Pancreatic ductal adenocarcinoma (PDAC) is one of the deadliest cancers, with a five-year survival rate around 12%. The main reason is late diagnosis — by the time most patients develop symptoms, the cancer has already spread beyond surgical cure. Catching it early, when surgery can still help, dramatically changes outcomes.
Currently, the only FDA-approved blood biomarker for PDAC is CA19-9, but it performs poorly for early detection. In high-risk individuals — such as those with genetic predispositions, a family history of pancreatic cancer, or pancreatic cysts — CA19-9 identifies only about half of early-stage cancers. A better blood test is urgently needed.
This study set out to discover a panel of blood proteins that, when measured together using multiplex technology and machine learning, could detect early-stage PDAC with much greater accuracy than CA19-9 alone.
Researchers collected blood serum from 352 patients across 13 sites in 6 countries. The patient groups included 75 early-stage PDAC patients (stage I or II), 50 healthy controls, 47 high-risk individuals (genetic or familial), 36 patients monitored for precancerous pancreatic cysts (IPMN), and additional groups with new-onset diabetes or chronic pancreatitis.
Two complementary technologies measured proteins in the blood: Olink multiplex proximity extension assays, which quantified roughly 2,261 individual proteins simultaneously, and 27 targeted immunoassays measuring specific candidates of interest including CA19-9. For 16 proteins, both platforms were used to assess agreement between them.
Machine learning — specifically elastic net regularization — then combined the most promising individual protein markers into 4- to 6-protein panel combinations. These combinations were evaluated by their ability to detect PDAC at 95% specificity across multiple patient subpopulations.
The best multi-protein combinations achieved 84% sensitivity at 95% specificity in the primary target population of high-risk controls and early-stage PDAC patients. This compares favorably to CA19-9's sensitivity of only 53% at the same specificity level — a dramatic improvement in detecting early disease.
Univariate analysis identified 41 individual proteins that showed statistically significant associations with PDAC. Proteins such as CTSD, ICAM1, MERTK, TIMP1, GPNMB, and PLA2G1B each showed sensitivity comparable to or slightly below CA19-9 alone, but when combined into panels, their complementary signals produced major improvements.
The multi-protein panels also performed better than CA19-9 in challenging subpopulations — including patients with new-onset diabetes and chronic pancreatitis, two conditions that commonly confuse CA19-9 readings. The AUC of the best combination reached 0.95 compared to 0.84 for CA19-9 alone.
Olink uses proximity extension assay (PEA) technology, where pairs of antibodies bound to DNA oligonucleotides attach to target proteins. When two antibodies bind their target simultaneously, the DNA strands hybridize and are amplified by PCR, creating a signal proportional to protein concentration. This allows hundreds to thousands of proteins to be measured from a small blood volume.
The platform demonstrated strong agreement with traditional ELISA immunoassays for 13 out of 16 proteins tested in parallel. Key biomarkers like LCN2, TIMP1, and CEA showed Spearman correlations of 0.8 to 0.96, confirming that Olink results reliably reflect true protein abundance in the blood.
Some proteins showed no correlation between Olink and immunoassay formats (IL1Ra, OPG, IL6), highlighting that platform-specific differences can occur and that biomarker validation must be performed in the intended assay format for clinical use.
Out of roughly 3,000 proteins measured, 41 emerged as promising biomarker candidates based on statistical significance across one or more patient subpopulations. These 41 proteins form the basis for a next-generation blood test that could be deployed in clinical settings using standard immunoassay platforms.
The clinical relevance of this test would be greatest in high-risk populations who currently undergo surveillance with imaging. A more sensitive blood test could guide the frequency and intensity of imaging surveillance, and might eventually reduce the burden of repeated MRI or endoscopic procedures.
The researchers note that the next step is independent validation of these 41 markers in a separate patient cohort not used for discovery. This is the critical step before any commercial or clinical adoption, and will confirm which markers truly generalize across diverse patient populations.
This study demonstrates that combining multiple protein biomarkers with machine learning can substantially outperform the current standard blood test for early PDAC detection. The 6-plex signatures achieved sensitivity of 84% versus CA19-9's 53% — a difference that could translate to thousands more patients being diagnosed at a surgically curable stage.
A key strength of this work is the multi-national, multi-center cohort design, which captures the biological diversity of real-world PDAC populations. The inclusion of difficult-to-distinguish control groups — chronic pancreatitis and new-onset diabetes — tests the panel's specificity under realistic clinical conditions.
If validated, this approach represents a significant step toward a routine surveillance blood test for pancreatic cancer in high-risk individuals, complementing existing imaging-based screening programs.