CASSIA: a multi-agent large language model for automated and interpretable cell annotation.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Cell type annotation is crucial for single-cell RNA sequencing (scRNA-seq) data analysis.
• Current methods often demand significant computational and domain expertise, leading to interpretation challenges and inconsistent results.
• Existing large language model approaches face issues like hyperconfidence, hallucinations, and a lack of reasoning.

Purpose of the Study:

• To develop an automated, accurate, and interpretable cell annotation tool for scRNA-seq data.
• To address the limitations of existing annotation methods, including accuracy, interpretability, and reliance on manual input.
• To leverage large language models for improved scRNA-seq analysis while mitigating common LLM pitfalls.

Main Methods:

• Development of CASSIA, a novel computational framework for automated cell type annotation.
• Utilized large language models with enhanced reasoning and quality assessment capabilities.
• Benchmarking against existing methods using diverse scRNA-seq datasets, including complex and rare cell populations.

Main Results:

• CASSIA demonstrated improved annotation accuracy across 970 cell types.
• The tool effectively analyzed complex and rare cell populations, outperforming existing methods.
• CASSIA provides users with reasoning and quality assessment for enhanced interpretability and confidence calibration.

Conclusions:

• CASSIA offers a significant advancement in automated cell type annotation for scRNA-seq data.
• The method enhances accuracy, interpretability, and accessibility in scRNA-seq analysis.
• CASSIA's built-in reasoning and quality assessment features help overcome limitations of previous approaches and LLMs.