rank_genes_groups one- or two-sided test

[arutik]

Hi!

Is it possible to do two-sided (or two-tailed) test in rank_genes_groups?

It seems like the tests you can choose from for now (t-test, wilcoxon etc.) are one-sided in that comparing group A to group B produces a list of DE genes that is not the same as in comparison of B to A.

Thank you.

Sincerely,
Anna Arutyunyan

[LuckyMD]

Hi @arutik,

The reason the two DE gene sets are not the same is that sc.tl.rank_genes_groups() only reports genes that are upregulated in one cluster (the one in specified in the group parameter) compared to the other. The test itself should be symmetric if I'm not mistaken.

[arutik]

Hi,

Actually that's not what I've experienced - if you compare with default rank_genes_groups test you get genes with positive and negative logFC, which means that the test reports both upregulated and downregulated genes in that comparison, but again, it's not symmetric - please try on a test dataset for yourself.. Also if I take lists produced by A vs B comparison and B vs A and filter the genes by a common adjusted p-value cutoff (0.05 let's say) I would get lists of different sizes, so all of that makes me think that the tests are not symmetric/two-sided.

[LuckyMD]

Do you have some example code to reproduce the behaviour you are seeing?

[arutik]

Hi, sorry for the delay, sure,

I have my adata object and let's say a key 'group' in the annotation that can either be A or B for each cell, and then I do:
sc.tl.rank_genes_groups(adata, groupby='group', n_genes=1000)

As a result, from the object attributes I can get the tables of DE genes, so I get a table for group A vs rest (which is only group B in this case) and for group B vs rest (which is only group A). Both of these lists contain both upregulated and downregulated DE genes, but are not symmetrical.

Let me know if this is unclear.

Thank you.

Sincerely,
Anna Arutyunyan