Thank you @PierreBoyeau and @adamgayoso for your replies!
An important note came up in another thread Feature selection and effects on DGE analysis? - #2 by romain_lopez
Related to this, the question is: what are the drawbacks of using all genes for scVI?
A follow-up question: do I get it right that values samples from the model to compare genes are the ρ values (expected frequencies)?