Harnessing genotype and phenotype data for population-scale variant classification using large language models and...

Area of Science:

• Genetics
• Bioinformatics
• Machine Learning

Background:

• Variants of Uncertain Significance (VUS) pose challenges in genetic testing for hereditary diseases.
• Underutilization of clinical data hinders VUS reduction due to a lack of scalable strategies.

Purpose of the Study:

• To assess a machine learning approach for improving variant classification and reducing VUS using genotype and phenotype data.
• To determine if machine learning can leverage underutilized clinical data for more accurate variant interpretation.

Main Methods:

• A multi-step machine learning model was developed using patient data from test requisition forms to generate 'patient scores' and 'variant scores' for pathogenicity inference.
• The study included 3.5 million patients, with models evaluated on discrimination, classification performance, and concordance with other pathogenicity measures.
• Clinical Variant Models (CVMs) were integrated into a classification framework, with expert review for high-confidence predictions.

Main Results:

• 595 out of 1,334 developed Clinical Variant Models (CVMs) demonstrated high performance (AUROCpatient ≥ 0.8 and AUROCvariant ≥ 0.8).
• High-confidence CVM predictions provided evidence for 5,362 VUS in 200,174 patients, addressing 23.4% of VUS observations in the studied genes.
• In 17 frequently tested genes, CVMs reclassified over 1,000 unique VUS, reducing VUS report rates by 9-49% per condition.

Conclusions:

• A scalable machine learning approach effectively uses underutilized clinical data to improve genetic variant classification.
• This method significantly reduces the rate of Variants of Uncertain Significance (VUS), enhancing the utility of genetic testing for hereditary diseases.