Clarification on PCs input data to generate cell types

[edg1983]

Hello,

Reading the documentation about cell type object generation, I understand that one should provide normalized counts as tpm and a table with principal components for each cell as X.

In the tutorial, it is reported "perform the initial dimensionality reduction when using only cells of a specific type or subtype to adequately capture the variation within that specific subset."

Does this mean that for each cell type, I need to re-compute PCs only for cells in that cell type? And in this case, do you suggest re-computing highly variable genes when doing that?

Thanks!

[edg1983]

I have another related question.

I have also run integration on my data using SCVI. Thus, I have SCVI latent that I use instead of PCs for neighbor computation, for example.

Should I use these latents or original PCs as the X object in the DIALOGUE cell type?
Or would it be better to use the original PCs or PCs re-computed for each cell type?