Evaluation of deep learning tools for chromatin contact prediction

Background Three-dimensional chromatin organization plays a central role in gene regulation and is commonly measured using Hi-C technology. Recently, deep learning models have been developed to predict Hi-C contact maps from genomic and epigenomic features, offering a computational alternative to costly experimental assays. However, the performance, robustness, and biological interpretability of these models remain unclear due to the absence of systematic benchmarking. Results We present a comprehensive benchmark to evaluate five Hi-C prediction models, C.Origami, Epiphany, ChromaFold, HiCDiffusion, and GRACHIP, across multiple evaluation criteria, including accuracy, visual fidelity and loop detection. Among all models, Epiphany achieved the strongest overall performance, combining high accuracy, cell-type generalization, realistic image quality and reliable loop detection. Moreover, we evaluated predicted contact maps using four different loop-callers to assess the impact of model choice on loop detection performance. Despite the coarse resolution, many models could recover biologically relevant interactions. Notably, structural map quality was more critical than the choice of loop-caller for reliable detection. Finally, ablation analyses revealed that epigenomic signals are influential features for accurate Hi-C prediction. Despite the use of multiple input modalities in many models, only a limited subset contributed substantially to predictive performance. Conclusions This study provides a systematic comparison of deep learning models for Hi-C prediction and highlights the importance of specific regulatory signals in reconstructing 3D chromatin organization. The proposed evaluation framework clarifies model behaviours and offers guidance for the development and interpretation of Hi-C prediction methods.