen (San Diego, CA). In anti-IL-10 mAb immunotherapeutic experiments, two doses of 200 µg antibody (clone JES052A5) or isotype control IgG1 (clone 43414, R&D Systems) in 100 µl PBS were injected i. p. daily, with the first dose starting on either day 2 or 4 post-WNV infection. To measure IFN-γ production, splenocytes from WNV-infected mice were isolated and cultured under one of two conditions. Half of the samples were incubated at 3×106 cells/tube at room temperature with no exogenous stimulation for 4 h and Golgi-plug (BD Bioscience Pharmingen, 1. 0 µg/ml) was added for the final 2 h (BD Bioscience Pharmingen). We used conditions previously described as optimal for spontaneous accumulation of cytokines after removal from the host [69]. The remaining samples were stimulated at 3×106 cells/tube with 50 ng/ml of PMA (Sigma-Aldrich) or with 1. 0 µg/ml of WNV NS4b peptide and 500 ng/ml of ionomycin (Sigma-Aldrich) for 4 h at 37°C and Golgi-plug (1. 0 µg/ml) was added during the final 2 h. Cells were then harvested and stained with the following cell surface antibodies (all were obtained from BD Biosciences Pharmingen): CD3 (clone 500A2), CD4 (clone RM4-5), CD8 (clone 53-6. 7), CD11b (M1/70), or CD19 (1D3) and fixed in 4% PFA. Cells were then permeabilized with 0. 5% saponin before adding FITC-conjugated anti-IFN-γ mAb (BD Bioscience Pharmingen, CA) or isotype-matched control FITC-conjugated rat IgG1 for intracellular staining. Splenic single cell suspensions were prepared from naïve C57BL/6 or IL-10−/− mice and CD4+ T cells were positively selected using anti-mouse CD4 (L3T4) -conjugated microbeads (Miltenyi Biotec, Auburn, CA) according to the manufacturer' s instructions and the purity of eluted cells was assessed by flow cytometry analysis. RAG1−/− mice received 1×107 CD4+ T cells (>98% purity) in 200 µl of PBS per mouse or PBS alone as control. We calculated standard errors of the means (SEM) and analyzed data by non-paired Student' s t-test for single mean comparisons. We performed survival curve comparisons using the log-rank test (GraphPad Prism 4. 0, La Jolla, CA). p<0. 05 was considered as statistically significant for all analyses. IL-10 (MGI 96537, PF 00726); WNVE (AF 206518); IFN-α (M 68944); IFN-β (NM 010510); IFN-γ (NM 008337); TNF-α (MGI 104798); IL-12/23 p40 (MGI 96540); Rag1 (MGI 97848); Actb (MGI 87904).

West Nile virus (WNV), a mosquito-transmitted RNA virus, is a worldwide cause of severe human and animal infection. Mammalian host immune responses to WNV infection are not completely understood and a vaccine or specific therapy is unavailable for use in humans. In the present study, we investigated the putative regulatory role of interleukin-10 (IL-10) during WNV infection in mice. We found that IL-10 signaling facilitates WNV infection and suppresses antiviral cytokine production in response to viral infection. Interestingly, blockade of IL-10 signaling by IL-10 neutralizing antibody increases survival of WNV-infected mice, suggesting a potentially novel therapeutic strategy to combat WNV infection. In addition, we found that CD4+ T cells produce a significant amount of IL-10 during WNV infection, providing a more accurate cellular target for IL-10 signaling inhibition. IL-10 also plays a critical role in suppression of excessive inflammation and immunopathology caused by autoimmune diseases or host immune system responses to infections; therefore, safety and efficacy of IL-10 signaling blockade as a therapeutic strategy against WNV infection deserves consideration.

Abstract Introduction Results Discussion Materials and Methods

virology/effects of virus infection on host gene expression immunology/immunity to infections virology/new therapies, including antivirals and immunotherapy virology/host antiviral responses

2009

IL-10 Signaling Blockade Controls Murine West Nile Virus Infection

The collective dynamics of multicellular systems arise from the interplay of a few fundamental elements: growth, division and apoptosis of single cells; their mechanical and adhesive interactions with neighboring cells and the extracellular matrix; and the tendency of polarized cells to move. Micropatterned substrates are increasingly used to dissect the relative roles of these fundamental processes and to control the resulting dynamics. Here we show that a unifying computational framework based on the cellular Potts model can describe the experimentally observed cell dynamics over all relevant length scales. For single cells, the model correctly predicts the statistical distribution of the orientation of the cell division axis as well as the final organisation of the two daughters on a large range of micropatterns, including those situations in which a stable configuration is not achieved and rotation ensues. Large ensembles migrating in heterogeneous environments form non-adhesive regions of inward-curved arcs like in epithelial bridge formation. Collective migration leads to swirl formation with variations in cell area as observed experimentally. In each case, we also use our model to predict cell dynamics on patterns that have not been studied before. Adhesive micropatterns (MP) determine the spatial distribution of the extracellular matrix (ECM) and therefore allow us to investigate and control cell shape, structure and function through experimental design. Over the last decade, they have emerged as an extremely versatile tool to investigate the inner workings of cells [1]. In particular, they are especially suited to achieve a quantitative understanding of how cells respond to external cues. Pioneering work with adhesive micropatterns has demonstrated the importance of the ECM-geometry for the survival of cells [2]. Later work showed how e. g. the organisation of the cytoskeleton [3,4] and of endomembranes [5] depend on ECM-geometry. Adhesive micropatterns have also been used to address the mechanical aspects of cells [6,7]. The versatility of adhesive micropatterns is further increased by combination with traction force microscopy on soft elastic substrates [8–10]. Although originally designed to immobilize single cells, micropatterns have also been extensively used to study their dynamic processes, including the different phases of cell spreading [3] or migration on stripe patterns with a focus on cell speed and persistency [11,12]. During recent years, the micropatterning approach has been increasingly applied also to multicellular systems. A first step towards multicellular systems is division of single cells which has been investigated with a focus on the central question how the cell division axis is determined by ECM-geometry [13,14]. It has been found that the statistical distribution of the direction of the cell division axis has a clear relation to the ECM-geometry. It has been argued that this relation is mainly provided by so-called retraction fibers that anchor the dividing cell to the adhesive micropattern [14,15]. The result of a division are usually two daughter cells that share one micropattern. Already this simple situation leads to very rich behavior. Cell pairs on square or circular micropatterns usually undergo a spontaneous symmetry break, assume a Yin-Yang shape and rotate persistently in the direction of the blunt cell sides [16]. Occasionally rotation can stop and cells rearrange, but eventually the cells resume rotation, either in the same or the opposite direction. Going beyond square or circular patterns, it has been found that the geometrical details of the pattern strongly influence the rotational behavior. In particular, it can be suppressed by using concave patterns that force the cells to deviate substantially from a circular trajectory [17]. Micropatterns are also increasingly used to study collective cell migration of large ensembles, mainly of monolayers of epithelial cells like MDCK-cells [18]. Collective cell migration is a hallmark of large scale cellular rearrangements during development, wound healing and cancer invasion [19–21]. Although cells moving collectively as sheets, strands, streams, clusters or tubes are often observed in three-dimensional situations, the two-dimensional setup with monolayers allows a more detailed analysis and in particular the use of micropatterns. The classical scratch-assay for two-dimensional wound healing assays has been replaced early by a microfabrication approach in which removable barriers allow to initiate collective cell migration into the gap without causing debris and ill-defined borders [22]. Over the last years, this assay has been adapted in many ways. One question of special interest are leader cells, that often emerge at the tips of protruding fingers [22,23]. Using microfabricated stencils to prepare monolayers of predefined shape, it has been shown that leader cells emerge from regions of high curvature [24]. Combination with the cell cycle marker Fucci in MDCK-2-cells has shown that removal of the barriers triggers release of the cells in the monolayer from cell cycle arrest [25]. Using adhesive micropatterns, it has been shown how topological defects can form when many cells have to self-organize in confined space [26]. The removable barrier approach is also increasingly combined with the micropatterning approach. A removable barrier has been used to trigger collective cell migration onto a stripe array, thus allowing the study of epithelial bridge formation as a function of ECM-geometry [27,28]. More recently, micropatterning has been used to study the closure of non-adhesive gaps in a monolayer [29,30]. In addition, these micropatterning techniques can also be combined with traction force microscopy, which allows us to understand the mechanical basis of wound healing [28,29,31,32]. Like cell pairs, cell collectives often show rotational motion which can be quantified best on adhesive micropatterns. The rotational motion of cell collectives on circular micropatterns is observed for systems up to 200 μm in diameter but eventually disappears for larger patterns [33]. Detailed analysis revealed a dependence of the persistence time of the rotation on the number of cells and their spatial coordination on the circular island [34]. In summary, adhesive micropatterns have become a standard method to study the shape, organisation and movement of cells bridging all the scales from single cells through small groups of a few cells up to large multicellular ensembles. Although highly relevant in the three-dimensional context of physiological tissue, most of these studies have been performed for a two-dimensional setup in which the full power of microfabrication and optical microscopy can be harvested. Despite the large biological relevance and the wealth of observations made in these experimental studies, however, a unifying model starting from the single cell basis and reaching up to the collective migration of large ensembles in the presence of geometrical constraints is still missing. Here we introduce such a comprehensive approach based on a two-dimensional cellular Potts model (CPM) that is particularly suited to describe cell dynamics on adhesive micropatterns. Motivated also by the power of quantification inherent in these experimental advances, the last decades have seen increased attention for modelling approaches to better understand cell shape, organization, mechanics and dynamics as a function of external cues [35]. An important focus of this work is the simultaneous prediction of cell shape and traction forces, using e. g. continuum mechanics approaches with contractility [9,36–41] or discrete contractile network models [6,42,43]. However, these approaches are usually restricted to static cells. To address dynamic processes, the CPM is much more suitable [44–47]. CPMs generate complicated shapes with high computational efficiency and propagate them under the action of a quasi-static energy function. They were originally designed to study cell sorting [48,49] motivated by the differential adhesion hypothesis [50]. Since then they have been widely used to address shape problems in biological systems, including e. g. packing in the Drosophila retina [51], gastrulation of zebrafish embryos [52] or cell packing in two- versus three-dimensional tissues [53]. They have also been adapted to describe collective cell migration, including streaming in cell monolayers [45,54,55], the formation of swirls [33], cell-ECM invasion [56], T-cell migration into lymph nodes [57] and the formation of endothelial networks on soft elastic substrates [58]. Closely related to CPMs are vertex models which have been used to study e. g. cell packing [59], the role of mechanical interactions in the Drosophila wing imaginal disk [60] and the role of contact inhibition in cell division [61]. However, vertex models can only describe cells with straight contours and are best suited to study large systems under high tension. Although they can be extended by viscoelastic elements to also describe dissipative processes, they are not suited well to describe subcellular processes such as protrusion and therefore they are usually not used to model the dynamics of single cells or small groups of cells. Another alternative to CPMs are phase field models [62–64], that recently have also been applied to cell ensembles [65,66]. Although well suited to study the effect of micropatterning [64,65], phase field models are computationally much more challenging than CPMs and due to their continuum character include non-local interactions. Here we introduce a two-dimensional CPM that allows us to bridge the gap between single cell and ensemble behaviour on adhesive micropatterns with a moderate computational effort. Although originally introduced for cell ensembles, CPM have been adapted before to describe also single cell migration [67,68] and single cell behaviour on adhesive micropatterns [69,70]. However, they have not been used yet to explain the whole range of cellular dynamics on micropatterns, ranging from single cells through pairs or small groups of cells to large communities with collective cell migration. In the following, we will introduce our computational framework in a step-wise manner, by considering experimental situations of increasing complexity, so that each new section requires the introduction of an additional set of rules. Because all rules are part of the same comprehensive framework, new sections will always include the rules of the previous sections. Our approach for single non-dividing cells spreading and moving on adhesive micropatterns has been described elsewhere [70]. Here we first extend this approach to dividing cells and show that we obtain very good agreement with experimental results for the orientation of the cell division axis on micropatterns of arbitrary shape. We then address the issue of cell-cell interactions and introduce a novel rule for formation and rupture of cell-cell bonds. Cell migration is introduced through a reduction of surface tension at the cell front and an increase at the back (graded tension model). With these ingredients, we can accurately predict the stability of cell pairs on micropatterns. In the same framework, we then also address collective cell migration for large ensembles, including epithelial bridge formation on comb patterns and swirl formation on circular and pacman patterns. Again we obtain very good agreement with experimental data. Our model gives important new insight into the way different cellular processes contribute to overall ensemble dynamics and in the future can be used to design micropatterns for desired cell dynamics. Our two-dimensional CPM for multicellular ensembles extends the CPM that we have developed earlier to predict single cell shape and spreading dynamics on an arbitrarily shaped micropattern (MP) [70]. The first row of Fig 1 shows the spreading process on a [L] shaped MP as it arises from our CPM (for a movie see S1 Movie). The driving force for spreading is an increase in adhesion energy as the cell (light gray) binds to more and more adhesive ligands on the adhesive substrate (dark gray). During adhesion and spreading on micropatterns, cells can dramatically change their projected area by transfering material from the third dimension into the adhesive region or back [71]. In our two-dimensional model, this corresponds to a reservoir for adhesive area that can be described by a chemical potential (with saturation at high adhesive area because adhesive area cannot grow without limits due to limitation in the influx of membrane and adhesion receptors). This approach is different from the classical CPM-formulation, that assumes that cells have a well-defined typical area (set e. g. by the density in a cell monolayer) and that deviations away from this average can be characterized by an elastic constant. Cell adhesion is only favorable as long as it does not lead to excessive mechanical energy that the cell has to spend to change its shape. The energy function and the model parameters describing this balance are given by Eq 1 and Table 1, respectively, in the Models section. The mechanical energy stored in the cell and its final shape is determined by different kinds of tension acting throughout the cell as illustrated in Fig 2a [6,35,70]. The surface tension σ mainly results from actomyosin contractility in the cell cortex and wants to reduce cell area. At the cell periphery, cortex and membrane fold back onto themselves, resulting in a simple line tension λs that wants to the draw the contour straight. Mathematically, the interplay between these two tensions leads to a Laplace-type law and circular contours [6]. Motivated by experimental results for cell shape on dot-like micropatterns [6,7], here we implement the tension-elasticity model that attributes an additional elastic energy to freely spanning arcs because they are reinforced by actin cables with a strong elastic response. The elastic arc rigidity k (compare Eq 1) leads to a line tension (defined as the derivative of the energy function with respect to the local cell contour length) that is increased from the simple line tension λs to a larger line tension λ along the arcs. The line tension λ is different for each arc and the resulting forces are shown schematically for the bottom arc in Fig 2a. Because it is the same for each point within a given arc, the Laplace law still holds and all arcs are circular, as also seen in the last image of the top row in Fig 1. Cell-cell junctions are contractile and pull the outer contour further inwards, compare the last snapshot in Fig 1 and the left and right arcs in Fig 2a. For simplicity and as explained in more detail below, here we assume that the resulting line tension at cell-cell boundaries equals the simple line tension λs of cell-matrix adhesion. Finally we account for an energy gain that comes with the formation of new adherens junctions. This corresponds to an additional line tension c pulling the endpoints outwards. Due to the force balance at these positions, the outer cell contour has a kink as revealed by the dashed lines fitted to the right circular arcs in Fig 2a. The different parameter values of our model can be fitted from experimental results for cell shape and forces for specific cell lines of interest [70], e. g. for breast epithelial MCF10A-cells on fibronectin patterns [8,17] or keratinocyte-derived HaCaT-cells that form epithelial bridges during wound healing [28]. These results are given in Table 1. In Fig 2b we show schematically how a cell pair is represented by two connected domains in the two-dimensional CPM. In our simulations, each cell is typically represented by 104 lattice sites. Using a refined marching square algorithm, cell area and perimeter can be calculated with very high accuracy (for a circular cell of this size, area and perimeter typically deviate by 0. 1% and 1. 5%, respectively). In contrast to other implementations of the CPM [53], we do not have to use a mapping between computed and actual perimeter values and therefore the line tensions can be compared directly to experiments. The orientation of the cell division axis is strongly influenced by the cell shape prior to division [13–15]. Previous models were able to reproduce division plane orientations correctly for adherent HeLa-cells [14] by relating the pattern geometry to a torque applied to the dividing cell. However, this model required an additional fit for each adhesion pattern and therefore lacks predictive capability for unknown patterns. Here, we introduce an alternative approach allowing predictions on arbitrary geometries without any parameter fit. Our approach mimics cell rounding and detachment during mitosis [72] and is motivated in detail in the Models section. Briefly, we let the cell contract by an increased line tension while all contacts between the cell and the matrix are disassembled. Inspired by the long axis rule [73,74], we fit an ellipse to the cell during the contraction when the cell area is contracted to r = 0. 36 of its original size. The major axis of this ellipse is chosen as the division axis. With this approach the complex cell shape is mapped onto a single direction. Noise inherent to our model through the Monte Carlo steps of the CPM result in statistical variations of this direction as we perform typically 20. 000 simulations for each pattern of interest. In experiments, the cell division orientation is typically determined by imaging the mitotic spindle and therefore we use the term spindle orientation synonymously for cell division axis. The middle row of Fig 1 shows snapshots of a dividing cell with an ellipse fitted to the contour once the prescribed area ratio r = 0. 36 has been reached. The shape transformation predicted by our model from a spread cell to a round one is very similar to the transformation observed for HeLa-cells [13]. Fig 3 shows that the resulting distribution of cell division axes (or, equivalently, spindle orientation, measured by the polar angle to the x-axis) agrees very well with the experimental results for HeLa-cells [14]. E. g. for a cell dividing on [L] it is most likely for the division axis to have an orientation of −45° and the two daughter cells are located on the two arms of the [L] shaped pattern as shown in Fig 1. Another simple example is the [bar] pattern, where the division axis is most likely at ±90°, corresponding to the two daughters being positioned at the top and bottom ends of the bar. Several key aspects are predicted correctly by our model. For [L dot] and [H], different distributions are predicted although both patterns have the same convex hull and similar cell shapes. This indicates that the location of the invaginated arcs influences the division plane. The model also predicts the rotation of the division axis by 90° correctly when the pattern is changed from [H] to [H3]. This rotation is shown for more aspect ratios of the [H] pattern in S1 Fig. The model also accounts for the strong broadening of the distribution when [L] is modified to [L dot]. The symmetry of the distribution is also predicted correctly. The patterns in the two central rows have a two fold mirror symmetry (except [square] and [X]) and one of the symmetry axis is selected as the main division axis. Patterns in the top and bottom row only have a one-fold mirror symmetry. However, the symmetry axis of the pattern does not set the main division axis as it can be seen for [pacman] and [star]. Both patterns have the same orientation of the symmetry axis but the most likely orientation is turned by 90°. We also note some small deviations between our predictions and the experiments. [H] and [two bar] differ only by the central bar and the distributions for our cells are very similar while experiments show a more strongly peaked distribution on [two bar]. The two patterns differ in size, [H] has a width of 36μm and [two bar] one of 30μm. It is not clear if the difference of the distributions comes from the size difference, i. e. that the [H] pattern is too large for the cells, or the absence of the central bar. For the simulations it was assumed that our cells can cover the complete [H] pattern. For the [arrow] (width 48μm) and [crossbow] (width 38μm) the size difference is even larger and the observed trend in the simulations to rotate the division axis by 90° is only due to this size difference. The [arrow] patterns requires cells with an area of 2000μm2 to be fully covered. Our cells have an area of 1300μm2 not fully covering the pattern and they localize to the tip of the arrow which elongates them perpendicular to the symmetry axis giving rise to the observed orientations. The distributions predicted by our model as shown in Fig 3 are almost independent of the parameters defining the cell shape, namely surface tension σ, simple line tension λs and arc rigidity k. The agreement between simulated and experimental distributions when the shape defining parameters are changed is quantified in subfigures (a) and (b) of S2 Fig. This figure shows that the optimal area ratio r = 0. 36 to fit the ellipse does not depend on the parameters. Only when the surface tension is large compared to the line tension do deviations occur. Cells are then no longer able to spread across the gaps on [L dot], [bar dot] and [two bar] and the different shapes before division result in poor agreement with experimental predictions for those patterns. Subfigure (c) of S2 Fig shows how the division plane is linked to shape fluctuations. If one allows for larger fluctuations, then one has to fit the ellipse earlier, because otherwise all information about the shape prior to division is lost. In general, fitting the ellipse early during contraction results in strongly peaked distributions and fitting it late in almost uniform ones. Subfigure (d) of S2 Fig shows that residual adhesion remaining during contraction worsens our predictions. After division two cells adhering to each other reside on one MP as shown in the bottom row of Fig 1. As described in detail in the Models section our approach to cell-cell adhesion differs from earlier versions of the CPM. Cell-cell adhesion is mediated by cadherins which are transmembrane proteins forming adherens junctions between cells [75]. Many models treat cell adhesion by a reduction of the line tension at cell-cell interfaces opposed to cell-medium interfaces [48,49,51,52,76]. The formation and breakage of adherens junctions releases and requires energy, respectively. Therefore, in an energy-based description cells can lower their total energy by increasing the length of cell-cell contacts. This is equivalent to assuming a reduced line tension at cell-cell boundaries. Earlier it has been argued that cortical contractility is equally important as is adhesion energy to determine cell-cell adhesion [52]. Cortical tension is reduced close to adherens junctions and this can have different effects than a change in adhesion energy only. The question then arises how to separate adhesion- and tension-based mechanisms in a energy-based description. Here we suggest that line tension should be kept high in order to avoid floppy configurations that are suppressed by contractility. In particular, negative values should not be allowed for the effective line tension λ. Further we suggest that the driving force for contact formation is mainly localized to the endpoints, as observed experimentally [17]. Thus energy is only gained when previously unconnected parts of two cells come into contact, but not if the existing contact line elongates by a pure deformation which does not form new contacts (compare S4 Fig). Our approach is equivalent to applying an outward directed force to the endpoints of a junction as shown in Fig 2a. Fig 4 shows two cells with different values for contractility and adhesion energy. The energy line density associated with the cadherin bonds is denoted by c. This quantity equals the number of cadherin molecules per length times the energy associated with each bound molecule. It is assumed to be of the order of several thermal energies per cadherin. Without an explicit treatment of cadherins (c = 0, top row of Fig 4), adhesion is only controlled by differences in the tension between cell-cell and cell-medium contacts. MCF10A-cell pairs on [H] pattern spread almost fully across the non-adhesive regions [17]. To achieve such spreading the tension at the cell-cell interface must be reduced to at least half the level of the cell-medium tension as shown in Fig 4a. This results in a strongly fluctuating interface between cells. By measuring the resistance to deformation it should be possible to estimate the line tension at the interface experimentally. Naively one would expect the tensions to be higher there because the cortices of both cells contribute to it. Setting the interface tension to the same value as the cell-medium tension results in no spreading across the non-adhesive areas (Fig 4b). The cell junction is too contractile in this case. With explicit treatment of cadherins at the endpoints (c < 0, bottom row of Fig 4) cells spread out more easily. A reduced tension still results in strong fluctuations at the interface (Fig 4c). Setting cell-cell and cell-medium tension equal now allows spreading (Fig 4d). The junction remains contractile and flat in this case. The contractility of junctions can also be seen by comparing the contour of the single cell and cell pair in Fig 1. The last frame shows two circles, the blue one is fitted to the contour of the cell pair and the green one is fitted to the contour of the single cell. The radius of the blue circle is smaller indicating a higher contractility of the cell. Migration of crawling animal cells is driven by actin polymerization pushing against the membrane at the front and myosin contractility pulling at the back. We describe both processes through a graded tension model described in detail in the Models section. Fig 5a demonstrates how the tension generated by the actomyosin activity is distributed spatially inside the cell. At the cell front actin polymerizes in a direction normal to the contour pushing it forward. The resulting forces are indicated by arrows and decrease gradually towards the sides of the cell because of decreasing polymerization activity. At the rear of the cell myosin contraction results in inward directed forces which are assumed to have the same magnitude as the actin generated forces. The combination of actin and myosin forces result in a reduction of the surface tension σ by σm = − μ cos (α). The parameter μ controls the strength of the migration machinery and the cosine the gradation towards the sides of the cell. The angle α is measured with respect to the polarization direction and defines the location of a point on the cell contour. As shown in Fig 5b the width of the gradation can be controlled by taking powers of the cosine function. For powers η < 1 the gradation is wide and protrusion forces at the center of the leading edge (α = 0) stay relatively constant and decay rapidly at the sides of the cell (α = ±π/2). For η > 1 the gradation is sharply peaked at the center of the leading edge. The width of the gradation has a strong impact on the cell shape of migrating cells. A broad gradation (η < 1) results in a keratocyte shaped cells while a narrow one (η > 1) results in cells resembling neutrophils (Fig 5c). The shape of the migrating cells is defined by the balance of forces arising from the line and surface tension and the migratory tension σm. At curved region of the contour the line tension results in a force perpendicular to the contour [70]. For concave shaped regions this force is directed inward. Its magnitude increases with increasing curvature. To maintain a stationary shape during migration, the forces generated through curvature, surface tension and migratory machinery must add up in a way which allows to translocate the cell without changing its shape. For η = 1 this shape is a circle as in the graded radial tension model [77]. For a circular shape, curvature induced forces are irrelevant because they always point to the center with constant magnitude along the contour. The resulting anisotropic tension responsible for translocation is the bare migratory tension σm = cos (α). It is proportional to the force acting on the contour which is, in turn, proportional to the velocity of the contour. Hence, the velocity is v (α) ∝ cos (α) which is the velocity required to propagate a circular contour without deforming it [78]. For gradation deviating from the simple cosine the shapes of Fig 5c arise. The broad gradation and strong protrusive forces for η < 1 are ultimately stalled by the high curvature at the sides of the lamellipodium. Viewed from a different perspective, strong changes of the gradation allow for changes of the curvature. Our gradation model assumes that actin polymerizes normal to the membrane with its activity decreasing towards the sides of the cell. If actin is assumed to polymerize uniformly in the polarization direction it pushes against the membrane with an angle reducing its force. The angle is given by the normal of the contour with respect to the polarization direction. The resulting shape of cells resemble fibroblasts as shown in Fig 5d for η < 1. We speculate that such an additional polarization effect might arise from the presence of stress fibers. For η > 1 this acting protrusion mechanism generates triangular cells which seem to have no biological counterpart. Single cells can maintain their polarity direction and move along it over a distance of several hundred of cell diameters [79]. Directionallity in cells migrating collectively in sheets can also stay correlated over several hours [18]. On longer time scales their movement becomes random. To control the polarity direction and persistence time we use the well established velocity alignment model [45,65,80–82] where cells try to align their polarity direction with their current movement direction. Polarity and movement direction do not need to be the same. Misalignment between them can occur through fluctuations of the center of mass movement, boundaries imposed by the MP or other cells which do not allow further movement along the polarity direction. The central parameter of the velocity alignment model is the lag time τ. It controls how fast the polarity direction aligns with the current movement direction. As discussed in the Models sections this model can be interpreted as cells keeping a memory of their past movement. The lag parameter τ sets the timescale for how long past steps are remembered. A typical cell trajectory obtained with this model is shown in subfigure (a) of S3 Fig. The implicit noise originating from the fluctuation allowance leads to a coupling between cell speed and persistence time as shown in subfigure (b) of S3 Fig. The persistence time first increases exponentially with speed. The increase saturates for higher speeds as observed for many different cell types [12]. We next turn to the migration of small groups of cells on MP. Special attention is needed when cells contact each other or encounter the border of an adhesive area of the substrate. Cells retract their protrusions and change directions when they encounter another cell [83]. This contact inhibition of locomotion helps cells to coordinate their migration collectively. We implement it by preventing any protrusive activity when a cell would invade another cell. Protrusions generated by polymerizing actin need the actin network to be anchored to the substrate which occurs a few microns behind the leading edge [78]. Above non-adhesive areas the anchoring support is missing and it is assumed that the polymerizing actin mesh can only reach a certain extension before polymerization just pushes the mesh backwards without extending the boundary. This is put into effect by reducing the migratory strength μ with distance from the adhesive substrate as described by Eq 4. With this model we can now predict behavior of cell pairs on MP. Snapshots of our simulation are shown in Fig 6 (also see S2 and S3 Movies) with the parameter values for MCF10A-cells. Indeed our results agree very well with the experimental observations for this cell line [17]. In Fig 6a a single cell initially spreads on a [square] pattern. The pattern is too large to be fully covered by the cell and it finds no stable position. It oscillate in a circular fashion between the corners of the pattern until it divides. The two daughter cells adhere to each other and start to rotate. At the leading edge a lamellipodium extending beyond the pattern is visible. The rotation direction is reversed from time to time as experimentally observed for MCF10A-cells [17]. In marked contrast, cell pairs on [H] patterns do not rotate. They remain in a stable configuration with the cell-cell junction positioned above the nonadhesive region. However, single cells on [H] patterns are highly mobile as shown in Fig 6b. They mainly oscillate from top to bottom (compare S3 Movie) but also from left to right. In general, the stability of a cell pair strongly depends on the size of concave regions. Cells on [C], [H] and [hourglass] pattern are stationary while they move on [square] and [C2] [17]. This suggests that rotation is larger if the convex hull of a pattern is adhesive [17]. In the framework of our model this can be understood by the requirement of an adhesive area near the leading edge to support actin polymerization and by the contractility of the cell-cell junctions. A minimal energy state for a cell pair is usually found when the length of the cell-cell junction is minimized. For the pair on the [H] pattern shown in Fig 6b this results in the junction being above the non-adhesive part of the substrate as every other configuration would make it longer. A minimal junction length minimizes the contribution of the line tension to the cell energy. On patterns with a continuously adhesive hull such as [square], junctions cannot be shortened by placing them above non-adhesive regions. For cell pairs to rotate junctions must be stretched and reach at least the diagonal extension of the patterns which is energetically unfavorable. The required energy comes from the migratory machinery. Large changes in junction length and therefore large energy differences make it unlikely that the cells maintain their polarity long enough to cross the energy barrier imposed by a diagonally oriented cell-cell junction. Persistent rotation can be promoted by increasing the migratory strength μ or the lag time τ which makes it easier to overcome the barrier. Alternatively the simple line tension λs can be decreased. Fig 7 shows in green distributions of the nucleus-nucleus axis orientations obtained for MCF10A-cells on various patterns [17]. On all patterns orientations with the shortest cell-cell junctions are most likely (the cell-cell junction is usually perpendicular to the nucleus-nucleus axis) but the degree of bias towards the shortest junctions varies. Among the shown patterns [H] offers the possibility of the shortest junction and has indeed the most sharply peaked nucleus-nucleus axis orientation. It is followed by [C] where the minimal junction length is slightly longer opposed to [H] because one more side is adhesive and the contour is not invaginated there. The longer minimal junction length results in a broader peak. This trend continues for [square] where now all sides are adhesive and the minimal junction length is set by the width of the pattern. The most relevant model parameters for persistent cell rotation are the shape-defining parameters: surface tension σ, simple line tension λs, arc rigidity k and the cadherin energy line density c. These parameters describe how strongly cells are invaginated at concave parts of the pattern. The migratory machinery strength μ and lag time τ act as a counterpart. They set how much force is available for the cells to extend their cell-cell junction and for how long this force is applied. From traction force measurements the values of the surface tension σ = 0. 83nN μm−1, the simple line tension λs = 2. 3nN and the arc rigidity k = 40nN are known for MCF10A-cells [70]. However, cells are too strongly invaginated to allow persistent rotational motion with these parameters. Agreement with experiments can only be obtained if the surface tension is lowered. We keep the values of the simple line tension and the arc rigidity fixed and fit the other model parameters to the experimental distributions for MCF10A-cells. The fit is performed by minimizing the least square deviation between experimental and predicted histograms with Powell’s method. Excellent agreement results as shown in Fig 7. Our model has no explicit description of a cell nucleus and we take the center of mass instead. The choice of the gradation width η has no influence on the agreement between simulation and experiments. It only influences the other fit parameters and we take η = 1/4 which results in cells with a broad lamellipodium. The optimal surface tension is given by σ = 0. 1 nN μm−1 which is significantly lower as previously reported [70]. The best fitting values for the migratory machinery are μ = 0. 19 nN μm−1 and τ = 185 MCS (Monte Carlo Sweeps) resulting in a persistence length of 40μm defined by angular correlations. The scale on which the protrusive activity decays is found to be d0 = 43 μm. Although this value is large, lamellipodia extend only about 5 μm beyond the adhesive substrate. The protrusive activity is mainly stalled by the line tension. For narrower gradations (η > 1) this decay length is more important and becomes shorter. The migratory tension is larger in this case. The optimal cadherin adhesion energy line density was found to be c = -6. 3 nN which results in an area density of 12. 6 nN μm−1 if the cell junction is assumed to have a height of 0. 5 nN. We can now investigate migration of cell pairs on arbitrary MP which have not been studied in experiments before. In Fig 8 we show some representative examples. This includes cells on a [2 circle] pattern which do not obey the symmetry axis of the pattern or completely asymmetric patterns as the [asymmetric] layout which results in a non-smooth asymmetric nucleus-nucleus axis orientation. We can also test layouts which further support our idea that the observed distributions can be explained by the contractility of the cell-cell junctions alone. One candidate is a Reuleaux triangle which is a triangle with circular edges. A Reuleaux triangle is a curve of constant width meaning that every line crossing the triangle through its center has the same length. For the rotating cell pair this means that the cell-cell junction always has the same length. The pair should rotate as cells on a circular pattern although the cells change their shape during rotation on the Reuleux triangle. The resulting nucleus-nucleus axis distribution shown in Fig 8 are almost uniform confirming the importance of the junction contractility. Deviations most likely arise because the junction is not always straight. The distribution on a regular [triangle] deviates significantly from a uniform distribution as also shown in Fig 8. The shapes of both cells are very similar, one cell is usually triangular having contact with one corner while the other is rectangular having contact with two corners (compare S4 and S5 Movies). Cells on the regular triangle also switch their rotation direction more frequently. The behavior of cells on the triangles strongly supports the importance of the junction contractility. The relative importance of shortening a junction by positioning it above a non-adhesive area or by selecting the shortest connection above adhesive parts can be investigated with [pacman] and [ellipse] patterns. On a [pacman] pattern the most likely orientation is with a horizontal oriented junction while the junction on a [ellipse] is oriented vertically along the shortest semiaxis (compare bottom row Fig 8). A combination of both patterns into an ellipse with a nonadhesive wedge ([e+p 1]) results in a superposition of the [pacman] and [ellipse] distributions. Increasing the aspect ratio further ([e+p 2]) shifts the distribution in favor of the ellipse. The reduction of myosin activity by blebbistatin revealed a stronger reduction in surface tension than in line tension for rat embryonic fibroblasts [7]. Thus, cells on concave patterns increase their size when myosin activity is reduced. A similar reduction in surface and line tension for MCF10A-cells described by our CPM results in broader distributions as shown in S5 Fig. The increased cell area make cell junctions above nonadhesive areas longer. They need to be extended less for rotation which results in the broader distributions. For [square] and [C2] the distributions stay the same because the junction length is not changed on these patterns by inhibition. For [hourglass] patterns the broader distribution has been observed before [17]. MP are increasingly used to investigate collective cell migration, often in combination with removable barriers. We first investigate the formation of epithelial bridges on a comb pattern when cells are allowed to migrate from a reservoir onto a stripe array such that they have to bridge over the intervening spaces in order to advance. Fig 9a shows cells initially confined to a rectangular reservoir. The cells spread, migrate and divide forming a monolayer. Contact inhibition decreases the proliferation rate in denser tissues [61] as explained in the Models section. With the high initial density shown in Fig 9a divisions are not frequent. The cells are highly mobile and form swirl-like structures (see S6 Movie) until the constraint confining them to the rectangular area is removed after 4. 000 MCS. Without confinement cells migrate fast to the newly available reservoir region until they reach the adhesive stripes. The movement of the cell front slows down and an epithelial bridge is formed between the stripes as observed experimentally for HaCaT-cells [28]. Extension of the bridge is driven by cell division inside the bridge, cells migrating into the bridge region and pulling of the leading cells on the bridge. Dividing cells loose contact to their neighbors which can form holes (e. g. at 12. 000 MCS). Migration into the bridge region is slower compared to migration on a homogeneously adhesive substrate because cells tend to keep contact with the substrate. Increasing the distance between the stripes from 120 μm in Fig 9a to 240 μm in Fig 9b prevents the formation of bridges as observed experimentally for HaCaT-cells [28] (see S7 Movie). The migration along the stripes is limited and the leading cells are closer to the reservoir. Inhibition of adherens junction by decreasing the cadherin line energy density from c = -3 nN to c = -1. 5 nN results in retraction of the bridge as shown in Fig 9c (see S8 Movie). With less cadherin activity balancing the contractile junctions the leading cells are too weak to pull the cell sheet outward. In addition, cells are less cohesive and cell division can result in larger holes as observed for HaCaT-cells [28]. Bridge extension can also be stopped when the stripes are not parallel but tilted as shown in S6 Fig (compare S9 Movie). When the cell sheet moves along the stripes the distance between stripes increases which slows down progress in the bridges. Our cellular Potts model allows us to separate the contributions of cell division and migration to bridge formation. Fig 9d shows the displacement of the cells along the stripes as a function of MCS averaged over 500 simulations. Without cell division (r0 = 0, green curve) cells advance very similar as the full system at first until the bridge formation starts at ≈ 6. 000 MCS. Without cell division in the bridges further progress is slightly slowed compared to the full model. Setting the migratory strength μ to zero has a much stronger effect. Without the active movement it is more difficult to pull cells into the non-adhesive regions slowing the overall progress. Additional removal of cell division further decreases the advance, but the effect is rather small. Compared to a freely migrating monolayer on a continuously adhesive substrate progress along the stripes is rather slow as observed experimentally [28]. Fig 9e quantifies the previously mentioned slowdown when cells start to move onto the stripes. The velocity oscillates for actively migrating cells as observed experimentally [28]. Responsible for this oscillation is the feedback mechanism in the velocity alignment model. Cells move along the stripes until further advancement is stalled when cells have to be pulled into the nonadhesive region. Without continuous progress in one direction the polarity vector becomes random and the migration direction can even be reversed. However, backward migrating cells turn around fast when they are stopped by the bulk and advance on the stripes again. Repetition results in the oscillations of the velocity. The amplitude decays because the oscillations get out of phase and 500 simulations are averaged. The cell density in the confined region also has an impact on the migration along the stripes as shown in Fig 9f. A higher initial number results in smaller cells which are biased more strongly towards free adhesive areas on the stripes making migration faster. The oscillatory behavior is also more pronounced for a larger number of initial cells and migration can even be reversed for a short time. Circular MP have been demonstrated to promote the formation of a single swirl [33]. We demonstrate here that modifications to the circular geometry can result in two swirls. With a low initial cell density on a [circular] pattern cells divide frequently and move in an uncoordinated fashion as shown in Fig 10d (see S10 Movie). When the cell density increases division slows down and the cells start to move in one large swirl as observed for MDCK-cells [33]. The effect of the population size for the formation of a swirl has been pointed out before [45]. The single swirl can be split into two swirls by the non-adhesive wedge on a [pacman] pattern as shown in Fig 10e (for movie see S11 Movie). At the left and right side of the [pacman] pattern cells are moving downwards and collide head on at the lower part from where they start to move to the center. The location of the collision fluctuates strongly (compare S11 Movie). The rotation can also occur with reversed direction. To quantify the swirls we calculate the vorticity which is a measure for the rotational tissue flow around a given point. The vorticity has been used before to characterize the impact of cell division on tissue dynamics [84]. A positive vorticity value means counterclockwise rotation of cells in the neighborhood of a central cell. This rotation is observed in the frame of the central cell. Fig 10c shows the vorticity field calculated by Eq 8 for cells on a [circle] pattern. A global counterclockwise rotating swirl of cells is indicated by the large area of positive vorticity at the center of the pattern. The vorticity at the border of the pattern is negative because cells have a lower angular velocity at the border compared to bulk cells. In the frame of bulk cells below the border layer the border cells move backwards resulting in the negative vorticity. The lower angular velocity is a consequence of the finite cell speed. On a circular pattern the outermost cells would need to be the fastest to achieve a constant angular velocity. For the [circular] pattern the cells in the layer below the border already move with the highest speed. Fig 10d shows the vorticity field for the [pacman] pattern. Two swirls rotating in opposite directions are indicated by the opposite signs of the vorticities in the left and right regions. Our work clearly demonstrates that adhesive micropatterns have a very strong influence on cell dynamics. This agrees well with experimental observations that have motivated our approach. Examples range from the location of daughter cells after cell division [14], cell pairs that are rotating on certain MP but not on others, and MP which guide collective cell migration [28,33]. By combining the CPM with various model approaches developed earlier for cell mechanics, cell division, cell migration, cell-matrix and cell-cell adhesion, we were able to generate a unifying yet computationally transparent and efficient framework which can explain a large range of experimental observations. This not only contributes to a better understanding of these situations, but in the future will allow us to design adhesive patterns that lead to desired cell dynamics. The classical CPM [48,49] describes cells by a phenomenological energy function motivated by the differential adhesion hypothesis (DAH) [50]. It has been extensively used to describe cell ensembles, but has been used only occasionally to describe single cell dynamics [67–70]. Our previous approach to single cells [70] has adjusted the formulation of the energy function for single cells to reflect that they show large differences in adhesive areas on different micropatterns. Mathematically, this is achieved by using an energy function Eq 1 that controls adhesive area by a chemical potential with saturation, rather than by an area-elasticity term that is appropriate when describing cell monolayers with constant thickness. Here we use this approach to first describe how the division plane of cells can be predicted from their shapes. The main idea behind our approach is to map the complex cell shape to a single direction by fitting an ellipse to it at an optimal time point. Fluctuations intrinsic to our modeling approach broaden this single direction into a distribution of directions. The division plane distributions obtained this way agree well with observations for HeLa-cells [14] (compare Fig 3). Our results allow us to speculate about the underlying mechanisms determining the orientation. During mitosis cells contract to a sphere with an intermediate ellipsoidal shape. During this contraction the mitotic spindle assembles which defines the orientation of the division axis. Our model suggests that the ellipsoidal shape sets the division axis and no direct interaction with the adhesive geometry is required during this stage. Support for this idea comes from the prediction that an ellipsoidal shape is sufficient to align a spindle through its growing astral MT [85] and that the division axis of non-adherent cells is also influenced by their shape [86]. In this view, experiments which speed up spindle assembly compared to cell contraction should result in sharply peaked distributions of the division plane because barely contracted cells are best fit by an ellipse with a high aspect ratio which is less susceptible to fluctuations. We do not model exponential growth in cell volume or projected area before division because we focus on the geometrical and mechanical aspects of cell dynamics and do not include any internal dynamics. In the future, this could be included, similar to the bidirectional coupling between signaling networks and cell shape as modeled before with the CPM [68]. For future studies it would also be interesting to combine division axis prediction and traction force calculation as it has been suggested that traction forces act as memory for the orientation information [87]. For multicellular systems cell-cell adhesion becomes relevant. Many models with an energy based description of cells implement cell-cell adhesion by a decreased line or surface tension at cell-cell interfaces [48,49,51,52,76]. There are different motivations for this approach. The first one is the DAH developed to explain cell sorting which treats cells similar to molecules in a liquid [50,76]. Differences in the attraction between cells can lower the energy when cells are rearranged and drives phase separation in this model. The lowered energy can be achieved by a decreased line tension between cells. The second approach is the differential interfacial tension hypothesis (DITH) [88] which explicitly takes into account that cortical tension of individual cells is reduced at cell-cell interfaces [52,89]. In an energy based description the DAH and DITH cannot be distinguished because they both result in a decrease of the energy when cell-cell interfaces are elongated. However, the response to deformation of an interface in both models is different. For the DAH deforming a cell-medium and cell-cell interface should be resisted in the same way and therefore yield the same line tension because the cortical tension is not changed. For the DITH deformations of the two interfaces should be resisted differently because the cortical tension at interfaces is lower. There is evidence that adhesion depends on both processes [76]. We therefore describe cell adhesion by a combination of DAH and DITH. At cell-cell interfaces the cortical tension is lowered and we also take the energy required for breaking and forming adherens junctions into account. With this approach we can control adhesion strength and contractility of intercellular interfaces independently. Our CPM can describe cells which require large forces to be separated while their interfaces are strongly contractile with a large tension. Our approach also allows a junction to remain stable for arbitrary adhesion strengths. A pure DITH model could be realized by a reduced fluctuation allowance at interfaces. Indeed actomyosin activity is known to be different at adherens junctions compared to the rest of the cell. Nevertheless this approach would still limit the maximum adhesion strength by the requirement of a positive line tension. Our approach to cell migration is inspired by the graded radial tension model [77]. Following our focus on cell tension, we addressed cell migration through a reduction of surface tension at the cell front and an increase at the back. This approach does not allow precise control of the underlying actin dynamics [67,68], however, the shape and dynamic behavior of several cell types are predicted correctly (compare Fig 5c and 5d and S3 Fig). The resulting predictions for the interaction and stability of cell pairs agree well with MCF10A-cells [17]. This requires a lower surface tension as previously reported for MCF10A-cells by means of traction force measurements [70]. A possible explanation is the effect of internal contractile stress fibers which contribute to the overall traction of a cell. These fibers usually do not connect to peripheral stress fibers and therefore do not contribute to the invagination of the contour. In the traction force evaluation [70] their contractile effect was averaged and integrated in the homogeneous surface tension most likely resulting in a higher value as obtained here. The parameter set we obtain for MCF10A-cells is not unique. The determining factor is the shape of the cells which is described by the dependence of arc radius R (d) on the spanning distance d[70]. Rescaling the shape defining parameters (σ, λs, k) and the migratory strength μ by the same factor and adjusting the cadherin line energy density accordingly leaves R (d) unchanged and shape and rotational behavior are the same as before. Only the ratios k/σ, λs/σ and μ/σ are relevant for the stability of MCF10A-cells. Absolute values for our parameters can be obtained when our model is compared to traction force measurements. However, the ratios obtained in the current work are roughly 8. 5 times larger than in our previous approach for single cells [70]. This indicates that the contributions of the surface tension is reduced for cell pairs compared to single cells. Consistent and unique values for all parameters could be obtained by combining traction force and rotation measurements for cell pairs in a single experiment. Combination with traction force microscopy could also be used to clarify the issue how cell division orientation correlates with shape, stress and strain [74]. In general, it would be very rewarding if one experimental group could measure all our model parameters for one cell type in one set of complementary experiments. Finally our approach can be applied to collective cell migration. We demonstrated that our CPM correctly predicts the formation of epithelial bridges [28] and swirls [33] (compare Fig 9 and Fig 10). Other multicellular systems that might be studied with this approach are e. g. leader cell formation triggered by curvature [24], closure dynamics above non-adhesive [29] or adhesive gaps [90]. Our implementation is fast enough (typical runtime of several minutes for systems consisting several hundreds of cells) to allow screening of pattern geometries and to address pattern optimization for specific tasks in future projects. Our cellular Potts model (CPM) was developed to predict the shape and forces of single cells on MP [70]. Motivated by experimental observations in this context, it pays particular emphasis to the variability of projected cell area and to invaginated shapes over non-adherent regions. Our effective energy function for single non-migrating cells is given by H = σ A + λ s l + ∑ arc i k 2 L 0, i (L i - L 0, i) 2 - E 0 A ref + A ad A ad. (1) The first term results from the surface tension σ generated mainly by actomyosin contractility in the cortex. The corresponding energy σA scales linearly with the projected cell area A. The second term results from the simple line tension λs acting throughout the cell contour. The corresponding energy λsl scales linearly with the cell perimeter l. The third term accounts for the reinforcement of the cell contour by elastic peripheral actin bundles (tension-elasticity model from [6]). It is only active in concave parts of the cell above non-adhesive areas of the substrate, compare e. g. the free spanning arcs of the cells on the [X] shaped pattern shown in Fig 2a. If the cell has several free spanning arcs, a sum is used in Eq 1. The strength of the arcs is controlled by the elastic arc rigidity k. Note however that the way we define it here, it has the physical dimension of a tension. Li is the contour length of an arc and Li, 0 its rest length. For the rest length we take the spanning distance of an arc which is the length of the straight connection between its endpoints (d in Fig 2). The last term in Eq 1 accounts for the gain in adhesive energy. The energy is lowered when the contact area Aad with the substrate is increased. The reduction in energy by covering more adhesive area saturates. This reflects the finite amount of adhesion molecules and membrane available to the cell. The ratio of W = E0/Aref defines the adhesive energy area density. The value of Aref is set through a target cell size A0 [70]. Note that in contrast to most other formulations of the CPM, we do not use an elastic (harmonic) energy term for the area, because this would constrain it much more than observed during cell adhesion. It is an essential ingredient of our model that cells can transfer material from the third dimension into the adhesive area [71]. Because we do not include an area-elasticity term, surface tension (defined as the derivative of the energy function for surface area) and surface energy area density are identical. Our implementation of the CPM uses a two-dimensional lattice to represent the cell as illustrated in Fig 2b. All lattice sites belonging to the same cell are marked by the same index. The perimeter of cells is calculated with a refined marching square algorithm [70]. To propagate the cell shapes we use Metropolis dynamics to minimize the cell energy Eq 1. During each step a random lattice site at the cell periphery or the sites surrounding it is selected. The energy Δ H = H new - H current (2) associated with changing it to the values of one of its neighbors is calculated. The step is accepted if ΔH < 0 and with the probability exp (−ΔH/T) otherwise. Here T is the fluctuation allowance following from the stochastic nature of the molecular processes underlying changes in cell shape and contractility. Low and high T-values correspond to small and large fluctuations, respectively. The time is measured in Monte Carlo sweeps (MCS), where 20. 000 MCS correspond roughly to 1 day in real time. During each sweep n inversion attempts are performed, where n denotes the number of lattice sites in the contour of all cells. For fitting circles to the invaginated arcs, we use a modified least square method [91]. During this fit the endpoints are weighted by a factor of 100 to ensure that the circles run through them. We implement cell division by closely following the experimentally observed sequence of events. On the onset of mitosis, stress fibers and focal adhesion disassemble and cells become spherical due to increased myosin contractility [72]. The mitotic spindle assembles and rotates driven by cortical cues. The cortical cues are established through the adhesive geometry prior to division [13]. Thus, the cell divison axis is predetermined during interphase by cell shape [72]. A mapping of the adhesive geometry controlled by MP has been demonstrated to reproduce division axis orientations for many MP geometries [14]. In our approach we put a strong emphasis on the transformation of the flat spread out cell into a sphere. We assume that during this process the spindle assembles and receives cues for its orientation. We let the cell contract and, inspired by the long axis rule [73,74], fit an ellipse to the cell contour during contraction. The major axis of the ellipse is then taken to be the cell division axis. Increased cortical tension driving a cell to become round during mitosis is achieved in the CPM by increasing the line tension. During rounding the sign of the surface tension is inverted, σ → σ ¯ = - σ. It takes the role of the osmotic pressure resisting contraction. Setting the simple line tension to λ s → λ ¯ s = - 2 σ A b / π ensures contraction, where Ab denotes the cell area prior to division. Focal adhesion and stress fiber disassembly are achieved by setting the adhesive energy density W and elastic arc rigidity k to zero. Rescaling the fluctuation allowance by T → T ¯ = T λ ¯ s / λ s ensures the same size of fluctuations before and during roundup. Although fluctuations are expected to decrease during division, this measure is important for the Metropolis algorithm not to get stuck in local minima. The moment of ellipse fitting is defined by the area ratio r = A/Ab, where A denotes the area during contraction. The value of r = 0. 36 is universal for all patterns. Cell division as simulated with the CPM on an [L] shaped MP is shown in Fig 1. Cell proliferation decreases with increasing cell density and we regulate it through the cell area with a Hill function of the form r (A) = r 0 A m / (A m + A 0 m) [61]. In this function r0 denotes the division rate without inhibition by other cells, A0 the typical cell size, and m is the Hill coefficient. Cell-cell interactions are mediated by cadherin proteins forming discrete connections between two cells similar to interaction between integrins and the extracellular matrix (ECM) [75]. In contrast to integrin-ECM interactions, however, which are fixed in space, an interface between two cells can deform. Many models [48,49,51,52,76] reduce the line tension at cell-cell junctions opposed to the cell medium interface based on the argument that surface contacts reduce the apparent intercellular surface tension [92]. This means that the tension acting at cell-cell junctions is set to λcc = ∂H/∂lcc = λs − γ, where λs denotes the line tension at a cell-medium interface, γ the reduction of it and lcc the cell-cell interface length. In this way, cells can decrease their energy by increasing the length of cell junctions. An increase of γ strengthens the adhesion between cells. The strength of cell-cell adhesion is limited in this approach by the necessity of a positive line tension λcc. With a very low tension interfaces fluctuate strongly and become unstable or completely dissolve for a negative tensions λcc. This can be prevented by restricting the cell perimeter to a default length l0 with a quadratic term (1/2) κ (l − l0) 2 in the energy function [51,76]. However, the resulting line tension λcc = ∂H/∂lcc = λs − γ + κ (l − l0) can still become negative, depends on the global perimeter of the cell and agrees only for γ/λs>1 with experiments [76]. Although cadherins are known to reduce the cortex tension to some extent [52,89], the tension generated by the cortices of both cells should add up resulting in an higher tension at the junction. Mimicking cell-cell adhesion by a reduced line tension does not allow independent control of cell-cell adhesion strength and contractility of the cell junction. In our work, we focus on the energy associated with the formation and breaking of discrete cadherin bonds. We picture the cell-cell contact as a zipper where cadherin bonds form and break only where the two parts of the zipper meet. Therefore, the cells energy is only changed when a cell junction is elongated or shortened at the termini of the junction. Elongation by deformation of a junction does not change the energy. As illustrated in the SI S4 Fig a deformation displaces existing cadherin bonds but does not generate new ones. The energy change associated with changes of the cell-cell interface is then Δ E c c = c Δ l c c δ (l ⇌ l c c), (3) where c denotes the cadherin energy line density, Δlcc the change of the cell-cell interface length and δ (l ⇌ l c c) is one if previously free cell interfaces come in contact to form a cell-cell interface (l ⇀ l c c) or break (l ↽ l c c). It is zero if lcc is changed by deformation. A detailed description for a lattice based implementation can be found in S7 Fig. Setting the cell-cell line tension to λs effectively means that we attribute a reduced line tension λs/2 to each of the two contacting cells. This accounts for the fact that cadherins reduce the cortex tension at cell-cell contacts [52,89]. An illustration of all tensions can be found in Fig 2a. Membrane displacement is driven by actin polymerization at the front and myosin contraction at the back [79]. Actin polymerization and retrograde flow are found to decrease gradually [93] towards the sides of the cell matching the graded radial extension found for keratocytes [77]. With the assumption of a graded filament density and a constant force originating from the membrane tension resisting polymerization, cell shape and motility can be related [94]. In turn, prescribing the cell shape allows to predict the correct actomyosin activities [93]. Coupling both shape and actomyosin can be achieved with a CPM [67,68]. A reaction-diffusion system for the small G-proteins which drive actin and myosin activity is solved in the domain predicted by the CPM model. Simplified reaction diffusion systems or orientation fields in combination with phase field models also predict correct cell shapes for keratocytes [62,64] and allow to study cell migration on micropatterns [65]. Because here we aim at a computationally efficient framework, we do not introduce internal fields. Consistent with our tension-based approach, we introduce a graded, direction-dependent tension originating from the polarized actomyosin activity. In its simplest form this tension is given by μ cos (α) [57], where μ is the force per unit length generated at the center of the leading edge and α the angle between a point on the membrane and the cell’s polarity direction as depicted in Fig 5a. Actin generated forces at the front (|α| < π/2) and myosin generated forces at the back (|α| > π/2) of the cell are equal but with opposite sign. We note that prescribing a force is different from prescribing an extension as in the graded radial extension model [77]. Extension is driven by the resulting force of actin polymerization resisted by membrane tension which depends on the cell shape. The width of the gradation is known to influence cell shape [93] and can be controlled by taking μ|cos (α) |η sign (cos (α) ) as actomyosin tension, where sign (x) denotes the signum function. The parameter η controls the width as shown in Fig 5b. The normal is defined as described in the supporting material of [70] in Fig S3c. In short, a circular mask is applied to each boundary pixel. The normal is defined as the line connecting the center of the circular mask and the center of mass of the occupied pixels in this mask. The radius of the mask is nine pixels. For a typical circular cell with radius of 56 lattice sites (area about 104 lattice sites) the mean absolute error of the theoretical and computed normal orientation is 1. 2%. Protrusions above non-adhesive areas lack the anchoring of adhesions which start to form in the lamella [78]. The polymerizing actin mesh can only reach a certain extension before further polymerization just pushes the mesh backwards. This is put into effect by reducing the protrusion strength by an exponential factor resulting in an effective strength μ eff = μ e - d / d 0, (4) where d is the shortest distance to the nearest adhesive island and d0 sets the scale on how far protrusions can extend beyond the edge of a MP. The final energy change associated with cell migration is ΔHm=−ΔA|cos (α) |ηsign (cos (α) ) μeffδ (invasion) ] (5) where ΔA denotes the area change. The whole expression is negative because the surface tension is reduced by this amount. The last term accounts for contact inhibition of locomotion. Cells have been demonstrated to coordinate their migration through cell-cell contacts [83]. They do not invade each other actively, which is implemented here by setting δ (invasion) to zero if a cell invades another cell. This only happens at the front of the cell (|α| < π/2). Retraction at the rear is not influenced by contact inhibition. A commonly used model to describe polarity is a velocity alignment model [45,65,80–82,95]. In this model cells tend to align their polarity with their current velocity. In our implementation of velocity alignment the change of the polarity direction p is defined as [80] p ˙ = v - p τ, (6) where v is the center of mass (COM) displacement between two Monte Carlo sweeps (MCS) and τ sets the time scale of alignment measured in MCS. In Fig 5a the polarity direction p points upwards. In general it changes according to Eq 6 after every MCS. The angle α used in Eq 5 is always calculated with respect to the current polarity direction. Some models [65,81,82] add a explicit noise term to Eq 6. In our approach noise enters implicitly through the finite fluctuation allowance. The COM displacement v is not necessarily aligned with the polarity direction due to membrane fluctuations. It certainly gets misaligned when a cell encounters an obstacle. If v is treated as a independent function a formal solution for the differential equation Eq 6 can be obtained in the form of p (t) = ∫ 0 t v (t ′) e - t - t ′ τ d t ′, (7) which allows to interpret the mechanism behind velocity alignment. The cell integrates its past displacements with an exponential memory kernel to arrive at its current polarity direction. Cells have been demonstrated to have a memory of the past movement [96]. We calculate the vorticity ω as defined in [84] by ω = (1 / A) ∑ r ∈ O v x (r) r y - v y (r) r x. (8) The region O denotes a circular area around a given point, r = (rx, ry) is the vector from the center of O to the center of a cell in O and v = (vx, vy) the velocity of this cell. The sum extends over all cells in O. A is the area of the cell at the center of O. This area is used to set the radius of O to R = 4 A / π. Hence, the movement of neighboring cells up to second order is used to calculate the vorticity. All of our simulation parameters are summarized in Table 1.

The collective dynamics of many cells is more than the sum of its parts. For example, large cell collectives often form streams, swirls or bridges that cannot be achieved by single cells. Yet the dynamic processes of single cells, especially their response to adhesive and mechanical cues, stays an essential element of the collective cell dynamics. Here we introduce a comprehensive modeling framework that allows us to predict cellular dynamics from the level of single cells up to the level of large cell collectives on the same footing. We focus on cellular dynamics on adhesive micropatterns as an especially successful approach to investigate and control complex cell behaviour. Our model successfully predicts a large range of experimentally observed phenomena, allows us to investigate the relative importance of the different cellular processes and in the future can be used to design new adhesive micropatterns that promote desired cell dynamics.

Abstract Introduction Results Discussion Models

cell physiology cell motility ellipses classical mechanics fluid mechanics cell cycle and cell division cell processes cadherins surface tension geometry cell polarity developmental biology mathematics materials science contractile proteins actins cell adhesion proteins materials by attribute continuum mechanics adhesives physics biochemistry cytoskeletal proteins cell biology cell migration biology and life sciences physical sciences

2016

Dynamics of Cell Ensembles on Adhesive Micropatterns: Bridging the Gap between Single Cell Spreading and Collective Cell Migration

Progeny capsids of herpesviruses leave the nucleus by budding through the nuclear envelope. Two viral proteins, the membrane protein pUL34 and the nucleo-phosphoprotein pUL31 form the nuclear egress complex that is required for capsid egress out of the nucleus. All pUL31 orthologs are composed of a diverse N-terminal domain with 1 to 3 basic patches and a conserved C-terminal domain. To decipher the functions of the N-terminal domain, we have generated several Herpes simplex virus mutants and show here that the N-terminal domain of pUL31 is essential with basic patches being critical for viral propagation. pUL31 and pUL34 entered the nucleus independently of each other via separate routes and the N-terminal domain of pUL31 was required to prevent their premature interaction in the cytoplasm. Unexpectedly, a classical bipartite nuclear localization signal embedded in this domain was not required for nuclear import of pUL31. In the nucleus, pUL31 associated with the nuclear envelope and newly formed capsids. Viral mutants lacking the N-terminal domain or with its basic patches neutralized still associated with nucleocapsids but were unable to translocate them to the nuclear envelope. Replacing the authentic basic patches with a novel artificial one resulted in HSV1 (17+) Lox-UL31-hbpmp1mp2, that was viable but delayed in nuclear egress and compromised in viral production. Thus, while the C-terminal domain of pUL31 is sufficient for the interaction with nucleocapsids, the N-terminal domain was essential for capsid translocation to sites of nuclear egress and a coordinated interaction with pUL34. Our data indicate an orchestrated sequence of events with pUL31 binding to nucleocapsids and escorting them to the inner nuclear envelope. We propose a common mechanism for herpesviral nuclear egress: pUL31 is required for intranuclear translocation of nucleocapsids and subsequent interaction with pUL34 thereby coupling capsid maturation with primary envelopment. Morphogenesis of herpesviral capsids is an intricate process initiated in the infected nucleus [1]. A fragile procapsid is formed and packaged with one copy of the viral genome that is generated by cleavage of replicated concatameric DNA molecules. During this process, the rather spherical procapsids change their conformation and mature into the icosahedral and more stable C capsids. These accumulate in large numbers in capsid assembly sites and in the nucleoplasm. Over time, the infected nuclei are enlarged, concurrently the capsids get dispersed, the host chromatin is marginalized, and the nuclear lamina is partially disintegrated [2–5]. How mature capsids are released from sites of assembly, and how they translocate from there to the nuclear envelope is not completely understood, and their mode of transport to the nuclear periphery is discussed controversially [5–9]. With a diameter of 125 nm, herpesviral nucleocapsids exceed the nuclear pore diameter forcing them to take a different route out of the nucleus. Nuclear egress involves primary envelopment of capsids at the inner nuclear membrane (INM) resulting in a transiently enveloped perinuclear particle followed by de-envelopment at the outer nuclear membrane (ONM) and release of capsids to the cytoplasm [10,11]. Nuclear egress of all herpesviruses is mediated by a group of conserved viral proteins. In Herpes simplex virus type 1 (HSV-1), pUL31, a nucleo-phosphoprotein [12], and pUL34, a type II membrane protein [13], are recruited to the INM where they form the nuclear egress complex (NEC; [13,14]). Both proteins are required for nuclear egress of capsids out of the nucleus since deletion of either NEC component leads to their nuclear retention concomitant with a defect in viral propagation [15,16]. Moreover, the NEC recruits several viral and cellular kinases to partially disintegrate the major host barriers, namely the chromatin and the nuclear lamina, and to provide access of capsids to the INM [17–21]. Current data on pUL34 and pUL31 interaction (s) support a temporally regulated and orchestrated sequence of events at the INM, e. g. docking of capsids at the nucleoplasmic face, initiation of membrane curvature, wrapping of capsids by the INM, completion of budding by membrane scission and release of enveloped capsids into the perinuclear space [22–27]. In vivo, co-expression of the two NEC proteins in absence of any other viral protein is sufficient to form and accumulate empty vesicles in the perinuclear space [28,29]. Recently, insights into the membrane-associated NEC activity have been obtained by in vitro systems [30,31]. Recombinant HSV-1 pUL31 and pUL34 form ordered coats on artificial membranes and can induce membrane curving, invaginations, and vesicle formation. Thus, the NEC represents the minimal virus-encoded membrane-budding machinery with an intrinsic activity to drive membrane budding and scission of vesicles [30,31]. During infection, the situation is more complex due to the presence of other viral and cellular factors and their spatio-temporal regulation. Among them are the nonessential HSV-1 protein kinase pUS3 [32], the viral proteins pUL47 [33] and ICP22/pUS1 [34] as well as numerous host factors [33]. In addition to its well documented role in primary envelopment of nucleocapsids [18,25,26,35], pUL31 may assist in viral genome cleavage/packaging [15,36–41] and thus link capsid maturation to nuclear egress. Several studies have reported a preferred nuclear egress of C capsids over A or B capsids ([10,11]; and references therein); however, the molecular mechanism of these sorting events is poorly understood. The minor capsid proteins pUL25 and pUL17 that physically interact with pUL31 [38,39,41,42] are candidates to contribute to this quality control of nuclear capsid egress [10,11,27,43,44]. Orthologous pUL31 proteins share several features. The larger C-terminal domain can be divided into four conserved regions CR1 to 4 [45,46] with CR1 of all pUL31 orthologs containing a binding site for the respective pUL34 ortholog (Fig 1A; [45–47], and references therein); however, additional binding sites are likely to exist in their C-terminal domain [25,26,48–50]. In contrast, the smaller N-terminal domains are variable and enriched in basic residues clustered in several patches (red in Fig 1B). Furthermore, a putative classical bipartite nuclear localization signal (NLS; [45,50–52]) has been identified by in silico analysis (Fig 1A, grey in Fig 1B). To characterize the functions of the N-terminal domain of pUL31 reported to be phosphorylated by the US3 protein kinase [18], we generated a series of HSV-1 mutants with a particular focus on the basic patches (Fig 1C). We identified a classical bipartite NLS embedded in the N-terminal domain that was however not required for nuclear import of pUL31 during HSV-1 infection. Furthermore, we show here that pUL31 and pUL34 entered the nucleus independently of each other via separate routes. pUL31 lacking the N-terminal domain associated with capsids in the nucleoplasm but was unable to support nuclear egress and viral replication. A considerable amount of pUL31ΔN was retained in the cytoplasm if co-expressed with pUL34 suggesting that these proteins had prematurely interacted, and that the N-terminal domain of pUL31 controls the interaction with pUL34. Interestingly, while the C-terminal domain of pUL31 was sufficient to interact with nucleocapsids, the N-terminal domain was required for translocation of capsids from the nucleoplasm to the nuclear envelope and for viral propagation. Together, our data suggest a highly regulated sequence of events during nuclear egress: pUL31 is initially targeted to nuclear sites of capsid assembly and then escorts the nucleocapsids to the nuclear envelope for primary envelopment, a process coordinated by the N-terminal domain of pUL31. Hep2 (ATCC-No. CCL-23), HeLa (ATCC-No. CCL-2) and Vero cells (ATCC-No. CCL-81) were cultured as described previously [53]. HSV1 (17+) Lox was used for all experiments [54,55]. The HSV-1 strain 17+ (kindly provided by D. J. McGeoch) and pHSV1 (17+) Lox [54–56] were used for PCR amplification. HSV-1 propagation, titration and kinetics were done as described previously [13,53]. Plasmid transfection was performed using Effectene Transfection Reagent (Qiagen), while BAC transfection was done using Lipofectamine 2000 (Invitrogen). The yeast 2-hybrid method (Y2H; [53,57]), the NEX-TRAP assay [58] and the LUMIER assay [59,60] were described previously. Cloning was performed by classical restriction or Gateway recombination according to the manufacturer’s protocol (Gateway, Invitrogen). Single base pair exchanges were introduced using the QuikChange Site-directed Mutagenesis Kit (Stratagene) and verified by sequencing. Constructs encoding pUL31 or mutants thereof were cloned into pCR3-N-myc destination vectors. Constructs encoding maltose-binding protein (MBP) -UL34 were cloned into the pCR3-MBP destination vector similar to the plasmid encoding Strep-pUL34 described previously [13]. The plasmids encoding bp1 (basic patch 1), bp2 or bp1bp2 were cloned using the plasmid EYFP (Clontech). Primers used to generate plasmids encoding EYFP-UL31-bp1bp2, -bp1, -bp2, and pUL31-mp1 (mutant patch 1), -mp2, -mp1mp2, -hbpmp1mp2, pSV40NLS-UL31-mp1mp2, pUL31ΔN, and pSV40NLS-UL31ΔN (Table 1) are described in Table 2. The plasmid EYFP-Nuc (Clontech) was used as control. Plasmids used for the yeast 2-hybrid (Y2H) and the LUMIER assays [59] and the plasmid EYFP-FRB-pUL31 [58] have been described before. The HSV-1 UL31 mutants were generated using pHSV1 (17+) Lox [54–56] and a modified galK positive counterselection scheme essentially as described ([13] and Striebinger et al. , in revision). First, the non-overlapping coding region of UL31 (Nucleotides 9 to 865) was replaced by a galK-kan cassette, which had been amplified using the pGPS-galK-kan plasmid and primers equipped with 50bp homologies flanking the UL31 locus (Table 3: H5-UL31/galk and H3-UL31/galk). In a second step, the galK-kan cassette was substituted with a UL31 region encoding either wild type (wt) pUL31, pUL31ΔN, pSV40NLS-UL31ΔN, pUL31-mp1 (mutant patch 1), pUL31-mp2, pUL31-mp1mp2, or pUL31-hyperbasic patch 1 (hbp1) mp1mp2 (Fig 1C; Tables 1,3 and 4). To rescue the ΔUL31/galk intermediate, the galK-kan cassette was replaced by the wt UL31 sequence. To reverse the pHSV1 (17+) Lox-UL31-mp1mp2 to a pUL31 wt sequence, a two-step recombination process was applied resulting in pHSV1 (17+) Lox-UL31-mp1mp2 revertant (rev). For PCR amplification, mutant plasmids that had been generated by site-directed mutagenesis using specific primers (Table 2) were used as templates. Details of BAC mutants are presented in Tables 1 and 4. Direct insertion of the SV40NLS coding sequence at the 5´end of UL31 would have perturbed the 3´coding sequence of UL32. To leave the UL32 coding sequence intact, a BAC was generated in which galK-kan was inserted into the UL31 locus while the original start site of UL31 was inactivated without changing the amino acid sequence of pUL32. Upon insertion of the coding sequence of pSV40NLS-UL31-mp1mp2, the overlapping 8 bp of UL31 and UL32 were duplicated. To reverse the pHSV1 (17+) Lox-SV40NLS-UL31-mp1mp2 to a UL31 wt sequence, a two-step recombination process was applied resulting in pHSV1 (17+) Lox-SV40NLS-UL31-mp1mp2 revertant (rev). This revertant still carries the 8 bp duplication of the 5´ UL31 region as well the mutated original start codon of UL31 (Tables 3–4). All BAC sequences were validated by sequencing of the DNA regions targeted by mutagenesis and by restriction pattern analysis of the entire BAC backbone. The pHSV1 (17+) Lox strains were reconstituted by transfecting BAC DNA into Vero cells using Lipofectamine 2000 according to the manufacturer´s instructions (Invitrogen). Hep2, HeLa or Vero cells grown on coverslips, either transfected or infected, were fixed with 2% formaldehyde/PBS (15 min, room temperature) and permeabilized with 0. 5% Triton X-100 (5 min, 4°C). Binding of antibodies to the HSV-1 Fc-receptor like proteins gE/gI was blocked with human blood sera of HSV-1 negative individuals/PBS for at least 3 h at room temperature [13]. Mouse monoclonal antibodies anti-myc (clone 9E10; kindly provided by J. von Einem), anti-MBP (NEB), anti-ICP0 (Santa Cruz), anti-ICP8 (kindly provided by R. Heilbronn) and anti-VP5 (clone 8F5; kindly provided by J. Brown) as well as rabbit anti-pUL31 and anti-pUL32 antibodies (kindly provided by B. Roizman and J. Baines [49]), anti-gM antibodies (kindly provided by T. Mettenleiter), and anti-pUL34 antibodies [13,53] were used. Anti-mouse and anti-rabbit fluorescently labelled secondary antibodies were from Invitrogen. Cells were examined using a confocal laser scanning microscope (LSM710; Zeiss, Oberkochen, Germany, or TCS SP5; Leica, Mannheim, Germany). Pictures were processed using Adobe Photoshop (Adobe) and Zen-Lite (Zeiss, Oberkochen, Germany). Fluorescence was measured along a 1 pixel thick and 6 μm long line using the" plot profile" tool of the software ImageJ (version 1. 48K) on 8 bit images (Zeiss LSM710) taken with a 63x objective, NA 1. 4, a pinhole aperture of 1 Airy unit, and a pixel size of 78 x 78 nm. To evaluate complex formation between pUL31 and pUL34,3. 5 x 106 HeLa cells were transfected with single plasmids encoding MBP-pUL34, myc-pUL31 or myc-pUL31ΔN, or a combination of a plasmid encoding MBP-pUL34 with one either encoding myc-pUL31 or myc-pUL31ΔN using Effectene according to the manufacturer´s protocol (Qiagen). Twenty-four hours post transfection (hpt), the cells were washed with ice-cold PBS, incubated for 20 min with ice-cold lysis buffer (20 mM Tris-HCl pH8,150 mM NaCl, 10% (v/v) glycerol, 0. 5% (v/v) Triton X-100,2 mM EDTA, with complete Protease-Inhibitor Cocktail (Roche) ). The lysates were pre-cleared by centrifugation (4°C, 12 000 rpm, 10 min) and incubation with Protein A Sepharose beads (GE Healthcare) for 10 min at 4°C. Following centrifugation (4°C, 5300 rpm, 10 min), the lysates were incubated with prewashed Amylose Resin (NEB). After incubation for 1 hour at 4°C on a rotating wheel, the supernatant was removed, and the beads were washed 3x using ice-cold lysis buffer. Proteins were released from the resin by incubation with 4x Lämmli buffer (room temperature, 15 min) and analyzed by SDS-PAGE followed by Western blotting using anti-MBP antibodies and anti-myc antibodies and peroxidase-conjugated secondary antibodies. Vero cells were seeded onto coverslips 1 day prior to infection. The cells were pre-cooled for 20 min on ice, and incubated with HSV-1 at 1 pfu/cell in CO2-independent medium containing 0. 1% (w/v) BSA for 2 h on ice on a rocking platform as described previously [54,61]. The cells were then shifted to regular growth medium at 37°C and 5% CO2 for 1 h. Non-internalized virus was inactivated by a short acid wash for 3 min (40 mM citrate, 135 mM NaCl, 10 mM KCl, pH 3), and the cells were transferred back to regular growth medium. After another 12 h, the cells were fixed with 2% (w/v) glutaraldehyde in 130 mM cacodylate buffer at pH 7. 4 containing 2 mM CaCl2 and 10 mM MgCl2 for 1 h at room temperature. Subsequently the cells were washed and postfixed for 1 h with 1% (w/v) OsO4 in 165 mM cacodylate buffer at pH 7. 4 containing 1. 5% (w/v) K3[Fe (CN) 6], followed by 0. 5% (w/v) uranyl acetate in 50% (v/v) ethanol overnight. The cells were embedded in Epon, and 50 nm ultrathin sections were cut parallel to the substrate. Images were taken with an Eagle 4k camera at a Tecnai G2 electron microscope at 200 kV (FEI, Eindhoven, The Netherlands). For quantitation, images were taken at low magnification (6000x) and merged (Adobe Photoshop) to cover the whole cell area. The capsids in the nucleus and in the cytoplasm were counted and the areas of the nucleus and the cytoplasm were measured (ImageJ). Capsid numbers were calculated per area in mm2. Bioinformatic analysis revealed two patches of positively charged residues composed of RRRSR (basic patch 1; bp1) and RRASRK (basic patch 2; bp2) separated by a linker region within the first 42 residues of HSV-1 pUL31 which resemble a classical bipartite NLS (http: //www. expasy. org/; Fig 1A and 1B; [45,50–52]). To be classified as an NLS, a given sequence has to target an unrelated cytoplasmic protein to the nucleus. In addition, it should mediate physical interaction with transport factors of the importin α/β family [62], and its mutagenesis should result in a cytoplasmic localization while re-addition should restore the nuclear residence [62]. EYFP-pUL31-bp1bp2 comprising only residues 21 to 42 of pUL31 (grey in Fig 1B) fused to EYFP was as efficiently targeted to the nucleus as EYFP-SV40NLS (Fig 2A). Both bp1 and bp2 of pUL31 were able to individually target EYFP to the nucleus although less efficiently than the combination of both (Fig 2A) while EYFP alone was located to both cytoplasm and nucleus. Yeast 2-hybrid (Y2H; Fig 2B) and LUMIER experiments (S1A Fig) furthermore demonstrated a physical interaction of pUL31 with transport factors of the importin α family [62]. While pUL31, pUL31-mp1 (mutant patch 1; Fig 1C) as well as pUL31-mp2 (mutant patch 2; Fig 1C) interacted with importins (Fig 2B), pUL31-mp1mp2 did not (Fig 1C; Fig 2B; S1A Fig). Thus, the residues 21 to 42 of pUL31 constitute a classical bipartite NLS that can mediate nuclear import. Its relevance for nuclear import of pUL31 was analyzed by transient expression of myc-tagged pUL31 or mutants thereof (Fig 2C and 2D). pUL31 was exclusively located to the nucleus (Fig 2C and 2D) consistent with previous results [25,35,49,63,64]. Mutant pUL31 with either the first (pUL31-mp1) or the second basic patch (pUL31-mp2) mutated were also located to the nucleus (Fig 2C). pUL31-mp1mp2, with three basic residues in each of the two patches being replaced by neutral residues showed a more pancellular distribution, while adding an SV40NLS to its N-terminus restored its nuclear localization (Fig 2C). An additional exchange of a single residue G10R generated a hyperbasic patch (hbp) identical to residues 21 to 25 (Fig 1C). The resulting pUL31-hbpmp1mp2 was located to both cytoplasm and nucleus, similar to pUL31-mp1mp2 (Fig 2C), indicating that such an artificial basic patch did not rescue nuclear import. pUL31ΔN that lacked the N-terminal 44 residues (Fig 1C) remained cytoplasmic; again its nuclear import was rescued by adding an SV40NLS (Fig 2D; [62]). To reveal any potential export activity of pUL31, we used the NEX-TRAP (nuclear export trapped by rapamycin) assay [58]. EYFP-FRB-pUL31 was exclusively located in the nucleus both in the absence or presence of rapamycin (Fig 2E). pUL31 was unable to reach the cytoplasmic gM-FKBP for rapamycin-induced dimerization at the TGN and thus lacked any export activity (Fig 2E), a finding further corroborated by the interspecies heterokaryon assay [58]. In summary, we conclude that HSV-1 pUL31 harbors an import activity within the N-terminal variable domain, but no export activity. The import activity of pUL31 is composed of a classical bipartite NLS and an unrelated import activity that together mediate the very efficient nuclear import of pUL31. Next we determined the subcellular distribution of the different pUL31 variants in the presence of pUL34. Strep-tagged pUL34 expressed alone was located in cytoplasmic structures and the nuclear envelope (Fig 3A, left; [13,35]). Upon co-expression with pUL31, pUL34 was exclusively targeted to the nuclear envelope while pUL31 was predominantly located in the nucleoplasm (Fig 3A, right) consistent with previous reports [13,35]. Co-expression of pUL31-mp1mp2 and pUL34 resulted in localization of both proteins in the cytoplasm (Fig 3A, right). pSV40NLS-UL31-mp1mp2 co-expressed with pUL34 however was targeted to the nucleoplasm indicating its nuclear import (Fig 3A, right). In contrast, upon co-expression of pUL31ΔN or pSV40NLS-UL31ΔN with pUL34, both proteins were predominantly located in the cytoplasm (Fig 3A, right). Thus, while the addition of an SV40NLS restored nuclear localization of pUL31ΔN in the absence of pUL34 (Fig 2D), this was not the case in the presence of pUL34 (Fig 3A, right). The nucleoplasmic distribution of pUL31 even upon co-expression with pUL34 (Fig 3A, right) suggested that the interaction of pUL31 and pUL34 might be regulated. pUL34 is a tail-anchored membrane protein while pUL31 per se is free to move between cytoplasm and nucleus. The current model for transport of integral membrane proteins to the INM [65] predicts that once anchored in the membrane of the endoplasmic reticulum (ER), pUL34 would be transported laterally along the ER membranes to the outer nuclear membrane (ONM), and the pore membrane (POM), eventually passing the peripheral nuclear pore channels to reach the INM. Transmembrane proteins with cytoplasmic domains above 60 kDa are too large to pass the peripheral nuclear pore channels [65]. To determine the mode of nuclear import of pUL34, pUL34 was fused N-terminally to the maltose-binding protein (MBP) thereby enlarging its cytoplasmic domain to about 60 kDa. Similar to Strep-pUL34 (Fig 3A, left), transiently expressed MBP-pUL34 was targeted to cytoplasmic structures resembling the ER (Fig 3B, left). As shown above, pUL31 expressed alone was exclusively located in the nucleus, pUL31ΔN essentially remained cytoplasmic while pSV40NLS-UL31ΔN was also nuclear (Fig 2D; Fig 3B, left). Upon co-expression of MBP-pUL34 and pUL31, MBP-pUL34 remained cytoplasmic whereas pUL31 was exclusively localized in the nucleus (Fig 3B, right). Thus, MBP-pUL34 was inserted into membranes already in the cytoplasm, but could not enter the nucleus due to its enlarged cytoplasmic domain. In contrast, pUL31 was efficiently imported into the nucleus. Interestingly, a different situation developed upon co-expression of MBP-pUL34 with pUL31ΔN or pSV40NLS-UL31ΔN (Fig 3B, right). With or without an NLS, a considerable amount of either pUL31 protein was retained in the cytoplasm, a finding reminiscent of the results obtained with Strep-pUL34 (Fig 3A, right). This suggested that in the wild type situation, the interaction between pUL34 and pUL31 is prevented in the cytoplasm. In absence of the N-terminal domain however, pUL31ΔN interacted prematurely with pUL34 and/or other components thereby retaining both proteins in the cytoplasm. To gain further insight into the interaction of pUL31 with pUL34, we used Y2H (Fig 3C) and LUMIER assays (S1B Fig). As expected, pUL31 physically interacted with pUL34 (Fig 3C; S1B Fig). The same was true for pUL31-mp1mp2 and pUL31ΔN (Fig 3C; S1B Fig) consistent with the notion that pUL31-CR1 and potentially other regions of the C-terminal domain contribute to the assembly of the NEC complex [25,26,48,49]. Interestingly, the N-terminal domain of pUL31 also interacted with pUL34, either alone or in co-operation with the neighboring CR1 of pUL31 (Fig 3C). To determine whether pUL31 and pUL31ΔN interacted directly with MBP-pUL34, co-affinity purification was performed. pUL31 or pUL31ΔN were transiently expressed either alone or together with MBP-pUL34. Both myc-pUL31 and myc-pUL31ΔN were co-purified with MBP-pUL34 but not with the Amylose resin alone (Fig 3D). Thus, both proteins had retained the ability to interact with MBP-pUL34 and did so in a specific manner. Interestingly and consistent with previous reports [66], in absence of pUL34 or if spatially separated from it, pUL31 appeared unstable (Fig 3D) while this seemed different with pUL31ΔN (Fig 3D). Taken together, these data show that the NEC proteins pUL34 and pUL31 utilize different transport routes to the nucleus. Most importantly, the presence of a functional N-terminal domain prevents pUL31 from interacting prematurely with pUL34 in the cytoplasm. Previous data suggested a role of the N-terminal domain of pUL31 in viral replication [18]. To analyze the function of the pUL31 N-terminal domain in the context of an HSV-1 infection, we generated pHSV1 (17+) Lox-ΔUL31, Lox-UL31ΔN, Lox-SV40NLS-UL31ΔN, Lox-UL31-mp1, Lox-UL31-mp2, Lox-UL31-mp1mp2, Lox-SV40NLS-UL31-mp1mp2, and Lox-UL31-hbpmp1mp2 using BAC mutagenesis (Fig 4A, 4C and 4E). A rescue mutant was generated for HSV1 (17+) Lox-ΔUL31/galK-kan (Fig 4B), and revertants were made for Lox-UL31-mp1mp2 as well as Lox-SV40NLS-UL31-mp1mp2 resulting in Lox-UL31-mp1mp2 rev (Fig 4D) and Lox-SV40NLS-UL31-mp1mp2 rev (Fig 4F), respectively. All mutations were verified by restriction digest and sequencing of the mutated regions of the BAC DNAs. Next, the BAC-DNAs of the respective mutants or the parental pHSV1 (17+) Lox were transfected into Vero cells (Fig 5A). pHSV1 (17+) Lox readily formed plaques surrounded by cells expressing the HSV-1 immediate early protein ICP0 (Fig 5A). Consistent with an essential function of HSV-1 pUL31 [15,34,67], transfection of pHSV1 (17+) Lox-ΔUL31 resulted in single cells expressing ICP0 while no plaques were formed (Fig 5A). Transfection of the pHSV1 (17+) Lox-UL31ΔN or Lox-UL31-mp1mp2 gave similar results (Fig 5A), and the N-terminal addition of an SV40NLS did not compensate the growth defect of either mutant (Fig 5A). In contrast, the revertants pHSV1 (17+) Lox-UL31-mp1mp2 rev and Lox-SV40NLS-UL31-mp1mp2 rev formed plaques as efficiently as the parental strain thus indicating the integrity of the BAC backbone (Fig 5A). These results furthermore demonstrate that the N-terminal addition of the SV40NLS had not impaired pUL31 function. The 3´coding region of the essential UL32 gene overlaps with the 5´coding region of UL31 (Fig 4A; [68]). When Vero cells had been transfected with pHSV1 (17+) Lox, Lox-ΔUL31, or Lox-UL31-mp1mp2, expression of UL32 was comparable and replication compartments appeared normal as indicated by the subcellular localization of ICP8, the HSV-1 single strand DNA binding protein (S2 Fig; [3,5, 8]). Thus, pUL31 mutagenesis had not affected pUL32 expression and function. Nevertheless, the UL31 mutant strains were unable to spread and to form plaques. Together these data show that the N-terminal domain of pUL31 and its basic patches are essential for plaque formation. Most importantly, the addition of the SV40NLS to pUL31-mp1mp2 or pUL31ΔN did not restore plaque formation, suggesting that the N-terminal basic patches of pUL31 convey additional functions beyond merely mediating nuclear import. To further analyze the role of the N-terminal basic patches the mutants pHSV1 (17+) Lox-UL31-mp1 or Lox-UL31-mp2 were generated (Fig 1C) and transfected into Vero cells; the resulting plaques were comparable to those of the parental BAC (Fig 5A). Thus, either of the authentic single basic patches was sufficient for virus replication. While the addition of the SV40NLS did not compensate the pUL31-mp1mp2 mutation (Lox-SV40NLS-UL31-mp1mp2 in Fig 5A), Lox-UL31-hbpmp1mp2 with the single mutation G10R that generated a sequence identical to basic patch 1 formed plaques (Fig 5A), although they were considerably smaller than those of the parental strain (Fig 5A). Thus, the G10R exchange partially complemented pUL31-mp1mp2 and restored function. Viral reconstitution showed that Lox-UL31-mp1 and Lox-UL31-mp2 replicated to parental titers, while the titers for Lox-UL31-hbpmp1mp2 were at least 2 logs lower (Fig 5B). Taken together, a single basic patch was sufficient to partially restore the crucial functions harbored within the N-terminal region of pUL31. Since the artificial SV40NLS did not compensate, the relative position within the N-terminal domain and the exact amino acid sequence are apparently important for the essential function of pUL31. To further decipher the function (s) of the pUL31 N-terminal domain, Vero cells transfected with the parental pHSV1 (17+) Lox or the mutant BACs (Fig 1C; Tables 1 and 4) were analyzed 20 hours post transfection (hpt) using monoclonal antibodies recognizing mature hexon capsid epitopes (mAb 8F5 [69,70] in combination with antibodies directed against pUL31 ([49]; Fig 6A; S3A Fig) or ICP8 (Fig 6B). Confocal fluorescence microscopy analysis showed that all forms of pUL31 were targeted to the nucleoplasm (Fig 6A; S3A Fig). Thus, during HSV-1 infection, both the N-terminal authentic and the SV40NLS were dispensable for nuclear targeting of pUL31. There was no labeling in Vero cells transfected with Lox-ΔUL31 demonstrating the specificity of the anti-pUL31 antibodies (S3A Fig). After transfection with parental pHSV1 (17+) Lox, Lox-UL31ΔN or Lox-UL31-mp1mp2 (Fig 6A), the subnuclear localization of pUL31 appeared punctuate and correlated with the capsid protein VP5 detected by antibodies to mature hexon epitopes (Fig 6A; [69]). While wt pUL31 was located to both nucleoplasm and nuclear envelope, pUL31ΔN or pUL31-mp1mp2 co-localized with capsids in the nucleoplasm, but not with the nuclear rim (Fig 6A). In contrast, ICP8, a marker of replication compartments [8], had a different subnuclear localization than the capsids (Fig 6B). Line histograms revealed that pUL31 and capsids largely co-localized while this was not the case for ICP8 and capsids (Fig 6A and 6B, right panels). Thus, pUL31, pUL31ΔN and pUL31-mp1mp2 could associate with capsids. Upon transfection with pHSV1 (17+) Lox-UL31ΔN or Lox-SV40NLS-UL31ΔN, pUL31ΔN or SV40NLS-UL31ΔN were partially retained in the cytoplasm (S4A Fig), reminiscent of their localization after co-expression of pUL31ΔN or pSV40NLS-UL31ΔN with pUL34 (Fig 3A and 3B). Thus, the absence of the amino-terminal domain conferred partial cytoplasmic retention of pUL31 while nuclear import and targeting to capsids still occurred. Taken together, pUL31 was targeted to capsids present in the nucleoplasm and the C-terminal domain of pUL31 was sufficient to mediate this association. To analyze the subcellular distribution of pUL34 upon mutagenesis of the N-terminal domain of pUL31, Vero cells were transfected with pHSV1 (17+) Lox or the UL31 BAC mutants (S3B Fig; S4B Fig). As expected, in cells transfected with the parental Lox, pUL34 was located to the nuclear envelope. In cells transfected with Lox-ΔUL31, Lox-UL31-mp1mp2, Lox-UL31ΔN, or Lox-SV40NLS-UL31ΔN, pUL34 was also targeted to the nuclear envelope although its distribution seemed more patchy (S3B Fig; S4B Fig). This suggested that in addition to pUL31, other viral and host factors contribute to targeting of pUL34 to the nuclear envelope, a finding also supported by other recent reports [9,33,34]. A single amino-acid exchange within pUL31-mp1mp2 resulting in pUL31-hbpmp1mp2 rescued the functions of the N-terminal domain (Fig 5B). To further define these functions, Vero cells were infected with HSV1 (17+) Lox or Lox-UL31-hbpmp1mp2 at an MOI of 1 (Fig 7A–7G). About 80% of the cells infected with either HSV1 (17+) Lox or Lox-UL31-hbpmp1mp2 expressed ICP0 at 4 hours post infection (hpi) (Fig 7C). Nevertheless, three phenotypes could be distinguished (Fig 7A and 7B): cells with capsids condensed in the nucleoplasm (category I), cells devoid of cytoplasmic capsids but with a dispersed and speckled appearance of nuclear capsids (category II), and cells with both nuclear and cytoplasmic capsids (category III). To quantify these phenotypes, cells were analyzed at 8 hpi (Fig 7D), 10 hpi (Fig 7E), 12 hpi (Fig 7A, 7B and 7F) or 16 hpi (Fig 7G) with a total of 60 infected cells for each condition (Fig 7D–7G). In the majority of cells infected with the parental strain, the nucleocapsids were dispersed throughout the nucleus with a considerable number of cytoplasmic capsids already at 8 hpi (Fig 7D). Only a few cells fell into category I or II while category III dominated, and this phenotype was further enhanced at later time points. In contrast, upon infection with Lox-UL31-hbpmp1mp2, the majority of cells belonged to category I at 8 hpi (Fig 7D). At 10 and 12 hpi, cytoplasmic capsids were detected in about 50 and 60%, respectively, of the cells (category III; Fig 7E and 7F). At 16 hpi, the percentage of cells in category III remained rather constant, at the same time, nuclei containing dispersed capsids (category II) increased, a phenotype rarely observed with the parental virus (Fig 7G). For closer inspection, Vero cells infected with the parental virus were compared to cells infected with Lox-UL31-hbpmp1mp2 (Fig 8A and 8B; S5 Fig). pUL31-hbpmp1mp2 had also been targeted to nucleocapsids (Fig 8A), whereas the nuclear replication compartments containing ICP8 did not co-localize with the nuclear sites of capsid assembly (Fig 8B). Line histograms clearly showed that the capsid protein VP5 co-localized well with pUL31-hbpmp1mp2 but not with ICP8 (Fig 8A and 8B, right panels). Depending on the stage of infection, pUL31-hbpmp1mp2 co-localized with capsids enriched in speckles in close association with the nuclear envelope (Fig 7A and 7B; S5 Fig). However, unlike in cells infected with the parental virus, a clear nuclear rim localization of pUL31 could not be detected [38,39]. The pattern of the subcellular localization of pUL25, a minor-capsid associated protein, was very similar to that of the major capsid protein VP5 and also co-localized with both wt pUL31 and pUL31-hbpmp1mp2 (Fig 8C). Upon infection with Lox-UL31-hbpmp1mp2, pUL34 was as efficiently targeted to the nuclear envelope as with the parental Lox (S6 Fig). To summarize, pUL31 was associated with nucleocapsids that had recruited pUL25 and escorted them to the nuclear periphery, a process that was delayed for Lox-UL31-hbpmp1mp2 concomitant with a defect in viral replication. Thus, basic patches within the N-terminal domain of pUL31 are required for efficient translocation of capsids from the nucleoplasm and to the sites of primary envelopment at the nuclear envelope. For high-resolution analysis, Vero cells were infected with Lox-UL31-hbpmp1mp2, fixed at 13 hpi (Fig 9) and analyzed by electron microscopy. Essentially all intracellular stages of virus maturation had been formed. These included mature capsids in the nuclear matrix (Fig 9A, arrowhead), capsids traversing the nuclear envelope (Fig 9B, arrowhead), early stages of secondary envelopment (Fig 9B, white arrow), and fully enveloped virions in vesicles (Fig 9A–9C, black arrows). Fully matured virions (Fig 9D, black arrow) as well as L-particles (Fig 9D, arrowheads) were located at the extracellular surface of the cells consistent with the production of infectious particles. Quantitation of the number of capsids in the nucleus and the cytoplasm revealed that although the cells contained similar amounts of capsids as after infection with the parental HSV1 (17+) Lox (Fig 9E), the number of cytoplasmic virus particles was significantly reduced after infection with Lox-UL31-hbpmp1mp2 (Fig 9F). Hence the ratio of nuclear to cytoplasmic capsids was significantly increased in cells infected with the mutant (Fig 9G). We thus conclude that essentially all steps of viral morphogenesis occurred in cells infected with the Lox-UL31-hbpmp1mp2. To summarize, HSV1 (17+) Lox-UL31-hbpmp1mp2 formed mature capsids, and pUL31-hbpmp1mp2 co-localized with mature VP5 hexon epitopes and pUL25. However, the escort of the capsids to the nuclear envelope seemed to be delayed for HSV1 (17+) Lox-UL31-hbpmp1mp2, consistent with a reduced production of infectious virions. Thus, the N-terminal basic patches of pUL31 were required for efficient translocation of capsids from the nucleoplasm to sites of primary envelopment. The NEC of the herpesviruses is composed of two conserved essential proteins, called pUL34 and pUL31 in HSV-1, that are required for primary envelopment at the INM and for nuclear egress of newly formed capsids [10,11]. In vivo, their co-expression leads to the formation of empty vesicles in the perinuclear space [28,29]. In vitro systems recently revealed that these two proteins represent the minimal virus-encoded membrane-budding machinery that contains intrinsic activity to drive budding and scission of membrane vesicles [30,31]. The observation that a complex formed between pUL34 and pUL31 exhibits membrane budding activity [30], instantly implies that the NEC activity needs to be spatially and temporally controlled and confined to sites of nuclear egress to enable efficient capsid nuclear egress and at the same time prevent perturbations on cytoplasmic membranes. This study shows that pUL34 and pUL31 utilized separate routes to the nucleus. pUL31 enters the nucleus through the central nuclear pore channels. The N-terminal domain of HSV-1 pUL31 contains a classical bipartite NLS composed of two basic patches that bound importin α and mediated nuclear import congruent with data on the pUL31 orthologs of HSV-2 [51], murine cytomegalovirus (MCMV; [45]), human cytomegalovirus (HCMV; [52]), and pseudorabies virus (PrV; [50]). The bipartite NLS of HSV-1 pUL31, however, was not essential for nuclear import as pUL31-mp1mp2 was also imported into the nucleus, both in absence and presence of other HSV-1 proteins, consistent with results obtained for pUL31 of HSV-2 [51]. Thus, auxiliary modes of nuclear import are likely to exist, in form of additional non-classical import sequences embedded in the amino-terminal domain [51] and likely also by piggy-backing with other viral partners, for example the pUL17/pUL25 complex [38,39,41,71] or ICP22 [34]. In contrast to soluble proteins like pUL31, integral membrane proteins such as pUL34 need to traverse the nuclear pore through its peripheral channels [65] that physically restrict the size of the cyto-/nucleoplasmically exposed domain [65]. pUL31 and pUL34 form a complex of about 60 kDa [30] that, if formed prior to nuclear import, would be too close to the size limitations of the peripheral nuclear pore channels [65]. Thus, a mechanism is required to prevent premature cytoplasmic association of pUL34 and pUL31. Data presented in this study show that pUL34 with its N-terminal domain enlarged by attaching MBP to mimic the size of a pUL34/pUL31 complex, was retained in the cytoplasm, while co-expressed pUL31 was still imported into the nucleus. Recent analysis of the HCMV NEC proteins also shows that pUL53, the ortholog of pUL31, precedes pUL50, the ortholog of pUL34, in nuclear import [52]. Together we conclude that the interaction of pUL31 and pUL34 and their orthologs is prevented in the cytoplasm to allow for their independent import into the nucleus. Strikingly, we observed that in absence of the 44 N-terminal residues of pUL31, both pUL31 and pUL34 were retained in the cytoplasm, and that the addition of the SV40NLS to pUL31ΔN was unable to confer its nuclear import in the presence of pUL34. Furthermore, we show that pUL31ΔN physically interacted with pUL34, and that both proteins co-localized in the cytoplasm consistent with earlier studies [25,48]. Together these data indicate that pUL31ΔN was retained in the cytoplasm most likely by an interaction with membrane-associated pUL34. A NEC complex preformed and retained in the cytoplasm would be unavailable for its essential nuclear functions. In addition, premature NEC formation in the cytoplasm would unleash the intrinsic activity of the NEC to drive budding and scission of membranes [30] with deleterious effects on membrane integrity and function. Our results here suggest that the N-terminal domain of pUL31 provides a mechanism to prevent premature cytoplasmic NEC activity, and that it is critical in modulating the interaction of pUL31 with pUL34. While pUL31ΔN and pSV40NLS-UL31ΔN were partially retained in the cytoplasm, significant amounts of them nevertheless had reached the nucleoplasm, most likely because they were expressed prior to pUL34. Despite the nuclear import of these pUL31 variants, there was a drastic reduction in replication for Lox-UL31ΔN and Lox-SV40NLS-UL31ΔN, implying that neither pSV40NLS-UL31ΔN nor pUL31ΔN were able to support capsid nuclear egress. Thus other functions than the NLS are encoded by the N-terminal domain of pUL31 and critical for viral propagation. We show that once imported into the nucleoplasm, pUL31 was primarily targeted to nucleocapsids consistent with previous reports [35,39,41]. While the nuclear capsids of Lox-UL31ΔN and Lox-UL31-mp1mp2 had recruited the mutated pUL31 proteins, plaque formation was still inhibited. Obviously, the C-terminal domain of pUL31 was sufficient to bind pUL31 to capsids whereas the N-terminal basic patches must contribute essential functions downstream of this event. The mutation G10R generated an artificial basic patch in pUL31-hbpmp1mp2 that partially restored its function: capsids translocated to the nuclear periphery although with lower efficiency than in the parental virus. Thus, a single N-terminal basic patch was critical in promoting capsid translocation and nuclear egress. This mutant thus unveiled a previously unanticipated sequence of events where pUL31 initially interacts with capsids at their assembly sites and then escorts them to the nuclear periphery, a process that is obscured during the fast progression of a natural HSV-1 infection. The basic patches may promote conformational changes of pUL31 and serve as a platform to recruit hitherto unknown viral and/or host proteins that rearrange the host chromatin and mediate capsid transport through the nucleoplasm to the INM [3,5]. Thereby, the capsids may be dispersed and translocated to sites of primary envelopment, as observed in this study and supported by previous findings [3]. The subnuclear localization of HSV-1 pUL31 correlated well with that of capsids harboring mature VP5 hexon epitopes and the minor-capsid associated protein pUL25, a finding further supported by biochemical evidence [38,39,41]. Thus, pUL25 and pUL17 that together form the capsid vertex-specific complex (CVSC) and promote cleavage and packaging of viral genomes into capsids [10,11], could link pUL31 with nucleocapsids to cooperate in maturation [15]. In such a scenario, the C-terminal domain of pUL31 composed of CR1-4 and sufficient for capsid binding may contribute to capsid maturation [38,39,41]. Interestingly, CR2 and CR4 of M53, the MCMV ortholog of pUL31 are also involved in DNA genome cleavage/packaging [37,40]. Furthermore, pUL31 physically interacts with a pUL25/pUL17 complex even in the absence of capsids [38]. Thus, a pre-formed complex of pUL31 and pUL25/pUL17 may bind to capsids during DNA cleavage and packaging. By such a mechanism, the capsid-associated pUL31 may contribute to completion of genome packaging [15,36–41] and selection for primary envelopment at the INM ([10,11], and references therein). The domains of pUL31 that interact with pUL34 seem to be initially masked but must eventually be exposed to drive NEC formation and activity. Upon infection with Lox-UL31-hbpmp1mp2, capsids with mature VP5 epitopes accumulated in close vicinity to the INM suggesting that primary envelopment was compromised if the authentic basic patches had been neutralized. Thus, the basic patches could serve two functions; to trigger the translocation of capsids from the nucleoplasm to the nuclear envelope, and to promote budding, but only in the presence of capsids. Indeed, in vitro data support an active role of the pUL31 basic patches in regulating the budding process; their deletion abrogated NEC activity but not the formation of the complex and its association with membranes [30]. Interestingly, cytoplasmic membranes harboring pUL34 and pUL31ΔN appeared undisturbed suggesting that a pUL31ΔN/pUL34 complex lacks NEC activity. An attractive scenario suggests that the basic patches in the N-terminal domain of pUL31 promote membrane budding by stabilizing a conformational switch within pUL34. In contrast, the non-essential US3 protein kinase seems not to be critical although it phosphorylates several sites within the N-terminal domain of pUL31 [18]. The severe phenotype of Lox-UL31ΔN and Lox-UL31-mp1mp2 contrasts that of HSV-1 strains lacking US3 [19,32] suggesting that pUL31 phosphorylation has modulatory or cell-type specific effects. We and others have shown that several regions of HSV-1 pUL31 interact with pUL34. The CR1 of all pUL31 orthologs interacts with pUL34 [45–47] probably involving pUL34 residues 137 to 181 [25,26,49,67]. As an extension of pUL31-CR1, the N-terminal domain could contribute to the interaction with pUL34. Furthermore, the C-terminal domain of pUL31 seems to be functionally linked to the N-terminal domain of pUL34 [25,48,50]. Upon infection with Lox-UL31-hbpmp1mp2, we observed capsid speckles with mature hexon epitopes that accumulated in close vicinity to pUL34 at the INM. These speckles resembled the phenotype of a C-terminal mutant of M53, the pUL31 ortholog of MCMV, that is also impaired in capsid egress [37] suggesting that the N- and C-terminal domains of pUL31 cooperate to coordinate the interaction with pUL34. Upon association with capsids in the nucleoplasm, pUL31 may undergo a conformational switch from a closed conformation with the N- and potentially the C-terminal domain covering pUL34-interacting regions to an open conformation that allows for interaction with pUL34 and thus capsid docking and budding at the INM. The interaction between pUL31 and pUL34 may even involve multiple and sequential conformational changes established during NEC formation and membrane budding, a process that appears to be highly orchestrated [22–27]. Our understanding of the precise molecular mechanism however must await the structural resolution of the NEC. Our data together with those of others suggest a highly regulated sequence of events during primary envelopment of HSV-1 (Fig 10; [10,11,25–27,35,38,39,41,44]: (A) Newly synthesized, cytoplasmic pUL31 is masked and inhibited by adopting a conformation that prevents premature interaction with pUL34, (B) pUL31 is imported into the nucleus independently of pUL34, but possibly in a complex with other viral and host proteins, (C) pUL31 associates with nucleocapsids and potentially contributes to genome packaging, (D) conformational changes in its N-terminal domain contribute to targeting of mature capsids to the INM, (E) capsid-associated pUL31 interacts with membrane-associated pUL34 to form the NEC, (F) further sequential interactions between several pUL31 and pUL34 molecules induce membrane wrapping until primary capsid envelopment is complete. In summary, our data not only provide insights into the molecular function of HSV-1 pUL31 during translocation of nucleocapsids from the nucleoplasm to sites of primary envelopment. Given the conservation of the many players involved, the sequence of events described here seems to be applicable to all herpesviruses. Transit of nucleocapsids from the nucleoplasm to the INM is undoubtedly a highly complex process [2–8]. Numerous cellular and viral factors are expected to assist and accompany the nucleocapsids decorated with pUL31 in order to pave their way from the nuclear interior to the INM [33,72]. Thus, the results as well as the tools presented are invaluable to identify proteins and to decipher their function involved in nuclear translocation of capsids, a step decisive for propagation of herpesviruses and potentially other viruses relying on nuclear morphogenesis.

Herpesviral capsid assembly is initiated in the host nucleus. Due to size constraints, newly formed nucleocapsids are unable to leave the nucleus through the nuclear pore complex. Instead herpesviruses apply an evolutionarily conserved mechanism for nuclear export of capsids called nuclear egress. This process is initiated by docking of capsids at the inner nuclear membrane, budding of enveloped capsids into the perinuclear space followed by de-envelopment and release of capsids to the cytoplasm where further maturation occurs. Two viral proteins conserved throughout the herpesvirus family, the membrane protein pUL34 and the phosphoprotein pUL31 form the nuclear egress complex that is critical for primary envelopment. We show here that pUL31 and pUL34 enter the nucleus independently of each other. pUL31 is targeted to the nucleoplasm where it binds to nucleocapsids via the conserved C-terminal domain, while its N-terminal domain is important for capsid translocation to the nuclear envelope and for a coordinated interaction with pUL34. Our data suggest a mechanism that is apparently conserved among all herpesviruses with pUL31 escorting nucleocapsids to the nuclear envelope in order to couple capsid maturation with primary envelopment.

Abstract Introduction Materials and Methods Results Discussion

2015

The Herpes Simplex Virus Protein pUL31 Escorts Nucleocapsids to Sites of Nuclear Egress, a Process Coordinated by Its N-Terminal Domain

Hospital-acquired infections (HAIs), including emerging multi-drug resistant organisms, threaten healthcare systems worldwide. Efficient containment measures of HAIs must mobilize the entire healthcare network. Thus, to best understand how to reduce the potential scale of HAI epidemic spread, we explore patient transfer patterns in the French healthcare system. Using an exhaustive database of all hospital discharge summaries in France in 2014, we construct and analyze three patient networks based on the following: transfers of patients with HAI (HAI-specific network); patients with suspected HAI (suspected-HAI network); and all patients (general network). All three networks have heterogeneous patient flow and demonstrate small-world and scale-free characteristics. Patient populations that comprise these networks are also heterogeneous in their movement patterns. Ranking of hospitals by centrality measures and comparing community clustering using community detection algorithms shows that despite the differences in patient population, the HAI-specific and suspected-HAI networks rely on the same underlying structure as that of the general network. As a result, the general network may be more reliable in studying potential spread of HAIs. Finally, we identify transfer patterns at both the French regional and departmental (county) levels that are important in the identification of key hospital centers, patient flow trajectories, and regional clusters that may serve as a basis for novel wide-scale infection control strategies. The emergence and spread of multi-drug resistant organisms threatens healthcare systems worldwide. [1] This is particularly true concerning methicillin-resistant Staphylococcus aureus, vancomycin-resistant enterococci, and multi-resistant gram-negative bacteria such as carbapenemase-producing Enterobacteriaceae (CPE). Spread of CPE is now a global public health problem associated with patient transfers between healthcare facilities within the same country as well as across national borders, as shown in many recent studies. [2–7] In recent years, patient transfer or referral data has been used to construct “healthcare networks” to propose innovative approaches for hospital infection prevention and control. Healthcare networks are cooperative healthcare systems where hospitals and other healthcare centers are linked by shared patients through secondary transfers or referral. [8,9] Rather than being exclusive to one sole hospital, as Ciccolini et al. argue, the extent of hospital-acquired infection (HAI) spread is dependent on the healthcare network connected by inter-institutional patient transfers. [8] Heterogeneous hospital patient populations and the interactions that occur between them and with the community are important in the understanding of the spatial spread of HAI between hospitals across geographic regions. [9] As early as 2007, studies applied more complex social network analysis approaches to reconstructed healthcare networks in order to demonstrate that infection control measures that take into account network properties can decrease the risk for outbreaks. [8,10] Lee et al. consider network properties to assess the individual influence of different hospitals and the impact of hospital proximities on HAI spread on a regional scale. [11] Many studies show that healthcare networks display a community structure. [8,12–14] Network analysis is especially effective in the identification of sensor hospitals for surveillance of HAIs. [15,16] In addition, mathematical models of healthcare networks may serve to inform decision-makers on enhanced coordinated regional and national approaches to infection control strategies, in a context where increasingly centralized healthcare systems favor the spread of HAIs. [8,15,17] Although national healthcare networks are informative regarding novel HAI control strategies, the impact of reconstructing these networks based on a general patient population rather than a HAI-diagnosed patient population has rarely been addressed. In this study, we assess and compare French healthcare networks based on either patients diagnosed with HAIs or the general patient population, in order to better understand the potential implications in terms of HAI spread predictions. To that aim, we perform social network analyses to describe the different patient flow patterns, network topology characteristics, and community clustering structure. More than 10 million hospital transfers were recorded in France in 2014, for a total of 2. 3 million transferred patients, creating a hospital network of 2063 hospitals (nodes) and 50026 patient trajectories (edges) linking them (Table 1). Patients with a HAI-specific diagnosis created a healthcare network of 1266 hospitals and 3722 connections for 13627 patient transfers. A larger population of patients suspected to have an HAI infection formed a healthcare network of 1975 hospitals and 18812 connections for a total of 128681 patient transfers. With the increasing number of patient transfers, the networks increased from an average 5. 88,19. 05, and 48. 05 average connections per hospital (average degree k¯) and an average 2. 31,4. 92 to 14. 02 transfers per connection (average strength s¯) for the HAI-specific, suspected-HAI, and general healthcare networks respectively (Table 1). Overall, the three networks displayed “scale-free” and “small-world” characteristics that indicated the presence of a small number of very highly connected hospitals with high degrees, referred to as “hubs. ” Analyses of the degree, strength, and shortest path length distributions in addition to the small-world characteristics of the healthcare networks are discussed in S1–S3 Texts and S1–S7 Figs. Compared to random networks, we also showed the general network was more clustered and efficient in transferring patients (S4 Text, S1 Table). We identified several high degree hospitals in all three networks with a consistent outlier–the Assistance Publique—Hôpitaux de Paris (AP-HP) –a conglomerate of 39 hospitals predominately in Paris and the Ile-de-France region represented as one hospital code in our database. [18] AP-HP also acted as the most important intermediary hospital system in the networks due to having the highest betweenness centrality measure. The hospitals involved in the patient transfers recorded in the three networks were of various types, including private rehabilitation and postoperative care facilities, acute-care hospitals or clinics, and hospital centers (Table 2). However, the majority of hubs, defined as the top 5% of hospitals by their degree, were large hospitals providing both acute and postoperative or rehabilitation care (67%, 65%, and 88% in the general, suspected-HAI, and HAI-specific networks respectively). In addition, in the general and suspected-HAI networks, hubs were mostly acute-care hospitals or clinics, hospital centers, or university hospitals centers, with many concentrated in the Ile-de-France, Marseille, and Lyon metropoles (31%, 33%, 28%, and 32%, 28%, 30% respectively). In contrast, university hospital centers rather than acute-care facilities dominated the hub hospitals of the HAI-specific network, representing 48% of hubs (Table 2). The hub university healthcare centers, which provided highly specialized services, included the AP-HP, Hospices Civils de Lyon, and the Assistance Publique—Hôpitaux de Marseille (AP-HM); among them there were also university hospitals of other major cities in France. To better understand the role of hub hospitals across the networks, the shared hospitals between the networks were ranked based on their degree, closeness, and betweenness (Fig 1). Overall, when comparing the degree, betweenness, and closeness, the hospital rankings did not differ between the complete set of 1266 HAI-specific network hospitals and these same hospitals in the general network (p = 0. 81, p = 1, p = 0. 99 respectively, Wilcoxon rank sum test), or between the 1975 suspected-HAI network hospitals and the same hospitals in the general network (p = 0. 99, p = 1, p = 0. 99, Wilcoxon rank sum test). For comparison and illustration purposes, we showed that random rankings for degree, betweenness, and closeness of all hospitals differed significantly between patient specific networks and the general network (p < 0. 05 respectively, Wilcoxon rank sum test) (Fig 1). Suspecting that the differences between rankings might exist between subsets of hospitals, we tested the differences between rankings on an increasing subset of shared hospitals, starting with the highest rank, adding the next ranked hospital, and testing for significant differences. As a result, we determined the range of hospital rankings across the networks where the rankings significantly differed. We defined significant differences as Wilcoxon rank sum test p-values under the 5% alpha risk which we represent as a grey area in Fig 1. Distributions of these p-values are provided in S8 and S9 Figs. For the HAI-specific network, the range of statistically significant degree ranking differences were observed between the 24th ranked hospital to the 1159th ranking hospital. For the suspected-HAI network, statistically significant degree ranking differences were observed between the 405th ranked hospital to the 1078th ranked hospital. For hospital rankings based on betweenness and closeness centrality measures, the hospitals ranked with highest and lowest centralities in the general network were also the hospitals ranked with highest and lowest centralities ranking in the HAI-specific and suspected-HAI networks. Even though hospital rankings of all hospitals did not differ, the majority did differ for betweenness ranks between the 33rd highest ranking to the 1183rd ranking in the HAI-specific network and the 71st highest ranking to the 1757th ranking in the suspected-HAI network (p < 0. 05, Wilcoxon rank sum test). Closeness rankings differences were observed for almost all rankings after the first 3 rankings in the HAI-specific network and after the first 6 in the suspected-HAI network. The lack of statistically significant differences for the highest rankings may have been only due to insufficient power and for lowest hospital rankings due to a series of repeating small closeness values. With this method, we highlight that differences do exist for subsets of hospitals, but we also observe that the most highly connected hub hospitals were consistently highly connected across the networks, irrespective of the different patient population that connected them. To further assess patient movement patterns in the networks, we investigated how our healthcare networks displayed “community” or hospital clustering structure. We compared hospital communities detected with two different community clustering algorithms: 1) the Greedy algorithm [19] that selected members of the communities to maximize the density of links between vertices as it reconstructed the network one vertex at a time and 2) the Map Equation algorithm [20], based on network structure-induced movement using a flow-based and information-theoretic method, detecting communities by measuring probability flows by taking into consideration the directionality and weight of the edges. In general, we detected fewer communities with the Greedy algorithm given that it seeks to maximize modularity–a value that measures the density of links inside communities by comparing the fraction of edges within the communities to the fraction in a random network; a maximum value of 1 corresponds to a network structure with the highest strength possible–as a result, the algorithm searched for the repartitions that maximized the density of the edges. [21–23] The Greedy algorithm considered pairwise interactions and the formation of the network whereas the Map Equation considered the interdependence of links and the dynamics of an already formed network. For each network, we calculated the modularity, the number of communities, community size, and average community clustering distance using the Greedy and Map Equation community detection algorithms (Table 3). For each community, the pairwise clustering distance was calculated as the average geographic distance between all pairs of hospitals of the same community in kilometers. Compared to the general healthcare network, the patient-specific networks had more communities. In the HAI-specific network, there were on average 35. 17 hospitals per community (SD = 44. 31) and 31. 40 kilometers between pairs of hospitals in the same community (SD = 25. 60 km). In the larger networks, the larger community sizes resulted in a higher average distance between community hospitals (41. 60 km (SD = 34. 71) and 39. 01 km (SD = 45. 63) for the suspected-HAI and general healthcare network respectively). For the Map Equation-based communities, as the number of communities decreased from the HAI-specific to suspected-HAI to the general healthcare network, the average community size and average community distance between hospitals of the same community increased (Table 3). Overall, the suspected-HAI network was more similar to the general network than the HAI-specific network in terms of community structure (S5 Text). The regional community clustering using the Greedy algorithm in the three networks are represented in Fig 2. The hospitals in communities were geo-localized, color-coded, and labelled across the networks according to the administrative region (s) they encompassed. We observed that the Greedy-based communities accurately reflected the French regional administrative structure (Fig 2). The identified community clusters formed hospitals communities in which most of the patients were shared between hospitals of the same region frequently centralized towards the hub acute-care centers, university hospital centers, and general hospital centers. On the other hand, the Map Equation-based communities displayed geographic community clustering at the French “departmental” or county level–the administrative division between the administrative region and the municipalities, similar to “counties” or “districts”; of which 96 departmental divisions are present in continental France. The vast majority of these departmental-level community clusters were acute-care centers followed by university hospitals centers and long-term care facilities. To further understand the community structure, we constructed intercommunity networks by combining patient flows between hospitals of the same community and across communities. The Greedy-based intercommunity network was composed of 18 nodes representing the sum of all patient transfers that occurred between hospitals of each community with 306 regional transfer trajectories (Fig 3A). Out of the 22 French metropolitan regions in 2014,4 pairs of 8 metropolitan regions were combined in this intercommunity network (Picardie and Champagne-Ardenne, Auvergne and Limousin, Aquitaine and Poitou-Charentes, and Bourgogne and Franche-Comté). The network was completely connected. All regional communities were connected to one another with an average of 4590 patients moving within these intercommunity trajectories over the year. Certain trajectories played a larger role in patient movement, notably Ile-de-France which admitted the largest number of patients from neighboring regions Picardie and Champagne-Ardenne (4772 transfers) and Centre (3205 transfers) where healthcare hubs were most concentrated. The subsequent largest traffic came from the Rhone-Alpes, the second largest regional center around the city of Lyon, which discharged patients to its neighboring regions (1482 transfers to neighboring Bourgogne and Franche-Comté and 1342 transfers to neighboring Provence-Alps respectively). Nonetheless, the greatest amount of transfers (93%) occurred within the communities themselves on average with up to 98% of transfers occurring within Ile-De France for instance. Although most of these transfers occurred within the communities, the regions remained highly interconnected and certain trajectories played an important role in the interregional and nation-wide movement of patients in France. Building the intercommunity network where community affiliation was determined by the Map Equation algorithm allowed us to consider communities based on the directionality of patient flow, which was overlooked by the Greedy algorithm. The intercommunity network was composed of 113 community nodes with 3215 trajectories with an average degree of 57 and an average of 2597 patients moving between these connections (Fig 3B). Map Equation-based intercommunity communities demonstrated more comprehensive department-level patient flow. Communities were composed of hospitals from many different departments within and across regions; however, the majority of communities were of hospitals within the same department where most of the patient exchange occurred. Concerning the most important transfer routes with the highest traffic, discharged patients coming from many neighboring departments were preferentially going to hospitals in one or a few number of departments, indicating that there was interdepartmental centralization of patient movement. For example, a community composed of 200 hospitals from 9 Ile-de-France departments sent the largest number of transfers (3137 patients) to 28 hospitals of which 24 were from one department (Val-d’Oise). In exchange, this 28-hospital community sent back 2772 patients to the larger community. Overall, patient transfers in Map-Equation communities displayed departmental clustering, but also demonstrated asymmetric movements of patients, concentrating towards small communities of hospitals usually in one department, illustrating the different nature of the communities. Patient sharing patterns and community clustering in the networks were also analyzed based on patient age groups in which new communities were identified (S6 Text, S10–S12 Figs). Moreover, analysis of monthly temporal dynamics of the networks showed that monthly communities may be less clustered and patients may not visit all of the hospitals each month but they still retained the same regional patient sharing patterns seen in the annual network (S7 Text). Having assessed the role of hospitals, hospital communities, and patient trajectories in each network, we considered if the differences in the patient-specific networks and the general networks are due to the number of patient transfers that could lead to structural differences between the specific patient population flows. We first compared the general patient network to two sets of 1000 networks built from a subset of randomly chosen patients: in the first set we selected the same number of patients as the HAI-specific network (21276 patients) at random and in the second set the same number as in the suspected-HAI network (394859 patients) at random 1000 times and reconstructed each network. Overall, both sets of random patient networks (RP) were smaller in size compared the general network in terms of the number of nodes, edges, edge weight, and as a result average degree (Table 4). In addition, most of the diameters and all average path lengths were larger in the RP networks. The diameters and path lengths of the RP networks are skewed and not normally distributed (p< 0. 001, Shapiro-Wilk normality test). As a result, the number of patients used to reconstruct the networks did have an impact of network characteristics. We then compared the characteristics of the HAI-specific and suspected-HAI networks to the RP networks with the same number of patients to assess if HAI patients modified network structure differently than other patients. Overall, the RP networks were larger than their HAI-specific and suspected-HAI healthcare networks analogues meaning that HAI patients were transferred to fewer hospitals than other patients (Table 4). However despite these differences, for some networks measures such as diameter, average path length, and global clustering coefficient, there was less of a difference between the RP networks and the HAI networks than the RP networks and the general network. For example, 63% of suspected-HAI-like RP networks had a diameter equal to or less than that of the suspected-HAI network (64) while 1. 9% of these networks had a diameter equal to or less than that of the general network (30). The average diameter (61. 59) and the average path lengths (3. 78) of these RP networks approached more of that of the suspected-HAI network than the general network. Thus having controlled for the number of patients and thus the size of the network, the differences observed between the suspected-HAI and the general network diameter and average path length may have been due to the suspected-HAI network being a subset of the healthcare network rather than due to differences between HAI patient transfer patterns and non-HAI patient transfers. In this study we show that the French healthcare networks have heterogeneous patient flows, demonstrate characteristics of small-world and scale-free networks, and are characterized with highly centralized movement of patients towards hub hospital centers. Hub hospitals are characterized as university hospitals and private hospitals in the large metropoles that dominate patient flow. The healthcare networks displayed two-level community clustering: regional community clustering reflecting the French administrative structure, and department or county-level clustering. Certain patient transfer trajectories play a more important role in transferring patients between the regional and departmental communities. Despite differences in the patient population and size, both the HAI-specific and suspected-HAI specific healthcare networks seem to rely on the same underlying structure as that of the general healthcare network. Due to weak sensitivity and specificity of the PMSI database to detect nosocomial infections with the sole ICD-10 Y95 diagnostic, the HAI-specific network is not reliable in demonstrating the real patient movement patterns for those infected with an HAI. [24–28] There was no confirmation if an infection was absent during admission and if an infection appeared during the first 48 hours of their stay. We suspect that the degree ranking differences and the low percent of acute-care facilities, notably private hospitals, in the HAI-specific network may be due to differences in coding practicing among hospitals rather than the epidemiology of HAIs. The suspected-HAI network reflects a standardized list of diagnoses used by the French HAI surveillance network which has been shown to be more specific and sensitive at detecting patients with HAIs. [28,29] Having considered the network size differences in the patient-specific networks and the general network, we show that despite the differences in size of the patient population, both the HAI-specific and suspected-HAI specific healthcare networks seem to rely on the same underlying structure as that of the general healthcare network. Indeed, patient-specific networks are a subset of the general patient network and are subject to the same network dynamics. Public university hospital centers and private hospitals in the main metropoles of France dominate patient flow. A study conducted in the Bourgogne region of France has shown that patient flow was centered towards the university hospital that admitted patients from the entire region and based on the regional proximity of the patients’ residence and patients also sought care in two of the closest main healthcare hubs for specialized care (Paris or Lyon). [30] Highly connected hospitals may harbor more MRSA and MRSA bacteremia cases and may have the most potential to transmit HAIs in the entire network. [12,13,31,32] HAIs may spread at a higher rate than expected at random due to the centralization of patient movement and due to the small average number of transfers required for patients to move throughout the network. A 2012 point prevalence study has shown that HAIs are most prevalent in cancers centers, university hospitals, and armed forces. [33] HAI prevalence was high in the Ile-de-France region which has many hubs, and the north-eastern regions which were not reflected by a higher number of transfers in the patient specific. [33] Albeit some difference in prevalence and patient transfer patterns, hubs should be proposed as targets for sentinel surveillance in addition to priority targets of HAI control strategies where HAI is most prevalent to achieve the most effective reduction in transmission across the country. [15] Regional community clustering patterns as a form of network connectedness are also important in the development of strategies for coordinated HAI control. [8,13] Our regional community clustering findings are consistent with that of the healthcare network of England in which communities tend to share more patients among clusters of hospitals in addition to patient flows centered towards a university hospital within the community. [13] Important intermediary trajectories may play a key role in the spread of HAI between hub hospitals and between communities. A study has shown that modifying the number of patients moving between communities may reduce the spread of MRSA. [34] Furthermore, we demonstrated that a two-tier hospital community exists. Depending on the clustering algorithm used, we identified clustering of healthcare communities at the regional level, consistent with the French administrative regions, and department-level communities and inter-departmental hospital clusters that took into account the directionality of patient flow. Coordinated department-level control such as screening of patients based on the identification of key department-level cluster admissions may be the first line of defense against HAI spread within the regions before spread reaches the hub university hospitals. We identified differences between department-level communities of the suspected-HAI and the general network that were overlooked at the regional community level. This may be important in distinguishing hospitals with higher potential to harbor HAI patients, with possible consequences in terms of HAI spread prediction. Studies have proposed reducing hospital connectedness in order to reduce the risk of epidemic spread of HAI in networks. [13,35] Decentralization of the healthcare system and more specifically human resource and specialized health services towards the regional and department levels may help reduce the high connectedness of hubs in the metropole centers and redirect patient transfers. France has moved towards regionalization strategies with the creation of regional hospital agencies, albeit not very effective. [36,37] In addition, the number of university hospitals may be insufficient, below that of the UK, a country with a similar population size. We recommend increasing the number facilities providing specialized services and distributing them at the local level to help redirect patient flow and potentially avoid large-scale HAI dispersal. We considered temporal dynamics, masked in a network constructed with data for the entire year of 2014, in which observed that monthly healthcare networks were smaller and displayed less centralized patient flow; hence, infection control strategies–for short-term control–should rely more on the local department-level dynamics to minimize hospital-level outbreaks and transmission to neighboring hospitals. In the long term, regional community dynamics may give us clues regarding the gradual propagation of specific HAI pathogens over time assuming HAI carriage patterns follow that of patient flow patterns in the healthcare networks. Further studies are required to assess the temporal dynamics of HAI spread in networks to identify any potential seasonality patterns of flow and how to prevent emerging multi-drug resistant bacteria from becoming endemic. Our study was subject to certain limitations which should be considered. Many of the university hospitals represent more than one public hospital or healthcare facility due to sharing the same identification number. For example, the largest outlier hub in Paris (AP-HP) represented 39 hospitals, 12 hospitals and 2 specialized health facilities constituted Hospices Civils de Lyon, 9 hospitals make up the university hospital of Toulouse, and 4 hospitals make up the APHM of Marseille. Consequently, university hospital centers accommodated a larger patient population than hospital centers or local hospitals, influencing the network characteristics, which may have led us to overestimate the specific patient movement patterns to and from these centers. However, the high concentration of other hospitals especially hub private hospital centers, armed forces hospitals, cancer centers, psychiatric hospitals, and private post-operative and rehabilitation centers within proximity of these public hospital hubs demonstrates that despite this, major cities such as Paris play the most important role in connecting patients in the national network and that the French healthcare network is a highly centralized system. The healthcare networks did not include patient flow from private nursing homes that have been shown to play an important role in HAI spread. [38–42] Without private nursing homes included in our study, our results only describe the network topology of hospital patient populations which may be both younger, have shorter duration stay, and may spread HAI differently than the complete nursing home population. As a result, transmission dynamics in our networks may over or underestimate average hospital centrality measures, the volume of patient movements, and the speed at which HAI can spread. By considering all HAIs as a whole, our networks and recommendations reflect action for a broad spectrum of HAIs; however, one should consider that specific HAIs can vary in terms of carriage and transmission patterns. In addition, recommendations based on our networks would overlook the potential exposure to community-acquired infections, although these may later spread in hospital settings, leading to healthcare-associated outbreaks. Future studies should consider all potential components of patient exposure to both community-associated and healthcare-associated infections and account for individual exposure histories to these infections. Despite these limitations, our study provides a first description and analysis of the healthcare networks in France. The identified characteristics and community structures may greatly improve future inter-hospital HAI control strategies. The general patient network responds best to informing regional control strategies targeting key patient trajectories and hub hospital centers. We show that the scale-free structure, the number of communities, and their distribution over the country remain qualitatively similar across all networks and that patient-specific networks rely on the underlying structure of the general patient network. Future studies should take into consideration network topology in the prediction of HAI spread and should consider the potential impact of different community definitions for multi-level infection control strategies. The Programme de Médicalisation des Systèmes d’information (PMSI) database, a comprehensive French medico-administrative database of hospital activity and patient discharge information, is used to construct the hospital networks. [24,25] The PMSI database has been used for epidemiological and medical research regarding HAIs. [24–28,43] A lack of sufficient specificity and sensitivity of the PMSI to detect HAIs is highlighted in these studies. Comparison between laboratory data and hospital data shows that the PMSI has limited coverage of detecting nosocomial conditions. [25–28] Hence, Gerbier et al. [28] use patient discharge summaries from the PMSI to detect nosocomial infections in the University Hospital of Lyon in 2006 and 2007 for the identification of HAIs in surgery, intensive care and obstetric units. They compare the PMSI data to a gold standard by systematic review of patient files for those classified under surgery, the Centre de Coordination de la Lutte contre les Infections Nosocomiales (CClin) Southwest surveillance network for ICU patients, and a combination of surveillance data from CClin and patient information data for obstetrics. The list of ICD-10 codes related to nosocomial conditions, which we entitle “suspected-HAIs, ” can be found in S1 Annex. Gerbier et al. find a sensitivity and specificity for case identification of nosocomial infections to be 26. 3% (95% CI 13. 2–42. 1) and 99. 5% (95% 98. 8–100. 0) for the identification of surgical site infections (78. 9% and 65. 7% by expanding the number of diagnostic codes) respectively; 48. 8% (95% CI 42. 6–55. 0) and 78. 4% (95% CI 76. 1–80. 1) in intensive care respectively, and 42. 9% (95% CI 25. 0–60. 7) and 87. 3% (95% CI 85. 2–89. 3) for identification of postpartum infections respectively. [28] Using patient transfer data from 2014, three healthcare networks are reconstructed based on the following criteria: Only direct transfers of patients who are discharged from a hospital and sent to another in another jurisdiction (“transfer”) or those who are discharged from one medical unit and move to another in the same hospital jurisdiction (“mutation”) are included. The hospital discharge summaries reflected the overall hospital stay of patients and a single diagnosis made them eligible without specification if it occurred during admission or at discharge. Patients who are discharged to their residence or deceased in a hospital are excluded. Patients hospitalized in non-continental European departments are also excluded. First, the networks of hospitals and healthcare centers are re-built in silico using patient transfer data to model the potential movement of patients with HAIs from one hospital to another. In the PMSI database, each patient discharge summary contains information on the hospital facility of stay. Each hospital facility is identified by its unique FINESS number (Fichier National des Etablissements Sanitaires et Sociaux). [44] For this study, the FINESS and stay number of each patient discharge summary are used to merge two PMSI databases: one for acute-care and one for long-term care hospitals. Each patient stay is also numbered by order of stay across different hospitals. To create the logical sequence of patient movement, we sort each discharge summary: by patient ID and patient stay number for all observations. The adjacency matrix [21], a graph of N nodes and E edges can be described by its’ N × N adjacency matrix A defined as: Aij={=1ifiandjareconnected=0otherwise In our patient transfer network, nodes (N) are defined as hospitals and edges (E) as the patient trajectories that connect hospitals. We computed origin i and target j hospitals for each patient stay by assessing if for each discharge the patient entered the hospital i as a transfer or mutation and left hospital i as a transfer or mutation. The same is computed for each j hospital. Using the iGraph package for R statistical software, we create the adjacency matrix of all i and j hospitals, including i and j if i did not transfer out any patients but received them and vice versa for j. [45] We also compute the number of patients moving between hospitals i and j, as wij. The sum of the edge weights of the adjacent edges, the weight strength, is given by: siw=∑j∈Γ (i) wij in which Γ (i) is the set of neighbor hospitals of i. [21] Edge weights represent the number of patients within the trajectories between two healthcare facilities. To identify the most important hospitals of a network, a series of centrality measures are calculated. The degree of a hospital, k, is the number of hospitals one hospital is connected to through its patient trajectories [21] defined as: ki=∑jAij The average degree of a network[21] is given by: 〈k〉=1N∑iki=2EN In addition, Aij is a directed graph in which the directionality of patient transfers from one hospital to another is taken into account. Consequently, we can calculate the indegree (deg-) and outdegree (deg+) of any given node in which the degree sum formula is given by: ∑n∈Ndeg+ (n) =∑n∈Ndeg− (n) =|E| Betweenness centrality measures the importance of hospital acting as an intermediary between other hospitals defined as: g (i) =∑s≠tσst (i) σst Where betweenness centrality g (i) is equal to the sum of the σst the number of shortest paths going from s to t through hospital i measuring the importance of hospital i to the organization of flow in the network. [21] The same measure is calculated for patient trajectories defined as: g (e) =∑e∈Eσst (e) σst where edge betweenness centrality g (e) is equal to the sum of the σst the number of shortest paths going from s to t through edge e measuring the importance of edge e to the organization of flow in the network. [21] Two community detection algorithms are used to assess community clustering for each network, which both take into account weighted graphs. [45] A common measure of the quality of partitions of a network into communities of densely connected nodes is modularity. Modularity is a scalar value between the vales of -1 and 1 that measures the density of links inside communities compared to links between them. [21,22] The modularity and different communities of our network are defined using a community detection algorithm. The Greedy algorithm developed by Clauset et al. [19] optimizes modularity as the algorithm relies on network formation and as a result, computes a smaller range of communities as modularity approaches 1; however, the Greedy algorithm does not take into account edge directionality and we detect communities for undirected graphs of the healthcare networks. On the other hand, the Map equation algorithm developed by Rosvall et al. detects communities based on patterns of flow and takes into account edge directionality and the directed graphs are assessed. [20] This algorithm detects communities based on network structure and how it influences the system’s behavior. Based on the community partitioning for each network, the mean geographic distance between hospitals of the same community is measured. To geo-localize hospitals, we used public government data on French hospital facilities and postal code addresses (https: //www. data. gouv. fr/). Using an online batch geocoding server (http: //www. findlatitudeandlongitude. com/), the hospitals’ addresses were converted to latitude and longitude coordinates. A distance matrix was calculated using the haversine formula to measure great-circle distances between all hospitals. [46] Two intercommunity matrices were developed to assess patient sharing between different communities 1) Greedy algorithm-based communities 2) Map Equation-based communities. Based on the algorithm, each hospital node is assigned a community number. A matrix summing the individual hospitals transfers for hospitals that share the same community is constructed and converted into a directed graph. In addition, the mean latitude and longitude are calculated for each community from individual geocodes of the member hospitals. For the Map Equation intercommunity network, the Greedy algorithm is applied to identify the number of communities present when modularity is maximized. Hospitals were ranked by their degree, betweenness, and closeness centrality measures for each network. When the centrality measures were equal, we replaced the rankings by the mean rankings. We tested the differences between rankings on an increasing subset of shared hospitals with the Wilcoxon rank sum test. The test was conducted as follows: starting with the highest ranked hospital in the general network, adding the next ranked general network hospital, and testing for significant differences between the general network rank and either the HAI-specific or suspected-HAI network rank of the same hospital until we compared all shared hospitals. As a result, we determined the thresholds where hospital rankings across the networks start to significantly differ which was defined as Wilcoxon rank sum test p-values under the 5% alpha risk. To compare the networks between each other, we built 1000 random patients networks from the general network. We selected the same number of patients as either the HAI (21276 patients) or suspected HAI networks (394859 patients) from the general patient network at random and reconstructed these networks using their hospital discharge summaries. We calculated various network measures and the proportion of random patient networks that had values greater than, equal to, or less than the general patient network and the respective patient-specific networks.

Hospital-acquired infections (HAIs), including emerging multi-drug resistant organisms, threaten healthcare systems worldwide. Efficient containment measures of HAIs must mobilize the entire healthcare network. Thus, to best understand how to reduce the scale of potential HAI epidemic spread, we explore patient transfer patterns in the French healthcare system. We construct and compare the characteristics of three different patient transfer networks based on data on transfers of patients with diagnosed HAIs, suspected HAIs, or of all patients. Our analyses show that these healthcare networks, the patient populations that comprise them and the patient movement patterns are heterogeneous and centralized. Despite the differences in patient populations, the HAI-specific and suspected-HAI healthcare networks have the same underlying structure as that of the general healthcare network. We identify key hospital centers, patient flow trajectories, at both the regional and department (county) level that may serve as a basis for proposing novel wide-scale infection control strategies.

Abstract Introduction Results Discussion Methods

medicine and health sciences sociology geographical locations social sciences health care ethnicities scale-free networks network analysis social networks nosocomial infections patients infectious diseases computer and information sciences hospitals french people centrality people and places france health care facilities population groupings europe

2017

Spread of hospital-acquired infections: A comparison of healthcare networks