Abstract
Understanding spatial organization of cells is critical for deciphering tissue function and disease. Single-cell RNA-sequencing (scRNA-seq) profiles transcriptomes at scale but loses spatial context, while spatial transcriptomics (ST) preserves spatial information but is constrained by cost and gene coverage. Here, we present REMAP, a deep learning framework that integrates gene expression with neighborhood-level gene-gene covariance to reconstruct multi-scale spatial organization of scRNA-seq data using one or multiple ST references. Across 2D and 3D mouse brain, human fetal cortex, and seven human cancer types, REMAP consistently outperformed existing approaches. Applied to a human multiple sclerosis atlas, REMAP resolved microglial neighborhood heterogeneity, and identified a rare pro-inflammatory microglia-astrocyte subpopulation. Across diverse cancers, REMAP recovered conserved spatially-defined cancer-associated fibroblast subtypes with known prognostic significance. By transforming cost-efficient single-cell datasets into spatially interpretable tissue maps, REMAP enables spatial hypothesis generation, microenvironment discovery, and population-scale inference of conserved and perturbed architectural principles in human disease.