HiCSR

HiCSR enhances low-resolution Hi-C contact maps using deep learning to recover high-resolution chromatin contact structures for studies of chromatin folding and architecture.

Key Features:

• Super-Resolution Capability: Produces high-resolution Hi-C contact maps that capture fine details, textures, and substructures such as Topologically Associating Domains (TADs) and chromatin loops.
• Machine Learning Integration: Integrates deep learning with a novel loss formulation including an adversarial loss from a Generative Adversarial Network (GAN), a feature reconstruction loss derived from the latent representation of a denoising autoencoder, and a pixel-wise L1 loss.
• Reproducibility and Accuracy: Outperforms HiCPlus, HiCNN, hicGAN, and DeepHiC on reproducibility metrics validated against benchmarks produced by the ENCODE Consortium.
• Versatility Across Conditions: Demonstrates robustness across different sequencing depths, cell types, and species and recovers biologically significant contact domain boundaries.

Scientific Applications:

• Genome-wide chromatin conformation analysis: Enables genome-wide studies of chromatin folding and architecture by producing high-resolution contact maps from low-resolution Hi-C data.
• Chromatin substructure characterization: Facilitates detection and analysis of TADs, chromatin loops, and contact domain boundaries relevant to genomic regulation and organization.

Methodology:

Implemented in Python, HiCSR optimizes a weighted combination of adversarial loss (from a GAN), pixel-wise L1 loss, and a feature reconstruction loss derived from the latent representation of a denoising autoencoder.