subsetting / subclustering, use raw

when I select a subset of cells using ad_sub=ad[ad.obs['louvain']=='subcluster_of_interest',:], and then re-apply preprocessing routines, this will use only the genes of ad.X (variable over the entire dataset), but not those that are variable only within the subcluster and might be informative for its substructure even if the variance doesn't pass the cutoff when evaluated over the entire dataset. basically, the set of variable genes can only shrink by subsetting..

I'd propose to either use

tmp=ad[ad.obs['louvain']=='subcluster_of_interest',:]
ad_sub=sc.AnnData(tmp.raw.X,obs=tmp.obs,var=tmp.raw.var)

to "reset" the .X matrix (maybe there's a better way?)
or to make sc.pp.highly_variable_genes work on ad.raw.X

scanpy==1.4.4 anndata==0.6.22.post1 umap==0.3.10 numpy==1.16.4 scipy==1.2.1 pandas==0.25.1 scikit-learn==0.20.3 statsmodels==0.10.1 python-igraph==0.7.1 louvain==0.6.1

Start at the sc.pp.highly_variable_genes entry point and inspect how a subsetted AnnData object and its .raw matrix are handled. Compare the proposed reset-of-.X approach with allowing highly variable gene selection from .raw.X; done means subcluster preprocessing can identify genes variable within the subset rather than only those selected for the full dataset.