Pancreatic cancer has a five-year survival rate below 12% largely because the vast majority of cases are diagnosed at an advanced stage when surgery is no longer possible. The anatomical location of the pancreas, the absence of early symptoms, and the lack of a reliable screening biomarker all contribute to this diagnostic delay. Detecting PDAC when it is still small and localized would dramatically improve survival outcomes.
CA19-9, the only widely used blood-based marker for pancreatic cancer, has poor sensitivity for early-stage disease and is elevated in several benign conditions. It cannot serve as a reliable screening tool for general populations. New proteomic approaches that measure panels of proteins in blood could capture a broader range of tumor-related signals, potentially detecting cancer at an earlier stage than CA19-9 alone.
Stacked machine learning classifiers that combine multiple individual protein signals through a supervised ensemble approach offer a mathematically principled way to integrate large numbers of protein features into a single classification decision. This study applied such an approach to identify a multi-protein serum panel for early PDAC detection.
The study used 539 serum samples from two prospective cohort studies: the ADEPTS study and the UKCTOCS cancer prevention study. The dataset included 46 PDAC cases and 493 healthy or benign controls, creating a highly imbalanced classification problem requiring careful handling to avoid bias toward the majority class.
Protein measurement was performed using the Olink Oncology II proximity extension assay panel, which quantifies 92 oncology-relevant proteins from minimal sample volumes with high sensitivity and specificity. This was supplemented with five in-house markers: CA19-9, IL6ST, VWF, PKM2, and THBS2, which were selected based on prior literature supporting their relevance to pancreatic cancer biology.
Sixteen diagnosis-specific base-learner classifiers were trained on different subsets of the protein features and then combined through a stacking ensemble, where a meta-classifier learns to optimally weight the predictions of the individual base models. This two-layer approach captures both the individual protein signals and higher-order interactions between them.
The stacked classifier achieved an AUC of 0.98 in the discovery cohort and 0.95 in the validation cohort. These values represent a marked improvement over CA19-9 alone, which achieved an AUC of only 0.79 in the same dataset. The improvement was statistically significant and consistent across both cohorts.
At 90% specificity - a threshold that would minimize false positives in a screening context - the panel achieved sensitivity of 0.99 in discovery and 0.86 in validation. This means approximately 86-99% of PDAC cases would be correctly identified while maintaining a false positive rate of only 10%, a substantial advance over current markers for early detection applications.
The 49 features selected by the model included proteins with known relevance to pancreatic cancer biology: ICOSLG (immune checkpoint ligand), GPNMB (expressed in aggressive tumors), ESM-1 (endothelial cell marker), VWF (vascular marker), ERBB2 (growth factor receptor), CEACAM5 (CEA family), and EGF. This biological coherence supports that the model is capturing genuine tumor biology.
CEACAM5 (carcinoembryonic antigen-related cell adhesion molecule 5) is a well-characterized cancer antigen elevated in multiple gastrointestinal malignancies. Its inclusion in the panel alongside CA19-9 suggests complementary rather than redundant signal, consistent with the substantial improvement over CA19-9 alone.
ERBB2 (HER2) is a growth factor receptor frequently overexpressed or amplified in pancreatic cancer. Its presence as a serum protein in PDAC patients may reflect shedding from tumor cell surfaces or elevated circulating levels from tumor-associated vasculature. ERBB2 is also a therapeutic target, making its elevation clinically significant beyond diagnosis.
The inclusion of vascular and immune-related proteins such as VWF, ESM-1, and ICOSLG suggests that the tumor microenvironment - including angiogenic and immune signaling - contributes detectable signals to the serum proteome even at early disease stages. This supports a model where the tumor alters its systemic environment in ways measurable in blood before symptoms appear.
A serum test achieving AUC 0.95 with 86% sensitivity at 90% specificity could transform surveillance programs for high-risk individuals, including those with familial pancreatic cancer syndromes, new-onset diabetes, or hereditary mutations in BRCA2, PALB2, or ATM genes. For these populations, even imperfect early detection would provide major survival benefit.
The blood-based format of the assay is a major practical advantage. Unlike EUS-based surveillance, which requires a specialist endoscopist and sedation, a serum protein panel can be administered in primary care settings and repeated periodically without significant patient burden. This could enable broad adoption in surveillance programs.
Integration with CA19-9 and other existing markers rather than replacement would likely be the optimal clinical deployment strategy, leveraging the new panel's sensitivity while retaining established infrastructure. Validation in stage-stratified cohorts would clarify whether the panel's advantage is concentrated in early-stage detection, which is the clinically critical question.
This study identified a multi-protein serum panel that achieves AUC 0.95 in validation for early pancreatic cancer detection, substantially outperforming CA19-9 alone. The stacked ensemble approach combining 16 base learners and 49 protein features demonstrated that complex multi-analyte patterns carry powerful diagnostic information.
The sample size, while among the larger available for pre-diagnostic PDAC serum studies, is still limited to 46 cases. Performance in truly early-stage asymptomatic patients detected through prospective screening programs would need to be confirmed, as pre-diagnostic samples differ from samples collected near the time of symptomatic diagnosis.
Future work should prioritize validation in a truly prospective screening cohort, development of a standardized clinical-grade assay format for the Olink panel markers, and cost-effectiveness analysis to determine whether screening performance justifies implementation in specific high-risk populations. If these steps are completed successfully, this panel could become a transformative tool for pancreatic cancer early detection.