DeepHiC

DeepHiC employs a conditional GAN framework where a generator network learns to upscale low-resolution Hi-C contact maps, while a discriminator network adversarially distinguishes generated high-resolution maps from real high-resolution experimental data.

• Generator: U-Net based architecture with skip connections for preserving fine-scale interaction features
• Discriminator: Patch-based classifier that evaluates local realism of contact patterns
• Adversarial loss combined with mean squared error and perceptual loss for structural fidelity
• Trained on paired low/high-resolution Hi-C data from the same cell type

DeepHiC incorporates genomic distance normalization into its training objective to account for the expected exponential decay of contact frequency with increasing linear genomic separation.

• Applies distance-stratified scaling to prevent the model from learning trivial distance-dependent background
• Preserves biologically meaningful long-range interactions while suppressing spurious noise
• Integrates stratum-adjusted correlation coefficient (SCC) as an evaluation metric
• Ensures enhanced maps maintain the characteristic power-law decay of real Hi-C data

The generator network processes Hi-C contact matrices at multiple resolutions simultaneously to capture both local chromatin loop structures and global A/B compartment organization.

• Encoder path: Progressive downsampling extracts hierarchical features from TAD-level to compartment-level
• Decoder path: Learned upsampling with skip connections reconstructs fine-scale interaction details
• Receptive field designed to span entire chromosomes for capturing long-range cis interactions
• Capable of enhancing maps from 50 kb to 5 kb resolution and beyond

DeepHiC employs a downsampling-based training paradigm that generates training pairs from high-quality experimental Hi-C datasets, eliminating the need for paired low/high-resolution data across multiple conditions.

• Random subsampling of high-coverage Hi-C reads simulates realistic low-coverage input
• Multiple downsampling ratios create a diverse training set from a single experiment
• Augmentation includes random genomic region cropping to increase sample diversity
• Enables transfer learning: models trained on one cell type can enhance data from related cell types

DeepHiC-enhanced contact maps are evaluated using multiple complementary metrics that assess both global structure and local interaction fidelity.

• Stratum-Adjusted Correlation Coefficient (SCC): Measures reproducibility while controlling for distance-dependent signal
• Pearson correlation: Genome-wide similarity between enhanced and ground-truth high-resolution maps
• TAD boundary recall: Accuracy in recovering topologically associating domain boundaries
• Chromatin loop detection: Precision and recall for identifying significant point-to-point interactions
• Outperforms bicubic interpolation and simple denoising baselines by substantial margins

Enhanced Hi-C maps produced by DeepHiC preserve and reveal biologically meaningful 3D genome features that are validated against orthogonal experimental data.

• Recovered chromatin loops validated against ChIA-PET and HiChIP data
• Enhanced TAD structures align with CTCF binding sites and cohesin localization
• A/B compartment calls from enhanced maps match those from high-coverage experimental data
• Enhancer-promoter interactions identified in enhanced maps correlate with gene expression
• Supports structural variant impact analysis by revealing disrupted folding patterns