#28 · Epigenomics

Deep learning models decode 3D chromatin architecture from Hi-C contact matrices

The common thread is the treatment of Hi-C contact matrices as two-dimensional inputs for task-specific deep architectures. scSTEC annotates chromatin subcompartments from individual single-cell Hi-C matrices through direct two-dimensional embedding and chromosome-wise consensus, without imputation. AxialTAD predicts TAD boundaries from Hi-C matrices using an Axial Contrast Module, reporting 76-80% fewer parameters than deepTAD. HiCP2GAN reuses a pretrained Vision Transformer Hi-C foundation model as the discriminator of a plug-and-play GAN that enhances low-resolution contact maps. What is new is matrix-native design: compactness in AxialTAD, per-cell analysis without imputation in scSTEC, and transfer of foundation-model representations into a generative enhancement setting in HiCP2GAN. These are three separate tasks rather than one shared model, so the shared element is the input representation and architectural adaptation to it.