MutQA: A Cross-Validated Q&A Dataset for Genetic Mutations
- 1University of New Mexico Cancer Center
Abstract
We introduce MutQA, a dataset of 330,387 citation-grounded question-answer (QA) pairs linking genetic mutations to their functional consequences as described in the biomedical literature. Each record pairs a protein-coding mutation with a natural-language question and an evidence-backed answer anchored to sentences in one of 27,442 open-access PubMed Central articles, spanning 6,620 genes and 125,593 gene-variant pairs. The dataset is constructed using CrossValQA, a framework that enforces information isolation between independent LLMs: one model generates QA pairs from an article, a second model answers the same question without access to the first model's response, and a third model judges whether each model's rationale supports the other's answer, so that retained QA pairs are reproducible from the source text rather than hallucinated. Of 401,406 audited QA pairs, 82.7% pass both directions of this cross-validation; those with a recoverable reference protein sequence form the released benchmark. In a 999-pair stratified audit, two expert annotators agree exactly on 86.4% of cross-validated QA pairs (95% CI 83.7-89.0; n = 638), a tier in which about 90% of ratings are "correct". Individual LLM judges agree with individual annotators at a comparable 87.0% on this tier but far less on QA pairs that fail cross-validation. Fine-tuning MutaPLM on MutQA nearly doubles its mean LLM-judge score for question-conditioned mutation-effect answering (1.45 to 2.82 on a 0-5 scale) and improves mutation-recovery accuracy from effect descriptions from 19.4% to 32.1%. Providing variant-matched MutQA retrieval context, without fine-tuning, improves strict answer accuracy across four LLM families by 20.0-28.6 percentage points, demonstrating the framework's ability to improve the reliability of LLM answers. We also apply CrossValQA to proteomics, providing a proof of concept for its use beyond mutation-specific QA. We release the dataset, construction framework, and evaluation code.
Related
- Handwriting-based Alzheimer’s disease classification on the DARWIN dataset: Repeated nested validation, task-level reduction, and cross-validated explainability
- Exploring Sleep Respiratory Endophenotypes in Prader-Willi Syndrome: Impact of Genetic Mutations
- Fine-grained sugarcane varietal classification using high-resolution PRSS-1 satellite imagery: a field-validated dataset and a spatial context-aware deep learning framework
- The Mottling Phenotype in Chickens Shows Genetic Heterogeneity and Is Caused by Mutations at the EDNRB2 Locus
- Impacts of different SETD1A mutations and genetic backgrounds on human induced pluripotent stem cell-derived hippocampal neurons
BibTeX
@misc{mccourt2026mutqa,
title = {{MutQA: A Cross-Validated Q&A Dataset for Genetic Mutations}},
author = {R. McCourt and D. Arredondo and O. Macaulay and S. Poshiya and O. Edeh and K. Virupakshappa and M. Nguyen Bach and E. Almeida Ruiz and Y. Hu and L. E. Tafoya and D. P. Hudgens and D. Kundnani and A. Sahu},
howpublished = {bioRxiv},
year = {2026},
month = oct,
doi = {10.64898/2026.10.27.762236},
url = {https://www.biorxiv.org/content/10.64898/2026.10.27.762236v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - MutQA: A Cross-Validated Q&A Dataset for Genetic Mutations
AU - R. McCourt
AU - D. Arredondo
AU - O. Macaulay
AU - S. Poshiya
AU - O. Edeh
AU - K. Virupakshappa
AU - M. Nguyen Bach
AU - E. Almeida Ruiz
AU - Y. Hu
AU - L. E. Tafoya
AU - D. P. Hudgens
AU - D. Kundnani
AU - A. Sahu
PY - 2026
DA - 2026/10/27
PB - bioRxiv
DO - 10.64898/2026.10.27.762236
UR - https://www.biorxiv.org/content/10.64898/2026.10.27.762236v1
AB - We introduce MutQA, a dataset of 330,387 citation-grounded question-answer (QA) pairs linking genetic mutations to their functional consequences as described in the biomedical literature. Each record pairs a protein-coding mutation with a natural-language question and an evidence-backed answer anchored to sentences in one of 27,442 open-access PubMed Central articles, spanning 6,620 genes and 125,593 gene-variant pairs. The dataset is constructed using CrossValQA, a framework that enforces information isolation between independent LLMs: one model generates QA pairs from an article, a second model answers the same question without access to the first model's response, and a third model judges whether each model's rationale supports the other's answer, so that retained QA pairs are reproducible from the source text rather than hallucinated. Of 401,406 audited QA pairs, 82.7% pass both directions of this cross-validation; those with a recoverable reference protein sequence form the released benchmark. In a 999-pair stratified audit, two expert annotators agree exactly on 86.4% of cross-validated QA pairs (95% CI 83.7-89.0; n = 638), a tier in which about 90% of ratings are "correct". Individual LLM judges agree with individual annotators at a comparable 87.0% on this tier but far less on QA pairs that fail cross-validation. Fine-tuning MutaPLM on MutQA nearly doubles its mean LLM-judge score for question-conditioned mutation-effect answering (1.45 to 2.82 on a 0-5 scale) and improves mutation-recovery accuracy from effect descriptions from 19.4% to 32.1%. Providing variant-matched MutQA retrieval context, without fine-tuning, improves strict answer accuracy across four LLM families by 20.0-28.6 percentage points, demonstrating the framework's ability to improve the reliability of LLM answers. We also apply CrossValQA to proteomics, providing a proof of concept for its use beyond mutation-specific QA. We release the dataset, construction framework, and evaluation code.
ER -