Large language model consensus substantially improves the cell type annotation accuracy for scRNA-seq data.

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Accurate cell type annotation is crucial for single-cell RNA sequencing (scRNA-seq) data analysis.
• Existing computational methods face limitations such as reference data dependency and model-specific biases.
• Current Large Language Model (LLM) approaches lack robust uncertainty quantification.

Purpose of the Study:

• To develop a novel framework, mLLMCelltype, that leverages collective intelligence from multiple LLMs for improved cell type annotation.
• To overcome the limitations of single-LLM approaches and reference-dependent methods.
• To enhance the reliability and interpretability of scRNA-seq data analysis.

Main Methods:

• Implemented a framework utilizing collective intelligence through an iterative deliberation process among multiple LLMs.
• Employed a consensus mechanism to aggregate insights from independent LLM agents.
• Validated the framework across 49 diverse scRNA-seq datasets.

Main Results:

• Achieved a mean accuracy of 77.2% in cell type annotation, a 15.7-percentage-point improvement over single-LLM baselines.
• Demonstrated high robustness to noisy input data.
• Showcased generalization capabilities on datasets released post-LLM training.
• Provided transparent reasoning chains and consensus-based confidence metrics.

Conclusions:

• mLLMCelltype significantly enhances the accuracy and reliability of cell type annotation in scRNA-seq data.
• The collective intelligence approach mitigates biases and improves uncertainty quantification compared to single LLMs.
• The framework reduces manual annotation effort and facilitates the interpretation of complex cellular landscapes.