GeneAgent: Self-verification Language Agent for Gene Set Knowledge Discovery using Domain Databases.

Area of Science:

• Genomics
• Bioinformatics
• Artificial Intelligence

Background:

• Gene set knowledge discovery is crucial for advancing human functional genomics.
• Large Language Models (LLMs) show promise but suffer from limitations like hallucinations.
• Existing methods require improvement for accuracy and reliability in biological data analysis.

Purpose of the Study:

• To introduce GeneAgent, a novel language agent with self-verification capabilities for gene set knowledge discovery.
• To improve the accuracy and reduce hallucinations in LLM-based biological data analysis.
• To demonstrate the practical utility of GeneAgent in uncovering novel gene functions and accelerating discovery.

Main Methods:

• Development of GeneAgent, a language agent with an integrated self-verification module.
• Autonomous interaction with biological databases to retrieve and validate information.
• Benchmarking GeneAgent against standard GPT-4 using 1,106 diverse gene sets.
• Manual review of GeneAgent's outputs to assess hallucination reduction and narrative reliability.

Main Results:

• GeneAgent significantly outperforms standard GPT-4 in gene set knowledge discovery tasks.
• The self-verification module effectively minimizes hallucinations and enhances analytical narrative reliability.
• Application to novel gene sets from mouse melanoma cell lines yielded expert-validated novel insights.

Conclusions:

• GeneAgent represents a significant advancement in LLM-based functional genomics, offering improved accuracy and reduced hallucinations.
• The self-verification mechanism is key to generating more trustworthy biological insights.
• GeneAgent has the potential to expedite gene function discovery and advance biological research.