I was wondering about two questions:
 Would that make sense to do log2 transformation on the denoised protein value (totalVI output)?
 It looks like the totalVI output denoised protein value are all non-negative, I wonder how this is guaranteed by the model?
Many thanks in advance,
yes as it’s on the scale of the original data (might need to add pseudocount)
The model uses transformations internally (exp or softplus) to guarantee non negativity
thanks so much for the prompt reply!
Would it make sense to further normalise the protein value by the ADT library size as well?
ADT library size is tricky – if you’re using the same exact antibody panel in each dataset it might make sense.