Abstract

4173 Background: Pancreatic ductal adenocarcinoma (PDAC) is one for the most lethal types of cancer. Nearly 80% of the PDAC patients are diagnosed at advanced unresectable stages. Therefore, an accurate early diagnosis assay is urgently needed for reducing PDAC related mortalities. In this study, we aim to utilize cfDNA fragmentomics and machine learning techniques for sensitively detecting PDAC patients. Methods: Total647 cases were prospectively enrolled in this study. Plasma samples were collected from each participant for shallow WGS (~5X). After quality control process, 166 PDAC patients and 167 healthy controls were enrolled as a training cohort. Additional 112 PDAC patients and 111 healthy controls were included as an independent validation cohort. Additional 67 cases with pancreatic benign cystic tumors were also enrolled for model validation including IPMN (Intraductal papillary mucinous neoplasms), MCN (Mucinous cystic neoplasms) and SCN (Serous cystic neoplasms). Four fragmentomics profiles, including copy number variation, fragment size, mutation signatures, and FRAGmentomics-based Methylation Analysis (FRAGMA) were utilized by a machine learning algorithm for developing a predictive model. Results: The predictive model showed an exceptional performance to distinguish PDAC patients from healthy controls, yielding Area Under the Curve (AUC) of 0.993 in the training cohort (5-fold cross validation) and 0.990 in the validation cohort. In the training cohort, our model was capable of detecting PDAC patients at sensitivity of 95.8% and specificity of 95.2% while using 0.49 as cutoff. The validation cohort achieved 99.1% sensitivity at 91.0% specificity while applying the same cutoff. Conclusions: Our model was able to accurately detect PDAC at early stages, by incorporating fragmentomics features through machine learning. It can potentially be used for PDAC early screening, and therefore reducing PDAC related mortalities.