Batch correction for joint modality umap

[GirayEryilmaz]

Hi Muon Team!

I have been using muon to analyze some cite-seq data. I wanted to make a joint umap of the two modalities. Which I can do with the pipeline I found in the citeseq data integration tutorial.

batch_key = 'batch'
sc.external.pp.bbknn(mdata['rna'], batch_key=batch_key)
sc.external.pp.bbknn(mdata['adt'], batch_key=batch_key)

mu.pp.neighbors(mdata, key_added='wnn')
mu.tl.umap(mdata, neighbors_key='wnn', random_state=10)

However what I get suffers from batch effects. I had expected the umap to be batch corrected since muon.pp.neighbors works on batch corrected neighborhoods of individual modalities but that does not seem to be the case. Am I missing something?

Notes:

1. For clustering I use muon's leiden which produces better results than scanpy's leiden in terms of batch effects.
2. The umaps of individual modalities are batch corrected successfully.

Thanks!

[ilia-kats]

The WNN algorithm requires both the neighborhood graphs of the individual modalities and the actual data as input, and uses both to calculate the integrated graph. In particular, it predicts a cell's data from its nearest neighbors (as given by the graphs) and compares this to the actual data to determine how much weight each modality should get in the construction of the integrated graph. See the original paper for details.

bbknn creates batch-corrected neighborhood graphs, but the actual data is not corrected/integrated. I'm guessing that this is the reason for your results.