batch correction on expression counts and embeddings

[zhangnan0107]

Thanks for sharing this dataset. I would like to ask for the data from cellxgene, is the batch correction applied for both the low-dimensional reduction embedding (e.g. UMAP) and the expression counts? Or it's just for the embedding. Thanks : )

[grst]

batch effect correction only applies to the low-dimensional embedding (adata.obsm["X_scANVI"]) and whatever is derived from it (e.g. neighborhood graph, UMAP).

For all downstream analyses, we accounted for batch effects independently by including covariates in the linear models used for comparison.

[zhangnan0107]

Thanks for your reply! I might have 2 follow-up questions about the expression counts in cellxgene data:

1. there are three layers - X, which looks like normalized data, layer count and counts_length_scaled. Not sure if I understood this correctly, count is the raw count from original studies, counts_length_scaled is scaled count for only Smart-seq2 platform data (so raw counts was kept for other platforms?), and may I ask which normalization method is used for X?

2. regarding batch effects, I think it can be added as cofactor in analysis like differential expression. I wonder for the dotplot of marker genes for cell-type annotation like in figure s1, did you also account batch effects in someway, or this is actually based on non-correction counts?

Thanks

[grst]

1. All you say is correct, X is simply scanpy.pp.normalize_total followed by scanpy.pp.log1p on the length-scaled counts

2. the dotplots showning the cell-type markers were not adjusted for batch effects (we are also not trying to make any quantitative claims here)