ctheodoris

/

Geneformer

like 205

Hi !
Thank you so much for your diligent work and patient reply all the time. I have a question about how state and goal cell embedding characterized in cell state mode .

For example, I have a number of state A cells, and a number of state B cells, and I perturbate genes to identify which gene will shift cell from state A to state B.

In my comprehension, state A embed and stated B embed is fixed, initially calculated by averaging their respective number of cell embeddings. So is it necassary to make sure state A cells and state B cells have considerable numbers in my input datasets?

Besides, it also bothers me when numbers of different state are imbalanced. For example, state A is more frequent than state B, so perturbation from ①state A to state B and from ②state B to state A seems not fair. In ①, state A got more cells to perturbate and results seem more convincing. In ②, state B got less cells to perturbate and results seem less convincing.

Except the questions above, I now try to explore the vs null mode. From your document and answer, if my comprehension is correct, I can choose single gene perturbation instead of all genes perturbation and finally obtain pvalue and FDR with vs null mode.

However there seems to be two different senario and I am not sure both of them can work with vs null mode.
①perturbate single gene one by one using in silico perturber, and put all the intermidiate results in the same directory, and then use vs null mode with in silico perturber_stas which will automically compare each gene with the other gene? (what should the parameter 'genes_perturbed=' be then ? 'all' or something else?)
②if perturbate a gene list, what will happen then ?

Beacuse some time I hope to perturbate interested genes to see their impact on cell state shift. The second question may be important cause 'all' genes run too slow.