Set n_pcs when use_rep is specified in sc.pp.neighbors

• Additional function parameters / changed functionality / changed defaults?

Sorry to file two issues in a row! I've been using Scanpy + scVI for my analysis recently, and am really enjoying it!

Currently, if use_rep is set in sc.pp.neighbors, then n_pcs is ignored. These seems like sane default behaviour to me - if we aren't using 'X_pca', then n_pcs doesn't really make sense.

This can actually be limiting, as I recently discovered. If you calculate a latent representation with scVI with n_latent = 50, you can't then do...

sc.pp.neighbors(adata, use_rep='X_scvi', n_pcs = 25)

...as the neighborhood graph is calculated on all 50 latent variables. If I want to use only 25 - say, as a point of comparison to see which best represents my data - I have to recalculate the latent representation with scVI with n_latent = 25. Basically, if you aren't using PCA, you have to calculate a full reduction for every number of dimensions you are interested in.

I don't think the fix would be that major. The source seems to be in the _choose_representation function, with the directly relevant snippet below:

https://github.com/theislab/scanpy/blob/5bc37a2b10f40463f1d90ea1d61dc599bbea2cd0/scanpy/tools/_utils.py#L48-L58

I think the only change needed would be to have the n_pcs is not None catch in all cases, not just when use_rep = 'X_pca'. Might also generalise the variable name to make it more clear it refers to any reduction, not just PCA. Admittedly, I only just started exploring the code base, so I'm not sure where else _choose_representation is called, or what other impacts this could have.

I have some blue sky time tomorrow, so I'll fork it then and see if I can whip up a pull request. Thanks for the excellent product!

Start in scanpy/tools/_utils.py at _choose_representation and inspect the other places that call it. Check how use_rep and n_pcs are currently handled, then verify that selecting fewer dimensions from a supplied representation works without changing existing PCA behavior. Done means the example using X_scvi and n_pcs=25 uses only 25 latent variables.