Learning the histone codes with large genomic windows and three-dimensional chromatin interactions using transformer.

Area of Science:

• Genomics
• Computational Biology
• Epigenetics

Background:

• Quantitative characterization of transcriptional control by histone modifications is computationally challenging.
• Existing methods often focus narrowly on linear genomic regions, limiting comprehensive analysis.

Purpose of the Study:

• To introduce Chromoformer, a deep learning architecture for quantitative deciphering of histone codes in gene regulation.
• To leverage 3D chromatin conformation for improved modeling of transcriptional control.

Main Methods:

• Developed a transformer-based deep learning architecture (Chromoformer).
• Incorporated three variants of attention mechanisms to model hierarchical transcriptional regulation.
• Utilized 3D chromatin interaction data to capture long-range dependencies.

Main Results:

• Chromoformer achieves state-of-the-art performance in quantitative histone code deciphering.
• The model effectively utilizes long-range dependencies between histone modifications for transcription initiation and elongation.
• Demonstrated the ability to capture quantitative kinetics of transcription factories and Polycomb group bodies.

Conclusions:

• Attention-based deep learning models offer significant advantages for analyzing complex interactions in epigenomes.
• Chromoformer provides a powerful framework for understanding gene regulation through 3D chromatin structure and histone modifications.