cfDNA片段组学可实现妇科肿瘤的灵敏早期检测和组织起源预测

Background: Early detection of gynecologic cancers remains challenging due to nonspecific symptoms and limited sensitivity of conventional biomarkers. We aimed to develop and validate cfDNA-based models for cancer detection and tissue-of-origin (TOO) classification. Methods: We prospectively enrolled 1,007 participants from two hospitals, of whom 763 passed eligibility and quality control. The training set (N=363; 173 cancer, 190 non-cancer) was used to develop models integrating four cfDNA features reflecting fragmentation, chromatin architecture, and epigenetic regulation via machine learning. The internal test set (N=158; 86 cancer, 72 non-cancer) and an independent external test set (N=242; 127 cancer, 115 non-cancer) were used for validation. Results: The diagnostic model achieved area under the curve (AUC) values of 0.974 (95% confidence interval [CI]: 0.954-0.994) and 0.975 (95% CI: 0.959-0.992) in the internal and external cohorts, with sensitivities of 83.7% and 82.7% at 98% specificity. High performance was observed across ovarian (AUC: 0.992 and 0.999), cervical (AUC: 0.972 and 0.989), and endometrial (AUC: 0.948 and 0.937) cancers, including stage I disease (AUC: 0.955 and 0.961). The model detected over 77% of cancers that were missed by CA125. Interception modeling projected a 26.4-68.9% increase in stage I diagnoses and 11.6-37.8% 5-year survival gains. The TOO model achieved >73% overall accuracy, with the highest accuracy for ovarian (81.3-86.7%), followed by cervical (70.7-73.3%) and endometrial (59.1-62.7%) cancers. Analytical validation demonstrated robust performance even at ultra-low sequencing depths of 1x, supporting scalability for population screening. Conclusions: cfDNA fragmentomics enables sensitive detection and tissue-of-origin classification of gynecologic cancers, complementing conventional biomarkers. These models hold promise for cost-effective, population-level early detection and risk stratification.