MIX-HIC

Deep learning on the 3D genome has largely been a collection of specialist models: one network calls chromatin loops, another imputes high-resolution Hi-C contact maps, a third predicts transcription from chromatin state. Each is trained from scratch on its own task, and each sees part of the picture. Hi-C measures how the genome folds; epigenomic assays measure how accessible it is. The two are tightly coupled — loop anchors sit at accessible, protein-bound regions — but models rarely learn that coupling as a shared representation.

MIX-HIC, from a team at the Hong Kong University of Science and Technology (Guangzhou) with collaborators at South China Normal University and City University of Hong Kong, is a foundation model that pretrains on both modalities jointly. It is the first such multimodal 3D-genome model, and it required building the corpus to train it: over one million paired samples of Hi-C contact maps and matched epigenomic tracks, assembled from the 4D Nucleome and ENCODE portals. Pretraining teaches two dual transformer encoders to align the modalities while preserving what is unique to each, after which a single set of weights is fine-tuned across structurally different downstream problems.

Posted to arXiv in April 2025 and accepted at NeurIPS 2025, MIX-HIC improves on task-specific state of the art in chromatin loop detection, Hi-C contact map prediction, and CAGE-seq expression prediction. Unlike sequence-first approaches such as Evo2HiC, it learns from the measured contact maps and epigenomic signal themselves.

#Key Features

• Paired multimodal pretraining corpus: 1,275,948 quality-filtered Hi-C/epigenomic pairs, retained from 1,828,401 candidates, drawn from 4DN Hi-C at 5 kb resolution and ENCODE ATAC-seq and DNase-seq across the HepG2, HCT116, IMR90, and WTC11 cell lines.
• Cross-modal interaction block: Four parallel dense networks split each input into modality-invariant and modality-specific representations, held apart by a contrastive loss and an orthogonality constraint so both shared structure and distinctive signal survive.
• Cross-modal mapping block: 1D adaptive pooling aligns sequence lengths across the two modalities and lets the model infer one from the other, so a missing Hi-C map can be filled in from epigenomic tracks alone.
• One backbone, three downstream tasks: The same pretrained weights are fine-tuned for loop detection, contact map prediction, and gene expression prediction rather than retrained per task.
• Released code and weights: The repository ships the full six-step pipeline — data normalization, pretraining-data generation, distributed pretraining, and all three fine-tuning and evaluation scripts — with checkpoints and processed downstream datasets distributed separately through Baidu Cloud.

#Technical Details

MIX-HIC uses two separate transformer encoders. The Hi-C branch takes 50×50 contact submatrices, splits them into 2×2 patches, and flattens them into a patch-embedding sequence processed through cascaded downsampling layers. The epigenomic branch takes 5,000-length track sequences through four convolutional layers with max-pooling before its transformer blocks. Both encoders and the matching decoders use a four-layer design with two transformer blocks per layer, at a feature dimension of 128 for the Hi-C and loop-detection settings and 256 for CAGE-seq expression prediction.

Across downstream benchmarks the model exceeds task-specific baselines on both GM12878 and K562, which are held out from pretraining. Chromatin loop detection reaches AUROC 0.9209 on GM12878 against 0.9046 for DLoopCaller, and 0.9194 on K562 against 0.8924, while recovering more ChIA-PET-validated loops at a higher validated proportion. Hi-C contact map prediction, run in the inference mode that reconstructs contacts from epigenomic tracks alone, reaches R² 0.8724 on GM12878 against 0.7993 for an LSTM-based EPCOT baseline, and 0.8001 against 0.7840 on K562. CAGE-seq expression prediction reaches R² 0.8833 on GM12878 (0.8578 baseline) and 0.9077 on K562 (0.8230). Pretraining runs under PyTorch distributed data parallel; the released environment specifies Python 3.8, PyTorch 2.3, and the usual genomics stack of hic-straw, pyBigWig, and deepTools.

#Applications

The model suits labs studying chromatin architecture in cell types where Hi-C is expensive, low-depth, or unavailable: the cross-modal mapping block turns comparatively cheap ATAC-seq or DNase-seq into a usable contact-map prediction, and the loop-detection head provides 5 kb annotations without a deeply sequenced library. Because a single pretrained backbone transfers across loop calling, contact imputation, and expression prediction, it also serves as a general starting point for new 3D-genome tasks rather than a bespoke architecture per question.

#Impact

MIX-HIC establishes multimodal pretraining as a workable strategy for the 3D genome, and the paired Hi-C/epigenomic corpus it introduces is a reusable asset independent of the model. Three caveats are worth knowing before adopting it. Pretraining covers four cell lines at a single 5 kb resolution, so behavior at finer scales or in distant tissue contexts is untested. The weights and processed datasets are distributed through Baidu Cloud rather than GitHub, HuggingFace, or Zenodo, complicating access outside mainland China. And the repository carries no license file, leaving reuse terms unstated even though the code and checkpoints are public.