Predicting Pathogenicity: Benchmarking SIFT, REVEL and AlphaMissense on ClinVar Missense Variants

This study compared SIFT, REVEL, and AlphaMissense using 5,000 ClinVar missense variants: an original sample of 1,000 variants and 4,000 additional variants. All variants were evaluated using the same fixed transcript-selection rule. The original sample yielded 752 complete cases, and the additional sample yielded 3,018. In the additional cases, ROC-AUC was 0.9035 for SIFT, 0.9794 for REVEL, and 0.9662 for AlphaMissense. The REVEL-AlphaMissense difference was small but clear at 0.0132 (95% paired bootstrap interval 0.0081-0.0187). Among 983 additional cases absent from a December 2023 ClinVar snapshot, the difference narrowed to 0.0045 (-0.0024 to 0.0115). Among selected-transcript missense variants in the expanded sample (n = 4,591), coverage was 97.84% for SIFT, 84.49% for REVEL, and 98.24% for AlphaMissense. A transcript audit found another missense transcript with a REVEL score for 109 of 155 original missing-score records. AlphaMissense placed 42 original complete cases in its ambiguous range. Both newer tools discriminated better than SIFT. REVEL showed a slight advantage in the additional sample, although this advantage was not evident in the historical snapshot-absent subset. These findings support the use of computational scores for variant prioritization, while emphasizing that clinical interpretation still requires calibrated evidence thresholds and other patient-level evidence.