ChromNet: A Multi‐Task Learning Framework for Cross‐Cell Type Prediction of 3D Chromatin Interactions Using Epigenetic Signals

Abstract The 3D organization of chromatin plays a fundamental role in gene regulation, cellular function, and disease mechanisms. However, current experimental techniques, such as Hi‐C, remain costly and labor‐intensive, limiting their application in large‐scale and disease‐related studies. To address this challenge, ChromNet is presented, a multi‐task learning framework that integrates epigenetic signals across diverse cell types to enable high‐precision prediction of chromatin architecture. By incorporating noise perturbation and auxiliary classification tasks, ChromNet improves the identification of topologically associating domains (TADs) and cell‐type‐specific chromatin structures, demonstrating superior generalization performance. Notably, ChromNet accurately predicts chromatin interactions in acute myeloid leukemia (AML) samples by leveraging epigenetic signals from both normal and diseased cells, highlighting its potential for studying disease‐associated chromatin remodeling. Across multiple key benchmarks, ChromNet consistently outperforms existing models, providing a robust and cost‐effective solution for large‐scale chromatin conformation studies. This framework enables the exploration of chromatin structural variations across both cell types and disease states, offering new insights into the relationship between 3D genome architecture and gene regulation.