MultiVI MuData setup rejects unpaired observations despite model supporting missing modalities

[catheriz]

When trying to use MultiVI with a MuData object containing unpaired observations (cells missing from some modalities), scvi-tools now raises:

ValueError: Detected unpaired observations in modality rna. Please make sure that data is fully paired in all MuData inputs. Either pad the unpaired modalities or take the intersection with muon.pp.intersect_obs().

This is surprising, as MultiVI is designed to handle missing modalities during modeling, yet the data validation enforces fully paired MuData objects.

Relevant code:

• src/scvi/data/_manager.py: validates MuData via _check_mudata_fully_paired
• src/scvi/data/_utils.py: raises error if any modality is structurally unpaired
• Recent commit possibly implicated: 1a2d127c1f7e364e21ea92e92c99ee844016a114 ("fix: AttributeError after rich upgrade and MultiVI errors after Mudata upgrade (fix: AttributeError after rich upgrade and MultiVI errors after Mudata upgrade #3776)")

# Prepare a MuData with unpaired observations
# This used to run (<=2024), but now fails immediately after upgrade
# ...
import muon as mu
# mdata: MuData with unpaired cells between rna & atac modalities
...
# Setup for MultiVI:
from scvi.model import MULTIVI
MULTIVI.setup_mudata(mdata)
# Exception raised here

[ori-kron-wis]

I think there is some confusion here with the definition of what is "unpaired".
From what I understand, MultiVI is designed to handle and integrate both paired multiomic data (like RNA and ATAC simultaneously) and single-modality unpaired data (only RNA or only ATAC) - but always when the same cells exist for all modalities. i.e, not in the case of missing cells from one of the datasets.

In such a case, you can easily impute the missing cells with 0's (manually), and it will still work for you. Previosuly, was done automatically with Anndata, which is no longer supported.

MultiVI can impute missing expression and accessibility for the modalities that miss them (see tutorial) - perhaps that's what causes the confusion with the work "unpaired".

Are you familiar with other things?

The latest commit fix you mentioned is about a recent upgrade with the MuData package.

[canergen]

I assume the confusion comes from originally using AnnData, where we padded missing observations with zeros. In MuData, obs can be missing in one modality. However, the dataloader expects all obs to exist and so just pad with zeros. There is no plan to change this behavior so we have consistent positions across modalities.
To clarify one point: MultiVI can handle mosaic integration (one modality is shared such as RNA+ATAC/RNA+protein/ATAC is fine) but not diagonal (no cells with overlapping modality).