Hi-C Resolution Enhancement

DeepHiC and similar frameworks employ generative adversarial networks (GANs) to upscale low-coverage Hi-C data. The generator produces high-resolution contact maps from downsampled input, while the discriminator distinguishes generated maps from real high-coverage data. This adversarial training pushes the generator to reconstruct biologically realistic fine-scale features:

• Reconstructs loops and sub-TAD boundaries invisible in sparse data
• Preserves distance-dependent contact probability decay
• Outperforms bicubic interpolation and simple smoothing methods
• Trained on matched high/low coverage pairs from the same cell type

Super-resolution models treat Hi-C enhancement as a matrix imputation problem on sparse, symmetric contact maps. Architectures leverage the intrinsic structure of chromatin interaction data:

• Convolutional layers capture local contact domain patterns
• U-Net style skip connections preserve both fine details and global structure
• Residual learning predicts the difference between low and high-resolution maps
• Distance-aware normalization accounts for the expected power-law decay of contacts with genomic separation
• Outputs maintain symmetry and non-negative constraints of valid contact matrices

Effective super-resolution requires learning features at multiple genomic scales simultaneously. Models process Hi-C maps through parallel pathways with different receptive field sizes:

• Fine-scale branches capture individual loop anchors and point interactions
• Intermediate branches resolve TAD boundaries and sub-domain structures
• Coarse-scale branches maintain A/B compartment integrity and chromosome territories
• Feature fusion layers integrate multi-scale information before final reconstruction
• This hierarchical approach prevents the loss of either local precision or global organizational patterns during upscaling

Training data is generated by computationally downsampling high-coverage Hi-C maps to simulate low-coverage inputs. This creates perfectly paired training examples:

• Original high-coverage map serves as the ground truth target
• Random subset of reads produces the low-coverage input
• Multiple downsampling ratios (e.g., 1/16, 1/32, 1/64) create a curriculum
• Models learn to generalize across varying degrees of sparsity
• Stratum-adjusted correlation coefficient (SCC) evaluates reconstruction quality at each genomic distance stratum independently

The critical metric for Hi-C super-resolution is not just visual similarity but preservation of biologically meaningful features after enhancement:

• Chromatin loops must be recovered with correct anchor positions
• TAD boundaries should align with CTCF binding sites and insulation score minima
• Enhancer-promoter interactions must be identifiable in the enhanced map
• Allele-specific folding patterns should remain distinguishable
• Validation against orthogonal assays like DNA FISH confirms that imputed contacts reflect genuine physical proximity, not computational artifacts

Super-resolution models trained on one cell type can transfer enhancement capability to other cell types with limited high-coverage data:

• Domain-invariant features like contact decay rates generalize across cell types
• Fine-tuning on small amounts of target cell type data adapts the model
• Domain-adversarial training encourages learning of cell-type-agnostic enhancement patterns
• Enables high-resolution reconstruction for rare or difficult-to-profile cell types
• Reduces sequencing costs by requiring high coverage only for a reference cell type

A genome-wide matrix quantifying the interaction frequencies between all pairs of genomic loci, derived from chromosome conformation capture assays. This matrix serves as the primary input for resolution enhancement models and the prediction target for sequence-to-structure algorithms. Contact maps are inherently sparse at fine scales due to sequencing depth limitations, motivating the need for computational super-resolution. Raw maps exhibit a strong distance-dependent background signal where proximal loci interact more frequently than distal ones.

A reproducibility metric specifically designed for Hi-C data that measures the similarity between two contact maps while accounting for the distance-dependent signal decay. Unlike naive Pearson correlation, SCC stratifies contact pairs by genomic distance before computing correlation, preventing the dominant proximal signal from masking differences in long-range interactions. This metric is the standard benchmark for evaluating Hi-C resolution enhancement models, ensuring that imputed fine-scale contacts reflect genuine biological structure rather than noise.

A high-resolution variant of chromosome conformation capture that uses micrococcal nuclease to fragment chromatin to the nucleosome level, providing finer detail of 3D genome folding than standard Hi-C. Micro-C resolves interactions at sub-kilobase resolution, revealing features such as individual nucleosome positioning and subtle enhancer-promoter contacts. This data serves as both a gold-standard training target for super-resolution models and a validation benchmark to confirm that computationally enhanced Hi-C maps recapitulate genuine fine-scale biology.

The systematic correction of systematic biases in Hi-C contact matrices before resolution enhancement. Key biases include GC content, mappability, and restriction fragment length. Methods like iterative correction and eigenvector decomposition (ICE) and matrix balancing ensure that observed contact frequencies reflect genuine biological proximity rather than technical artifacts. Proper normalization is a critical preprocessing step—super-resolution models trained on uncorrected data will amplify biases rather than recover true chromatin interactions.