Benchmarking BRCA2 Missense Variant Pathogenicity Predictors Across Functional Domains and in a Subset of Variants Observed in Ashkenazi Jewish Individuals

Pathogenic variants in BRCA2 substantially increase hereditary cancer risk, yet many identified missense variants remain classified as variants of uncertain significance (VUS). This study evaluated SIFT, PolyPhen-2, and REVEL for BRCA2 missense single-nucleotide variants using ClinVar classifications as the reference standard, with gene-wide, domain-specific, and Ashkenazi Jewish (AJ)-observed subset analyses. After missense-only and review-status filtering, 9,077 variants remained; 651 had definitive ClinVar classifications (96 Pathogenic/Likely pathogenic and 555 Benign/Likely benign). All three predictors returned scores for these 651 variants. Because the predictors can return intermediate calls, primary sensitivity and specificity were calculated among confident binary predictions and intermediate-call rates were reported separately. Gene-wide, REVEL had the highest accuracy (92.8%) and specificity (96.0%), whereas SIFT had the highest sensitivity (98.9%) but low specificity (48.7%); PolyPhen-2 showed 81.8% accuracy, 88.2% sensitivity, and 80.5% specificity. In the OB Folds, REVEL showed the best balanced binary-call performance (90.1% accuracy, 90.9% sensitivity, 89.7% specificity), although 93 of 194 variants received intermediate REVEL calls. The BRC Repeats contained no definitively pathogenic missense variants, so sensitivity was not estimable. Using an exact genomic merge and AF_ASJ > 0, 35 missense variants were observed in the AJ subset (27 Benign and 8 VUS/conflicting; no definitively Pathogenic variants). Consequently, AJ sensitivity was not estimable; specificity among confident benign calls was 100% for REVEL (95% CI 85.7%–100%), 100% for PolyPhen-2 (81.6%–100%), and 42.3% for SIFT (25.5%–61.1%). These results support context-aware use of in-silico predictions as supporting evidence, while emphasizing predictor abstention, domain-level uncertainty, and the limits of population-subset inference.