We investigated whether DNA methylation and cell-free DNA (cfDNA) fragmentation patterns can improve circulating tumor DNA (ctDNA) detection in advanced pancreatic cancer. In a cohort of 33 patients, ctDNA detection was performed in a tumor-agnostic fashion using DNA methylation, cfDNA fragment lengths, and 4-mer 5′ end motifs. Machine learning models estimating ctDNA levels were built for each individual detection method and their combination. All models significantly differentiated ctDNA levels in patients from healthy individuals (P < 0.001). Using the highest estimated levels in healthy volunteers as cutoffs, ctDNA was detected in 79%, 67%, 67%, and 55% of patients using methylation, fragment length, end motifs, and the combined model, respectively. Univariable Cox regression showed that all ctDNA level estimates were associated with increased hazard ratios (HR, all P < 0.001) for progression-free survival (PFS) and overall survival (OS). Multivariable Cox regression confirmed ctDNA levels as an independent predictor of PFS (HR = 1.9, P < 0.001) and OS (HR = 2.7, P < 0.001). Our findings suggest that machine learning models based on DNA methylation, cfDNA fragment lengths, and cfDNA end motifs can estimate ctDNA levels and predict clinical outcomes in advanced pancreatic cancer.

https://febs.onlinelibrary.wiley.com/doi/10.1002/1878-0261.70116