Integration of cfDNA fragmentomics for early biliary tract cancer detection.

4131 Background: Biliary Tract Cancer (BTC) is a highly aggressive malignancy with poor survival outcomes, primarily due to the lack of effective early detection methods and late-stage diagnoses. Current diagnostic tools, including imaging and invasive endoscopic procedures, are limited in their sensitivity and specificity for identifying early-stage disease. This study addresses this critical gap by developing a novel, non-invasive approach for BTC detection using circulating cell-free DNA (cfDNA) fragmentomics features. Methods: The study cohort included 163 patients diagnosed with BTC and 165 healthy individuals, divided equally into training and validation cohorts. All participants’ plasma samples were collected for a low-depth whole genome sequencing (WGS) process to extract three key cfDNA fragmentomics features: Copy Number Variation (CNV), Fragment Size Distribution (FSD), and Promoter Fragmentation Entropy (PFE). These features were utilized to develop a machine learning model, which was trained and validated through 5-fold cross-validation. An external cohort of 55 patients with benign diseases and 18 Tis/High-grade cases was used to further evaluate the model robustness. Results: The stacked ensemble model reached an Area Under the Curve (AUC) of 0.96 in the validation cohort, showing excellent performance in identifying BTC from healthy participants. At an 86% training specificity cutoff, sensitivity achieved 90.91% (95% CI: 81.26% - 96.59%) and specificity 87.88% (95% CI: 77.51% - 94.62%). While PFE performed as a strong single feature with an AUC exceeding 0.92. The model demonstrated its effectiveness in early-stage detection, with the sensitivity increasing from 80% in stage I to 95.65% in stage II. The model surpassed traditional biomarkers (AUC > 95% compared to ~75% for CA19-9) and demonstrated consistent performance across subgroups. External validation revealed 89% sensitivity for early lesions and 89% specificity for benign cases, highlighting its potential for non-invasive early detection of BTC. Conclusions: This study demonstrates a reliable and non-invasive strategy for early BTC detection, leveraging cfDNA fragmentomics features and a robust machine learning framework. The model’s high accuracy and reproducibility in both internal and external cohorts highlight its potential for clinical implementation, offering a transformative approach for BTC screening. Early diagnosis enabled by this method may significantly improve patient outcomes and survival rates, marking a major advancement in clinical practice.