EnHiC: learning fine-resolution Hi-C contact maps using a generative adversarial framework

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• High-throughput chromosome conformation capture (Hi-C) enables genome-wide chromatin interaction mapping.
• Generating high-resolution Hi-C data requires expensive deep sequencing, limiting its application.
• Machine learning, particularly neural networks, offers a potential solution for enhancing Hi-C data resolution.

Purpose of the Study:

• To introduce EnHiC, a novel method for predicting high-resolution Hi-C matrices from low-resolution input data.
• To leverage generative adversarial networks (GANs) and non-negative matrix factorization principles for Hi-C data enhancement.
• To improve the accessibility and utility of high-resolution Hi-C data for biological research.

Main Methods:

• Developed EnHiC, a GAN-based framework for Hi-C matrix resolution enhancement.
• Incorporated non-negative matrix factorization-inspired techniques to extract rank-1 features from multi-scale low-resolution matrices.
• Validated EnHiC using three human Hi-C datasets.

Main Results:

• EnHiC accurately and reliably enhanced the resolution of Hi-C matrices.
• EnHiC outperformed existing GAN-based models in Hi-C data enhancement.
• Predicted high-resolution matrices from EnHiC facilitated accurate detection of topologically associated domains and fine-scale chromatin interactions.

Conclusions:

• EnHiC provides an effective computational approach to generate high-resolution Hi-C data from low-resolution inputs.
• The method significantly improves the analysis of chromatin architecture and interactions.
• EnHiC is publicly available, promoting wider adoption and further research in the field.