Hi all,
i am trying to perform scVI integration on a merged dataset processed in R and normalized with scTransform. I want to use native Python implementation rather than Seurat wrapper, since i want to play around with the actual model, so I converted my object to AnnData.
Can I use scVI on SCT-corrected ‘counts’ (not log1p-ed) as input? There is little information around: Seurat purposely allows for SCT assays as scVI input (see this PR), while reading from here it seems the exact opposite.
Theoretically, it probably won’t break, but SCVI is not designed to have corrected SCT counts, but only to use raw counts for integration (there is an a priori assumption on the input data distribution, and SCVI does its own normalization). You can try, of course, and compare results.
The 2 sources do not really contradict each other; it’s just that the Seurat wrapper can’t really give what you need (raw counts) to you - it doesn’t mean that forcing it with its approx counts is the way to go.
Hi @ori-kron-wis ,
Out of curiosity, I tried using scVI with SCT-corrected counts and it doesn’t break indeed. We are currently comparing performance to other integration methods, together with a scVI run on raw counts.
The 2 sources do not really contradict each other; it’s just that the Seurat wrapper can’t really give what you need (raw counts) to you - it doesn’t mean that forcing it with its approx counts is the way to go.
I understand what you are saying. FYI, I also found this issue in which a contributor admittedly use scVI on SCT. However, regardless of what other people are doing out there, i think that the reason you gave me (unmet data distribution assumptions post-SCT) is enough to resort to the raw counts.