Clustering with scVI latent space

I was just checking out the sc best practices:
Is it feasible to use the scVI latent space as base for computing the neighbor graph?
Or are there any drawbacks using this representation instead of the PCA?

Yes, you can use the scVI latent space to compute the neighbor graph.
latent = model.get_latent_representation()
adata_subset.obsm[“X_scVI”] = latent
sc.pp.neighbors(adata_subset, use_rep=“X_scVI”)

I haven’t had issues using the scVI representation, but I have noticed that the scVI latent space can vary a lot based on the genes that you select to make the model from.