ChatTogoVar: a TogoVar-based retrieval-augmented generation system for precise genomic variant interpretation.

Area of Science:

• Genomics
• Artificial Intelligence
• Bioinformatics

Background:

• Large language models (LLMs) are increasingly used for human genomic variant interpretation.
• General-purpose LLMs exhibit 'hallucinations,' impacting reliability in specialized genomic queries.
• Accurate variant interpretation is crucial for genomic medicine and personalized healthcare.

Purpose of the Study:

• To develop a reliable LLM-based system for human genomic variant interpretation.
• To mitigate LLM hallucinations in genomic variant analysis.
• To enhance the accuracy and practical utility of AI tools in genomic medicine.

Main Methods:

• Developed ChatTogoVar, a retrieval-augmented generation system querying the TogoVar database.
• Integrated retrieved variant information (allele frequency, clinical significance) into LLM prompts.
• Constructed and utilized a benchmark of 1500 template-variant combinations for assessment.

Main Results:

• ChatTogoVar outperformed a general-purpose LLM and an existing specialized system on a 150-question benchmark.
• Automated LLM-based scoring of 1500 questions confirmed ChatTogoVar's superior performance.
• The system demonstrated improved accuracy in genomic variant interpretation.

Conclusions:

• Integrating a reliable variant database with LLMs enhances genomic variant interpretation accuracy.
• ChatTogoVar shows promise as a practical tool for supporting genomic medicine.
• The approach mitigates LLM hallucinations, improving reliability for clinical applications.