Precision Grounding: augmenting large language models with evidence-based databases for trustworthy genetic variant summarization.

PURPOSE: To propose a novel method that augments LLMs with evidence-based, variant-specific information to improve summarization accuracy and support clinical decision making. METHODS: We proposed Precision Grounding which uses a query tool that integrates domain expert-selected resources and allows users to query relevant factual information using identifiers in databases. In our case, which is genetic variant summarization, we developed CATT (ClinGen website: https://shorturl.at/pw81X; GitHub: https://github.com/mgbpm/clingen-ai-tools; Zenodo: https://doi.org/10.5281/zenodo.18896080), an open-source tool integrating publicly available ClinGen, ClinVar, and GenCC databases. Users can query and retrieve curated evidence via Variation IDs to ground LLM outputs. We compared our approach with baseline methods, including web-search grounding and retrieval-augmented generation (RAG; specifically, MedRAG) using 50 expert-selected variants. RESULTS: GPT-4o was selected due to its good performance on our task during a pilot test. Using GPT-4o, we found web-search grounding performed better than MedRAG since MedRAG failed to generate clinically useful summaries due to limitations of relevant information in its databases. Precision Grounding outperformed web-search grounding, achieving significantly higher accuracy and completeness scores, which were based on a 5-point Likert-Scale of 4.76 (+0.74) and 4.94 (+0.84), respectively. Error analysis revealed that Precision Grounding reduced clinically significant hallucinations, such as incorrect pathogenicity classification and summarizing the wrong variant. CONCLUSION: Precision Grounding outperformed existing grounding approaches for genetic variant summarization. Our open-source tool, CATT, enables integration of curated, domain-specific genetic variant knowledge and significantly reduces hallucinations in LLM outputs. By enhancing the accuracy and completeness of variant interpretation, this framework holds strong potential for real-world implementation in clinical decision-support systems. Its modular, interoperable design allows for seamless integration into clinical genomics workflows, supporting more efficient, trustworthy, and scalable variant review processes.