Scanvi best practices

Help scvi-tools
1

Hi,

I’ve been reviewing the latest changes in SCANVI. I noted the removal of pretraining scvi and that the option to do pretraining was moved to the from_scvi_model() method.

Could you elaborate on whether it is still best practice to pretrain with scvi, or whether just using scanvi on its own is better?

Thanks
Charlotte

1 Like
2

It is still best practice to pretrain a SCVI model and then instantiate SCANVI with the from_scvi_model class method. We moved this around for API reasons.

1 Like
3

Thanks for the quick response!

1 Like
4

Does it make a difference if I already add the labels_key when training the SCVI model, i.e.

scvi.data.setup_anndata(adata, batch_key='batch', labels_key="seed_labels")
scvi_model = scvi.model.SCVI(adata)
scvi_model.train()
scanvi_model = scvi.model.SCANVI.from_scvi_model(scvi_model, 'Unknown')
scanvi_model.train()

or if I add the labels only when running SCANVI?

scvi.data.setup_anndata(adata, batch_key="batch")
scvi_model = scvi.model.SCVI(adata)
scvi_model.train()
scvi.data.setup_anndata(adata, batch_key="batch", labels_key="seed_labels")
scanvi_model = scvi.model.SCANVI.from_scvi_model(scvi_model, "Unknown", adata=adata)
scanvi_model.train()

The former method is used in the “seed labelling” tutorial, the latter in the “atlas-level integration” tutorial.

5

@grst either way will work, setup anndata just creates a dictionary in adata.uns["_scvi"] and the labels won’t do anything to SCVI. This will be more explicit in a future release.

1 Like
6

Hi,

I have a similar question. If I have a scvi model trained with one labels_key, but then later want to use this scvi model to create a scanvi model but predicting a different set of labels (a new labels_key) - is this possible? Can I replace the original labels_key?

7

In this case, it’s better not to provide the labels key to scvi, as the only thing this enables is gene-label specific dispersion parameters.

If you do provide the labels_key, what you want will not be possible; however, if you only provide the labels at the time of scanvi initialization it is possible