MutQA: A Cross-Validated Q&A Dataset for Genetic Mutations

MutQA provides 330,387 citation-grounded QA pairs linking protein-coding mutations to functional consequences, anchored in 27,442 PubMed Central articles across 6,620 genes and 125,593 gene-variant pairs.

• Why it matters: Reliable mutation-effect question answering is limited by hallucinations and poor reproducibility from biomedical text, creating a need for cross-validated, evidence-anchored QA benchmarks.
• What they did: Using CrossValQA, three LLMs generate, answer, and judge QA pairs with information isolation; 82.7% of 401,406 audited pairs pass both directions, and 999-pair audits show 86.4% exact annotator agreement.
• The result: Fine-tuning MutaPLM on MutQA nearly doubles mean judge score (1.45 to 2.82) and improves mutation-recovery accuracy (19.4% to 32.1%), while variant-matched retrieval boosts strict accuracy by 20.0–28.6 points across four LLM families.