Downstream applications that use count matrices or embeddings from multiple modalities frequently operate on tensors.
For instance, in the current muon.tl.mofa() interface, there is a way to use or discard observations that are present only in some modalities.
We could consider refactoring that code and implementing these strategies (e.g. obs union with missing data or imputation, obs intersection) as part of the data loader that will make it accessible for all the downstream applications.
Downstream applications that use count matrices or embeddings from multiple modalities frequently operate on tensors. For instance, in the current
muon.tl.mofa()
interface, there is a way to use or discard observations that are present only in some modalities.We could consider refactoring that code and implementing these strategies (e.g. obs union with missing data or imputation, obs intersection) as part of the data loader that will make it accessible for all the downstream applications.