Multi-Sample Pipeline and Cell Type Annotation

[flcello]

First of all, thanks a lot for the great package! It's really very helpful.

I need to do cell type annotation for a multi-sample experiment, where I have ATAC and RNA data for the same samples (but not the same cells), with the cells in the RNA data already being annotated. I followed your tutorial for cell cluster annotation, computing gene scores for the AnnDataset that contains all ATAC samples, concatenating with the annotated RNA counts of all samples, and then fitting SCVI / SCANVI models with the modality as batch variable. However, the training curve is jumpy and training always errors out after a while. I also tried using the sample as an additional categorical covariate, but that does not solve the problem. It works quite well though when I do the cell type annotation separately per sample, but I wonder whether I loose some power here for samples with lower number of cells. Do you have any suggestions what would be the best approach here? Thanks!