Unraveling Epigenomic Landscape: Benchmarking Machine Learning Methods for Expression Prediction

Abstract

Histone modifications and enhancers play crucial roles in gene regulation by modulating the accessibility of genomic regions and influencing gene expression patterns. The interplay between histone modifications and enhancers provides a dynamic regulatory framework that orchestrates diverse cellular processes, including development, differentiation, and disease progression. Understanding the complex relationships between these epigenetic mechanisms and gene regulation is essential for unraveling the intricacies of cellular identity and function.Within this work, we extensively benchmark numerous established machine learning methods in a gene-specific manner, harnessing the EpiATLAS dataset of nearly a thousand cell types compiled by the International Human Epigenome Consortium (IHEC). As part of our benchmarking efforts, we have developed a novel CNN architecture that exploits H3K27ac histone mark signatures measured in a megabase genomic window around each gene. With the exceptional ability that CNNs demonstrate in unlocking complex relationships between regulatory elements and gene expression, we are able to meticulously analyze these models to uncover fascinating genomic attributes linked to gene expression profiles.