Single-cell data integration across weakly linked modalities

Rapid advancements in technology enables the measurement of multimodal data at single-cell resolution, but with emerging modalities that are characterized by weak correlations with other modalities. Several computational approaches attempt to integrate these weakly linked multimodal data, but face challenges regarding accurate modeling relationship between cells and learning meaningful cell representation. In this study, single-cell MultiModal data Integration through Hypergraph Contrastive Learning (MMIHCL), a deep learning-based framework that leverages an optimized adaptive k-nearest neighbor graph to model single cell pair-wise relationships for multimodal data integration is presented. MMIHCL uses hypergraph contrastive learning to capture the high-order information of a graph to produce cell representations. Comprehensive benchmarking using a multi-dimensional evaluation framework demonstrates that MMIHCL consistently delivers high-quality integration across diverse weakly linked datasets and maintains high accuracy in strongly linked scenarios. Crucially, MMIHCL exhibits versatile utility in downstream applications: it enables accurate cross-modality feature prediction via explicit cell matching, and empowers robust disease classification and drug target discovery by leveraging optimized joint embeddings. A python implementation of MMIHCL is publicly available at https://github.com/SundayChou/MMIHCL.Author summary: The rapid advancement of single-cell sequencing technology enables us to obtain information such as transcriptomics, epigenomics, and proteomics of individual cells. However, different omics data often “have their own stories”, some of which show loose correlations. We present MMIHCL, a deep-learning framework that first lets each cell elect its most trustworthy neighbors to build a flexible network, then upgrades this network into a hypergraph where local “friend circles” contrast and learn from each other, thereby extracting more robust “cell identity cards”. Beyond achieving superior technical alignment, MMIHCL demonstrates versatile practical utility for biological research. By establishing explicit cell-to-cell matches, it allows researchers to accurately predict missing data modalities, effectively serving as a cross-omics translator. Furthermore, the high-quality cell representations empower the discovery of subtle biological signals, such as distinguishing pathological states in type 1 diabetes and identifying potential drug targets under immune stimulation. This method provides a robust tool for deciphering cellular heterogeneity and disease mechanisms in the era of multi-omics.