JSNMF: Joint analysis of scRNA-seq and epigenomic-seq data via semi-orthogonal nonnegative matrix factorization.

This is the Python implementation of the JSNMF algorithm. Note that this implementation can employ GPU to speed up the code, just simply submit the script to the GPU clusters will work.

You can use the following command to install JSNMF:

pip install JSNMF-py

The main class JSNMF needs to be initialized with at least two anndata.AnnData object, forRNA, and ATAC data, or data from two modalities, respectively. The preprocessed data is stored in RNA.X and ATAC.X. And the true cell labels should be kept in RNA.obs['celltype']. Note that the data preprocessing process is done with R. The default number of maximum epochs to run, i.e. the max_epochs parameter, is set as 200. So it is quite simple to initialize a JSNMF model with the following code:

from JSNMF.model import JSNMF
test_model = JSNMF(rna,atac)

After initializing, run the model is also quite easy:

test_model.run()

The result is saved in test_model.result, which is a dict, and the major output of JSNMF, the complete graph S, can be get with easy access:

S = test_model.result['S']

JSNMF class also has other methods, you can use the help or ? command for more details explanations of the methods.

Please refer to here for a simple illustration of the use of JSNMF model, with clustering and visualization results shown. The guide for obtaining the gene-peak association can be found here. And the tutorial of using some extensions or variants of JSNMF model is here.

If you need the data for the gene-peak association tutorial, please contact me directly.