All genes or highly variable genes?

[adamgayoso]

@Valentine_Svensson

Maybe there’s some way to create a model structure where you have two representations? One Z Z Z that has a representation based on HVG’s / panel genes, and another Z’ Z′ Z’ that describes variation in the other genes / all genes or something like this?

Yeah you could probably allocate some dimensions of the latent space to decode HVGs and then remaining dimensions decode non HVGs + HVGs? Maybe? Could easily run into identifiability issues with this type of thing.