Beyond a single threshold: calibration of CADD v1.7 and AlphaGenome for ACMG/AMP computational evidence across variant classes, model versions, and reference data

Abstract

Clinical laboratories use computational predictors as supporting evidence for or against pathogenicity under the ACMG/AMP guidelines, but this evidence is reliable only when score thresholds are calibrated. Genome-wide predictors such as CADD are often applied using a single cutoff across all variant types and model versions, and the new AlphaGenome Variant Impact score (AVI) has not been calibrated. Following a prespecified plan and the ClinGen calibration method, we calibrated CADD v1.7 and AVI on rare ClinVar missense variants. Both predictors provided moderate evidence of pathogenicity and strong evidence of benignity, and their thresholds held for variants classified after calibration. AVI performed better than CADD in each comparison and approached the missense-specific predictors. Thresholds available in 2024 gave AVI evidence in the final direction for 72.5% of the variants of uncertain significance (VUS) resolved by 2026, compared with 60.5% for CADD. Applied to 1.2 million current missense VUS, the calibrated predictors agreed on pathogenic or benign evidence for four in ten, and on strong pathogenic evidence for 186 VUS reviewed by expert panels, which can be reassessed first. Outside missense variants, the same thresholds failed, producing false-benign evidence for pathogenic noncoding variants. A change in CADD version altered the evidence level for a third of variants without improving discrimination. Our results suggest that computational evidence is most reliable within the variant class and model version in which it was calibrated. In practice, laboratories may prioritize missense-specific predictors, use AVI when a genome-wide predictor is needed, and reassess thresholds after model updates.