Finding spatially variable ligand-receptor interactions from spatial transcriptomes

Abstract Spatial transcriptomics has emerged as a groundbreaking tool for studying ligand-receptor interactions between cells, where such interactions can exhibit spatial variability. To identify spatially variable ligand-receptor interactions (SVIs), we present SPIDER, which constructs cell-cell interaction interfaces with minimized Dirichlet energy, models interface profiles with knowledge-graph-informed interaction signals, and identifies spatially variable signals with multiple probabilistic models. We applied SPIDER to twelve datasets from five platforms of various tissues and obtained intriguing insights. First, applying SPIDER to mouse-developing embryo data identifies distinct SVIs that potentially drive the early dorsal-ventral separation of esophageal tracheal progenitors, and applying SPIDER to lung samples suggests the proximal-distal differentiation axis of lung development. Then, from the pancreatic ductal adenocarcinomas dataset, SPIDER identified two SVI-defined clusters at the tumor boundary and an immune niche of mixed immune cells. Furthermore, on a curated breast cancer database of seventeen samples from three cancer subtypes, SPIDER identified transient SVIs along the tumor-microenvironment trajectory that are invariant or specific to cancer subtypes. Finally, in mouse and human cortex samples with manual annotations, SPIDER identified SVIs that improved spot clustering, trajectory inference, and single-cell reconstruction.