73 / 2026-04-05 16:06:28

ChromMamba: A generalizable pre-training model for chromatin interaction prediction from cell-type-specific ATAC-seq

3D genome organization,AI,pre-training,zero-shot,structural variant,generalization

We present ChromMamba, a generalizable framework for long-range Hi-C prediction from DNA sequence and chromatin accessibility. By combining fine-grained sequence–accessibility modeling with long-context learning, ChromMamba recovers chromatin interaction patterns and higher-order structural features across scales. Increasing the diversity of training cell types improves de novo prediction in unseen cell types, suggesting that the model learns transferable rules of chromatin folding. Scaling ChromMamba to larger genomic windows enables modeling of long-range chromatin interactions and in silico simulation of megabase-scale structural variants. Application to yeast further shows that the framework generalizes to genomes lacking canonical CTCF-mediated organization. Overall, these results demonstrate that 3D genome architecture can be predicted from DNA sequence and chromatin accessibility across species and scales.