Benchmarking of variant pathogenicity prediction methods using a population genetics approach.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Variant pathogenicity predictors are vital for rare disease research.
• Current predictors face challenges due to ascertainment bias and data circularity in training/testing datasets.
• A need exists for reliable benchmarking methods independent of predefined 'ground truth' variant sets.

Purpose of the Study:

• To benchmark commonly used variant pathogenicity predictors using an orthogonal approach.
• To identify the most reliable predictors for distinguishing deleterious genetic variants.
• To assess the utility of the Context-Adjusted Proportion of Singletons (CAPS) metric.

Main Methods:

• Benchmarking pathogenicity predictors using population-level genomic data from gnomAD.
• Utilizing the Context-Adjusted Proportion of Singletons (CAPS) metric for variant analysis.
• An orthogonal approach was employed, avoiding reliance on curated disease or mutagenesis variant sets.

Main Results:

• CADD and REVEL were identified as the best-performing predictors for differentiating variant deleteriousness.
• REVEL showed superior calibration among the evaluated predictors.
• CAPS demonstrated utility as a meta-analysis tool and highlighted biases in ClinVar-based training.

Conclusions:

• The study provides a robust method for evaluating variant pathogenicity predictors.
• CADD and REVEL are recommended for identifying deleterious variants, with REVEL offering better calibration.
• CAPS offers a valuable tool for population-based variant interpretation and bias detection in predictor training.