This GitHub repository contains the neural network tool scover associated with a publication: https://genomebiology.biomedcentral.com/articles/10.1186/s13059-023-03021-9.
Q: What is this?
A: scover is a convolutional neural network (CNN) for de novo inference of cis-regulatory motifs from single-cell data. It finds weights for these motifs across pseudo-bulks and also reports the 'influence' of each motif. The network is written in pytorch.
I have attached a few example jupyter notebooks for in the example_notebooks directory: 1) data generation using scRNA-seq data, 2) extracting promoter sequences, 3) running scover. Note that for the first notebook, the dataset is gzipped (otherwise it would exceed github file size limits), running gunzip Marrow.h5ad.gz will give the file necessary for that notebook.
For the pooling step, we have added an additional notebook with the option to run SEACells instead of our pooling algorithm, but we have not tried to train the model using data generated with SEACells yet.
For plotting, see the figures github.
For the datasets used in the manuscript, please see our Zenodo.
MIT