Integrating Structural Genomic Features for Ovarian Cancer Risk Prediction in Diverse Populations

Abstract Ovarian cancer remains a leading cause of cancer-related mortality due to late-stage diagnosis and the lack of effective screening tools. While polygenic risk scores (PRS) offer some discriminatory value, they predominantly rely on linear SNP effects. We hypothesized that integrating genome-wide structural features could capture an inherited baseline burden of copy number variations. Specifically, we used Log R Ratio (LRR) and B Allele Frequency (BAF) to improve risk prediction. Using the All of Us Controlled Tier, we constructed a demographically balanced cohort of 1,440 ovarian cancer cases and 4,320 matched controls. LRR and BAF metrics were extracted from genotyping arrays for 17 key DNA repair genes and segmented into 4 equal-length bins per chromosome (88 bins total). We evaluated predictive performance across seven feature sets using the H2O AutoML framework, validating with 30 repeated random splits. The integration of structural genomic features yielded modest predictive performance (median AUC = 0.573, SD = 0.046). LRR-containing feature sets consistently outperformed BAF-only and gene-only configurations. A label-permutation negative control confirmed that these results exceeded chance-level performance (p = 1.0 × 10⁻⁵). The models achieved extremely high recall (> 0.95) at the cost of limited precision (~ 0.50), and performance was stable across the demographically diverse cohort (~ 50% non-White). The modeling of germline structural variation provides a verifiable biological signal independent of traditional PRS. Current predictive accuracy limits immediate clinical utility for positive diagnosis. However, such models may hold future utility as negative predictive tools to rule out extreme low-risk populations.