DeepHiC: A generative adversarial network for enhancing Hi-C data resolution.

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Hi-C is crucial for studying 3D genome organization but often yields low-resolution data due to high costs and technical limitations.
• Low-resolution Hi-C data leads to information loss and reduced biological interpretability.
• Existing methods struggle to accurately reconstruct high-resolution contact maps from sparse data.

Purpose of the Study:

• To develop DeepHiC, a novel deep learning approach for generating high-resolution Hi-C contact maps from low-coverage sequencing data.
• To enhance the resolution and biological insights obtainable from Hi-C experiments.
• To provide a user-friendly tool for researchers to improve their Hi-C data.

Main Methods:

• Development of DeepHiC, a generative adversarial network (GAN) specifically designed for Hi-C data enhancement.
• Utilizing adversarial training to restore fine-grained details in Hi-C matrices.
• Downsampling high-resolution Hi-C data to simulate low-coverage scenarios for model training and validation.

Main Results:

• DeepHiC successfully reproduces high-resolution Hi-C data from as little as 1% of downsampled reads.
• The method significantly improves the accuracy of chromatin loop identification and topologically associating domain (TAD) detection.
• DeepHiC outperforms existing state-of-the-art methods in prediction accuracy for Hi-C contact maps.
• Application to mouse embryonic development Hi-C data facilitated enhanced chromatin loop detection.

Conclusions:

• DeepHiC offers a powerful solution for overcoming the resolution limitations of traditional Hi-C sequencing.
• The tool enables researchers to extract more detailed information about 3D genome organization from cost-effective, low-coverage data.
• DeepHiC has the potential to advance research in various fields, including developmental biology and disease studies.