TM-Loop: Transformer multi-omics hierarchical detection of chromatin loop

Abstract

Abstract Chromatin loops form the hierarchical 3D genome architecture and act as critical regulatory hubs, bringing distant cis-elements together to precisely control target gene transcription. Disruption of these spatial interactions is closely linked to human diseases including cancer. Hi-C provides genome-wide interaction maps, yet robust identification of chromatin loops remains a major bottleneck in 3D genomics, especially for contact matrices with low signal-to-noise ratio and high sparsity. We propose TM-Loop, a framework for accurate chromatin loop detection that integrates multi-omics data, Transformer deep learning, and hierarchical multi-scale clustering. Using 10 kb Hi-C matrices as core data, it combines ATAC-seq and CTCF ChIP-seq signals to build a weighted feature system and reduce sample imbalance. The Transformer’s multi-head attention captures global and local feature dependencies, while dual-threshold filtering and anchor-guided clustering effectively remove false signals and lower false discovery rate. Source code: https://github.com/yimuhuashui/TM-Loop. Experiments show TM-Loop outperforms existing methods in APA, protein enrichment, and 3D structural consistency, providing a new tool for high-precision genome-wide chromatin loop analysis.