Hi! I have a sort of theoretical question about the totalVI denoised protein measurements. From the paper, it seems like for the RNA modality, there is a specific library size-based scaling factor, but it doesn’t seem like there’s something equivalent for the protein modality in the model. I was wondering if the resulting denoised protein measurements take some metric of per cell library size into account, or if they should be treated more like raw counts and if I should normalize them by some size factor?
There’s no protein scaling library size analogous to RNA, as normalization is learned internally, by tuning a mixture factor of the foreground and background proteins, per cell, in the latent space.
You should not apply additional size factor normalization then.