TissueNarrator: Generative Modeling of Spatial Transcriptomics with Large Language Models.

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Spatial transcriptomics (ST) provides gene expression data with spatial context, crucial for understanding cellular interactions.
• Current computational methods for ST often lack generative capabilities and struggle to incorporate biological knowledge for accurate interpretation.
• Simulating cell behavior and predicting intercellular communication in situ remains a challenge.

Purpose of the Study:

• To introduce TissueNarrator, a novel framework for spatial omics analysis.
• To leverage large language models (LLMs) for understanding spatially conditioned gene expression patterns.
• To develop a generative model for simulating cellular behavior and predicting intercellular interactions.

Main Methods:

• Representing tissue sections as 'spatial sentences' using rank-based gene lists and spatial coordinates.
• Applying pretrained LLMs to learn gene expression patterns conditioned on spatial context.
• Utilizing the framework for generating cellular profiles, predicting interactions, and performing in silico perturbation analyses.

Main Results:

• TissueNarrator demonstrated superior quantitative performance across various ST technologies (MERFISH, Perturb-FISH, CosMx SMI).
• The model successfully generated realistic, context-aware cellular profiles and predicted biologically meaningful intercellular interactions.
• It identified key ligand-receptor and signaling pathways involved in tissue organization.

Conclusions:

• TissueNarrator establishes a scalable, generative paradigm for spatial omics data analysis.
• The framework integrates biological knowledge with spatial context, enabling advanced modeling and simulation of tissue systems.
• A conversational inference mode allows for natural-language querying of tissue organization, enhancing accessibility.