HiCFoundation

Hi-C measures which pairs of genomic loci are in physical contact, and the resulting contact matrix is the standard readout of three-dimensional genome architecture. Working with it is awkward. The data are pairwise rather than linear, so the toolchain built for ATAC-seq, DNase-seq and ChIP-seq tracks does not apply. Resolution is expensive — doubling it requires roughly four times the sequencing depth — which puts confident loop detection out of reach for rare cell types. And each analysis task has historically had its own bespoke method with its own assumptions.

HiCFoundation treats the contact matrix as an image and pretrains one encoder on it. The model was built by Xiao Wang, Yuanyuan Zhang, William Stafford Noble, Sheng Wang and colleagues at the University of Washington and Purdue University, posted to bioRxiv in December 2024 and published in Nature Methods. Submatrices of 224×224 bins are divided into non-overlapping 16×16 patches; a masked-autoencoder objective, in the style of vision transformers, asks the model to reconstruct the full submatrix from the visible patches. The key adaptation is a patch-wise contrastive loss added to the reconstruction and SSIM terms, which keeps training signal alive in the near-empty patches far from the diagonal where a naive reconstruction loss would be trivially satisfied by predicting zero.

Where catalog entries such as Evo2HiC, Mix-HiC and TruHiC each target one Hi-C problem, HiCFoundation's premise is that a single pretrained representation can serve all of them.

#Key Features

• Frozen encoder, task-specific decoders: After pretraining, the encoder is held fixed and shared; only a lightweight decoder is fine-tuned per task, giving roughly ten times faster task training than training the same architecture from scratch.
• Multi-scale embeddings: The model emits patch, submatrix and chromosome-level embeddings, so it can be used as a feature extractor as well as through its task heads.
• Low-coverage loop detection: Separate high- and low-coverage loop decoders let the same model call loops in shallow experiments, where pixel-level predictions are clustered by mean shift into final calls.
• Hi-C to epigenome: One decoder predicts six epigenomic assays at 1 kb resolution from a 128 kb Hi-C input, inferring regulatory activity from conformation alone.
• Single-cell Hi-C: A dedicated decoder enhances sparse single-cell contact maps, trained against 4× down-sampled data with a loss that damps gradient sensitivity to large counts.

#Technical Details

Pretraining drew on 1,015 Hi-C experiments from over 300 species, sourced from ENCODE, 4D Nucleome and other collections — 521 human biosamples plus 494 experiments from non-human species — yielding 118 million contact submatrices and 23 billion patches, with 404 experiments in the training split. The architecture is deliberately asymmetric: a 304 million-parameter encoder paired with a 26 million-parameter decoder, so capacity concentrates where it is reused. Reconstruction improves on the strongest baseline by 17.7% SSIM, 15.2% Pearson correlation and 19.1% Spearman correlation. On loop detection the fine-tuned model reaches mean F1 scores of 81.6% on high-coverage and 75.1% on low-coverage Hi-C, beating HiCCUPs and Chromosight respectively at p < 0.001. On resolution enhancement, scored across SSIM, PSNR, Pearson, Spearman, TAD F1 and loop F1 against retrained HiCARN, HiCNN and HiCSR, it achieves a mean rank of 1.08 on human and 1.05 on mouse against 2.38 for the next best method. A triplet-margin-trained decoder separates biological replicates from non-replicates with a 110% larger margin on human and 158% larger on mouse than the second-best method.

#Applications

Groups working with limited material get the most direct benefit: enhancing low-coverage Hi-C recovers structure that would otherwise require deeper sequencing. Applied to hematopoietic stem and progenitor cells differentiating into neutrophils, the enhanced maps revealed genome-wide loop loss while loops around genes essential for neutrophil function were maintained. TAD F1 improvements across 316 species from the DNA Zoo collection make it usable on genomes far from human without retraining, and the epigenomic decoder is useful wherever Hi-C exists but matched ATAC-seq or ChIP-seq does not.

#Impact

HiCFoundation demonstrates that the masked-autoencoder recipe transfers to contact matrices once the loss accounts for their sparsity, and that a frozen general encoder plus small task decoders is a workable pattern for a field that had accumulated one specialized tool per task. Release practice is unusually complete: source code on GitHub under Apache-2.0, a mirrored Zenodo deposit carrying the code together with the pretrained and fine-tuned models, a separate Apache-2.0 repository holding the data-processing and reproducibility code for the paper, seven checkpoints on Hugging Face covering pretraining and each fine-tuned task, a documented inference.py with per-task commands, a Docker build, and a Colab notebook for users without a local GPU. Constraints are real — a CUDA GPU with at least 12 GB is required, the Hugging Face card is a pointer back to the repository rather than a full model card, and the epigenomic decoder covers a fixed set of assays at fixed resolution rather than arbitrary tracks.