Evaluation of deep learning tools for chromatin contact prediction

Three-dimensional chromatin organization is essential for gene regulation and is commonly measured using Hi-C contact maps. Recent deep learning models have been developed to predict Hi-C maps from genomic and epigenomic features. However, their relative performance and biological interpretability remain poorly understood due to the lack of systematic evaluation. Here, we present a comprehensive benchmarking framework that evaluates five Hi-C prediction models: C.Origami, Epiphany, ChromaFold, HiCDiffusion, and GRACHIP, across predictive accuracy, visual fidelity, and downstream biological analyses. Among them, Epiphany consistently achieved the best overall performance, combining high accuracy, cross-cell-type generalization, realistic map quality, and reliable loop recovery. The framework further shows that epigenomic features, particularly CTCF binding and chromatin co-accessibility, are the primary drivers of accurate Hi-C pattern prediction. Notably, although many models incorporate multiple omics inputs, only a limited subset substantially contributes to performance. This manuscript clarifies model behaviour and provides guidance for developing and interpreting Hi-C prediction methods.