Doublet filtering function

Hi,

I tried the DoubletDetection Python library on my data and got some interesting result. As it can be run directly on a numpy array of count matrix (adata.X), I thought it would be an interesting feature for scanpy.

clf = doubletdetection.BoostClassifier() 
doublets = clf.fit(adata.X).predict()
adata.obs['doublet'] = pd.Categorical(doublets.astype(bool))

[falexwolf]

Hi @JonathanShor,

you don’t need to create a custom API. One point of Scanpy is to provide convenient access via anndata to many single-cell packages around. The only thing needed for that is to provide a very simple interface like this or this or several of the other tools… Simply click on the GitHub links in the Scanpy docs…

If your package works reliably, both the restrictions you mention should in principle not prevent adding your package. Of course, in the future, we want all elements of Scanpy to scale to millions of cells, not just the core tools. But for a lot of people, it’s right now helpful to have a large number of tools available also for relatively small datasets.

The only problem is to avoid cluttering the Scanpy API with virtually any tool there is. Tools in the API should have passed a certain quality check.

Doublet detection is a difficult problem. Already last autumn, we played around with @swolock 's tool but didn’t end up using it - it was good, but in our situation, it didn’t seem to apply (are you eventually going to distribute a package for it @swolock ?). I myself quickly wrote a tool, too, but it didn’t work well. Just yesterday, this appeared. Then there is also this on “empty cell detection”. There are more tools out there, I think…

What I mean is: computationally detecting doublets is still something where the field has not agreed on a consensus. Just like batch correction. Therefore, I would not add a tool tl.doublet_detection or tl.detect_doublets to the API at this stage.

There are two options. Either we create a .beta module of the API for tools that don’t even have a preprint and add your tool and similar cases in the future there. We could make a separate page for that entitled Cutting Edge Beta Tools which advertises these tools for people to try out and play around with it. When you have a solid preprint and/or publication or if you think that your tool should go in the main API anyways 😄, we should add your package as tl.detect_doublets_ONEWORDDESCRIBGINGYOURALGORITHM…

@flying-sheep @gokceneraslan @fidelram @dawe anyone opinions on such cases?

[pinin4fjords]

Thanks @fidelram, that will run the whole Scrublet workflow so will certainly do the trick. But I’d prefer a more Scanpy-integrated approach, which I think I can see how to do from @swolock’s fork.