Project description:Spatially resolved transcriptomics technologies have significantly enhanced our ability to understand cellular characteristics within tissue contexts. However, current analytical tools often treat cell type inference and cellular neighbourhood identification as separate and hard clustering processes, resulting in models that are not comparable across tissue feature scales and samples, thus hindering a unified understanding of tissue features. Our computational framework, SPARROW, addresses these challenges by representing cell types and cellular organization patterns as latent embeddings learned through an interconnected neural network architecture. SPARROW integrates clustering directly into the learning of these latent embeddings, enabling feature extraction specific to clustering while ensuring comparability across samples through shared latent spaces. When applied to diverse datasets, SPARROW outperformed state-of-the-art methods in cell type inference and microenvironment zone delineation and uncovered microenvironment zone-specific fine cell states that reveal underlying biology. Furthermore, SPARROW algorithmically achieves single cell spatial resolution and whole transcriptome coverage---an experimental challenge---by integrating spatially resolved transcriptomics and scRNA-seq data in a shared latent space. This formulation enabled SPARROW to uncover both established and novel microenvironment zone-specific ligand-receptor interactions in human tonsils---discoveries not possible with either data modality alone. Overall, SPARROW provides a comprehensive characterization of tissue features across scales, samples and conditions.
