A machine learning framework combining cfDNA fragmentomics and serum biomarkers for early ovarian cancer detection

Ovarian cancer (OC) is a leading cause of cancer-related mortality in women, largely due to the lack of effective strategies for early detection. Here, we aimed to develop a liquid biopsy assay integrating cell-free DNA (cfDNA) fragmentomic features with serum biomarkers for sensitive OC detection.

Plasma cfDNA from training (n = 91) and independent validation (n = 46) cohorts comprising patients with OC, benign ovarian diseases, and healthy controls, underwent low-coverage whole-genome sequencing to extract copy number variation, fragment size distribution, and Neomer features. Fragmentomic features were first integrated using a stacked machine-learning model and subsequently combined with serum biomarkers CA125 and HE4 to construct the final diagnostic model. Model performance was evaluated in the overall cohort and stratified by disease stage, histological subtype, and tumor grade. An external validation cohort (n = 58) was further used to assess model generalizability.

The combined model integrating cfDNA fragmentomic features and serum biomarkers demonstrated superior diagnostic accuracy compared with all alternative approaches. In the independent validation cohort, the model achieved an AUC of 0.968 (95% CI: 0.896-0.996), with 85.7% sensitivity and 96.0% specificity. For early-stage OC (FIGO stage I and II), the model yielded an AUC of 0.938 (95% CI: 0.864-0.988), achieving 72.2% sensitivity at a specificity of 96%. Robust performance was observed across histological subtypes (AUC: 0.925-0.991) and tumor grades (AUC: 0.976-0.977). Stratified analyses further confirmed strong discrimination between OC and healthy controls (AUC: 0.995, 95% CI: 0.980-1.000) as well as benign ovarian diseases (AUC: 0.963, 95% CI: 0.921-0.993). In the external validation cohort, the combined model maintained robust diagnostic performance, achieving an AUC of 0.962 (95% CI: 0.898-0.991), with 86.2% sensitivity at 96% specificity.

Integrating cfDNA fragmentomics with CA125 and HE4 via machine learning demonstrates strong potential for ovarian cancer detection and clinicopathological subtyping, supporting future evaluation for clinical translation in population screening and preoperative risk assessment.