rank genes groups errors on less than 2 cells in a category

Hello,

I am using scanpy rank genes groups, and rank genes group filter for differential expression analysis after using a classifier. I often receive errors because statistics cannot be calculated on these types of low count groups. The workaround I have found is to drop these cells from the adata object, and then continue with differential expression.

Is there an existing solution for this that is better? Could we consider adding this as a flag to the function call? What I have in mind is a flag like "ignore_low = True". The flag would operate by taking the passed adata object, applying the 2 cell filtration internally, and performing differential expression as normal on this internal object. It would then append the relevant uns categories to the original adata object before exiting. The threshold could even be passable to make this more general.

What do we think? Is this too niche for this scale of a repository? In principle, I think that forcing these observations to be dropped is not best practice.

Start by locating the rank_genes_groups and rank_genes_groups_filter APIs and reviewing how low-count categories currently reach the statistical calculation. The issue proposes an optional low-count threshold, internal filtering, and preservation of relevant uns categories; done requires an agreed design for that behavior.