Multi-modal Weighted Nearest Neighbors

This is a python implementation of Weighted Nearest Neighbors with some added features. WNN was introduced by Hao et al. in 'Integrated analysis of multimodal single-cell data' as a method to integrate multi-modal single-cell data (CITE-Seq, ATAC-Seq, scRNA-Seq...) into a single space. I did my best to reimplement the method in the pre-print but keep in mind that the original method may change from now and and the final publication, that may create some discrepencies.

Differences between MWNN and WNN

• Support for an arbitrary number of modalities, at the moment WNN supports only two
• Possibility to use radius nearest neighbors instead of KNN

How to use it

  from mwnn.mwnn import MWNN
  
  rna_adata = sc.read("scRNASeq.h5ad")
  prot_adata = sc.read("CITESeq.h5ad")

  sc.pp.pca(rna_adata, n_comps=30)
  sc.pp.pca(prot_adata, n_comps=18)

  mwnn = MWNN()
  #we are using 20 neighbors for both modalities
  mwnn.add_modality(rna_adata.obsm["X_pca"], "rna", 20)
  mwnn.add_modality(prot_adata.obsm["X_pca"], "protein", 20)
  mwnn.fit()

  prot_adata.obsm["mwnn"] = wnn.weighted_similarities
  sc.pp.neighbors(prot_adata, use_rep="mwnn")
  sc.tl.umap(prot_adata)

Installation

  git clone git@github.com:tariqdaouda/MWNN.git
  cd MWNN
  python setup.py

Alternatively you can just copy the mwnn.py in your current folder, altought I do not condone copy / pasting as it a sure way to maintenance hell.

References

• WNN paper: Integrated analysis of multimodal single-cell data
• R implementation from the authors