SpaLLM: a general framework for spatial domain identification with large language models.

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Spatial transcriptomics (ST) technologies reveal gene expression within tissue context.
• Current ST analysis often ignores gene functional information, limiting biological insights.
• Identifying distinct spatial domains is crucial for understanding tissue organization.

Purpose of the Study:

• To introduce SpaLLM, a framework enhancing spatial domain identification in ST data.
• To integrate large language model (LLM) gene embeddings with traditional ST analysis.
• To leverage biological knowledge from gene descriptions for improved spatial analysis.

Main Methods:

• Utilized pre-computed GenePT embeddings from NCBI gene summaries.
• Combined LLM-derived gene features with cell-gene expression matrices.
• Integrated enriched cell representations with graph-based spatial analysis methods.

Main Results:

• SpaLLM consistently improved spatial domain identification across 12 Visium datasets.
• Validation on an independent osmFISH dataset confirmed SpaLLM's effectiveness.
• The framework successfully integrated gene expression patterns with functional knowledge.

Conclusions:

• SpaLLM offers a generalizable framework for enhancing spatial transcriptomics analysis.
• The integration of LLM embeddings provides biologically informed gene representations.
• SpaLLM is broadly applicable and can be seamlessly integrated into existing pipelines.