Leveraging Gene Regulatory Networks for Single-Cell RNA-Seq Imputation

Aarthi Venkat, Mamie Wang, Sabrina Su, Jeremy Gygi

See Project Outline.
See Lab Notebook.

Each cell represents a graph of genes. The task we want to perform is node prediction for each cell - predicting the gene expression value of a gene that does not have a value due to sc-RNASeq dropout. Features are expressions of other genes for each cell, where the set of other genes can be defined as:

1. Genes considered first-degree neighbors in the gene network graph
2. All genes except target gene
3. A set of random genes of size N, where N is the number of first-degree neighbors the gene has in the gene network graph.

If the gene network graph is a "good" graph, the first-degree neighbors should provide as much or more information than all the genes, and more information than the random set.

Initial code from Dutil et al and Bertin et al Github, altered for the mouse genome, single-cell RNASeq data, and regression task. Primarily using data/, models/, and notebooks/.

Most up-to-date notebook is notebooks/1.2_MLP_Week8_LN_Regression.ipynb.