#34 · Epigenomics

Predicting and reconstructing 3D genome organization from chromatin and Hi-C data

The common thread is computational inference of genome organization from inputs other than direct high-resolution contact maps. Hi-cGAN uses conditional generative adversarial networks to predict Hi-C interaction matrices at 2-25 kb bins from chromatin factor tracks, positioned against sequence-based Akita and C.Origami. HiC-LEGO, a hierarchical graph-based framework, reconstructs 3D genome structures from Hi-C and Micro-C at kilobase resolution while preserving TADs. The single-cell HiRES study supplies biological grounding: CTCF loss makes TAD-like boundaries more variable and reduces A-compartment interactions, and its SALTAFinder method identifies domain features. What is new is the use of chromatin factor tracks as predictive inputs for matrix generation, alongside explicit comparison against existing methods. This describes the dominant thread rather than a shared architecture, since HiRES is experimental profiling rather than a modeling framework.