Cancer genomic profiling predicts pathogenicity of BRCA1 and BRCA2 variants.

Accurate classification of BRCA1 and BRCA2 variants is essential for cancer risk assessment and therapy selection, yet over one-third remain variants of uncertain significance (VUS). Here, using 120,660 real-world cancer genomic profiles with BRCA1 or BRCA2 variants from a >800,000-sample cohort, we develop machine learning models that predict pathogenicity using clinical and tumor-derived features, including a pan-cancer homologous recombination deficiency signature, co-mutated genes, zygosity, and cancer type. Trained on classified variants from ClinVar, our models achieved near-perfect performance, with validation ROC-AUC of 1.000 for BRCA1 and 0.989 for BRCA2 variants with ≥5 observations, translating to strong benign or pathogenic evidence for VCEP classification. Applying these models to 1,073 BRCA1 and 1,639 BRCA2 VUS, we strengthened or enabled classification of 39.48% BRCA1 and 50.52% BRCA2 assessable variants. This approach transforms underutilized tumor profiling data into evidence that can be directly integrated into variant classification, providing a scalable framework for other tumor profiling datasets and cancer genes associated with defined tumor genomic features.