Overcoming artificial structures in resolution-enhanced Hi-C data by signal decomposition and multi-scale attention

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Chromosome conformation capture (Hi-C) data provides insights into 3D genome organization but often has limited resolution.
• Computational enhancement is crucial for inferring high-resolution genomic features from Hi-C data.
• Existing deep learning methods can introduce artificial structures and focus on local patterns, limiting their effectiveness.

Purpose of the Study:

• To develop a novel computational method, SHARP, for accurate high-resolution enhancement of Hi-C data.
• To address limitations of current deep learning approaches, such as patch-induced artifacts and insufficient capture of global patterns.
• To improve the identification of significant genomic interactions and their relationship with chromatin states.

Main Methods:

• SHARP decomposes Hi-C data into three signal types: 1D proximity, contiguous domains, and fine structures.
• Deep learning is selectively applied to the fine structure signals to avoid patch-related artifacts.
• Incorporates both local and global attention mechanisms to capture multi-scale contextual information for enhanced pattern recognition.

Main Results:

• SHARP demonstrates superior performance in resolution enhancement accuracy compared to state-of-the-art methods.
• The method effectively avoids the creation of artificial structures often seen in deep learning-enhanced Hi-C data.
• SHARP shows improved identification of significant interactions and enrichment in relevant chromatin states across different datasets and species.

Conclusions:

• SHARP offers a robust and accurate approach for enhancing Hi-C data resolution, overcoming limitations of previous methods.
• The signal decomposition strategy and attention mechanisms in SHARP enable more reliable inference of 3D genome organization.
• This method advances the analysis of Hi-C data, facilitating a deeper understanding of genome structure and function.