Multidimensional 5-hydroxymethylcytosine features in cell-free DNA enable the detection, staging and subtyping of pancreatic ductal adenocarcinoma.

Pancreatic ductal adenocarcinoma (PDAC) is a highly lethal malignant cancer with limited biomarkers for early detection and disease stratification. Here, we investigated whether multidimensional 5-hydroxymethylcytosine (5hmC) features in plasma cell-free DNA (cfDNA) could support the noninvasive detection, staging, and subtyping of PDAC. We performed a genome-wide cfDNA 5hmC analysis in 274 individuals, including 204 patients with PDAC and 70 non-PDAC controls, and extracted seven categories of features covering both coverage-based and fragmentomic signals. PDAC was characterized by widespread and structured 5hmC alterations across multiple genomic and fragment-level feature classes, and these signals reflect widespread multitissue perturbation rather than pancreatic tissue contribution alone. Stage-related analyses revealed a progressive shift from early developmental and metabolic programs toward later immune- and stroma-associated programs. Pathological subtype analysis further suggested progression-associated ordering defined by lymph node metastasis and vascular invasion, with partially distinct molecular features associated with different invasive patterns. Motivated by these findings, we developed a two-level machine learning framework that integrates multiple 5hmC feature types. The final stacked model achieved strong performance for PDAC detection (ROC-AUC = 0.952), while the staging model showed moderate discrimination (macro-AUC = 0.721), and the subtyping model demonstrated good performance (micro-AUC = 0.831; macro-AUC = 0.818). These findings suggest that multidimensional cfDNA 5hmC profiling provides a promising noninvasive framework for PDAC detection, stage assessment, and pathological subtyping.