Abstract 4933: Leveraging chromosome-scale length variants for ovarian cancer risk prediction

Abstract Introduction: Polygenic risk scores are tools used to predict outcomes based on germline genetics. Typically, human germline genetics are described by millions of SNPs. These scores often rely on linear combinations of SNPs to estimate risk, though nonlinear interactions may also contribute. Advanced machine learning techniques are capable of detecting non-linear patterns between groups, but they require a dataset with far more subjects than features, making it infeasible to analyze genetic differences between populations without millions of participants. To address this, we developed a genome representation using a small set of values that summarize chromosome-scale length (CSLV) variations. Using this method and machine learning, we tested whether there are genetic distinctions between women who develop ovarian cancer and those who do not. Methods: To test our hypotheses, data from the NIH All of Us Research Program was used to determine whether CSLV can predict the occurrence of ovarian cancer among women from diverse racial backgrounds. CSLVs were computed from germline DNA for each participant, and we constructed a dataset comprising women diagnosed with at least one of the following conditions: malignant neoplasm or tumor of the ovary, malignant epithelial tumor of the ovary, cell adenocarcinoma of ovary, ovarian cystadenocarcinoma, or endometrioid carcinoma of the ovary (TRUE). An equal number of matched controls without any of these conditions (FALSE) were randomly selected using an exact matching method based on age, race, and sex at birth. Machine learning models were developed using the H2O AI platform to differentiate TRUE/FALSE groups. Model performance was evaluated by calculating the area under the receiver operating characteristic curve (AUC) on unseen test data, with scrambled labels as controls. Results: CSLV parameters were calculated for 312,944 patients from the NIH All of Us V7 controlled tier with genetic microarray data. We then identified a study cohort that included 1,120 patients diagnosed with ovarian cancer at least once during the study period. The average age at the time of first diagnosis with ovarian cancer within the cohort is 55 years. Most ovarian cancer patients self-identified themselves as White (736), None Indicated (168), Black or African American (125), Asian (29), and multiracial (14). Using CSLV values, we evaluated the ability of our model to differentiate between patient groups by assessing the AUC. The AUC analysis provided insight into whether CSLV-based models could reliably predict ovarian cancer risk and occurrence, offering a potential tool for stratifying patients based on genetic risk. Conclusion: A CSLV can serve as an effective genomic representation for predicting ovarian cancer risk and occurrence. By leveraging CSLV and machine learning, this approach has the potential to improve early detection and personalized risk assessment for ovarian cancer patients. Citation Format: Sarvenaz Fatapour, James P. Brody. Leveraging chromosome-scale length variants for ovarian cancer risk prediction [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 4933.