Hi-C Data Normalization

A fast algorithm that finds a doubly stochastic normalization of a contact matrix by minimizing the sum of squared errors. It scales the matrix so that rows and columns sum to a constant, effectively decoupling bias correction from the distance-dependent signal.

• Advantage: Computationally efficient for sparse matrices
• Application: Standard preprocessing in Hi-C analysis pipelines
• Property: Preserves the symmetry of the contact matrix
• Comparison: Often converges faster than vanilla ICE for high-resolution data

A regression-based approach that models raw contact counts as a function of known biasing factors including GC content, mappability, and restriction fragment length. Each bias source is estimated independently and then divided out.

• Model: Observed = True Signal × Bias_GC × Bias_Mappability × Bias_Length
• Strength: Interpretable and biologically grounded
• Weakness: Cannot account for unknown or latent biases
• Tools: Often implemented via HiCNorm or fit-hi-c frameworks

Generates an expected contact probability as a function of genomic distance, then normalizes observed contacts by this expectation. This transforms the matrix into an observed/expected (O/E) ratio that highlights biologically significant interactions.

• Process: Fit a power-law decay curve P(s) ~ s^(-1) to the data
• Output: A Z-score or fold-change matrix over the genomic distance baseline
• Use Case: Essential for identifying Topologically Associating Domains (TADs) and significant loops
• Integration: Often combined with ICE or KR balancing in a two-step pipeline

A strategy for normalizing Hi-C matrices across multiple experimental conditions or time points to a common reference distribution. It uses quantile normalization or loess smoothing to remove systematic technical variation between samples.

• Goal: Make contact maps directly comparable across different experiments
• Method: M-A plots borrowed from microarray normalization
• Application: Differential chromatin interaction analysis
• Challenge: Requires careful handling of biological variability vs. technical noise

Specialized methods for normalizing extremely sparse single-cell Hi-C data, where over 99% of possible contacts are missing. Techniques include random walk with restart, network propagation, and low-rank matrix completion to impute missing values before balancing.

• Core Challenge: Severe sparsity violates the equal-visibility assumption
• Approaches: scHiCNorm, Higashi, and DeepHiC-based imputation
• Validation: Often benchmarked against DNA FISH distance measurements
• Emerging Trend: Deep learning models that jointly impute and normalize

The foundational matrix balancing algorithm that produces a doubly stochastic matrix from a raw Hi-C contact map. ICE iteratively adjusts row and column sums to equal 1, ensuring each genomic bin has equal total visibility. This corrects for GC content, mappability, and restriction fragment length biases without requiring explicit bias models. The algorithm converges when the matrix becomes symmetric and balanced, making it the standard preprocessing step before TAD calling or A/B compartment analysis.

A fast, non-iterative algorithm for transforming a raw contact matrix into a doubly stochastic form. It directly computes bias vectors by solving for diagonal scaling matrices that equalize row and column sums. Key advantages over ICE include:

• Deterministic convergence in a single pass
• Lower computational complexity for large genomes
• Preservation of symmetry in the corrected matrix This method is the default normalization in the cooler and HiC-Pro pipelines, providing a scalable solution for high-resolution contact maps.

A statistical model that captures the background contact probability as a function of linear genomic distance. In Hi-C data, contact frequency decays approximately as a power law with distance: P(s) ∝ s^(-1). This expected model is computed by averaging all interactions at each genomic distance stratum. Normalization divides observed contacts by this expected value, converting raw counts into observed/expected (O/E) ratios. This step is critical for revealing biologically meaningful loops and domains that deviate from the polymer physics baseline.

A regression-based approach that directly models the contribution of known technical biases to contact probability. The model fits: C_ij = B_i * B_j * f(d_ij), where B_i and B_j are bias factors for bins i and j, and f(d_ij) is the distance-dependent expected value. Bias factors are estimated from:

• GC content of each restriction fragment
• Mappability scores from alignment uniqueness
• Fragment length affecting ligation efficiency This approach provides interpretable bias estimates and allows for the removal of specific artifact sources.

The simplest normalization method that divides each contact count by the total number of reads in the experiment, converting raw interactions to contacts per million (CPM). While computationally trivial, it fails to account for systematic biases like GC content or mappability. This method is suitable only for:

• Initial data exploration and quality control
• Comparing replicates with similar bias profiles
• Low-resolution analyses where fine-scale biases average out For publication-quality analysis, ICE or Knight-Ruiz balancing is strongly preferred.

The gold-standard metric for evaluating Hi-C normalization quality and reproducibility. SCC computes the Pearson correlation between two contact maps separately for each genomic distance stratum, then aggregates these stratum-specific correlations using a weighted average. This approach prevents the dominant signal from short-range contacts from masking differences in long-range interactions. SCC values range from 0 to 1, with >0.9 indicating excellent reproducibility. It is the primary benchmarking metric in tools like HiCRep and HiC-Spector.