VISTA uncovers missing gene expression and spatial-induced information for spatial transcriptomic data analysis

Understanding cell activities in their spatial context is crucial for revealing spatially influenced cellular states. While single-cell RNA-seq (scRNA-seq) provides comprehensive gene expression profiles, it lacks spatial resolution. Subcellular spatial transcriptomics (SST) captures spatial information but measures only limited genes. To overcome this problem, we present VISTA, a model that predicts unmeasured gene expression in SST data by integrating scRNA-seq and SST through variational inference and geometric deep learning with uncertainty quantification. Across four datasets, VISTA achieves superior imputation accuracy, scalability, and efficiency. Its accurate imputation supports diverse downstream analyses, including identifying spatially variable genes, uncovering ligand-receptor interactions, inferring spatial RNA velocity, performing in-silico perturbations, and disentangling spatial versus intrinsic expression variations. By bridging comprehensive molecular coverage with spatial precision, VISTA enhances the interpretability and utility of SST data, advancing our understanding of tissue organization and cellular microenvironments. VISTA is a scalable, uncertainty-aware model that imputes missing genes in spatial transcriptomics by integrating scRNA-seq and spatial data. It improves accuracy, enables discovery of spatial patterns, CCIs, RNA velocity, and supports perturbation transfer.