STAGE: A Foundation Model for Spatial Transcriptomics Analysis via Graph Embeddings with Hierarchical Prototypes

Spatial transcriptomics offers an unprecedented opportunity to elucidate the spatial organization of tissues by capturing gene expression profiles while preserving tissue architecture. This enables the identification of spatial niches and deepens our understanding of tissue function and disease-associated microenvironments. However, consistent identification of spatial domains across samples, tissues, and even technological platforms remains a formidable challenge, due to low-dimensional and heterogeneous gene panels across platforms, pronounced batch effects, and substantial biological variability between samples. To address these limitations, we propose STAGE, a generalizable foundation model for spatial transcriptomics via graph embeddings. At its core, STAGE introduces a hierarchical prototype mechanism to capture global semantic representations of spatial niches, alongside an efficient online expectation-maximization algorithm to enable scalable learning from large-scale heterogeneous data. Pretrained on a large dataset comprising 32 million cells from 18 tissue types, STAGE learns robust cell representations within their neighborhood graphs and supports niche inference for domain recognition. Comprehensive evaluations on multiple benchmark datasets demonstrate that STAGE substantially enhances domain consistency in cross-platform, cross-sample, and cross-tissue spatial domain identification tasks, outperforming existing state-of-the-art methods. Furthermore, STAGE supports critical downstream biological analyses, highlighting its strong potential as a powerful tool in biological research.