Sequence-to-Contact Prediction

These models learn a direct mapping from raw DNA sequence to Hi-C contact matrices without intermediate hand-crafted features. The entire pipeline—from sequence encoding through spatial proximity prediction—is trained jointly via backpropagation.

• Input: One-hot encoded DNA sequence or k-mer tokenized representations spanning megabase-scale loci
• Output: Predicted contact probability matrix at multiple resolutions (e.g., 1 kb, 5 kb, 10 kb bins)
• Loss functions typically combine mean squared error on log-transformed contacts with Stratum-Adjusted Correlation Coefficient (SCC) optimization
• Enables in silico perturbation experiments: mutate a sequence and observe predicted folding changes without wet-lab assays

Chromatin folding is governed by signals operating at vastly different length scales—from individual transcription factor binding motifs (tens of base pairs) to topologically associating domains (hundreds of kilobases). Sequence-to-contact models employ hierarchical architectures to capture this spectrum.

• Convolutional layers with dilated kernels detect local motifs (CTCF, cohesin loading sites)
• Transformer attention heads model long-range dependencies between distal regulatory elements
• U-Net style encoder-decoder structures preserve spatial resolution while expanding receptive fields
• Multi-scale inputs may include epigenomic tracks (chromatin accessibility, histone marks) as auxiliary channels

Leading models incorporate biophysical constraints to ensure predictions are physically plausible. Rather than treating chromatin as an arbitrary graph, they respect the polymer nature of DNA.

• Distance-dependent decay priors: Contact probability decreases with genomic distance following power-law relationships observed in real Hi-C data
• Excluded volume constraints: Two loci cannot occupy the same spatial position
• Loop extrusion simulation: Some architectures explicitly model cohesin-mediated extrusion dynamics as a differentiable module
• Polymer physics-informed loss terms penalize structures that violate known folding principles
• This integration dramatically improves generalization to unseen genomic contexts

Modern sequence-to-contact architectures leverage pre-trained genomic foundation models as feature extractors. These DNA language models, trained on massive genome-wide datasets via self-supervision, encode rich evolutionary and regulatory information.

• Nucleotide transformer and Enformer-style backbones generate contextualized sequence embeddings
• Pre-training tasks include masked nucleotide prediction and next-k-mer prediction across species
• Transfer learning from these backbones dramatically reduces the labeled Hi-C data required for fine-tuning
• Embeddings capture splice sites, promoter architecture, and conserved non-coding elements without explicit annotation
• Enables cross-species generalization: models trained on human data can predict folding in mouse

Advanced models predict 3D folding separately for maternal and paternal chromosomes, revealing how genetic variation influences chromatin architecture. This capability is critical for understanding the regulatory impact of heterozygous variants.

• Requires phased genotypes or haplotype-resolved reference genomes as input
• Models learn to distinguish folding patterns driven by single nucleotide polymorphisms (SNPs) in CTCF binding sites
• Enables prediction of allele-specific enhancer-promoter loops that explain imbalanced gene expression
• Critical for interpreting non-coding variants from genome-wide association studies (GWAS)
• Validated against allele-specific Hi-C and DNA FISH experiments

Sequence-to-contact models function as generative engines for creating realistic Hi-C maps in silico. This capability addresses the scarcity of experimental 3D genome data across diverse cell types and conditions.

• Generative adversarial networks (GANs) produce high-fidelity contact maps indistinguishable from experimental data
• Variational autoencoders (VAEs) enable latent space interpolation between cell states
• Models can simulate the effects of structural variants (deletions, inversions, translocations) on genome folding
• Synthetic maps augment training datasets for downstream models (e.g., enhancer-promoter link predictors)
• Enables virtual knockout experiments: predict folding changes after removing specific regulatory elements

A genome-wide matrix quantifying interaction frequencies between all pairs of genomic loci. It serves as both the primary training target and the ground-truth benchmark for sequence-to-contact models. Each entry in the matrix represents the probability that two DNA fragments are physically ligated, reflecting their spatial proximity in the nucleus. Deep learning models are trained to predict these maps directly from DNA sequence, bypassing the need for costly experimental assays.

A reproducibility metric specifically designed for Hi-C data that measures the similarity between predicted and experimental contact maps. Unlike standard correlation, SCC accounts for the distance-dependent background signal by stratifying contacts by genomic distance. This prevents models from achieving artificially high scores by simply learning the generic decay of contact probability with linear distance, forcing them to capture biologically meaningful loop and domain structures.

A mechanistic model of chromatin organization wherein cohesin complexes actively translocate along DNA, extruding progressively larger loops until blocked by CTCF boundary elements. Sequence-to-contact models implicitly learn the sequence determinants of this process, including CTCF motif orientation and cohesin loading sites. Understanding loop extrusion physics informs the architectural priors embedded in polymer physics-informed neural networks for 3D genome reconstruction.

A statistical correction applied to Hi-C contact maps to account for the expected background contact frequency decay as a function of linear genomic distance. The probability of contact between two loci decreases rapidly as their separation increases. Normalization removes this power-law distance dependence, revealing specific interactions like loops and domains. Sequence-to-contact models must learn to predict both the background decay and the deviations that indicate functional structures.

The use of generative models, such as GANs or VAEs, to create artificial but realistic Hi-C contact maps. These synthetic maps augment limited training datasets and enable the simulation of genetic perturbation effects on genome folding. By generating contact maps for hypothetical sequence variants, researchers can predict how mutations might disrupt normal chromatin architecture without performing experiments, accelerating the functional interpretation of non-coding variants.