Integrative analysis of cfDNA features from ultra-low coverage whole genome sequencing enables robust detection of ovarian cancer

Background: Despite advances in circulating tumor DNA analysis, reliable detection of oncological disease from ultra-low coverage whole genome sequencing (ulcWGS) remains challenging, particularly at low tumor fractions. This study leverages cell-free DNA (cfDNA) characteristics to develop and evaluate a robust, integrative binary predictive model for ovarian cancer (OC) status screening. OC represents a growing global burden and is often diagnosed at advanced stages due to the lack of specific early symptoms and effective screening strategies, highlighting the need for sensitive and broadly applicable early detection approaches.Methods and Findings: We analyzed plasma cfDNA from OC patients (N = 85) and cancer-free controls (N = 41) using ulcWGS (~1x). Participation in the study was voluntary, and all participants provided written informed consent before any study-related procedures under study approval No. 16119-8/2022/EUIG. Within an integrated workflow combining standardized laboratory processing, bioinformatic pipelines, and machine learning (ML), we extracted 21 features capturing copy number variations (CNVs) and fragmentomic characteristics to identify complementary signatures distinguishing OC from controls. Predictive models were developed using XGBoost with hyperparameter optimization and evaluated on an independent test set (n = 25% of the cohort). A dual-threshold classification strategy was applied to define an uncertainty zone and optimize screening performance.CNV-derived and fragmentomic features assessed in exploratory analysis on the training-validation set showed moderate discriminative power (AUC 0.569 - 0.946) but substantial overlap between groups. On the test set, the CNV-only model achieved an AUC of 0.855 (sensitivity 85%, specificity 50%), while the fragmentomics-only model reached an AUC of 0.8825 (sensitivity 95%, specificity 30%). Both feature domains captured complementary aspects of tumor-derived cfDNA, with fragmentomics favoring sensitivity and CNV-derived metrics improving specificity. Integration of both feature classes improved performance, yielding an AUC of 0.900, sensitivity of 85.00%, and specificity of 90.00%. SHAP analysis confirmed contributions from both feature types without a single dominant predictor.Conclusions: We present an integrative cfDNA framework for OC detection based on ulcWGS that combines CNV and fragmentomic signals to improve diagnostic performance over single-feature approaches. By enabling robust detection of tumor-associated patterns at ultra-low sequencing depth, this approach demonstrates that meaningful cancer discrimination can be achieved without reliance on deep sequencing. This highlights the potential of cost-effective and scalable liquid biopsy strategies for population-level cancer screening their integration into personalized and preventive oncology. Keywords: Liquid biopsy, ovarian cancer, ultra-low coverage whole genome sequencing, cell-free DNA, cell-free tumor DNA, cancer detection, copy number variations, insert size, fragmentomics, machine learning