Machine learning, network-based models for gene expression, activity, function

Division Of Basic Sciences - Nci

Investigators

Linked publications, trials & patents

Abstract

We have several subprojects to report on here: 1. We have developed a machine learning tool -- miRSCAPE -- to infer miRNA expression in a sample from its RNA-seq profile. We establish miRSCAPE's accuracy separately in 10 tissues comprising 10,000 tumor and normal bulk samples and demonstrate that miRSCAPE accurately infers cell type-specific miRNA activities (predicted vs observed fold-difference correlation 0.81) in two independent datasets where miRNA profiles of specific cell types are available (HEK-GBM, Kidney-Breast-Skin). When trained on human hematopoietic cancers, miRSCAPE can identify active miRNAs in 8 hematopoietic cell lines in mouse with a reasonable accuracy (auROC = 0.67). Finally, we apply miRSCAPE to infer miRNA activities in scRNA clusters in Pancreatic and Lung cancers, as well as in 56 cell types in the Human Cell Landscape (HCL). Across the board, miRSCAPE recapitulates and provides a refined view of known miRNA biology. miRSCAPE is freely available and promises to substantially expand our understanding of gene regulatory networks at cellular resolution. This work is now published at iScience. We are collaborating with Dr. Chi-Ping Day to apply miRSCAPE to single cell data in mouse melanoma model derived cells to investigate the miRNAs mediating the resistance to immunotherapy. 2. We have been extending miRSCAPE to infer transcription factor (TF) activity in single cells and our preliminary results show substantial improved over widely used tools. 3. We have applied our Network-based transcriptome mining tool PathExt to understanding breast cancer heterogeneity and TNBS therapy resistance. Our study shows greater commonality among BRCA subtype than what the conventional transcriptome-based approach may suggest. We also identify FOXA1 as a key mediator of neo-adjuvent chemotherapy resistance in triple negative breast cancer. This work has now been submitted. 4. We are extending our PathExt tool (originally designed for bulk transcriptomic data) to single cell transcriptomic data as an alternative tool to identify cell type/state markers.