Tissue Reconstruction with Generative AI

The spatial arrangement of cells is fundamental to their function, but single-cell RNA sequencing dissociates cells from their native spatial context. We present LUNA, a generative AI model that reconstructs tissue architecture from dissociated gene expression profiles by learning spatial priors from spatially resolved transcriptomics datasets. We validate LUNA by reconstructing the MERFISH mouse brain atlas comprising over 1.2 million cells, inferring the spatial locations of dissociated cells in a mouse central nervous system single-cell RNA-sequencing atlas, and enriching the Slide-tags data by imputing the locations of nuclei lost during profiling. LUNA generalizes to previously unseen cell types in the MERFISH mouse brain atlas and to pathological samples, identifying regions of disease-associated alterations in a mouse model of Parkinson's disease. Our results establish generative tissue reconstruction as an approach for overcoming current technological limitations in spatial biology and advancing our understanding of tissue organization and function.