vitkl
commented
3 months ago We generally don't recommend doing gene imputation because you cannot add information to the data that isn't there ("False signals induced by single-cell imputation" Andrews et al https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6415334/). There are very few valid use cases. I would generally not use imputed counts for analysis - only for visualisation. You need to consider this carefully.
You could decompose spatial counts into likely contributions of several cell types - see tutorial.
You can compute imputation as follows:
# read per cell type averages for more genes and make sure that order of cell labels matches adata_vis.uns['mod']['factor_names']
more_genes_g_fg = pd.read_csv(csv_file, index_col='Unnamed: 0').reindex(columns=adata_vis.uns['mod']['factor_names'], fill_value=0.0).T
# export cell abundance
adata_vis.obs[adata_vis.uns['mod']['factor_names']] = adata_vis.obsm['q05_cell_abundance_w_sf']
# compute average per gene & location weighted by abundance of cell types
more_genes_x_sg = np.einsum('sf,fg->sg', adata_vis.obs[adata_vis.uns['mod']['factor_names']], more_genes_g_fg)
Hi, Can we perform gene imputation using cell2location ? (This is what can do with Tangram)