HBOCpred: An Ensemble Framework for Prioritisation of Germline Missense Variants in Hereditary Breast and Ovarian Cancer

Abstract

HBOCpred was developed to support research prioritisation of unresolved germline missense variants in a defined 17-gene study panel. Four calibrated learners provide vote states (V0–V4) and a continuous modified adaptive classifier score (SMAC). The same directional descriptors apply to labelled and unresolved records: unanimous benign/likely benign-directed (V0), mixed (V1–V3), and unanimous likely pathogenic/pathogenic-directed (V4) votes. Development used 2160 anchors and five repeats of nested five-fold cross-validation, with preprocessing fitted within training partitions. The annotation-completeness rule permitted evaluation of 2046 anchors per repeat. Mean balanced accuracy for the secondary ≥3-of-4 endpoint was 0.9604; mean SMAC area under the receiver operating characteristic curve (AUROC) and Brier score were 0.9909 and 0.0294. In the first repeat, unanimity covered 1968/2160 anchors (91.11%), with source-label agreement in 1920/1968 unanimous outputs (97.56%). Leave-one-gene-out AUROC was 0.9812 among 2045 evaluable anchors, with substantial gene-specific variation. Among 43,679 unresolved records, V0, V1–V3, and V4 contained 33,414, 4491 and 5774 variants, respectively. HBOCpred makes learner agreement and disagreement available for expert review and functional follow-up. Independent validation is needed to assess its performance and utility in unresolved variants.