Abstract / Summary
Background Pancreatic ductal adenocarcinoma (PDAC) is highly lethal, largely because diagnosis typically occurs at advanced, incurable stages. Early detection via blood-based biomarkers could improve risk stratification. We investigated whether whole-blood DNA methylation patterns can identify individuals at elevated PDAC risk years before clinical diagnosis, using a deep learning-based anomaly detection framework applied to methylation array data from prospective cohorts. Methods Autoencoder models were trained on healthy control samples (N=27,185) from diverse sources available in the EWAS Data Hub and applied to Michaud’s prospective cohorts (blood collected between 1982-89), comprising 393 future PDAC cases stratified by cumulative time-to-diagnosis cutoffs (N=26 ≤ 3 y; N=57 ≤ 5 y; N=89 ≤ 7 y; N=138 ≤ 1 0 y) and 431 controls. Autoencoders were trained on windows of 10,000 consecutive CpGs with reconstruction error (REC) scores quantifying methylation anomalies. Predictive value was evaluated via stratified cross-validation using ROC AUC [95% CIs] and benchmarked against raw β-value-based classifiers. External validation was performed in the Framingham cohort (N=14 cases; N=1,331 controls, recruited between 1971-75). Functional enrichment of REC-flagged CpG regions was assessed using gene-set over-representation analysis against KEGG terms. Results In the prospective cohorts, REC-based classification achieved its highest macro AUC of 0.758 [0.696–0.820] in the ≤ 3-years-to-diagnosis subset, versus 0.547 [0.445-0.649] for the 𝛽 -value benchmark (ΔAUC = 0.221, p = 0.010). External validation yielded an AUC of 0.614 [0.452–0.771]. REC-flagged regions were enriched in PDAC-relevant pathways, including p53 signaling and pancreatic cancer. Conclusion Whole-blood DNA methylation harbours prediagnostic PDAC-associated anomalies that are detectable through deep learning-based reconstruction error modelling but not through conventional β-value analysis.