sciLaMA: A Single-Cell Representation Learning Framework to Leverage Prior Knowledge from Large Language Models

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Single-cell RNA sequencing (scRNA-seq) offers high-resolution cellular insights but faces analytical challenges.
• Existing deep learning models struggle to integrate biological knowledge or handle tabular gene expression data efficiently.
• Large Language Models (LLMs) present computational and applicability limitations for scRNA-seq data.

Purpose of the Study:

• To introduce sciL-aMA, a novel framework for representation learning in scRNA-seq data analysis.
• To bridge the gap between task-specific models and LLMs by integrating gene embeddings with gene expression data.
• To provide a computationally efficient and interpretable method for single-cell data analysis and gene module discovery.

Main Methods:

• Developed sciL-aMA, a framework combining multimodal LLM gene embeddings with scRNA-seq data.
• Utilized a paired-Variational Auto-Encoder (VAE) architecture for integrated representation learning.
• Generated context-aware representations for both cells and genes.

Main Results:

• sciL-aMA outperforms state-of-the-art methods in key downstream scRNA-seq tasks.
• Demonstrated superior performance in batch effect correction and cell clustering.
• Achieved effective identification of cell-state-specific gene markers and modules.

Conclusions:

• sciL-aMA provides a computationally efficient, unified framework for comprehensive single-cell data analysis.
• The model enables biologically interpretable gene module discovery.
• This approach enhances the utility of LLMs in the analysis of scRNA-seq data.