Chen, Y.; Fayer, S.; Jain, S.; Benazouz, M.; Sverchkov, Y.; Stone, J.; Sharma, H.; Bergquist, T.; Stewart, R.; Mooney, S. D.; Craven, M.; Radivojac, P.; Starita, L. M.; Fowler, D. M.; Pejaver, V.

Approximately 90% of missense variants in ClinVar are variants of uncertain significance, limiting clinical utility of genetic testing. Variant effect predictors (VEPs) generate scores for any missense variant, providing massive potential to empower classification. Realizing this potential requires calibration to translate VEP scores into evidence. However, current genome-wide calibration masks predictor heterogeneity across genes, causing evidence misassignment. We developed an automated and flexible calibration framework performing gene-specific calibration when control variants are sufficient. For genes with limited control variants, we aggregate protein domains with similar VEP score distributions to enable robust calibration. Applying this framework to three VEPs across 2,769 genes, gene-specific and domain-based calibration increase variants with assigned evidence and improve evidence accuracy versus genome-wide calibration. These calibrations are available through PredictMD (igvf.mavedb.org), providing clinicians with calibrated computational evidence. Together these calibration strategies substantially increase the clinical utility of VEPs for variant classification.