Bridging Large Language Models and Single-Cell Transcriptomics in Dissecting Selective Motor Neuron Vulnerability

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Interpreting single-cell RNA sequencing (scRNA-seq) data for cell identity and function is challenging.
• Existing methods may not fully capture the biological context of gene expression.
• Leveraging diverse data sources is crucial for advancing single-cell analysis.

Purpose of the Study:

• To develop a novel framework for generating biologically contextualized cell embeddings.
• To integrate gene-specific textual annotations with scRNA-seq data.
• To enhance the interpretability of downstream single-cell analyses.

Main Methods:

• Ranking genes by expression level within each cell of an scRNA-seq dataset.
• Retrieving gene descriptions from the NCBI Gene database.
• Transforming gene descriptions into vector embeddings using large language models (LLMs) including OpenAI models, BioBERT, and SciBERT.
• Computing expression-weighted average embeddings across top-N genes per cell.

Main Results:

• Generation of compact, semantically rich cell embeddings.
• Successful integration of structured biological data with LLM-based text representations.
• Demonstration of a multimodal strategy for enhanced data interpretation.

Conclusions:

• The proposed framework offers a novel approach to understanding cell identity and function.
• The method provides biologically contextualized cell embeddings for scRNA-seq data.
• This approach facilitates more interpretable downstream applications like cell type clustering and trajectory inference.