Enhancing Hi-C contact matrices for loop detection with Capricorn: a multiview diffusion model.

Area of Science:

• Genomics
• Computational Biology
• Biophysics

Background:

• High-resolution Hi-C contact matrices are essential for understanding 3D genome architecture but are costly to generate.
• Existing computational methods for enhancing sparse Hi-C data often lack biological specificity, treating all contacts similarly.

Purpose of the Study:

• To develop a machine learning model, Capricorn, for enhancing low-coverage Hi-C data.
• To improve the accuracy of identifying biologically significant chromatin structures, such as loops, from sparse Hi-C matrices.

Main Methods:

• Capricorn utilizes a machine learning approach incorporating small-scale chromatin features as additional views.
• A diffusion probability model backbone is employed to generate high-coverage Hi-C matrices.
• The model's performance is evaluated in cross-cell-line, cross-chromosome, and combined cross-setting scenarios.

Main Results:

• Capricorn significantly outperforms existing state-of-the-art methods, reducing mean squared error by 17% and improving F1 score for loop identification by 26%.
• The method demonstrates robust performance across different experimental settings, enhancing downstream loop identification by 14%.
• The core 'multiview' concept improves existing methods like HiCARN and HiCNN, validating its broad applicability.

Conclusions:

• Capricorn provides a powerful and cost-effective solution for Hi-C resolution enhancement.
• The method enables the discovery of chromatin features previously undetectable in low-coverage data.
• Capricorn facilitates more accurate 3D genome structure analysis and chromatin loop identification.