scCotag: Diagonal integration of single-cell multi-omics data via prior-informed co-optimal transport and regularized barycentric mapping

Recent advances in high-throughput single-cell technologies have enabled characterization of cellular states across distinct omics layers, yielding complementary insights into the organization of biological systems. To jointly leverage these disparate modalities, computational methods have been developed to align non-overlapping cell populations profiled in different omics layers, a task known as diagonal integration, thereby facilitating more fine-grained biological interpretations. However, existing integration approaches unrealistically assume that all cells are alignable and treat prior-knowledge-derived feature correspondences as fully reliable, retaining them unrefined throughout the alignment process. In this work, we introduce scCotag, a co-optimal transport (COOT)-based deep learning frame-work for diagonal integration of single-cell multi-omics data, to address these limitations. scCotag first infers cell alignment and feature correspondence with prior-informed COOT in an iterative manner, and then leverages the resulting transport plans to jointly learn cell and feature embeddings via regularized barycentric mapping and graph reweighting. Systematic benchmarking on human brain, bone marrow, and blood single-cell RNA-seq and ATAC-seq datasets demonstrates the overall superior performance of scCotag over state-of-the-art methods in both cell alignment accuracy and embedding accuracy. In particular, scCotag excels in simulated imbalance scenarios where not all cells are alignable. Applying scCotag to data in postmortem brains of Alzheimers disease (AD) and non-AD (NoAD) patients further yields more fine-grained biological interpretation of regulatory mechanisms. Together, these results demonstrate that scCotag provides a robust framework for single-cell diagonal integration and regulatory inference.