Abstract
Decoding gene expression from epigenomic landscapes remains a fundamental challenge in genomics. We introduce EpiExpr, a flexible deep learning framework that predicts gene expression from 1D epigenetic tracks (EpiExpr-1D) and integrates 3D chromatin interactions (EpiExpr-3D) to capture distal regulatory effects. Leveraging residual convolutional networks and graph neural networks, including graph attention and graph transformer models, EpiExpr models both local and long-range regulatory influences. Applied to GM12878 and K562 cells, EpiExpr-1D and 3D improve gene expression prediction relative to reference approaches. Analysis using CRISPRi-FlowFISH validated enhancers confirms that EpiExpr-3D accurately prioritizes regulatory elements, compatible with activity-by-contact scores. Remarkably, EpiExpr achieves performance comparable to DNA sequence-based transformer models without requiring sequence embeddings, offering a computationally efficient alternative. This approach provides a scalable, multi-resolution framework (https://github.com/souryacs/3CExpr) for dissecting the contributions of epigenetic modifications and 3D genome organization to gene regulation, enabling broader application across cell types and experimental settings.