Hi-Compass: a depth-aware deep learning framework for predicting cell-type-specific 3D genome organization from single-cell to spatial resolution

Abstract

Abstract Three-dimensional genome organization controls cell-type-specific gene expression through chromatin interactions, yet systematic analysis across diverse cellular contexts remains limited by experimental constraints. Here we present Hi-Compass, a depth-aware deep learning framework that predicts cell-type-specific chromatin organization using only chromatin accessibility data as cell-type-specific input. By dynamically accommodating variability in sequencing depth, Hi-Compass enables robust predictions across the full spectrum of data scales, from sparse single-cell to high-coverage bulk profiles. Benchmarking shows that Hi-Compass achieves superior concordance with experimental Hi-C data compared to existing methods, with particularly strong recovery of high-confidence chromatin loops. Applied to peripheral blood and embryonic heart datasets, Hi-Compass resolves cell-type-specific chromatin interactions and systematically links disease-associated variants to putative target genes. The framework further enables spatially resolved chromatin interaction prediction in hippocampal tissue and demonstrates cross-species applicability through fine-tuning to mouse systems. Hi-Compass expands the capacity to study three-dimensional genome regulation across biological scales and species.