Hi, I have a question regarding RNA-seq sample visualization using PCA. In the DESeq2 vignette, counts data that has been transformed using rlog or vst are directly fed into prcomp for PCA, without further scaling/standardization. What is the reasoning for not scaling/standardizing the transformed counts?

One thought I had, in argument against further standardization, is that there is no distortion of the variance due to unit differences since all gene expression levels are in counts.

Out of curiosity, I ran PCA on the rlog transformed counts as well as the (rlog + standardized) counts, and the overall sample clustering patterns remain the same, but the loadings change. I'd like to do some functional analysis of the loadings, so I just want to make sure I'm analyzing the correct PCA plot. Thanks!

VST and rlog contain scaling for sequencing depth.

I guess you mean scaling in terms of Z-scoring? There are plenty of posts towards scaling in PCA both here on Bioconductor and in other communties such as StackExchange and CrossValidated. There is no general rule or right/wrong here, scaling (or not) simply weights genes differently in PCA. For example, the Bioconductor world (for example the single-cell infrastructure here) typically does not scale data before PCA, while alternative frameworks such as Seurat for sure does, and iirc ScanPy does as well. Choice is yours.