Multimodal 3D Genome Pre-training

TL;DR

MIX-HIC addresses the need for holistic 3D genome understanding by integrating 3D structure (Hi-C) and epigenomic context into a multimodal foundation framework. It uses dual encoders, a cross-modal interaction block, and a cross-modal mapping block trained on the largest publicly available paired dataset to learn both modal-invariant and modal-specific knowledge via contrastive and orthogonal losses. The model achieves state-of-the-art performance on Hi-C contact map prediction, chromatin loop detection, and CAGE-seq expression prediction across GM12878 and K562, and demonstrates strong few-shot performance and cross-cell-type generalization. This work provides a scalable resource and a framework for imputing missing modalities, with potential impact on understanding gene regulation and disease.

Abstract

Deep learning techniques have driven significant progress in various analytical tasks within 3D genomics in computational biology. However, a holistic understanding of 3D genomics knowledge remains underexplored. Here, we propose MIX-HIC, the first multimodal foundation model of 3D genome that integrates both 3D genome structure and epigenomic tracks, which obtains unified and comprehensive semantics. For accurate heterogeneous semantic fusion, we design the cross-modal interaction and mapping blocks for robust unified representation, yielding the accurate aggregation of 3D genome knowledge. Besides, we introduce the first large-scale dataset comprising over 1 million pairwise samples of Hi-C contact maps and epigenomic tracks for high-quality pre-training, enabling the exploration of functional implications in 3D genomics. Extensive experiments show that MIX-HIC can significantly surpass existing state-of-the-art methods in diverse downstream tasks. This work provides a valuable resource for advancing 3D genomics research.

Figures (8)

• Figure 1: Pre-training stage of MIX-HIC. MIX-HIC employs a dual-encoder architecture to extract refined features from both Hi-C contact maps and epigenomic tracks. The modal-specific and modal-invariant representations are learned via contrastive learning and orthogonal constraints within the cross-modal interaction block. A cross-modal mapping block is developed to further regularize the bimodal representations and facilitate cross-modal complement. $\phi_M$ and $\phi_E$ denote the Hi-C contact map encoder and epigenomic track encoder, respectively.
• Figure 2: Fine-tuning stage of MIX-HIC. Epigenomic features are captured from the pre-trained encoder, while Hi-C contact map features are obtained either directly from the pre-trained encoder or through feature mapping based on the epigenomic features. MIX-HIC incorporates a modality fusion block to integrate the bimodal representations, followed by a task-specific decoder for final predictions of downstream tasks.
• Figure 3: Comparison of the number of predicted loops with the number of corresponding ChIA-PET-supported loops across various deep learning methods on GM12878 cell line.
• Figure 4: Few-shot chromatin loop classification performance across different training data ratios. Mean values and standard errors are calculated over five independent runs with varying random seeds.
• Figure 5: Methods comparison for cross-cell-type evaluation.