Robust characterization and interpretation of rare pathogenic cell populations from spatial omics using GARDEN

Spatial omics links molecular measurements to their positions in tissue, revealing cellular organization and interactions. Yet most computational tools highlight common cell types and overlook rare populations that can drive disease. Here we show GARDEN, a computational framework that identifies and characterizes these pathogenic cells or regions in spatial omics by embedding graph-based dynamic attention into a spatially-aware graph fusion contrastive model. GARDEN works consistently across tissues, species and resolution scales, and aligns consecutive sections to reconstruct 3D anatomy. In an Alzheimer's disease model, GARDEN localizes C1qa/C1qb-marked microglia in amyloid-β regions and reveals key immune pathways. In nasopharyngeal carcinoma it identifies tiny tertiary lymphoid structures, and in breast cancer it uncovers inflammatory M1-like macrophages near ductal carcinoma in situ and links them to pro-metastatic signaling. An interpretation module pinpoints key immune signatures, and GARDEN extends to spatial chromatin accessibility, providing insight into epigenetic regulation and informing diagnostics and therapeutic targeting.