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

Area of Science:

• Genomics
• Epigenetics
• Computational Biology

Background:

• The 3D genome architecture is crucial for biological processes.
• Chromatin A/B compartments reflect genomic activity and epigenetic networks.
• High costs and complexity of Hi-C limit cross-cell-type analyses.

Purpose of the Study:

• To develop an efficient method for predicting genome A/B compartments.
• To overcome the limitations of Hi-C technology for large-scale genomic studies.

Main Methods:

• Proposed THC-Net, a multimodal deep learning architecture.
• Integrated Transformers, Hyena operator, and CNNs for A/B compartment prediction.
• Utilized six histone modification marks as input features.

Main Results:

• THC-Net achieved an average AUROC of 93.1% in cross-cell-type validation across six cell lines.
• Demonstrated high statistical redundancy among input histone modification features.
• Identified reliance on active enhancer/promoter signals for A compartment definition.

Conclusions:

• THC-Net outperforms existing methods in chromatin compartment classification.
• The model shows robust performance and versatility across various cell lines.
• Provides a novel computational tool for precise chromatin compartment prediction.