spaLLM: enhancing spatial domain analysis in multi-omics data through large language model integration.

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Spatial multi-omics technologies offer insights into tissue organization by integrating gene expression data with spatial information.
• Deciphering distinct spatial domains in tissues is hindered by the inherent sparsity of gene expression data in these technologies.

Purpose of the Study:

• To develop a novel computational method, spaLLM, for enhanced spatial domain analysis in multi-omics data.
• To leverage large language models to improve data representation and overcome limitations of sparse gene expression data.

Main Methods:

• spaLLM integrates a pre-trained single-cell language model (scGPT) with graph neural networks and multi-view attention.
• The approach compensates for sparse gene expression data, enhancing sensitivity and resolution across different omics modalities.
• The method is designed to process multiple spatial modalities, including RNA, chromatin, and protein data, with adaptability to new technologies.

Main Results:

• Benchmarking against eight state-of-the-art methods across four datasets and platforms demonstrated spaLLM's superior performance.
• spaLLM consistently outperformed existing methods in multiple supervised evaluation metrics.
• The model effectively improves data representation for spatial domain identification.

Conclusions:

• spaLLM represents a significant advancement in spatial multi-omics data analysis, particularly for identifying spatial domains.
• The integration of large language models offers a powerful approach to address data sparsity and improve analytical resolution.
• The method's flexibility and demonstrated superior performance position it as a valuable tool for biological research.