Calibration and evaluation of machine-learning algorithms for missense variant classification under ACMG/ClinGen...

Missense variants represent a large proportion of variants of uncertain significance (VUS) in clinical genetics. The ClinGen framework now enables quantitative use of in silico predictors for variant classification (PP3/BP4), but newly developed machine-learning-based tools such as AlphaMissense (AM) and PrimateAI-3D (PAI3D) have not been evaluated in this context using real-world clinical data.

We calibrated AM and PAI3D using ClinGen's quantitative framework and assessed their performance alongside REVEL, a widely used meta-predictor, across three datasets: ClinVar_2023, ClinGen, and Columbia Diagnostic Exome Sequencing (CDEX). Predictive performance was evaluated using area under the curve (AUC), F1 scores, and classification impact on VUS.

REVEL demonstrated the highest overall performance, particularly for autosomal recessive variants and autosomal dominant variants tolerant to loss-of-function. AM outperformed other tools for autosomal dominant variants in haploinsufficiency genes. Applying updated PP3/BP4 criteria to ClinGen VUS resulted in over 10% being upgraded to likely pathogenic and less than 3% downgraded to likely benign, with REVEL yielding the most LP upgrades and AM the most LB downgrades.

Our findings highlight the complementary strengths of AM, PAI3D, and REVEL under the ClinGen framework, supporting more accurate and mechanism-aware missense variant classification in clinical settings.