Seurat uses log-transformed and scaled data for analysis, Scanpy uses raw, which method is better?

… Hello Scanpy, In @LuckyMD 's amazing paper (https://www.embopress.org/doi/full/10.15252/msb.20188746), Table 1 shows that using raw data to calculate the maker genes of clusters is the appropriate way. But the raw data was not regressed out with mitochondrial genes, gene counts, cell cycle scores…So there will be so many mito genes ranked on the top of the marker gene list. What shall we do with these mito genes, because usually they represent the dead cell-released RNA contaminations?

In Seurat, they did every downstream analysis and plotting by using the log-transformed and scaled data (see below, the scaled dots in Seurat violin plot). Scanpy draws all plots by setting use_raw=True. I’m wondering which method is better?

BTW, logFC will become negative and disappear for the marker genes of clusters when we set use_raw=False in sc.tl.rank_genes_groups(adata, 'leiden', method='wilcoxon'. Please check this https://github.com/theislab/scanpy/issues/2057.

Thanks! Best, YJ

[hyjforesight]

Hello @LuckyMD Thanks for the response! Could you please also check why the logFC becomes negative and disappear for the marker genes of clusters? #2057 Thanks! Best, YJ

Because adata was regressed, gene expression will become negative, cannot be loged.

[LuckyMD]

If you have to regress out covariates, maybe you could do it after log transformation? I’m not 100% sure about this approach either though.