DeepGEP: Deep learning for gene expression prediction from multi-omics in mammals

Abstract

Deep neural networks offer great potential for integrating multi-omics data to predict gene expression and uncover regulatory mechanisms. Here, we developed DeepGEP, an attention-based Long Short-Term Memory model trained on 228 datasets, including RNA-seq, ATAC-seq, and ChIP-seq of four histone modifications (H3K4me3, H3K4me1, H3K27ac, H3K27me3) from humans, pigs, and cattle. DeepGEP outperformed several other machine learning methods, achieving Pearson correlation coefficients (PCCs) of 0.70–0.82, with accuracy improving up to 0.93 after K-means clustering. Attention weight analysis highlighted regulatory regions within ±1000 bp of transcription start sites and revealed that H3K4me3 and chromatin accessibility contributed most strongly to gene expression prediction, while H3K4me1, H3K27ac, and H3K27me3 played less prominent roles. Our study demonstrates the value of integrating chromatin accessibility and histone modifications for accurate cross-species gene expression prediction, providing a versatile framework for multi-omics modeling and advancing understanding of mammalian gene regulation.