Pancreatic ductal adenocarcinoma (PDAC) remains one of the deadliest solid cancers, with only about one in eight patients alive five years after diagnosis. The central challenge is timing. Early tumors often cause no symptoms, and there is no clinically established blood test for early detection. As a result, most patients are diagnosed only after the disease has reached an advanced, often incurable stage.
For years, biomarker development has often started with a seemingly straightforward comparison: how do patients with pancreatic cancer differ from healthy individuals? That comparison is essential for discovery, but it does not fully reflect the question clinicians face in surveillance clinics. The practical question is harder: among people already known to be at increased risk, who is developing a dangerous pancreatic lesion and needs more intensive evaluation now?
Define, enrich and find: starting with the people at greatest risk
Because PDAC is relatively uncommon in the general population, population-wide screening is neither practical nor currently recommended. A more realistic approach is to begin with individuals whose risk is already elevated - those with a strong family history, inherited cancer-predisposition variants, pancreatic cysts, or chronic pancreatitis.
Many of these individuals undergo repeated magnetic resonance imaging (MRI) or endoscopic ultrasound (EUS). These tools are important, but they are resource-intensive and, in the case of EUS, invasive. Our study therefore began with a different question: could a blood test help identify which high-risk individuals should move to more intensive investigation? This is the logic of the Define-Enrich-Find (DEF) framework: define risk, enrich the surveillance population, and then use increasingly sensitive tools to find disease.
A molecular window into the pancreas
Our approach grew from a biological observation: microRNAs can circulate in blood in different forms, and those forms may carry complementary information. Some microRNAs circulate freely in plasma; others are packaged inside exosomes, small membrane-bound vesicles that protect their molecular cargo and can retain features of the cells from which they originated.
We reasoned that examining both compartments might provide a more complete signal than either one alone. Cell-free microRNAs can contribute to sensitivity, while exosome-associated microRNAs may add biological specificity. In this sense, exosomes offer a small molecular window into what is happening inside the pancreas.
In our earlier work, a combined cell-free and exosomal miRNA signature showed promising performance for detecting early-stage PDAC. But a promising discovery signal is not the same as a clinically credible test.
To take the next step, we undertook a prospective, multicenter study that began with 1,785 plasma samples collected across 12 institutions in the United States, Japan, Italy and South Korea. After RNA quality control, 1,757 samples from 1,649 unique participants were analyzed across separate training, validation, external testing, longitudinal treatment and cross-reactivity cohorts.
Building for robustness, not just a high AUC
One of the less visible challenges in developing a miRNA assay is normalization: which stable reference should be used to measure expression across samples reliably? Rather than relying on a single conventional control, we searched large sequencing datasets and identified three stable endogenous reference miRNAs.
We retained all three. For each reference, we trained separate XGBoost models using cell-free and exosomal miRNAs, then combined those outputs into a single miRNA signature. The aim was to reduce dependence on any one normalizer and favor signals that remained consistent across several ways of measuring them.
We took the same approach to model selection. Instead of choosing the configuration that produced the single highest validation AUC, we prioritized the model that behaved most consistently between the training and validation cohorts at clinically important specificity thresholds. Once selected, the model parameters and diagnostic cutoffs were locked before they were applied to subsequent cohorts.
In the independent validation cohort, the locked miRNA signature achieved an AUROC of 96.9%. That was encouraging, but the most important test was still ahead: performance in the people for whom surveillance decisions are actually difficult.
The real test: people at high risk of PDAC
We next evaluated the locked miRNA signature in an independent external testing cohort of 440 participants.
This cohort was deliberately more challenging than the classic cancer-versus-healthy-control setting. It included patients with PDAC or high-grade dysplasia (HGD), average-risk controls, and high-risk controls with familial or hereditary predisposition, chronic pancreatitis, or pancreatic cystic lesions. These were not simply healthy volunteers; they were people in whom a blood test could have the greatest clinical value - and face the greatest biological complexity.
In this setting, the miRNA signature achieved an overall AUROC of 88.6%. The AUROC was 93.7% when PDAC/HGD was compared with average-risk controls and 86.6% when compared with high-risk controls.
For us, this distinction mattered. We were no longer asking whether the assay could separate cancer from health. We were asking whether it could help identify PDAC among people whose genetics, inflammation, or pancreatic abnormalities had already placed them in a surveillance pathway.
Longitudinal sampling added another dimension. MiRNA signature levels generally declined during neoadjuvant chemotherapy and after surgery, with larger decreases among patients who remained recurrence-free. In patients who later relapsed, the signal rose again as recurrence approached.
These longitudinal findings were exploratory, and the study was not designed as a recurrence-monitoring trial. Still, they suggested that the signature may reflect changes in disease burden, an observation that will need dedicated prospective evaluation.
Putting the pieces together: PANXEON
The next question was whether the miRNA signature could provide additional useful information beyond CA19-9, the blood marker most commonly used in pancreatic cancer care. CA19-9 can be helpful for disease management, but its performance is less reliable for early-stage detection, and some patients do not express it at all.
In the external testing cohort, CA19-9 performed better for advanced PDAC than for stage I-II disease. Importantly, among participants whose CA19-9 levels were below the conventional clinical cutoff, the miRNA signature still distinguished early-stage PDAC/HGD from controls.
That finding suggested that the two markers were capturing partly different biology. Rather than asking one marker to do everything, we combined their complementary information.
The resulting composite score is PANXEON - PANcreatic cancer eXosome Early detectiON.
In the external testing cohort, PANXEON achieved an AUROC of 91.7% for distinguishing PDAC/HGD from controls. For stage I-II PDAC, sensitivity was 86.8%, with specificity of 96.8% in average-risk controls and 84.4% in high-risk controls. Compared with CA19-9 alone, PANXEON reduced both false-positive and false-negative classifications.
That trade-off is especially important in surveillance. A false-positive blood test can lead to additional imaging, invasive procedures and considerable anxiety; a false-negative result can delay evaluation of a potentially curable cancer. The goal is therefore not maximal sensitivity at any cost, but a test that can help make the next clinical decision more precise.
From early detection to disease interception
We then asked an even more consequential question: could PANXEON identify advanced precursor disease before an invasive cancer had formed?
We examined 67 patients who underwent surgery for pancreatic cysts considered clinically high risk. Fourteen cysts contained HGD, a pre-invasive lesion sometimes described as stage 0 disease. PANXEON was positive in nine of these 14 cases and remained negative in 40 of the 53 cysts without HGD.
The numbers are small, and no blood test should determine whether a patient undergoes pancreatic surgery on its own. These findings therefore require validation in larger prospective cyst-surveillance cohorts.
Still, this was perhaps the most compelling conceptual shift in the study. The question was no longer only whether we could find pancreatic cancer earlier, but whether molecular surveillance might eventually help identify the disease at a pre-invasive stage, when intervention has the potential to prevent cancer from developing.
How early is early enough?
That possibility raises the question that now matters most: how early does the signal appear?
Our current study shows that PANXEON can detect early-stage PDAC and HGD in blood samples collected around the time of clinical diagnosis or intervention. It cannot tell us how long before diagnosis those molecular changes become detectable. A cross-sectional snapshot can show that a signal is present; it cannot reveal when that signal first emerged.
We are now collecting serial samples from high-risk individuals to trace these molecular changes over time. The next challenge is to define the lead-time window during which HGD or PDAC can be detected early enough to prompt management changes, while maintaining the specificity required for safe surveillance.
PANXEON is not intended to replace MRI or endoscopic ultrasound, and it is not ready for population-wide screening. Its potential role is as a complementary blood-based layer within high-risk surveillance - helping clinicians decide who needs closer investigation and when. Ultimately, the goal is not simply to find pancreatic cancer earlier. It is to find it early enough to change what happens next.