THC-net: an attention-based deep learning model for chromatin compartment prediction from histone modifications

The three-dimensional architecture of the genome plays a central role in fundamental biological processes. Chromatin compartmentalization into A compartments (active transcription domains) and B compartments (repressive chromatin domains) not only visually represents genomic functionality but also provides a molecular anatomical perspective for deciphering cell-type-specific epigenetic regulatory networks. However, the inherent high cost of Hi-C technology-including experimental complexity, sequencing depth requirements, and data analysis barriers-has become a significant challenge in resolving cross-cell-type dynamics of compartmentalization.

To address this, we propose THC-Net, a chromatin compartment prediction method based on a multimodal deep learning architecture. This model integrates the self-attention mechanism of Transformers, the long-sequence modeling capability of the Hyena operator, and the local feature extraction advantages of convolutional neural networks to predict genomic A/B compartments. Across six cell lines (IMR90, HMEC, K562, GM12878, HUVEC, NHEK), THC-Net achieved an average AUROC of 93.1% in cross-cell-type validation. We also performed feature sufficiency and redundancy analysis, which demonstrates that the six histone modification input features exhibit a high degree of statistical redundancy, and the model primarily relies on the strong signals generated by active enhancers or promoters to define A compartments.

Experimental results show that THC-net achieved higher average AUROC compared to other methods in predicting chromatin compartment classification. The model exhibits robust performance and versatility across cell lines including GM12878, K562, and IMR90, providing a novel tool for precise chromatin compartment prediction.