The three-dimensional organization of the genome orchestrates cell-type-specific gene regulation, yet generating high-resolution Hi-C data remains resource-intensive and is particularly challenging for rare cell populations, single cells, and spatially resolved tissues. Computational prediction of Hi-C contact maps offers a scalable alternative, but existing methods typically depend on multiple experimental input modalities and are sensitive to data sparsity and sequencing-depth variation, limiting their applicability across biological scales.

Here we present Hi-Compass, a depth-aware deep learning framework that predicts cell-type-specific Hi-C contact maps using only ATAC-seq signal, DNA sequence, and a generalized CTCF binding profile as input. The model adopts a CNN-Transformer hybrid architecture: three parallel convolutional encoders extract features from each input modality, a Transformer decoder fuses the multi-modal representations to generate 2 Mb contact maps at 10 kb resolution, and a cell-type discriminator is jointly optimized to enforce cell-type specificity. A key innovation is the depth-adaptive module in the ATAC-seq branch, which explicitly conditions the model on input sequencing depth and enables robust predictions across depths spanning four orders of magnitude. To improve training stability and enhance the visibility of fine-scale structural features, we further introduce a contrast stretching normalization step that preprocesses raw Hi-C matrices prior to supervision.

We benchmarked Hi-Compass against state-of-the-art methods including Akita, C.Origami, Epiphany, and ChromaFold. Hi-Compass consistently achieved superior concordance with experimental Hi-C data across insulation score correlation, SSIM, and distance-stratified correlation metrics, and generalized robustly to held-out cell types and chromosomes. Downstream loop detection on predicted maps preferentially recovered high-confidence chromatin loops, and aggregate peak analysis against orthogonal cohesin HiChIP and RAD21 ChIA-PET datasets confirmed that Hi-Compass predictions faithfully capture protein-mediated chromatin contacts.

To demonstrate its utility across biological scales, we applied Hi-Compass in three settings. First, by integrating single-cell ATAC-seq profiles into meta-cells, Hi-Compass resolves cell-type-specific chromatin interactions in complex tissues and systematically links non-coding GWAS variants to putative target genes, with concordance to tissue-matched eQTLs comparable to that achieved using experimental bulk Hi-C. Second, by aggregating spatial ATAC-seq signals into meta-spots, Hi-Compass reconstructs spatially resolved chromatin interaction maps in intact tissue sections, revealing region-specific loops that correspond to local gene expression patterns—providing a computational bridge between spatial epigenomics and 3D genome architecture. Third, through lightweight fine-tuning, Hi-Compass adapts to mouse data and achieves prediction accuracy comparable to that in human samples, supporting cross-species investigations of chromatin topology.

Collectively, Hi-Compass extends cell-type-specific 3D genome analysis from bulk cell lines to single-cell and spatial resolution, and from human to mouse, offering a versatile computational framework for exploring chromatin organization in contexts where experimental Hi-C is impractical.