Optimal transport for label transfer in single-cell multi-omics integration.

Abstract

Single-cell multi-omics datasets are rapidly expanding, and integrating complementary modalities can provide a more comprehensive view of the molecular mechanisms underlying biological processes. However, cross-modality alignment remains challenging due to modality-specific measurement differences and mismatches in cell-type proportions. Here, we present single-cell Optimal Transport-based Label Transfer (scOT-LT), a semi-supervised label-transfer framework that aligns single-cell RNA sequencing (scRNA-seq) and scATAC-seq data using label-aware unbalanced optimal transport, which tolerates compositional mismatch while favoring label-consistent correspondences. scOT-LT learns a shared embedding through unbalanced optimal transport-guided alignment and transfers cell-type labels from the annotated scRNA-seq reference to unlabeled scATAC-seq via entropic OT coupling. Evaluations on multiple real-world datasets show that scOT-LT achieves strong modality mixing and high label-transfer accuracy, remains robust under downsampled scRNA-seq annotations, and can reliably detect novel cell types. Thus, scOT-LT not only improves integration and label-transfer performance but also yields explicit, interpretable cross-modality coupling, providing a practical approach for multimodal integration and annotation.