Circulating Protein Biomarkers for Use in Pancreatic Ductal Adenocarcinoma Identification

Pancreatic ductal adenocarcinoma (PDAC) is a dismal disease, with less than 10% of patients surviving 5 years after diagnosis (1, 2). PDAC is currently the fourth leading cause of cancer-related deaths, but is expected to be the second leading cause by 2030 (1, 3). Symptoms of PDAC are initially vague, and 80% of patients are diagnosed at an advanced stage (2, 4). This leaves less than 20% of patients eligible for tumor resection, the only potentially curative treatment (4, 5). PDAC arises from premalignant lesions, including pancreatic intraepithelial neoplasia, intraductal papillary mucinous neoplasms (IPMN), and mucinous cystic neoplasms (2).

Detection of PDAC at an early stage while the tumor is still potentially curable is, therefore, critical. Despite the quality of today's imaging modalities, their accuracy for early detection of PDAC remains imperfect (6), precluding them as screening tools. Carbohydrate antigen 19-9 (CA19-9) is the only biomarker with clinical usefulness in PDAC, namely monitoring patients after surgery and during oncologic treatment (7). However, elevated CA19-9 is not specific for PDAC and can also be found in other types of cancer and benign conditions (e.g., cholestasis). Furthermore, between 10% and 15% of patients with PDAC are Lewis blood group negative, and therefore lack the enzyme necessary to synthesize CA19-9 (2, 8). The sensitivity and specificity of CA19-9 for discriminating patients with PDAC from healthy controls are 68% and 95%, respectively, with a ROC AUC value of 0.82 (9). For discriminating PDAC from benign pancreatic diseases, CA19-9 has a sensitivity of 81.7% and specificity of 82% with an AUC of 0.88 (7). Hence, CA19-9 cannot be used alone for early diagnosis of PDAC.

Inflammation is a hallmark of cancer (10) and a known risk factor for PDAC (11–14). Chronic pancreatitis is often present alongside PDAC (15). Furthermore, a specific hallmark of PDAC is the dense desmoplastic stroma in the tumor composed of both cellular and noncellular components (2, 16). In a typical PDAC tumor, cancer cells may comprise only 5%–20% of the cells (16). Other cell types include immune cells, endothelial cells, and different stromal cells, mainly modified fibroblasts called stellate cells (16). The noncellular components include extracellular matrix, enzymes, growth factors, and cytokines (17). High plasma concentrations of several cytokines are associated with poor survival in PDAC, in particular IL6, IL8, and IL10 (18). IL6 promotes an antiapoptotic and proliferate state in tumor cells, enabling resistance to both chemotherapy and immunotherapy (19, 20). The activation of intracellular pathways by IL6 leads to secretion of IL8 and IL10. IL8 is associated with chemoresistance toward gemcitabine (21). Furthermore, IL6 leads to secretion of other proteins, including matrix metalloproteinases (MMP), TGFβ, and VEGF (22).

The discovery of new biomarkers for diagnostic workup of PDAC is needed. Several attempts have been made in proteomics, but none have yet translated into daily clinical practice. One of the great challenges in proteomics is capturing low-abundance disease-associated proteins in the huge pool of plasma proteins found in healthy individuals (23). An antibody-based proximity extension assay (PEA) provides sensitive and specific detection of low-abundant proteins in blood (24, 25). However, data are missing regarding the performance of these markers in discriminating between patients with PDAC, benign pancreatic disease, and healthy individuals.

We examined circulating protein levels in patients with PDAC (stages I–IV), healthy blood donors, and patients with nonmalignant pancreatic diseases with the Multiplex Immuno-Oncology Assay from Olink to develop protein biomarker panels determined in serum samples. Such a panel would represent a major step in diagnosing patients earlier than today and thereby facilitate early surgery, ultimately leading to an overall improvement in patient survival.

The study was conducted according to the REMARK (26) and TRIPOD guidelines (27), see the Supplementary Data for checklists.

This prospective, multicenter study included 985 serum samples from five different groups: (i) 701 patients with PDAC stage I–IV, (ii) 26 patients with chronic pancreatitis, (iii) 38 patients with IPMN, (iv) 38 patients with benign pancreatic diseases, and (v) 180 healthy blood donors. The patients in groups (i)–(iv) were included in the Danish multicenter open cohort study: the biomarkers in patients with pancreatic cancer study [“biomarkers in patients with pancreatic cancer (BIOPAC) – can they provide new information of the disease and improve diagnosis and prognosis of the patients;” ClinicalTrials.gov ID: NCT03311776; www.herlevhospital.dk/BIOPAC; refs. 28, 29]. The BIOPAC study was approved by the Institutional Review Board of the Danish Ethics Committee (VEK, j.nr. KA-20060113) and the Danish Data Protection Agency (j.nr. 2012-58-0004, HGH-2015-027, I-Suite j.nr. 03960). Inclusion took place at the Department of Surgery, Rigshospitalet, Copenhagen University Hospital (Copenhagen, Denmark), and five oncology departments in Denmark. Of the patients with locally advanced or metastatic PDAC, 75.5% were included at Herlev Hospital (Herlev, Denmark). Staging was updated according to tumor–node–metastasis classification version 8. For the patients undergoing surgery, the staging was done by pathologists, otherwise the staging was done on the basis of imaging modalities (typically CT scans).

Patients from the BIOPAC study were selected according to available serum samples and classified according to the histologic or cytologic pathology results as PDAC, chronic pancreatitis, IPMN, or benign pancreatic diseases (e.g., serous cystadenomas and pancreatic heterotopia). Patients with IPMN, chronic pancreatitis, and benign pancreatic diseases were included in the BIOPAC study as suspected cases of PDAC prior to surgery at the Department of Surgery, Rigshospitalet, Copenhagen University Hospital (Copenhagen, Denmark).

All patients included in the BIOPAC study provided written informed consent. The serum samples were collected before the first treatment according to national guidelines (surgery or first-line palliative chemotherapy) in the period between July 1, 2008 and November 11, 2018.

Samples from the healthy individuals were collected in January and February 2019 from the Danish corps of volunteer blood donors at Aalborg University Hospital (Aalborg, Denmark). We sought to have an equal distribution between males and females and preferably included healthy individuals aged 60 years and older. After the age of 70 years, it is no longer possible to donate blood in Denmark.

Up to 11 years of follow-up data were available for patients (PDAC and non-PDAC) included in the BIOPAC study. No follow-up data were collected for the healthy blood donors. The study was conducted in accordance with the Declaration of Helsinki.

All samples were centrifuged at 2,300 × g at 4°C for 10 minutes, and serum was then aliquoted in Greiner Tubes (Cryo.s Freezing Tubes, 2 mL, GR-121280, Greiner Bio-One GmbH). The serum was subsequently stored at −80°C. Upon collection for this study, samples were thawed at room temperature, mixed using a vortex mixer, and centrifuged at 3,800 rpm for 10 minutes. Then, 250 μL was aliquoted to tubes (2.0 mL Graduated w/o Ribs Screw Tubes, Natural from SSIbio), labeled with an individual number, and stored at −80°C until analysis.

Serum levels of CA19-9 were determined using the Immulite 2000 GI-MA Assay (Siemens, catalog no., L2KG12), which is a solid-phase, two-site sequential chemiluminescence immunometric assay. Imprecision was monitored with two internal controls at 16 and 83 kU/L, with coefficients of variation of 8% and 9%. Accuracy was monitored within the standard UK NEQAS program. Elevated CA19-9 was defined as >37 kU/L. Serum CA19-9 was not available for 40 samples (PDAC, n = 36; benign pancreatic diseases, n = 1; and healthy individuals, n = 3).

Serum samples were analyzed for 92 proteins (for the full list of proteins, see Supplementary Table S1) using the Proseek Multiplex Immuno-Oncology Assay from Olink Proteomics (www.olink.com; ref. 25). The immuno-oncology panel was chosen a priori because of the known linkage between inflammation and the development of pancreatic cancer, and furthermore, PDAC is known for an immunosuppressive microenvironment (13, 14, 30).

The measurements were based on the PEA technology, in which 1 μL of serum is mixed with a set of 92 pairs of antibodies that are linked to oligonucleotides (probes). Upon binding to the target antigen, the probes are brought into proximity, which leads to the oligonucleotides being extended by a DNA polymerase. This then acts as a surrogate marker for the specific antigen and is quantified by real-time PCR. Data are presented as normalized protein expression (NPX) values, which is an arbitrary unit on log₂ scale to normalize data and minimize intra- and interassay variation. A high NPX value corresponds to a high protein concentration (24).

Samples were randomized across Olink PEA plates and normalized for any plate effects using the built-in interplate controls according to the manufacturer's recommendations. In addition, eight patient samples from the first five plates analyzed before the rest of the plates were included on all subsequent plates for bridging purposes. For details on the assay limit of detection, see Supplementary Materials and Methods.

The analyses were performed according to the manufacturer's instructions at BioXpedia. BioXpedia was blinded to the study endpoint as no research questions or clinical data were passed on before all samples had been analyzed.

The statistical analyses and model building were exploratory because no single validated model exists for these types of analyses. The same data and research questions were given to two bioinformaticians, who independently, and blinded from each other's work, developed a prediction model using slightly different approaches. In the first approach, values for missing CA19-9 analyses in 40 samples were imputed, and no samples were excluded. In the second approach, three samples (2 patients with PDAC and one healthy participant) were excluded because of outlier analyses, and the 40 samples with missing CA19-9 values were excluded. All analyses were done with data of the 92 proteins from the Olink immuno-oncology panel and CA19-9.

The dataset was split into discovery and replication cohorts. With the first approach, 66% of the samples were assigned to the discovery cohort, and with the second approach 70% of the samples were assigned to the discovery cohort. The remaining samples were assigned to the respective replication cohorts. In the first approach, both the discovery and the replication cohorts were further split into two equally sized training and test cohorts.

A comparison of protein levels according to diagnosis (PDAC vs. non-PDAC, meaning all patients with nonmalignant pancreatic diseases and healthy individuals) was made with a t test or Wilcoxon rank-sum test, where appropriate. Box plots were created to visualize abundance values of each protein stratified by diagnosis. Volcano plots were made to describe the relationship between the P values and the log₂ fold changes of the proteins for the comparison of patients with PDAC versus non-PDAC individuals. Principal component analyses were performed to see how well the diagnostic groups were separated. In the first approach, LASSO and Ridge regression models were used to explore the performance and stability of the differentially expressed proteins. The signatures were identified by fitting a model on the training set and thereafter employing this on the test set, repeating the process 100 times. The 92 proteins + CA19-9 were each given a proportion score according to the number of times the protein was included in a model.

For each model fitted this way, another model was fitted using the protein signature and the age of the patient. A DeLong test was used in the replication cohort to compare the AUC of each protein signature, with age added as a predictor with the same signatures without age added as a predictor.

In the second approach, LASSO regression and elastic net models were used to explore the performance and stability of the differentially expressed proteins. With both statistical approaches, panels of proteins were reported for the differentiation of patients with PDAC from all non-PDAC individuals, and for the differentiation of patients with PDAC versus healthy individuals, and patients with PDAC versus patients with nonmalignant pancreatic diseases. Results are presented as ROC curves and AUC values. For further details, see Supplementary Materials and Methods.

Statistical analyses were conducted in R version 3.6.1 (first approach) and R version 3.6.3 (second approach), and P < 0.05 was considered statistically significant. The P values resulting from the first statistical approach were corrected for multiple testing using the Benjamini–Hochberg method. No correction for multiple comparisons were performed in the second approach. Models were fitted and evaluated using the R packages glmnet and pROC. Plots were generated using the R package ggplot2. Imputation was done using the R package impute.

No power calculations for appropriate study size were made because this was a biomarker discovery study. If not stated otherwise, we considered all patients with PDAC as one group regardless of stage.

Patient characteristics are shown in Table 1. The median age of patients with PDAC was 68 years. There was a significant difference in age (P < 0.01) between patients with PDAC and patients with chronic pancreatitis and benign pancreatic diseases, and healthy individuals, with mean differences of −6.70 years [95% confidence interval (CI), −10.16 to −3.23], −3.86 years (95% CI, −6.75 to −0.97), and −6.98 years (95% CI, −8.43 to −5.53), respectively. The median age in the IPMN group was comparable with the age of the PDAC group [mean difference, 0.21 years (95% CI, −2.67 to 3.11)]. A total of 284 patients with PDAC (40.5%) underwent surgery, including 151 (97.4%) of the 155 patients with stage II disease, and 150 (72.8%) of the 206 patients with stage III disease.

The 92 proteins in the Olink immuno-oncology panel and CA19-9 were measured for differential expression between patients with PDAC and non-PDAC individuals combined (all patients with nonmalignant pancreatic diseases and healthy individuals). The NPX values of 78 proteins were significantly different for patients with PDAC compared with non-PDAC individuals (P values from 0.047 to 5.26 × 10⁻⁵⁵; see Supplementary Table S2). Three proteins (CA19-9, IL8, and MMP7) also had a more than log₂ fold change of NPX value. A volcano plot illustrating the relationship between P values and log₂ fold changes is presented in Fig. 1.

The first statistical approach identified 19 candidate protein signatures for separation of patients with PDAC from non-PDAC individuals. The 19 signatures had a decreasing number of proteins and a resulting increased stability of the composition of included proteins. Of the 19 different protein signatures, 13 signatures contained less than 25 proteins (for details on the signatures, see Supplementary Table S3). Age did not elevate the AUC of any of the protein signatures when added to the models (DeLong test: all P = 1.0).

A signature containing a low number of proteins is desirable. The signature including nine proteins + CA19-9 (named index I) was deemed the best performing among the 19 candidates in terms of AUC, sensitivity, positive predictive value (PPV), the stability of the included proteins, and the number of proteins. Index I consisted of the following proteins: T-cell surface glycoprotein CD4 (CD4), cytotoxic and regulatory T-cell molecule (CRTAM), Fas ligand (FASLG), IL8, IL10, monocyte chemotactic protein 3 (MCP-3), MMP-7, TNF-related apoptosis-inducing ligand (TRAIL), VEGFC, and CA19-9. A higher expression of the proteins in patients with PDAC compared with non-PDAC individuals was found for most of the proteins. However, for FASLG, TRAIL, and VEGFC, a lower expression of the proteins was found in patients with PDAC compared with non-PDAC individuals. Index I in the discovery cohort, comparing all stages of PDAC versus non-PDAC individuals, gave an AUC of 0.93 (95% CI, 0.89–0.96), sensitivity of 0.86 (95% CI, 0.82–0.98), specificity of 0.85 (95% CI, 0.71–0.92), and PPV of 0.93 (95% CI, 0.89–0.96). In the replication cohort, all stages of PDAC versus non-PDAC individuals gave an AUC of 0.92 (95% CI, 0.87–0.97), sensitivity of 0.95 (95% CI, 0.77–1.0), specificity of 0.76 (95% CI, 0.68–0.94), and PPV of 0.90 (95% CI, 0.87–0.97). Details on performance of all the 19 signature candidates to index I are presented in Supplementary Table S4.

Each signature was evaluated in the discovery and replication cohorts for its ability to separate patients with PDAC stages I/II and stages III/IV from non-PDAC individuals. The models were trained and tested in the training and testing sets of the discovery cohort and the replication cohort where patients with stage III/IV were left out in the case of stage I/II analysis, and stage I/II patients were left out in the case of stage III/IV analysis. This gave AUCs of 0.92 (95% CI, 0.87–0.98; for PDAC stage I/II vs. non-PDAC individuals) and 0.96 (95% CI, 0.93–0.99; for PDAC stage III/IV vs. non-PDAC individuals) in the replication cohort.

When comparing patients with PDAC to the healthy individuals alone, the AUC was 0.99 (95% CI, 0.98–0.99) and 0.99 (95% CI, 0.99–1.0) in discovery and replication cohorts. Patients with PDAC compared with patients with nonmalignant pancreatic diseases combined (chronic pancreatitis, IPMN, and benign pancreatic diseases) gave AUC values of 0.85 (95% CI, 0.80–0.90) and 0.80 (95% CI, 0.71–0.90) in discovery and replication cohorts.

Figure 2A–C shows the ROC curves with corresponding AUC values for index I for all patients with PDAC compared with all non-PDAC individuals (A), and for the patients with PDAC divided by stage into two groups: stage I/II (B) and stage III/IV (C). Table 2 lists all AUC values from the different comparisons. Supplementary Fig. S1 shows the ROC curves for the comparisons of PDAC versus healthy and PDAC versus nonmalignant pancreatic diseases.

In the second statistical approach, the training of the model was performed with a cross-validation step repeated 500 times. During this process, 23 of the proteins + CA19-9 had regression coefficients (weights) consistently larger than zero: caspase-8 (CASP-8), C-C motif chemokine 3 (CCL3), CCL20, CCL23, CD4, CD40 ligand (CD40L), CRTAM, macrophage colony-stimulating factor 1 (CSF-1), C-X-C motif chemokine 1 (CXCL1), FASLG, IL1α, IL5, IL8, IL10, IL12, IL33, lysosome-associated membrane glycoprotein 3 (LAMP3), MCP-3, MHC class I polypeptide-related sequence A/B (MIC-A/B), MMP-7, programmed cell death 1 ligand 2 (PD-L2), TRAIL, VEGFC, and CA19-9, resulting in the second signature, designated index II.

When comparing patients with PDAC versus non-PDAC individuals, index II gave an AUC of 0.96 (95% CI, 0.95–0.98), sensitivity of 0.87 (95% CI, 0.84–0.91), specificity of 0.91 (95% CI, 0.87–0.95), and a PPV of 0.96 (95% CI, 0.94–0.98) in the discovery cohort. In the replication cohort, patients with PDAC versus non-PDAC individuals gave an AUC of 0.93 (95% CI, 0.90–0.96), sensitivity of 0.91 (95% CI, 0.87–0.85), specificity of 0.81 (95% CI, 0.72–0.89; Table 2), and a PPV of 0.92 (95% CI, 0.88–0.96; Fig. 2, bottom).

When split into two groups according to stage (I/II and III/IV), PDAC compared with non-PDAC gave AUCs of 0.90 (95% CI, 0.85–0.94; for PDAC stage I/II vs. non-PDAC individuals) and 0.95 (95% CI, 0.93–0.98; for PDAC stage III/IV vs. non-PDAC individuals) in the replication cohort. Comparing patients with PDAC with the healthy individuals alone gave AUC values of 0.996 (95% CI, 0.994–0.999) and 0.998 (95% CI, 0.996–1.0) in the discovery and replication cohorts, respectively. Patients with PDAC compared with patients with nonmalignant pancreatic diseases combined (chronic pancreatitis, IPMN, and benign pancreatic diseases) gave AUC values of 0.89 (95% CI, 0.86–0.92) and 0.84 (95% CI, 0.78–0.90) in discovery and replication cohorts, respectively.

Figure 2D–F gives the ROC curves with corresponding AUC values for index II for all patients with PDAC compared with all non-PDAC individuals (D), and for the patients with PDAC divided by stage into two groups: stage I/II (E) and stage III/IV (F). Table 2 lists all AUC values from the different comparisons.

All nine proteins + CA19-9 included in index I were also found in index II (see Supplementary Fig. S2). When ranking the proteins in index II according to contributed weight to the model (regression coefficients), six of the nine proteins included in index I (MMP-7, TRAIL, FASLG, VEGFC, CD4, and IL10), as well as CA19-9, were among the top 10 ranking in index II (see Supplementary Table S5). Looking at the top ranking 15 proteins with the largest weight contributing to the model in index II, eight of the nine proteins + CA19-9 in index I were found among them. All proteins in index I were among the top 16 performing proteins in index II (see Supplementary Table S5; Supplementary Fig. S3).

Figure 3 shows box plots illustrating differential expressions of each protein and CA19-9 in index I and index II stratified by diagnosis. Details are presented in Table 3.

When leaving CA19-9 out of the analyses, the indices resulted in AUCs of 0.90 (95% CI, 0.86–0.93) and 0.88 (95% CI, 0.82–0.95) for the discovery and replication cohorts in index I, and 0.94 (95% CI, 0.92–0.96) and 0.89 (95% CI, 0.85–0.93) for the discovery and replication cohorts in index II. See Supplementary Fig. S4 in the Supplementary Data for further details.

Only 1 patient with chronic pancreatitis and 1 patient with IPMN were diagnosed with PDAC during the follow-up period. The median follow-up period for the chronic pancreatitis and IPMN patients was 4.4 years. The 2 patients were diagnosed 35 months (patient with chronic pancreatitis) and 14 months (patient with IPMN) after inclusion in the BIOPAC study. The patient with chronic pancreatitis had NPX values above the median for patients with PDAC for the following proteins from the indices: CCL20, CXCL1, FASLG (below), TRAIL (below), and VEGFC (below). For FASLG, TRAIL and VEGFC, patients with PDAC had a median NPX value lower than that for non-PDAC, and the patient with chronic pancreatitis had NPX values below the median for patients with PDAC.

The patient with IPMN had NPX values above the median for patients with PDAC for the following proteins from the indices: CCL23, CCL3, CD4, CD40L, IL10, IL33, MIC-A/B, and PD-L2.

Here, we describe two independently developed novel protein panels for discriminating patients with PDAC from non-PDAC individuals, using the combination of CA19-9 with nine or 23 circulating proteins. All nine proteins in index I were included in index II. At replication, both panels produced AUCs at 0.90 or above. These are novel findings.

The aim of this study was to explore whether a panel of circulating proteins could identify PDAC. To test this, we examined pretreatment serum samples from patients with stage I–IV PDAC, patients with premalignant lesions (IPMN), patients with benign pancreatic diseases (including chronic pancreatitis), and healthy blood donors. We included blood samples from patients with IPMN and chronic pancreatitis, as these conditions are precursors and differential diagnoses of PDAC (2). As IPMN per definition is a premalignant lesion without invasion histologically, we would not expect to find cancer-related proteins in the blood of patients with IPMN.

A high-throughput low volume PEA technology from Olink was used to measure the expression of 92 proteins using the Olink immuno-oncology panel. This panel was chosen because inflammation is a key player in the development of PDAC (11, 13, 14).

Two independent bioinformaticians each identified a panel of proteins from this dataset using slightly different approaches: index I (nine proteins + CA19-9) and index II (23 proteins + CA19-9). AUC results were slightly better for index II than index I in the discovery cohort, but almost the same in the replication cohort. Index I, however, had a slightly higher sensitivity than index II in the replication cohort. The contribution of the additional 14 proteins included in index II was not substantial, and four of the proteins exclusive to index II (IL1α, LAMP3, IL5, and IL33) had either nonsignificant differences between PDAC and non-PDAC cases in univariate analyses (IL1α and LAMP3) or had many samples with NPX values below the lower limit of detection (IL1α, IL5, and IL33; see Supplementary Fig. S5). These four proteins were not among the top ranking proteins in index II (see Supplementary Table S5). When removing CA19-9 from the analyses, both indices still performed at a high level with AUC values between 0.88 and 0.95. However, it is evident that the contribution of CA19-9 to the models raises the performance and remains important in estimating risk.

The proteins included in the two indices can be grouped according to the biological function they are involved in: apoptosis (CASP-8, FASLG, and TRAIL), immunosuppression (CCL20, CD4, IL10, MMP-7, MIC-A/B, and PD-L2), inflammation (CRTAM, IL8, CXCL1, IL33, CD40L, IL1α, CSF-1, IL5, and IL12), hypoxia (LAMP-3), chemotaxis (MCP-3, CCL23, and CCL3), and vascular remodeling (VEGFC; Supplementary Table S6). Several of the proteins have been described previously in preclinical or clinical studies involving PDAC cell lines or patients: CCL20 (31), CD4 (30, 32), FASLG (33), IL8 (18, 34, 35), IL10 (18, 36), MCP-3 (37), MMP-7 (38, 39), and TRAIL (40).

IL8 and IL10 are elevated in patients with PDAC and correlated with each other (18) and with IL6, a well-known prognostic biomarker in PDAC (41). IL8 is a proinflammatory chemokine, and its expression is stimulated by various cytokines (including IL6), hypoxia, and reactive oxygen species (34). Activated intracellular pathways leading to IL8 expression include NF-κB, PI3K-AKT, and p38 MAPK (34). These are also known downstream signaling pathways of activated KRAS, the oncogene activated in more than 90% of PDAC tumors (2). IL10 is secreted by regulatory T cells in the tumor microenvironment, contributing to a local immunosuppressive environment in PDAC tumors, which promotes angiogenesis, matrix remodeling, and tumor metastasis (36).

Other research groups have explored differentially expressed circulating protein profiles in patients with PDAC compared with healthy individuals and/or with benign or premalignant pancreatic diseases. To our knowledge, 15 studies have been published since 2000 exploring circulating proteins for diagnosis of PDAC (Supplementary Table S7; for a full reference list of these studies and details on the PubMed search, see the Supplementary Materials and Methods). We chose to include only the studies including ≥100 patients with PDAC and with validation in the same article. The largest studies (42–44) included >400 patients with PDAC divided into discovery and validation cohorts. Two of the studies (43, 44) reported panels of three markers each: CA19-9, transthyretin, and leucine-rich alpha-2-glycoprotein 1; and CA19-9, apolipoprotein A-IV, and tissue inhibitor of metalloproteinase 1. The first study found an AUC of 0.93 in a comparison of patients with PDAC and patients with benign pancreatic diseases, with a sensitivity of 0.82, specificity not reported (43). The second study reported an AUC of 0.934 for the comparison of patients with PDAC and patients with pancreatitis, with a sensitivity of 0.86 and a specificity of 0.90 (44). Even though healthy individuals were included in the second study, no comparison was made between patients with PDAC and all non-PDAC controls (44).

The third study (42) compared a diagnostic protein signature of 29 proteins identified from circulating protein levels of 443 patients with PDAC in the discovery cohort with protein levels in 8 patients with IPMN and 888 healthy individuals. In the validation cohort, protein levels from patients with PDAC were compared with protein levels in 57 patients with chronic pancreatitis, 20 patients with IPMN, and 219 healthy individuals. The signature included VEGF and revealed an AUC of 0.96 for patients with PDAC (stage I/II) versus normal controls, sensitivity 0.95 and specificity 0.96. The AUC value for discriminating chronic pancreatitis from PDAC stage I/II was 0.84. No comparison was given for the ability of the signature to distinguish PDAC from all non-PDACs (42).

Most of the proteins included in these different studies (Supplementary Table S7) are not a part of the Olink immuno-oncology panel. The proteins found in our two indices are, therefore, different from the ones reported previously.

Two studies have sought to combine proteins with genetic alterations in cell-free DNA [circulating tumor DNA (ctDNA)] for the early detection of cancers, including pancreatic cancer (45, 46). The first study reported a five-marker panel consisting of CA19-9, carcinoembryonic antigen, hepatocyte growth factor, osteopontin, and KRAS mutations in ctDNA. Using this panel, they found 64% of the patients with PDAC in a cohort of patients with PDAC (n = 221) and healthy controls (n = 182). No AUC values were reported (45). The second study reported a panel comprising of eight proteins and ctDNA, termed CancerSEEK (46). This panel was tested in several cancer types, including 93 patients with pancreatic cancer. The panel correctly identified 72% of the patients with pancreatic cancer. AUC values were not reported for the specific diagnoses. Further studies are needed to evaluate whether the combined detection of mutant KRAS ctDNA and circulating protein levels can be used for early diagnosis of PDAC.

The main strengths of our study are the inclusion of more than 700 patients with all stages of PDAC and 280 individuals from appropriate control groups. A further strength is the robustness of the findings because two bioinformaticians independently worked on the same dataset and reached similar results.

The two protein signatures were derived from a cohort of mixed stages of PDAC, reflecting protein expressions of a wide range of tumor burdens. Both indices showed lower AUC values (0.90) when patients with stage I/II PDAC were compared with non-PDAC individuals than when patients with stage III/IV PDAC were compared with non-PDAC individuals. Biologically, this makes the indices credible, but clinically this also means that the two panels were less likely to distinguish earlier stages. However, our serum samples from patients with PDAC were not only compared with healthy blood donors, but also with high-risk patients with differential and predisposing diagnoses for PDAC (chronic pancreatitis and IPMN). Our goal was to discover new biomarkers to be used in a setting of a population above a certain a priori risk. Thus, another strength of our study is the inclusion of more than 100 patients with nonmalignant pancreatic diseases. Furthermore, when comparing patients with PDAC with only healthy blood donors, we found very high AUC values (0.99 and 1 for index I and II), supporting the strength of both our indices compared with protein panels reported previously.

Patients with PDAC were significantly older than the individuals included in three of the four groups of controls (chronic pancreatitis, benign pancreatic diseases, and healthy individuals). However, age was added to the models in the process of developing index I and was not found to contribute additional information. This suggests that the signatures might be useful independent of age under more realistic clinical conditions in a population of elderly people, with or without inflammatory diseases, either seen by a general practitioner or referred to a hospital for pancreatic cancer diagnostic workup.

Our study has several limitations. We had a limited number of patients with less advanced disease stage, the most relevant group to identify, because the majority of our patients (74.5%) had stage III or IV disease. Furthermore, our cohort is from a relatively homogenous Scandinavian population, and extrapolation to more heterogenous populations may be limited. Our results are also limited by the preselection of proteins by Olink, and proteins found in other studies could not be examined. In addition, our study was exploratory and will not alter clinical practice on its own.

A next step should consist of a formal prospective test of the clinical utility of the panels among patients in the initial phase of diagnosis alongside the normal diagnostic work-up for pancreatic cancer. Such a study would ideally confirm our findings and determine cutoffs, which could be used in a subsequent more heterogenous study of a clearly defined population at risk, in which clinical decisions are taken on the basis of the protein index. Only then can any conclusions about clinical validity be drawn. It would also be interesting to do tissue analyses of the surgical specimens available to further explore the tissue origin of the proteins included in the two indices.

PDAC remains a cancer disease with one of the highest mortalities, and game-changing breakthroughs in the treatment and prognosis of these patients remain to be seen. If, however, more patients are diagnosed at an early stage, a larger proportion of patients will undergo surgery, which will result in a survival benefit. This study identified two circulating protein indices with the potential to discriminate between individuals with and without PDAC.

S.C. Lindgaard reports grants from Celgene, VELUX Foundation, The Danish Cancer Society, Harboe Foundation, Beckett Foundation, Foundation of Merchant M. Kristjan Kjær and wife Margrethe Kjær, and Memorial Fund of Carpenter Holm during the conduct of the study. Z. Sztupinszki reports grants from Velux Foundation during the conduct of the study. No disclosures were reported by the other authors.

S.C. Lindgaard: Conceptualization, data curation, funding acquisition, investigation, visualization, writing–original draft, project administration, writing–review and editing. Z. Sztupinszki: Formal analysis, visualization, methodology, writing–review and editing. E. Maag: Data curation, formal analysis, visualization, methodology, writing–review and editing. I.M. Chen: Conceptualization, resources, supervision, writing–review and editing. A.Z. Johansen: Data curation, writing–review and editing. B.V. Jensen: Resources, writing–review and editing. S.E. Bojesen: Resources, writing–review and editing. D.L. Nielsen: Conceptualization, supervision, writing–review and editing. C.P. Hansen: Resources, writing–review and editing. J.P. Hasselby: Resources, writing–review and editing. K.R. Nielsen: Resources, writing–review and editing. Z. Szallasi: Formal analysis, supervision, writing–review and editing. J.S. Johansen: Conceptualization, resources, data curation, supervision, funding acquisition, writing–review and editing.

Many thanks to the biomedical laboratory scientists Charlotte Falk, Vibeke H. Holm, Sarah D. Jensen, and Helle B. Sørensen for their assistance with processing the blood samples. Molecular biologist Mie B. Krüger is thanked for the assistance in managing with the biobank. The nurses Lars M. Larsen, Lene T. Voxen, Finn Hejlesen, Signe Gamborg Nielsen, Jette L. Hansen, Mette Tholstrup Bach, Betina Nielsen, Nina Spiegelhauer, and Marianne Melton, and the physicians Dan Calatayud, Nicolai A. Schultz, Louise Skau Rasmussen, Mette Karen Yilmaz, Niels Henrik Holländer, Per Pfeiffer, Jon Kroll Bjerregaard, Fahimeh Andersen, Svend Erik Nielsen, Camilla Palmquist, and Karin Bagni are acknowledged for their contribution to inclusions of patients, data registration, or sample collection. Thank you to all the funding parties: Celgene (salary to S.C. Lindgaard, salary to E. Maag, and Olink analyses), VELUX Foundation (salary to Z. Szallasi), The Danish Cancer Society (salary to S.C. Lindgaard), Harboe Foundation (Olink analyses), Beckett Foundation (Olink analyses), Foundation of Merchant M. Kristjan Kjær and wife Margrethe Kjær (Olink analyses), and Memorial Fund of Carpenter Holm (Olink analyses).

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.