Identifying Cellular Niches in Spatial Transcriptomics: An Investigation into the Capabilities of Large Language Models

Spatial transcriptomic technologies enable measuring gene expression profile and spatial information of cells in tissues simultaneously. Clustering of captured cells/spots in the spatial transcriptomic data is crucial for understanding tissue niches and uncovering diseaserelated changes. Current methods to cluster spatial transcriptomic data encounter obstacles, including inefficiency in handling multireplicate data, lack of prior knowledge incorporation, and producing uninterpretable cluster labels. We introduce a novel approach, LLMiniST 1 , to identify spatial niche using a zero-shot large language models (LLMs) by transforming spatial transcriptomic data into spatial context prompts, leveraging gene expression of neighboring cells/spots, cell type composition, tissue information, and external knowledge. The model was further enhanced using a two-stage fine-tuning strategy for improved generalizability. We also develop a user-friendly annotation tool 2 to accelerate the creation of well-annotated spatial dataset for fine-tuning. Comprehensive method performance evaluations showed that both zero-shot and fine-tunned LLMiniST had superior performance than current non-LLM methods in many circumstances. Notably, the two-stage fine-tuning strategy facilitated substantial crosssubject transferability. The results demonstrate the feasibility of LLMs for tissue niche identification using spatial transcriptomic data and the potential of LLMs as a scalable solution to efficiently integrate minimal human guidance for improved performance across large-scale datasets.