Pathogenicity Prediction of Missense Variations in Hereditary Cancer Genes

Abstract

HerCanPred, a machine-learning-based pathogenicity classifier specifically optimized for 63 cancer-predisposition genes, was developed to improve the interpretation of missense variants in hereditary cancer syndromes. This model integrates sequence conservation with structural features derived from AlphaFold2 (AF2) structures. HerCanPred achieved a strong performance, outperforming 23 established predictors. SHAP analysis identified AF2-derived structural features, specifically local pLDDT confidence scores and relative solvent accessible area, as the strongest predictors of variant impact. Benchmarking the strengths and limitations of HerCanPred against existing methods showed that misclassification of pathogenic variants was concentrated in disordered and surface-exposed regions, whereas benign failures were more broadly distributed. HerCanPred and three established predictors were also applied to over 57,000 variants of uncertain significance (VUS) from the same gene set. Notably, 166 VUS were reassigned as pathogenic and 75 as benign, with an enrichment of the NF1, FH, and MLH1 genes. By combining gene-specific training with 3D structural information, HerCanPred provides a robust framework for reducing diagnostic uncertainty. Our findings demonstrate that targeted, structure-aware tools can contribute to resolving VUS, providing a rational basis for systematic variant reinterpretation and more informed medical management in hereditary cancer care.