satijalab
commented
6 years ago Thanks for the question!
Unfortunately there is not a single answer. In some cases, cell-based normalization fails. This is because cell-normalization makes an assumption that the total ADT counts should be constant across cells. That can become a significant issue if you have cell populations in your data, but did not add protein markers for them (this is also an issue for scRNA-seq, but is significantly mitigated because at least you measure many genes).
However, gene-based normalization can fail when there is significant heterogeneity in sequencing depth, or cell size. The optimal strategy depends on the AB panel, and heterogeneity of your sample.
So I think you are correct to try both on your system, but this highlights the need for new normalization approaches for antibody seq data (stay tuned)
Hi,
in the method section of the Stoeckius et. al. CITE-Seq paper, the ADT CLR normalization is performed cell-wise, while the Seurat genesCLR normalization is performed gene-wise. Is this the 'slight improvement' you mention in the multimodal vignette?
When creating *.fcs files (to identify cells by their classical cell surface markers) from the normalized ADT data I received feedback from our wet-lab scientists that the cell-wise normalization consistently looks much closer to Cytometry data you would expect from identically stained cells while the gene-wise normalized data makes it harder to gate by standard definitions of positive/negative marker expression, e.g. see plots below.
Did the cell-wise/gene-wise switch show better results for downstream analysis?