Abstract
Abstract
Foundation models offer a promising paradigm for modeling spatial transcriptomics, but capturing tissue context over cellular graphs makes training at scale challenging. We introduce spaGFM, a graph foundation model that serializes cellular neighborhoods through random walks to generate transformer-compatible representations of tissue organization. By replacing classical graph message passing with augmented self-supervised neighborhood reconstruction, spaGFM captures higher-order spatial context while allowing scaling at the atlas level. Pretrained on 43.9 million cells from 132 image-based spatial transcriptomics datasets, spaGFM produces robust representations at both the neighborhood and cell level that transfer across tissues, disease contexts, and spatial technologies. SpaGFM identifies tertiary lymphoid structures across cancer cohorts and spatial platforms, captures cell-level transcriptomic perturbation responses associated with T cell vicinity in spatial CRISPR experiments, and characterizes glomerular organization associated with pathological grade in diabetic kidney disease biopsies. Overall, spaGFM establishes a graph foundation-model framework for learning cellular organization and its functional consequences.