TissueNarrator: Generative Modeling of Spatial Transcriptomics with Large Language Models

The intricate spatial organization and molecular communication among cells are fundamental to multicellular systems. Spatial transcriptomics (ST) enables gene expression profiling while preserving spatial context, providing rich data for studying cellular interactions and tissue dynamics. However, most existing computational approaches focus on embedding-based tasks and provide limited generative capacity for simulating cell behavior in situ. Moreover, accurately interpreting spatial interactions requires extensive biological knowledge, which current models do not incorporate. Here, we introduce TissueNarrator, a framework that reformulates spatial omics analysis as a language modeling problem. By representing tissue sections as spatial sentences - rank-based gene lists augmented with spatial coordinates and metadata - TissueNarrator leverages pretrained large language models (LLMs) to learn spatially conditioned gene expression patterns. The model generates realistic, context-aware cellular profiles, predicts intercellular interactions, and performs in silico perturbation analyses. Across multiple ST technologies (MERFISH, Perturb-FISH, and CosMx SMI), TissueNarrator achieves superior quantitative performance and recovers biologically meaningful ligand-receptor and signaling pathways. Furthermore, a conversational inference mode enables natural-language querying of tissue organization. By integrating pretrained biological knowledge with spatial context, TissueNarrator establishes a new, scalable generative paradigm for modeling, simulating, and reasoning about tissue systems.