Learning common and specific patterns from data of multiple interrelated biological scenarios with matrix factorization

Abstract

High-throughput biological technologies (e.g., ChIP-seq, RNA-seq and single-cell RNA-seq) rapidly accelerate the accumulation of genome-wide omics data in diverse interrelated biological scenarios (e.g., cells, tissues and conditions). Data dimension reduction and differential analysis are two common paradigms for exploring and analyzing such data. However, they are typically used in a separate or/and sequential manner. In this study, we propose a flexible non-negative matrix factorization framework CSMF to combine them into one paradigm to simultaneously reveal common and specific patterns from data generated under interrelated biological scenarios. We demonstrate the effectiveness of CSMF with four applications including pairwise ChIP-seq data describing the chromatin modification map on protein-DNA interactions between K562 and Huvec cell lines; pairwise RNA-seq data representing the expression profiles of two cancers (breast invasive carcinoma and uterine corpus endometrial carcinoma); RNA-seq data of three breast cancer subtypes; and single-cell sequencing data of human embryonic stem cells and differentiated cells at six time points. Extensive analysis yields novel insights into hidden combinatorial patterns embedded in these interrelated multi-modal data. Results demonstrate that CSMF is a powerful tool to uncover common and specific patterns with significant biological implications from data of interrelated biological scenarios.