PromptSTG: prototype-guided prompting for few-shot spatial transcriptomics annotation

Single-cell resolution spatial transcriptomics (scST) simultaneously captures gene expression and spatial coordinates at an unprecedented scale, which provides powerful opportunities to dissect tissue architecture and cell-cell interactions. However, accurate cell type annotation remains challenging, particularly in scenarios where reliable labels are scarce and manual annotation is costly. These challenges are further amplified in tissues characterized by complex spatial dependencies and pronounced cellular heterogeneity, especially within tumor microenvironments. To address these issues, we propose Prompt-guided Spatial Transcriptomics Graph (PromptSTG), a graph-based few-shot learning framework for robust cell type annotation in scST data. PromptSTG integrates spatial information and transcriptomic features to model biologically coherent cellular neighborhoods and enables accurate label propagation from a small set of labeled cells to large unlabeled populations. Across extensive benchmarks spanning multiple spatial transcriptomics platforms and tissue types, PromptSTG consistently outperforms existing methods in annotation accuracy, robustness, and scalability under few-shot settings. Moreover, PromptSTG reconstructs spatially coherent tissue organization and effectively identifies rare yet biologically important cell populations, such as immature oligodendrocytes and ependymal cells, which together account for less than 7% of total cells. In complex tissue environments, the method preserves both global tissue structure and fine-grained cellular boundaries, yielding biologically meaningful and spatially consistent annotations. All source codes are available at https://github.com/KEAML-JLU/PromptSTG.