Some suggestion about GNN

[HuangDDU]

Is your feature request related to a problem? Please describe.
I want to develop a new model with Pytorch Geometric which is a GNN library based on Pytorch. I hope you can develop some api to make it more easier.

Describe the solution you'd like

1. Please develop a obsm_field for AnnDataManager to register obsm(n_obs*n_obs) attribute. Because scanpy neighbors save the graph adjacency matrix in obsm as key "connectivities".
2. DataSplitter in scvi.dataloaders maybe also should be updated to support minibatch operation. This minibatch operation is similar to subgraph extraction.

[adamgayoso]

Thank you for the request.

We are interested in this but have no estimated timeline for when this would be available.

[HuangDDU]

Is there anyone involved in solving this problem? If not, can I join this project and solve this problem? I am interested in graph deep learning and hope to integrate this technology into scvi-tools.

[canergen]

Hi, happy to discuss further. Especially with spatial data graphs are critical but also for single-cell it would be cool. I think best way to get started would be to implement a model in external that uses graph. E.g. https://brbiclab.epfl.ch/projects/stellar/ could be a nice start.