Iterative Random Forests to detect predictive and stable high-order interactions

Genomics has revolutionized biology, enabling the interrogation of whole transcriptomes, genomewide binding sites for proteins, and many other molecular processes. However, individual genomic assays measure elements that operate in vivo as components of larger molecular machines that regulate gene expression. Understanding these processes and the high-order interactions that govern them presents a substantial statistical challenge. Building on Random Forests (RF), Random Intersection Trees (RIT), and through extensive, biologically inspired simulations, we developed iterative Random Forests (iRF). iRF leverages the Principle of Stability to train an interpretable ensemble of decisions trees and detect stable, high-order interactions with same order of computational cost as RF. We demonstrate the utility of iRF for high-order interaction discovery in two prediction problems: enhancer activity for the early Drosophila embryo and alternative splicing of primary transcripts in human derived cell lines. In Drosophila, iRF re-discovered the essential role of zelda (zld) in early zygotic enhancer activation, and novel third-order interactions, e.g. between zld, giant (gt), and twist (twi). In human-derived cells, iRF re-discovered that H3K36me3 plays a central role in chromatin-mediated splicing regulation, and identified novel 5th and 6th order interactions, indicative of multi-valent nucleosomes with specific roles in splicing regulation. By decoupling the order of interactions from the computational cost of identification, iRF opens new avenues of inquiry in genome biology, automating hypothesis generation for the discovery of new molecular mechanisms from high-throughput, genome-wide datasets.