CLRm1 across proteins vs cells?

[jfreimer]

When using the analyze function, is there a way to specify if CLR should be run across cells vs across proteins?

Thanks,
Jake

[LTLA]

analyze() just calls computeClrm1Factors(), which will compute a CLRm1 size factor for each cell. Specifically, it computes the geometric mean across proteins within each cell, which is then used to define a size factor that aims to remove biases between cells.

No attempt is made to scale each protein; the scaling is only applied on a per-cell basis, much like how library size normalization is performed. I vaguely remember reading somewhere that people would scale the rows of the ADT matrix so that each protein's abundance was of similar magnitude for easier comparison. I don't think this is necessary:

• It's fine for different proteins to have different abundances. This won't interfere with most downstream analyses that only care about log-fold changes between cells for the same gene. I can imagine some steps that are more dependent on the absolute abundance (e.g., cell type annotation) but I don't think protein-specific scaling would help here.
• It's easier to interpret values without protein-specific scaling, as a reader can easily recover the size of the original counts by just doing 2^[log2-normalized abundance], rather than worrying about some extra scaling factor.
• If the protein-specific scaling is applied before the log-transformation, it won't have much effect on the log-fold changes, and so is harmless/useless outside of altering the effective shrinkage from the pseudo-count.
• If the protein-specific scaling is applied after the log-transformation, this will mess up the interpretation of the log-fold changes. I also think this is counterproductive as it will force a protein with high heterogeneity to "be similar" to a protein with background levels of abundance; for example, if our scaling forced all proteins to have the same variance, we would effectively down-weigh the interesting variance in the former and up-weigh the irrelevant variance in the latter. This is similar to my arguments against scaling in PCA of gene expression data.

Of course, if you disagree, you can just take the log-normalized ADT matrix produced by analyze() and do your own thing. Personally, I would suggest limiting any protein-specific scaling to the runPca() step; this ensures that any scaling will affect distance/neighbor calculations without interfering with the interpretation of log-fold changes in scoreMarkers(), etc.