Genomic Distance Normalization

The core principle of genomic distance normalization is modeling the power-law relationship between contact probability and linear genomic separation. In Hi-C data, the probability that two loci interact decreases predictably as the number of base pairs between them increases, following a decay curve approximated by P(s) ∝ s^(-1) for mammalian genomes. This expected background signal must be removed to reveal biologically meaningful interactions such as chromatin loops and TAD boundaries. The decay function is typically estimated empirically from the contact matrix itself by averaging interaction frequencies across all locus pairs at each genomic distance bin, creating a distance-dependent expected contact profile.

A direct normalization approach that divides each observed contact count by the expected count for that genomic distance. The expected matrix is constructed by computing the average contact frequency for all locus pairs separated by the same linear distance. The resulting O/E matrix highlights interactions that are enriched or depleted relative to the genomic distance baseline:

• O/E > 1: Enriched interactions, indicating structural features like loops
• O/E < 1: Depleted interactions, often at TAD boundaries
• O/E = 1: Interactions consistent with random polymer behavior This transformation is essential for visualizing and comparing contact maps across different genomic regions and experimental conditions.

A reproducibility metric specifically designed for Hi-C data that measures the similarity between two contact maps while stratifying by genomic distance. Unlike standard correlation coefficients that can be dominated by the strong distance-dependent signal, the SCC computes the correlation within each distance stratum separately and then aggregates these stratum-specific correlations using a weighted average. This ensures that the metric reflects the reproducibility of fine-scale structural features rather than simply recapitulating the shared distance decay profile. The SCC is the standard benchmark for evaluating the accuracy of 3D genome folding predictions against experimental Hi-C data.

Multiple systematic biases confound raw Hi-C contact frequencies and must be addressed through normalization:

• Restriction fragment length: Longer fragments generate more ligation junctions
• GC content: Affects amplification efficiency during library preparation
• Mappability: Repetitive regions have ambiguous read alignment
• Copy number variation: Amplified or deleted regions show altered contact frequencies
• Chromatin accessibility: Open chromatin is more susceptible to digestion Normalization methods must account for these covariates to ensure that residual contact enrichment reflects genuine 3D proximity rather than technical artifacts.

In sequence-to-contact prediction models such as Akita and DeepHiC, genomic distance normalization is integrated directly into the training objective. Models learn to predict normalized contact probabilities rather than raw counts, with the loss function often incorporating distance-stratified weighting to prevent the abundant short-range contacts from dominating gradient updates. Some architectures explicitly condition predictions on genomic distance as an input feature, allowing the model to learn the distance-dependent prior implicitly. This integration ensures that predicted contact maps are directly comparable to experimentally normalized Hi-C data and that evaluation metrics like the SCC reflect genuine structural prediction accuracy.

A genome-wide matrix quantifying the interaction frequencies between all pairs of genomic loci. This is the primary data structure that genomic distance normalization is applied to, correcting for the inherent bias where loci that are linearly close on the chromosome interact more frequently due to random polymer dynamics rather than functional looping.

• Raw maps show a dominant diagonal signal
• Normalization reveals biologically relevant chromatin loops and TADs

A reproducibility metric specifically designed for Hi-C data that measures the similarity between two contact maps while explicitly accounting for the distance-dependent signal. Unlike standard correlation, the SCC stratifies interactions by genomic distance before calculation, ensuring that a model is penalized for getting the fine-scale looping wrong, not just for matching the background distance decay curve.

• The gold standard for benchmarking sequence-to-contact prediction models
• Ranges from -1 to 1, where 1 indicates perfect agreement

The systematic correction of systematic biases in Hi-C contact matrices prior to distance normalization. This includes correcting for GC content, mappability, and restriction fragment length. Methods like Iterative Correction and Eigenvector decomposition (ICE) perform matrix balancing to ensure that every row and column has equal visibility, removing technical artifacts from the biological signal.

• Explicit normalization: models the bias directly
• Implicit normalization: uses matrix balancing algorithms

A quantitative metric calculated from normalized Hi-C data that measures the degree to which a genomic locus is insulated from interactions with neighboring regions. It is computed by sliding a square window along the diagonal of the contact map and summing the interactions crossing the window. TAD boundaries are identified as local minima in the insulation score profile.

• Directly dependent on accurate distance normalization
• Used to validate CTCF binding site prediction

A metric that quantifies the upstream or downstream bias of chromatin interactions at a given genomic bin. It infers the directionality of loop extrusion by calculating the difference between contacts with downstream and upstream regions. A high positive value suggests a downstream-oriented domain, while a negative value indicates an upstream bias.

• Complements the insulation score for domain boundary detection
• Sensitive to artifacts if distance normalization is poorly calibrated

A mechanistic model of chromatin organization wherein cohesin complexes actively reel DNA to form progressively larger loops until blocked by CTCF boundary elements. This process explains the formation of TADs and the observed power-law decay of contact probability with genomic distance, which is the very signal that distance normalization seeks to factor out.

• The theoretical foundation for why distance normalization is necessary
• Predicts a characteristic 'corner peak' pattern in Hi-C maps