Agent SPI-WSI: In context learning for computationally spatial pathwayinferring on whole slide histopathology images conditioned on bulk RNAsequencing using pathologist in the loop.

Bulk RNA sequencing, while cost-effective compared to high-resolution spatial transcriptomics, averages gene expression across heterogeneous cell populations and thus lacks spatial context. To address this limitation, we introduce a structured, human-guided, multi-stage computational AI agent SPI-WSI, that iteratively generates, evaluates, and refines biologically informative natural-language prompts, thereby localizing bulk-derived pathway activity within histopathology slides. Our pipeline uses in-context prompting in large language models (LLMs) to adapt dynamically to the task of prompt generation. Candidate prompts are first produced by the LLM and then subjected to a secondary pathologist critiquing by the same LLM that cross-references PubMed to ensure both biological plausibility and specificity. Each approved prompt against their image tile is scored using the vision language foundation model (CONCH). We benchmarked different LLMs, Gemini 2.0, Gemini 2.5, Claude 3.7 and Claude 4.0, and found that Claude 4.0 achieves the highest cosine similarity (0.7) between image and prompt embeddings. Pathologist-driven scoring and manual segmentation confirm that our method accurately identifies clusters of spatial pathological morphologies. In addition, we have validated the method against ground-truth spatial transcriptomic spots using inhouse and public datasets. Overall, the trend emphasized by ground-truth spatial RNA sequencing prompts is closely aligned with those from bulk prompt. This pathologist-in-the-loop workflow enables large-scale, reproducible tissue profiling and grounds AI-driven spatial annotations.