Research map / Vision-language and multimodal / Human and embodied modeling

Multimodal 3D Genome Pre-training

NeurIPS 2025 · Vision-language and multimodal › Human and embodied modeling · approach: Medical vision-language models · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes MIX-HIC, a dual-encoder Transformer foundation model pre-trained on over 1.2M paired Hi-C contact maps and epigenomic tracks (ATAC/DNase-seq), with cross-modal interaction and mapping blocks, evaluated on Hi-C contact prediction, loop detection and CAGE-seq expression prediction.

Problem

3D genome analysis tools are mostly single-task and single-modality, and paired Hi-C plus epigenomic data is scarce and costly, so a unified multimodal model is needed for transfer across 3D genomics tasks.

Approach

Combine a dual-encoder Transformer with a cross-modal interaction block that splits features into modal-invariant (contrastive-aligned) and modal-specific (orthogonality-constrained) parts, plus a cross-modal mapping block that infers missing Hi-C features from epigenomic tracks.

Key results

• Hi-C contact map prediction R^2 of 0.8724 on GM12878 (+9.3% over runner-up) and 0.8001 on K562 (+2.1%) using MIX-HIC-Infer.
• Chromatin loop detection AUROC 0.9209 on GM12878 and 0.9194 on K562 for MIX-HIC-Bimodal, above DLoopCaller and Peakachu.
• CAGE-seq expression R^2 of 0.8833 on GM12878 and 0.9077 on K562 (+3.0% and +4.3% over best baseline).

Measured effects reported in the paper

Surprising findings

• Without pre-training, bimodal input underperforms Hi-C-only input for K562 loop detection, attributed to heterogeneity between modalities.
• Predicting Hi-C contact maps from epigenomic tracks alone (MIX-HIC-Infer) scores higher than some baselines that use full inputs, despite Hi-C being the target.

Limitations

• Pre-training uses only ATAC-seq and DNase-seq epigenomic tracks and omits DNA sequence.
• Downstream evaluation limited to two cell lines (GM12878, K562) and three tasks.
• Dataset release is not explicitly stated in the text.

Experimental setup

Datasets: 4DN Data Portal Hi-C contact maps ENCODE ATAC-seq and DNase-seq tracks ENCODE CAGE-seq ENCODE CTCF ChIA-PET loops

Benchmarks: GM12878 Hi-C contact prediction K562 Hi-C contact prediction GM12878 chromatin loop detection K562 chromatin loop detection GM12878 CAGE-seq expression K562 CAGE-seq expression

Metrics: R^2 precision recall F1 AUROC proportion of validated loops

Baselines: Epiphany C.Origami EPCOT-LSTM EPCOT-Transformer Peakachu DLoopCaller ChromoSight HiCExplorer EPI-CNN EPI-Graph

Builds on: Vision Transformer EPCOT contrastive learning

Closest papers by meaning

• Structure-Aware Fusion with Progressive Injection for Multimodal Molecular Representation Learning (NeurIPS 2025 · similarity 0.84)
• ULIP-2: Towards Scalable Multimodal Pre-training for 3D Understanding (CVPR 2024 · similarity 0.83)
• MM-Mixing: Multi-Modal Mixing Alignment for 3D Understanding (AAAI 2025 · similarity 0.83)
• Multimodal Lego: Model Merging and Fine-Tuning Across Topologies and Modalities in Biomedicine (ICLR 2025 · similarity 0.83)
• 3D-MolT5: Leveraging Discrete Structural Information for Molecule-Text Modeling (ICLR 2025 · similarity 0.83)
• A Cross Modal Knowledge Distillation & Data Augmentation Recipe for Improving Transcriptomics Representations through Morphological Features (ICML 2025 · similarity 0.82)