SPELL: Spatial Prompting with Chain-of-Thought for Zero-Shot Learning in Spatial Transcriptomics

Zero-shot learning (ZSL) for cell-type classification in spatially resolved transcriptomics remains underexplored, particularly when integrating spatial context with marker gene semantics. Here, we introduce SPELL (Spatial Prompt-Enhanced Zero-Shot Learning), combining graph autoencoder (GAE)-derived spatial embeddings with chain-of-thought (CoT) prompting for zero-shot classification. SPELL uses a spatial k-nearest neighbor graph to encode local cellular neighborhoods and generates interpretable prompts that integrate marker gene expression and the spatial embedding norms. We evaluated SPELL across five state-of-the-art zero-shot LLM classifiers on MERFISH, MIBI-TOF, and Stereo-seq datasets for cell-type classification. Guided by only expression values and spatial context, the two BART models solved the classification task surprisingly well (distilbart-mnli-12-1i 64\% accuracy on the MERFISH, bart-large-mnli achieved 52\% accuracy on MIBI-TOF dataset). Interestingly, removing the spatial context from the CoT prompt-revealed a significant performance drop (20 – 24 \% drop in accuracy), underscoring spatial information's critical role in zero-shot learning. Our work bridges spatial omics with LLM reasoning, enabling flexible adaptation to unseen cell types while maintaining biological interpretability.