A new AI framework can classify hundreds of genetic variants as accurately as a human expert in a fraction of the time, research suggests.
Combining AI-assisted CURAtor (AI-CURA) with the latest large language models (LLMs) could streamline the diagnosis of rare genetic diseases.
The workflow system, described in Science Translational Medicine, performed as well as clinical experts in classifying 150 variants, while adhering to complex expert guidelines.
It was also able to categorize 150 variants with conflicting classifications.
“This study pushes the boundaries of fully automated variant interpretation,” commented senior journal editor Catherine Charneski, PhD, from the University of Bath.
Whole genome sequencing (WGM) has proven pivotal in ending the prolonged diagnostic odyssey of many patients with rare genetic disorders.
To manage the huge number of variants identified through WGS, attempts have been made by expert associations and working groups to establish guidelines and recommendations.
These now form a widely adopted classification system that categorizes variant-associated evidence into distinct rule-based categories.
But while some rules can be readily automated, most evidence still needs manual interpretation of the literature. This requires variant curators to possess broad knowledge across various aspects of molecular biology and genetics, as well as a deep understanding of expert recommendations to accurately interpret and score variants.
Wei Ma, PhD, and colleagues from the Hong Kong Genome Project (HKGP) therefore developed AI-CURA, a fully automated framework for variant classification that integrates LLMs to handle both literature-independent and literature-dependent evidence.
The tool integrates the assessment of evidence for non–literature-based criteria, which can be automated using standard bioinformatic tools, with a separate LLM-supported assessment of literature-based evidence.
Two state-of-the-art LLMs—DeepSeek-R1 and o3-mini-high—were tested for their ability to summarize literature-derived evidence relevant to variant classification.
The team found that the open-source DeepSeek-R1 outperformed o3-minihigh and had high sensitivity and 100% specificity in interpreting rules from the American College of Medical Genetics and Genomics (ACMG) that require understanding literature-based evidence.
They then tested it using 150 variants curated by ClinGen experts, with 150 expert-curated variants and 150 variants with conflicting classifications from the Clinical Genome Resource.
The open-source LLM DeepSeek-R1 showed high concordance with ClinGen experts in establishing a final diagnosis.
“In this study, DeepSeek-R1 demonstrated high accuracy (89.3 to 100%) in determining the application of seven literature-dependent ACMG rules,” the authors reported.
They added: “Our use of LLMs substantially streamlined the variant analysis and interpretation process. LLMs can finish summarizing the literature evidence in minutes.
“In comparison, curators in the HKGP typically spend around four hours per patient on WGS curation, with most of this time dedicated to reviewing literature.”
The post AI-CURA Automates Genetic Variant Classification appeared first on Inside Precision Medicine.