Using CellBender on sc/sn-RNA-seq and CITE-seq samples part of multi-sample cohort setting

[browaeysrobin]

Dear @sjfleming,

Thank you for developing this very useful tool and the clear descriptions in the paper. We are looking forward to using it ourselves but I have some questions concerning the use of CellBender on samples (sc/sn-RNA-seq & CITE-seq) in the context of a multi-sample cohort setting.

According to Supplementary Section 2.3 of the paper, it is recommended to set nFPR = 0 "In a cohort setting, it is important to set nFPR = 0 to avoid over-correction beyond the expected noise budget. Using larger values of nFPR naturally imparts a bias on the output by preferentially keeping only the most certain cell counts, which is unsuitable when aggregating data from many samples."

General questions:

• 1) Do you have experience yourself with setting nFPR to 0 compared to the recommended default for non-cohort datasets?
• 2) Will this not retain too much background noise, and thus not entirely solve the issue of spurious DE results due to differences in background noise?
• 3) Would it be really problematic to use the same but low nFPR value (eg 0.01) for all your samples in a cohort if nFPR would not provide sufficient denoising?

Questions concerning CITE-seq data:

• 4) What are your recommendations for CITE-seq samples as part of a multi-sample cohort setting? We were wondering about this since the recommended nFPR value for CITE-seq data is quite high, and thus the "opposite" recommendation than for cohort studies. Our CITE-seq data typically provides ADT data for > 100 antibody features.
• 5) How strongly are the cell calling and denoising of the RNA counts affected by whether CellBender is run on the RNA assay alone or on both RNA+ADT? Will the corrected RNA count matrix be the same?

Thanks a lot!

[benduc]

Dear @sjfleming ,

Thanks for developing Cellbender!

Did you have the opportunity to have a look at this query? I have the same kind of setting and would appreciate some feedback.

Thanks a lot!