Use ontological labels

[jkobject]

(ok the title is a bit strong)
Free form text is not the best way to convey cell types. Ontologies exist for these labels which is used by many in the community (cellxgene) and allow a reproducible analysis of cell types.

For foundation models like scPRINT it would allow better comparison of its performance than the current linear sum assignment method (check the latest merged PR about adding scPRINT)

Also it would make all the datasets use similar labels and allow for label transfer interrogations

Tools like lamindb's bionty make conversion easy and all datasets on cellxgene have this format already it shouldn't take more than a few hours to do the transfer.

[mxposed]

Hi Jérémie. Yes, the title is a bit strong : )
In the general case, this task is for label projection, where we assume there is no ontology, and we want to manually annotate 1 dataset and apply these labels to another.

However, you're right, for some cases ontologies exist and people already did the unifying work of applying them to multiple datasets.

What would you say if we added a separate dataset to label projection task with the ontology labels you're using in scPRINT?

[jkobject]

Hello @mxposed

Ahah, sorry about that :)

I think that would be great if we added one such dataset, especially since there are so many coming from the CZI's cellxgene with such ontological labels.

[jkobject]

Something else, if you have some people labelling datasets, it would be very interesting to see the agreement between biologists (thus defining an upper bound on accuracy). Both might also be given different subsets of the dataset

[mxposed]

I think that would be great if we added one such dataset, especially since there are so many coming from the CZI's cellxgene with such ontological labels.

Great! I opened a new task for that (#27), please feel free to add details there. I think we should keep this issue open for now.