Abstract
Accurate pathogenicity classification of BRCA1 and BRCA2 variants is critical for personalised risk management in hereditary breast and ovarian cancer. To benchmark current practice, we systematically evaluated 18 widely used in-silico predictors—SIFT, PolyPhen-2, MutationTaster, CADD, REVEL, and 13 others—on a balanced set of 1,803 pathogenic and benign variants curated from the BRCA Exchange database. Nine performance metrics (accuracy, sensitivity, specificity, PPV, NPV, FPR, FNR, Matthew’s correlation coefficient and AUC) were examined. We further integrated these metrics into a single composite score to facilitate ranking. ClinPred outperformed all other tools, achieving an AUC of 0.792, an MCC of 0.421, and the highest composite score (0.81). CADD (raw and phred-scaled), MetaRNN, VEST4, and AlphaMissense formed a robust second tier, whereas SIFT and LRT showed high missing-value rates and low sensitivity. These findings offer practical guidance for the interpretation of clinical variants and establish an up-to-date benchmark for future algorithm development in precision oncology.