Regressing out genes #492

Activity

1. LuckyMD commented on Feb 19, 2019

   

   LuckyMD

   on Feb 19, 2019

   Contributor

   More actions

   Hi,
   I'm not entirely sure how Seurat does it, but I assume you could take the mean expression level (or mean z-score) of a couple of genes, store that in a .var column, and regress that out by sc.pp.regress_out(adata, var_col)?

   Something like this:

   adata.var['genes_of_interest'] = adata.X[:,gene_list].mean(0)
   sc.pp.regress_out(adata, genes_of_interest)
   

   If you want to ensure an equal contribution of all the genes to the gene score without weighting by mean gene expression, you could first use sc.pp.scale() on a copy of the adata object like this:

   adata_tmp = adata.copy()
   sc.pp.scale(adata_tmp)
   adata.var['genes_of_interest'] = adata_tmp.X[:,gene_list].mean(0)
   sc.pp.regress_out(adata, genes_of_interest)
   

   Note that I have not tested this code... so no guarantees ;).
2. ivirshup

   added

   enhancement

   on Jan 24, 2021
3. ivirshup commented on Jan 24, 2021

   

   ivirshup

   on Jan 24, 2021

   Last edited by ivirshup

   Member

   More actions

   This would be very easy to implement with: sc.get.obs_df(adata, keys).

   That would not solve the problem of (I believe) constant genes giving errors.
4. ivirshup

   added

   good first issueeasy first issue to get started in OSS community contribution!easy first issue to get started in OSS community contribution!

   on Jan 24, 2021
5. flying-sheep

   removed

   Enhancement ✨

   on Dec 16, 2024
6. chenyangsu commented on Nov 10, 2026

   

   chenyangsu

   on Nov 10, 2026

   Last edited by chenyangsu

   More actions

   This looks covered now. Since 1.13.0b1 (#4359), regress_out's keys became covariates, which takes anndata.acc references, so genes can be mixed with obs columns:

   from anndata.acc import A
   sc.pp.regress_out(adata, [A.obs["n_counts"], A.X[:, "MALAT1"]])
   sc.pp.regress_out(adata, [A.obs["n_counts"], A.layers["counts"][:, "MALAT1"]])  # from a layer

   The reference says which matrix it reads, so a name that is both an obs column and a gene isn't ambiguous, and an unknown gene raises a KeyError. I checked this on 1.13.0b1 (Python 3.13): regressing on A.X[:, "MALAT1"] gives the same result as copying that gene into .obs first and regressing on the copy, which is still the workaround on 1.12.x.

   Could this be closed? If it helps, I can open a small PR adding a gene example to the covariates docstring and a test for it.

   I used an AI coding assistant to help prepare this; I reviewed and tested it.