Stratum-Adjusted Correlation Coefficient (SCC)

SCC computes the Pearson correlation coefficient between two contact maps independently for each genomic distance stratum. This stratification prevents the natural power-law decay of contact frequency with distance from dominating the similarity score. By grouping locus pairs by their linear separation, SCC ensures that short-range and long-range interactions contribute equitably to the final metric, rather than being swamped by the high signal at close distances.

The final SCC value is calculated as the geometric mean of the per-stratum correlation coefficients. This aggregation method penalizes models that perform well on one distance scale but poorly on another. Unlike an arithmetic mean, the geometric mean approaches zero if any single stratum has a correlation near zero, enforcing consistent performance across all interaction ranges. This property makes SCC particularly sensitive to structural failures in specific folding regimes.

SCC is designed to be robust to differences in sequencing depth between the two contact maps being compared. Because the metric operates on normalized contact values within each distance stratum, it measures the concordance of relative interaction patterns rather than absolute read counts. This allows fair comparison between deeply sequenced reference maps and shallower experimental replicates or computationally predicted outputs.

SCC has become the de facto evaluation standard for sequence-to-contact prediction models such as Akita and DeepHiC. When benchmarking a model's ability to predict 3D genome folding from DNA sequence alone, SCC quantifies how well the predicted contact map reproduces the structural features of the experimental Hi-C map. Values typically range from 0 to 1, with higher scores indicating superior structural fidelity.

SCC incorporates a smoothing parameter (h) that controls the neighborhood aggregation of contacts before correlation is computed. This parameter defines a window around each genomic locus over which interaction frequencies are averaged. A larger h suppresses noise and emphasizes domain-scale structures like TADs, while a smaller h preserves fine-scale features such as individual chromatin loops. The choice of h directly influences which structural scales dominate the evaluation.

SCC is often used in conjunction with insulation score profiles to validate predicted TAD boundaries. While SCC measures global map similarity, the insulation score quantifies the degree of interaction isolation at each genomic locus. A high SCC between predicted and experimental maps, combined with concordant insulation score minima at known TAD boundaries, provides strong evidence that a model correctly captures domain-level chromatin organization.