Hi, I have a quick question about the principle of scVI and totalvi. When considering the decoder of totalvi, it seems that the paper claimed that their proposed method can denoise the rna and protein dataset. I wonder how to interpret the principle of such design, as the reconstruction is still between the generated data and observed data. Furthermore, can we also use the outputs from scVI as denoised feature?
Hi, totalVI models the protein as a mixture of foreground and background. To denoise the expression, we just use the foreground signal so it derives from the best reconstruction. Similarly in scVI and totalVI for the RNA part we sample only the estimated rate of the decoder (see definition of negative binomial in our webpage). For a ZINB reconstruction loss, this estimated rate can strongly differ from the reconstructed signal (removal of zero inflation).
1 Like