HiCARN

HiCARN enhances the resolution of Hi-C contact matrices to enable improved inference of three-dimensional chromatin architecture from low-resolution Hi-C data.

Key Features:

• Deep learning-based resolution enhancement: Predicts high-resolution Hi-C matrices from low-resolution Hi-C contact matrices using deep learning models.
• Cascading Residual Networks: Implements two cascading residual networks, HiCARN-1 (a convolutional neural network, CNN) and HiCARN-2 (a generative adversarial network, GAN).
• Cascading connections: Employs cascading connections throughout the models to improve prediction accuracy.
• Quantitative evaluation: Evaluated using image evaluation and Hi-C reproducibility metrics.
• Performance on downsampled data: Demonstrates superior predictive accuracy relative to existing algorithms on downsampled high-resolution Hi-C datasets at 1/16, 1/32, 1/64, and 1/100 levels.
• Biological output validation: Enables extraction of topologically associating domains (TADs), chromosome 3D structures, and chromatin loop predictions from enhanced matrices.

Scientific Applications:

• TAD identification: Detection and extraction of topologically associating domains from enhanced Hi-C matrices.
• Chromosome 3D reconstruction: Reconstruction of chromosome three-dimensional structures from enhanced contact maps.
• Chromatin loop prediction: Prediction of chromatin loops from enhanced Hi-C data.
• Regulatory interaction inference: Improved resolution for inferring enhancer-promoter interactions and sub-domain structures.

Methodology:

Uses deep learning to predict high-resolution Hi-C matrices from low-resolution inputs via two cascading residual networks (HiCARN-1 as a CNN and HiCARN-2 as a GAN) with cascading connections; trained and validated on downsampled high-resolution Hi-C datasets at 1/16, 1/32, 1/64, and 1/100 and evaluated using image evaluation and Hi-C reproducibility metrics, with validation through extraction of TADs, chromosome 3D structures, and chromatin loop predictions.