Release

[Zethson]

I'd like to write an annbatch use case which will link to or even use cellink. Could we please have a release on pypi? Note that https://pypi.org/project/cellink/ exists so you might need to rename the project.

[LArnoldt]

Hi @Zethson, please note that we were already developing this for a specialized use case since a few weeks with @selmanozleyen, just merged today, please see here: https://github.com/theislab/cellink/blob/main/src/cellink/tl/external/_livi_annbatch.py

In my mind the MILDataloader could technically also use annbatch, both for single-cell MIL cases and genotype + single-cell. However, the latter is not very common and more model bound than IO-bound since not a lot of people train models with more than a few thousand SNPs. For DNA language models specialized dataloaders exist, and they don't use single-cell in these models yet. Single-cell alone quite a few people do, but this is often with a single patient per batch and limited data sizes, so I'm not sure, this would be worth it.

What was the scenario you were envisioning?

[Zethson]

I am basically interested in the simplest way to enable https://annbatch.readthedocs.io/en/stable/tutorials/genetics.html which currently requires an unreleased version (because there still isn't a release) of celllink. If you think livi_annbatch is a better alternative than what I've cooked up there, I'll have a look and try it out. But even this would of course require a release of cellink.

If you feel like this whole use case that you had done in your paper isn't valid in the first place and I should instead look at different dataloaders, I'm also happy to learn.

Thanks!

[LArnoldt]

Hi @Zethson

The dataloader use cases I mentioned are all for paired genetics + single-cell use cases (So is the LIVI dataloader I shared). For the paper we added a use case using genetics only, e.g. for neural-network based polygenic risk scores or rare variant association testing. Here the restrictions I mentioned (e.g. limited number of samples) don't apply and annbatch is useful. cellink also provides the preprocessing for these genetics formats, which is then converted to a standard AnnData format and the data processing is following exactly the annbatch standard processing tutorials. Happy to release cellink to pypi in the next days (will comment here when done) so you can have a look into the preprocessing and get the annbatch tutorial running. I think there is no necessity to add anything in the cellink repo on annbatch?

[Zethson]

Nope, this sounds good to me. Thanks!