Early detection of pancreatic cancer remains one of oncology's greatest unmet needs - the majority of patients are diagnosed at late, incurable stages because there are no validated early detection methods. This pilot study explores a liquid biopsy approach using circulating cell-free DNA (cfDNA) methylation patterns detected in blood to identify pancreatic cancer patients.
The study enrolled 7 pancreatic cancer patients and 12 healthy controls, totaling 19 participants. While small, pilot studies are valuable for establishing proof-of-concept and identifying the most promising biomarker candidates before investing in larger, more resource-intensive validation cohorts.
DNA methylation - the addition of a methyl group to cytosine bases, particularly at CpG sites - is an epigenetic modification that is frequently altered in cancer. Tumor-derived cfDNA carrying cancer-specific methylation patterns can be detected in blood, offering a non-invasive window into tumor biology without requiring biopsy.
cfDNA was isolated from plasma samples and subjected to Illumina Infinium EPIC BeadChip methylation profiling, which simultaneously measures methylation levels at over 850,000 CpG sites across the genome. This comprehensive, unbiased approach captures the full epigenomic landscape of cfDNA rather than targeting predefined loci.
The analysis identified 4,556 differentially methylated CpG sites that distinguished cancer patients from controls. These sites underwent feature selection and dimensionality reduction to identify the most discriminatory subset for downstream AI classification modeling.
The use of EPIC arrays - rather than targeted bisulfite sequencing of known cancer methylation markers - provides an agnostic discovery platform that can identify novel cancer-specific methylation patterns. This discovery-oriented approach is particularly valuable in pancreatic cancer where well-validated methylation biomarkers are still limited.
A total of six different AI/machine learning platforms were applied to the methylation feature set, including traditional machine learning algorithms and deep learning approaches. Remarkably, all six platforms achieved AUC values between 0.90 and 1.00, demonstrating robust and consistent discrimination between cancer patients and controls across diverse modeling approaches.
The top-performing model was a Deep Learning approach that achieved a perfect AUC of 1.00 with 100% sensitivity and 100% specificity in cross-validation. This means the deep learning model correctly classified all 7 cancer patients as cancer and all 12 controls as cancer-free.
The consistency of high performance across multiple modeling approaches suggests that the underlying methylation signal is strong and robust - not an artifact of any specific algorithm - and that cfDNA methylation profiling captures a genuine biological difference between pancreatic cancer patients and healthy individuals detectable in peripheral blood.
Pathway enrichment analysis of the 4,556 differentially methylated CpG sites revealed consistent overrepresentation of several cancer-relevant signaling pathways. The top enriched pathways were Phospholipase D signaling, AMPK signaling, MAPK signaling, and Notch signaling - all of which have established roles in pancreatic cancer biology.
MAPK pathway activation is a key downstream effector of KRAS - the most commonly mutated gene in PDAC - and its epigenetic dysregulation may reflect the widespread transcriptional changes driven by oncogenic KRAS signaling. Notch signaling is implicated in pancreatic cancer stem cell maintenance and treatment resistance.
The identification of these pathways through an unbiased epigenomic approach validates that the differentially methylated CpGs are not random noise but reflect biologically meaningful cancer-related gene regulation changes, lending mechanistic credibility to the methylation biomarker findings and suggesting potential therapeutic targets.
If validated in larger cohorts, cfDNA methylation profiling could serve as a blood-based screening or early detection test for pancreatic cancer - a disease where even a modest improvement in early detection rates could translate to large survival benefits given the dramatic stage-dependent survival differences. Unlike CA19-9 (the current clinical biomarker), methylation patterns are not elevated in non-malignant conditions like pancreatitis.
The primary limitation is the very small sample size (7 cases, 12 controls). Perfect classification performance in such small datasets is prone to overfitting and may not reflect true generalizability. The 100% AUC from deep learning, while exciting, should be interpreted as preliminary evidence requiring substantial validation rather than a clinical-grade result.
Future studies will need to address whether cfDNA methylation signals are detectable at early disease stages (Stage I/II), whether signals can be detected in patients with other pancreatic conditions (pancreatitis, cysts), and whether the approach performs adequately in the at-risk populations (familial pancreatic cancer, new-onset diabetes) most likely to benefit from screening.