Scalable integration and prediction of unpaired single-cell and spatial multi-omics via regularized disentanglement

Understanding cellular states urgently requires methods capable of integrating large-scale, heterogeneous single-cell and spatial omics data. However, these data are often completely unpaired due to destructive assays and suffer from technical noise, variable feature coverage, and immense scale. We present scMRDR, a scalable computational framework leveraging regularized disentangled representation learning to integrate multiple, completely unpaired single-cell omics datasets with heterogeneous resolutions and coverages. scMRDR overcomes common data-pairing requirements and computational bottlenecks by learning a unified, structure-preserving latent embedding, efficiently scalable to large-scale multi-omics data. This integrated representation further enables robust cross-modal translation like predicting chromatin accessibility from gene expression and, critically, allows for the imputation of spatial coordinates onto non-spatial single-cell modalities using a reference atlas. This spatial mapping capability provides the necessary input for sophisticated, spatially-aware statistical models, enabling the identification of novel spatially variable genes and the dissection of epigenetic regulatory programs within their native tissue context.