Transfer learning of multicellular organization via single-cell and spatial transcriptomics

Abstract

Spatial tissues exhibit complex gene expression and multicellular patterns that are difficult to dissect. Single-cell RNA sequencing (scRNA-seq) provides full coverages of genes, but lacking spatial information, whereas spatial transcriptomics (ST) measures spatial locations of individual or group of cells, with more restrictions on gene information. To integrate scRNA-seq and ST data, we introduce a transfer learning method to decipher spatial organization of cells named iSORT. iSORT trains a neural network that maps gene expressions to spatial locations using scRNA-seq data along with ST slices as references. iSORT can find spatial patterns at single-cell scale, identify key genes that drive the patterning, and infer pseudo-growth trajectories using a concept of SpaRNA velocity. Benchmarking on simulation data and comparing with multiple existing tools show iSORTs robustness and accuracy in reconstructing spatial organization. Using our own new human artery datasets, iSORT shows its capability of dissecting atherosclerosis. Applications to a range of biological systems, such as mouse embryo, mouse brain, Drosophila embryo, and human developmental heart, demonstrate that iSORT can utilize both scRNA-seq and ST datasets to uncover multilayer spatial information of single cells.