Quantitative Characterization and Interpretation of Rare Spatial and Transcriptomic Heterogeneity from Spatial Omics using GARDEN

Recent advances in spatial transcriptomics have enabled comprehensive molecular profiling with unprecedented resolution and sensitivity, revolutionizing our understanding of tissue microenvironments and intercellular communication at spot and single cell resolutions. However, existing computational methods are adept at capturing common spatial expression patterns but struggle to differentiate cellular intrinsic variability, often failing to identify and quantify rare cell populations and their unique spatial transcriptional signatures, thus hindering the ability to capture spatial regulation for further biological insights. To address these challenges, we present GARDEN, a novel computational framework that leverages the synergy between dynamic Graph Attention mechanism and a spatially-aware adversarial training strategy to jointly identify and characterize Rare cell types or regions DEtectioN in spatial omics.