MANGO: Inferring gene regulatory networks from single cell multiomics

Abstract

Abstract Transcription factors (TFs) and their target gene regulatory networks (GRNs) control immune cell fate decisions during differentiation and responses to environmental cues and signals. To better understand how GRN control immune outcomes and are dysregulated in disease settings, we developed MANGO (Multiomics Aided Neural Graph Ontology), a computational method that integrates single cell gene expression and chromatin accessibility data to infer genome-wide, cell-type specific GRN. MANGO harnesses graph neural networks to generate representations of gene and regulatory link behavior and uses existing interpretable methods, like UMAPs, to visualize the GRN and gather insights about functional neighborhoods of regulators or programs of regulation. By scoring each regulatory interaction and aggregating those with common sources, TF activity can be quantified and compared between networks. We applied MANGO to a human PBMC dataset to define GRN and identify key regulatory, cell-type specific TFs, despite sometimes having low changes in expression. Additionally, we show that MANGO can distinguish cells by regulation network at single cell resolution by feeding each cell's expression and accessibility through its respective cell type's regulation network and projecting the resulting embedding in lower dimensions with a UMAP. Finally, MANGO was used to uncover memory B and T cell GRNs that prime rapid secondary immune responses. Given the increase in single cell multiome datasets, we anticipate MANGO will provide unique insights into GRN in normal and diseased immune responses, and ultimately give rise to patient specific understanding of the immune system and precision medicine. Supported by grants from NIH/NIAID (R01 AI148471) to CDS