SpaIM: single-cell spatial transcriptomics imputation via style transfer

Abstract

Abstract Spatial transcriptomics (ST) technologies have transformed our understanding of cellular organization but are limited by sparse signals and restricted gene coverage. To address these challenges, we introduce SpaIM, a style transfer learning model that leverages single-cell RNA sequencing (scRNA-seq) data to predict unmeasured gene expressions in ST profiles. By disentangling shared content and modality-specific styles, SpaIM effectively integrates scRNA-seq’s rich gene expression with the spatial context of ST. Evaluated across 53 datasets spanning sequencing- and imaging-based spatial technologies in various tissue types, SpaIM consistently outperforms 12 state-of-the-art methods in improving gene coverage and expression accuracy. Furthermore, SpaIM enhances downstream analyses, including ligand-receptor interaction inference, spatial domain characterization, and differential gene expression analysis. Released as open-source software, SpaIM expands accessibility and utility in ST research. Overall, SpaIM represents a robust and generalizable framework for enriching ST data with single-cell information, enabling deeper insights into tissue architecture and cellular function.