Development and Multiple Validation of the Protein Multi-marker Panel for Diagnosis of Pancreatic Cancer

Pancreatic ductal adenocarcinoma (PDAC) is one of the most lethal cancers, responsible for 432,242 patient deaths worldwide annually (1). PDAC is the fourth leading cause of cancer-related mortality, with a 5-year survival rate of approximately 10% (2). Because of its poor prognosis, mortality (n = 432,232) coincides with incidence (n = 458,918; ref. 1). According to recent studies, the incidence of new PDAC has been gradually increasing over time, and PDAC is projected to become the second leading cause of cancer-related mortality by 2030 (3). Although surgical resection remains the most effective treatment, only 15% to 20% of diagnosed patients are eligible to undergo this process because the vast majority of patients are diagnosed at later stages (locally advanced or metastatic PDAC; ref. 4). Consequently, improvements in early detection are integral to allowing patients with PDAC to access existing curative therapies.

Current efforts to develop a diagnostic test for PDAC have primarily focused on blood or urine biomarkers due to their minimally invasive natures (5). Carbohydrate antigen 19-9 (CA 19-9), which is the sole blood-based biomarker approved by the FDA for the clinical management of PDAC, has been thoroughly studied and utilized in clinical applications. However, CA 19-9 demonstrates inadequate sensitivity and specificity, which has limited its diagnostic performance for general screening (6). Previous studies have reported that the sensitivity and specificity of CA 19-9 levels for PDAC diagnosis ranged from 70% to 80% and 80% to 90%, respectively (7). Moreover, CA 19-9 is irrelevant to the approximately 10% of patients with PDAC who do not express Lewis blood group antigens (7). Therefore, the development of a biomarker with increased sensitivity and specificity that is applicable to all patients with PDAC remains necessary.

Previous studies have suggested the addition of alternative biomarkers to compensate for the limitations of CA 19-9 for PDAC diagnosis (8). Considering the heterogeneity of cancer development and progression and the complexity of biofluids, such as serum or plasma, single biomarkers are likely to be insufficient to achieve the sensitivity and specificity required for cancer diagnosis (9). Therefore, recent biomarker studies have shifted to a multiparametric analysis. The combination of multiple biomarkers has been reported to be a useful strategy in several studies, providing increased diagnostic accuracy, sensitivity, and specificity (10). Using multiple markers, both the sensitivity and specificity reached relatively high values for the discrimination of patients with PDAC from healthy individuals. Recent studies have reported the notable diagnostic performance of various combinations of proteins (11), metabolic biomarkers (8), and miRNAs (12) for PDAC diagnosis.

Protein biomarkers have been utilized in numerous clinical applications, including diagnosis, disease prognosis, therapeutic evaluations, and drug efficacy assessments (13). The evaluation of human plasma proteins can reflect various aspects of patient physiology, including tumor growth and disease conditions. Consequently, blood-based protein biomarkers represent a promising source of disease indicators that can be evaluated through less-invasive diagnostic tests (14). Proteomic analysis by mass spectrometry (MS) has been widely applied to biomarker development in recent years. In particular, MS-based, targeted proteomic approaches [e.g., multiple reaction monitoring-mass spectrometry (MRM-MS)] have been applied for their robust and high-throughput capabilities, which enable protein measurements in the range of hundreds of proteins while retaining high sensitivity, selectivity, and reproducibility (15). Furthermore, MRM-MS is largely unaffected by matrix effects and other contaminants because it specifically detects peptides of interest and their corresponding fragments. These characteristics of the MRM-MS assay indicate its potential to serve as a complementary method to antibody-based assays for biomarker development (16).

Using a large-scale dataset containing protein biomarker candidates that were discovered and verified in our previous study (17), we aimed to develop a protein-based, multi-marker panel for PDAC diagnosis and to compare the performance of this panel against that for CA 19-9, both alone and in conjunction with other biomarkers. In addition, the protein-based, multi-marker panel was validated using multiple independent datasets.

This study used a targeted proteomic (MRM-MS) dataset that was obtained from a previous study (17) and contained quantitative data from 1,008 plasma samples. Briefly, clinical plasma samples were obtained from recruited patients with PDAC, benign pancreatic disease, other cancers, other benign diseases, and healthy controls. Patients were recruited from five medical centers in Seoul, Republic of Korea (South) [National Cancer Center (NCC, n = 128), Seoul National University Hospital or Seoul National University Hospital Healthcare System Gangnam Center (SNUH, n = 553), Samsung Medical Center (SMC, n = 131), Asan Medical Center (AMC, n = 122), and Yonsei Severance Hospital (YSH, n = 74)], between January 2011 and December 2013. The total cohort comprised 401 patients with PDAC, 109 patients with benign pancreatic disease, 300 healthy controls, 149 patients with other cancers, and 49 patients with other benign diseases. The distribution of tumor stages in patients with PDAC was 248 early-stage cases [stage IA (n = 8), IB (n = 12), IIA (n = 86), and IIB (n = 142)] and 153 late-stage cases [stage III (n = 31) and IV (n = 122)]. Tumor stages for PDAC were classified according to the 7th edition of the American Joint Committee on Cancer (AJCC; ref. 18). Detailed information regarding patient characteristics is presented in Table 1 and Supplementary Table S1.

To discriminate PDAC, the dataset was reconfigured to define PDAC as the case group and healthy controls and benign pancreatic diseases as the control groups. The total clinical samples were divided into four sets for developing and validating the multi-marker panel. Samples from four medical centers (NCC, SNUH, SMC, and YSH) were integrated and randomly divided to form a training set and a validation set, with the training set consisting of 261 PDAC cases (148 early-stage and 113 late-stage) and 290 controls and the validation set consisting of 65 PDAC cases (27 early-stage and 38 late-stage) and 72 controls. An independent validation set, consisting of 75 PDAC cases (73 early-stage and two late-stage) and 47 controls, was obtained from the AMC. To perform subgroup analyses, an additional dataset was constructed containing other cancers cases (breast cancer, n = 52; colon cancer, n = 45; and thyroid cancer, n = 52), in addition to PDAC cases. The additional set consisted of 50 PDAC cases and 149 other cancer cases, which were all obtained from SNUH. To minimize any variations that might have resulted from using samples that originated from different institutions, several PDAC cases from the training and validation sets were reused. Because PDAC was set as the case group while healthy controls or benign pancreatic diseases were set as the control group, the other benign disease group was not used in this study.

This study was conducted in accordance with the guiding principles of the Declaration of Helsinki. The study protocols were approved by the corresponding institutional review boards of all participating institutions (SNUH surgery H-0901-010-267, SNUH internal medicine H-0412-138-005 and H-0412-138-006, SNUH HSGC H-1305-573-489 and C-1301-095-458, YSH 4-2013-0725, NCC NCCNCS13818, SMC 2008-07-065, AMC 2013-1061), and written informed consent was obtained from all participants who contributed biospecimens.

In previous studies, we identified 54 proteins (68 peptides) as biomarker candidates using a large-scale, systematic process. Through the biomarker discovery phase in the previous study, several sources, including databases, traditional cancer markers, and microarray analysis data from resected PDAC tissue, were used to obtain 1,000 protein biomarker candidates for the diagnosis of PDAC using proteomic analysis (database and literature: 508 proteins; traditional cancer markers: 22 proteins; microarray analysis: 456 proteins; known mutated protein: 14 proteins). Of 1,000 initial candidates, 205 proteins with AUC values > 0.6, when evaluated as single markers, were identified as targets that can be detected by MS, based on 134 plasma samples (50 PDAC cases and 84 controls) obtained from SNUH. Among 205 proteins, 176 proteins (corresponding to 217 peptides) were selected as unique. MRM-MS quantitative analysis using stable isotope-labeled standard peptide resulted in the identification of 54 differentially expressed proteins between PDAC and healthy controls, which were chosen as the final MRM-MS quantification targets. During the verification phase, a total of 54 proteins (68 peptides) were measured by MRM−MS in 1,008 plasma samples. The biomarker discovery, verification, and sample preparation methods and the MRM-MS procedure were detailed in a previous study (17).

The data preprocessing consisted of two steps. The first step was the selection of high-intensity proteins to reduce analytic variations. Using raw data from the MRM-MS analysis, which was imported into Skyline (McCoss Lab, University of Washington, Seattle, WA), those with peak intensities below the following criteria were removed (peak intensity < 1,000). A total of 11 biomarker candidates were removed, leaving 57 biomarkers. The second step was to transform the data type, to improve the normality of the distribution. Many statistical procedures assume that the variables are normally distributed. If this assumption is unmet, the development of an erroneous multi-marker panel will occur. To ensure the normal distribution of each marker, three different types of transformation were performed: (i) raw; (ii) log (x + 10^–10); and (iii) square root. Then, the skewness value for each marker was calculated using the skewness function in the e1071 package of R (19). When the skewness value was calculated following data transformation for 57 biomarker candidates, including CA 19-9, the skewness values ranged from 0.18 to 29.89.

The type of data transformation was selected, in which the average of the absolute values of the skewness of the case group and the control group is least. The data type for each marker was transformed using the selected transformation method: 11 log (x + 10^–10), 42 square roots, and 5 raw. The skewness value of these optimized transformations ranged from 0.1 to 1.96, and the normality of all 58 biomarker candidates was improved (20). After transformation, all biomarker candidates were standardized to a normal distribution, with a mean of 0 and a SD of 1. From proteins associated with multiple peptides, only a single peptide was selected on the basis of AUC and P values, to ensure that each peptide corresponded to a single protein. Throughout this process, 13 peptides were removed and 44 biomarker candidates remained. Data that were pertinent to the 58 biomarker candidates are included in Supplementary Table S8.

Differences in the protein expression levels between two groups were compared using the Mann–Whitney U test. A logistic regression analysis with stepwise selection was performed to build a multi-marker panel using the 44 biomarker candidates. During the stepwise selection procedure, a P value of < 0.05 was required to allow a biomarker candidate into the model and remain in the model. The multi-marker panel was trained to classify patients with PDAC against healthy controls or patients with benign pancreatic disease. To provide a more accurate indication of how well the multi-marker panel generalizes to unseen data, the multi-marker panel was validated using the corresponding validation set and an independent validation set. The diagnostic performance of the multi-marker panel for the discrimination of patients with PDAC from healthy controls or patients with benign pancreatic diseases was assessed using ROC curve analysis, particularly AUC and 95% confidence intervals (CI). The AUC CIs were calculated using the DeLong method and integrated within the pROC package (21). Sensitivity and specificity were calculated at the optimal cut-off value, which was determined by the Youden index (22) for the training set. The threshold for statistical significance was set at 0.05. All reported P values are two sided. Statistical analyses were performed using R (version 3.6.0).

We previously identified and verified 68 peptides (corresponding to 54 proteins) as PDAC biomarker candidates from 1,000 PDAC-related initial protein candidates (17). The 1,000 candidate proteins were selected as potentially PDAC-related proteins by compiling databases, traditional cancer markers, and microarray analysis data from resected PDAC tissue. Of these 1,000 candidate proteins, 205 were detectable by MS in plasma samples. Among these individual 205 protein markers, 68 peptides associated with 54 proteins had AUCs greater than 0.60 for the discrimination of PDAC cases from healthy controls and were subsequently selected for the quantification analysis using a large-scale cohort. These selected targets were then quantified by MRM-MS in a large-scale, plasma sample cohort, comprising patients with PDAC (n = 401), patients with benign pancreatic disease (n = 109), healthy controls (n = 300), other patients with cancer (n = 149), and other patients with benign disease (n = 49), which were obtained from five major medical centers (NCC, SNUH, SMC, YSH, and AMC). Using this large-scale, quantification dataset, we developed a proteome-based predictive model.

To improve the protein-based predictive model, data preprocessing procedures were applied, considering two contributing factors: (i) the robustness of quantified markers and (ii) the distribution of the dataset. Regarding the robustness of the markers, 11 peptides with low intensities (lower than 1,000) in the chromatogram were excluded from the list of 68 biomarker candidates because low intensities may affect analytic reproducibility and variable predictive performance.

In addition, large-scale quantitative data from biological sources tend to be skewed in distribution. To reduce these differences and strengthen the classification of PDAC cases, the distribution of each biomarker candidate was explored to calculate its skewness. The skewed distribution of peptides was addressed by transforming the distribution of each peptide using common data transformations: raw, square root, and logarithmic transformations [log (x +10^–10)]. Skewness values were calculated for each case and control group because differences in the expression levels of these markers were observed for each case (PDAC) and control (healthy control and benign pancreatic disease) group. The optimal transformation was performed for each peptide to minimize the average absolute values of skewness for both groups (case and control). A total of 53 skewed candidates were transformed with log or square root transformations to mimic normal distribution. No transformation was required for five peptides with minimally skewed raw data. The square root transformation was performed for 42 peptides, and 10 peptides and CA 19-9 were transformed by log (x +10^–10) transformation. After performing data transformations, the skewness values for all protein candidates were lower than 2, and the distribution largely resembled a normal distribution, resulting in more symmetrical data, suitable for subsequent statistical analyses (Supplementary Fig. S2). After data transformation, all biomarker candidates were standardized to a mean of 0 and a SD of 1.

Prior to the development of the multi-marker panel, proteins associated with multiple peptides were matched by selecting the best peptide so that single peptides corresponded to single proteins. To match a single peptide to a single protein, the P values and AUC values of all peptides associated with a single protein were compared. A single peptide was selected for 13 proteins that were each associated with two peptides. Consequently, 44 peptides (corresponding to 44 proteins) were selected as biomarker candidates to establish a predictive model. The overall study design is presented in Fig. 1 and Supplementary Fig. S1.

We constructed a multi-marker panel from selected biomarker candidates to improve the diagnostic performance for PDAC. To distinguish PDAC from healthy controls and noncancerous benign pancreatic diseases, the case group was set to patients with PDAC, and the control group was set to healthy controls and patients with benign pancreatic disease. In a large-scale dataset, samples from four medical centers were integrated to form training and validation datasets, consisting of 261 PDAC cases (148 early-stage and 113 late-stage) and 290 controls in the training set and 65 PDAC cases (27 early-stage and 38 late-stage) and 72 controls in the validation set (Table 1; Supplementary Table S1). A multi-marker panel was developed using the 44 selected biomarker candidates described in the preprocessing procedures. To construct a multi-marker panel with optimal predictive performance, we applied logistic regression analysis to the training set, with stepwise selection, using P values. Using stepwise selection, a multi-marker panel, consisting of 14 protein markers [clusterin (CLU), complement C5 (C5), plasma kallikrein (KLKB1), platelet basic protein (PPBP), IFN-related developmental regulator 1 (IFRD1), IGF-binding protein 2 (IGFBP2), intercellular adhesion molecule 1 (ICAM1), C4b-binding protein alpha chain (C4BPA), receptor-type tyrosine-protein phosphatase eta (PTPRJ), extracellular matrix protein 1 (ECM1), vimentin (VIM), C4b-binding protein beta chain(C4BPB), plasma serine protease inhibitor (SERPINA5), transthyretin (TTR)], was selected, and the predicted probability of PDAC cases according to the multi-marker panel was calculated, as follows: Logit (p) = 0.1789 + (–1.6768 × CLU) + (1.5854 × C5) + (–1.9541 × KLKB1) + (2.0778 × PPBP) + (0.7881 × IFRD1) + (1.2657 × IGFBP2) + (1.3406 × ICAM1) + (1.5931 × C4BPA) + (–0.5218 × PTPRJ) + (–1.14 × ECM1) + (0.8387 × VIM) + (–1.2183 × C4BPB) + (0.6271 × SERPINA5) + (–1.0129 × TTR). All markers in the multi-marker panel selected by the stepwise selection procedure were significantly associated with the prediction of PDAC (P < 0.05; Supplementary Table S2).

We evaluated the diagnostic performance when distinguishing PDAC cases from controls using the multi-marker panel and compared the results of the multi-marker panel against the diagnostic performance of CA 19-9 alone. The optimal cut-off value that the multi-marker panel predicted for patients with PDAC was 0.576, as determined by the Youden index. In the training set, the developed multi-marker panel yielded an AUC of 0.977 (95% CI = 0.967–0.987), which was significantly greater than the AUC for CA 19-9 alone (AUC = 0.872, 95% CI = 0.838–0.907; P < 0.001, DeLong test). The panel yielded a sensitivity of 90.4% and a specificity of 95.9%, which was higher than the sensitivity of CA 19-9 alone (69.7%) but with slightly lower specificity (98.6%; Table 2; Fig. 2A).

In the validation set, the panel also returned a significantly greater AUC value than CA 19-9 alone (0.953, 95% CI = 0.914–0.991 vs. 0.832, 95% CI = 0.754–0.910; P < 0.01, DeLong test). The sensitivity was 84.6%, and the specificity was 94.4%, demonstrating a great improvement in sensitivity for the multi-marker panel compared with CA 19-9 alone (66.2%) while retaining slightly lower levels of specificity (98.6%; Table 2; Fig. 2B). These results indicated that the multi-marker panel achieved improved diagnostic performance in the new set.

The diagnostic performance of CA 19-9 showed lower sensitivity and higher specificity than the multi-marker panel for both the training set and the validation set. Considering these characteristics, we examined the ability of the multi-marker panel to complement CA 19-9 by investigating whether the combination of the multi-marker panel and CA 19-9 could improve the diagnostic performance. The classification performance for the combination of the multi-marker panel and CA 19-9 achieved the best AUC value for both the training and validation sets (AUC = 0.989, 95% CI = 0.981–0.996 in training set, AUC = 0.972, 95% CI = 0.944–0.999 in validation set). The sensitivity values were 92.7% and 86.2%, respectively, and the specificity values were 98.3% and 97.2%, respectively. These results indicated that the inclusion of CA 19-9 in the multi-marker panel improved the AUC and sensitivity compared with using CA 19-9 alone for distinguishing between patients with PDAC and healthy controls or patients with benign pancreatic disease (Table 2; Fig. 2C and D).

The further validation of the multi-marker panel was performed using an independent validation set, consisting of 75 PDAC cases (73 early-stage and two late-stage) and 47 controls obtained from a single medical center (AMC; Table 1; Supplementary Table S1). In the independent validation set, the multi-marker panel had an AUC of 0.928 (95% CI = 0.885–0.971) for the discrimination of PDAC cases from controls, with a sensitivity of 81.3% and a specificity of 89.4%. The AUC for the multi-marker panel was significantly greater than that for CA 19-9 alone (0.771, 95% CI = 0.687–0.856; P < 0.001, DeLong test). The multi-marker panel outperformed CA 19-9 for sensitivity, although specificity was lower (Table 3; Fig. 3A). Moreover, the multi-marker panel that included CA 19-9 yielded an AUC of 0.952 (95% CI = 0.918–0.985), which was significantly higher than that of CA 19-9 alone (P < 0.001, DeLong test). The multi-marker panel combined with CA 19-9 also had higher sensitivity than CA 19-9 alone (89.3% vs. 61.3%) but had lower specificity (87.2% vs. 95.7%; Table 3; Fig. 3B).

The major limitation of CA 19-9 is that it may be elevated in patients with other malignancies, such as ovarian, liver, and colorectal cancers (23). This nonspecific expression of CA 19-9 makes CA 19-9 inadequate for the reliable diagnosis of PDAC. Therefore, we evaluated the ability of the multi-marker panel to distinguish PDAC cases from other cancer cases. We procured an additional dataset consisting of PDAC cases and other cancer cases (50 PDAC cases and 149 other cancer cases) from SNUH. The other cancer cases included 52 breast cancer cases, 45 colon cancer cases, and 52 thyroid cancer cases (Supplementary Table S3).

The results from this additional dataset indicated that the multi-marker panel had a higher AUC of 0.946 (95% CI = 0.916–0.976) compared with an AUC of 0.886 (95% CI = 0.820–0.952) for CA 19-9 alone, with sensitivity improving from 76% to 92%. However, specificity was lower for the multi-marker panel than for CA 19-9 alone (79.2% vs. 92.6%; Supplementary Table S4; Supplementary Fig. S3A). The combined panel, including CA 19-9, showed greater improvements than the multi-marker panel without CA 19-9, as demonstrated by an AUC of 0.968 (95% CI = 0.945–0.991). The sensitivity was 98%, and the specificity improved to 82.6% (Supplementary Table S4; Supplementary Fig. S3B). Details on the subgroup analysis between PDAC and cancer type can be found in Supplementary Materials and Methods (Supplementary Fig. S5; Supplementary Table S5).

The development of an effective biomarker that can be used to screen and provide an earlier diagnosis of PDAC will allow more patients with PDAC to be identified early enough to undergo resection and ultimately improve survival rates (24). The ideal diagnostic biomarkers for effective screening are those that can be obtained in a minimally invasive manner and can distinguish PDAC from healthy individuals or other benign pancreatic diseases with satisfactory sensitivity and specificity.

In this study, a 14-protein multi-marker panel was developed to diagnose PDAC and tested in two validation sets to investigate the clinical applicability of the MRM-MS assay. The multi-marker panel demonstrated significant improvements over the CA 19-9, as evidenced by an AUC value of 0.977, with a sensitivity 90.4% and a specificity 95.9% for the training set. The multi-marker panel showed a high diagnostic performance AUC value of 0.953, with a sensitivity of 84.6% and a specificity of 94.4% for the validation set. When evaluated using the independent validation set, the multi-marker panel displayed a high AUC value of 0.928, with a sensitivity of 81.3% and a specificity of 89.4%. Although CA 19–9 alone consistently had high specificity, the multi-marker panel outperformed CA 19-9 for both AUC and sensitivity. The discrimination performance of the multi-marker panel for the training set was consistent with the results of the validation set and the independent validation set. The multi-marker panel yielded significantly improved discrimination performance relative to CA 19-9 alone and showed high diagnostic discrimination power when used for three different sets. In particular, the multi-marker panel showed high diagnostic performance for the independent validation set that consisted of early-stage PDAC, indicating the possibility of achieving early PDAC diagnosis. Moreover, adding CA 19-9 to the multi-marker panel increased the performance, suggesting that the multi-marker panel can complement CA 19-9 and result in the increased power to distinguish PDAC cases from controls. The combined panel showed a better AUC value than the multi-marker panel alone, for all three datasets, and showed improvements in both sensitivity and specificity. These results indicated that the AUC and sensitivity of the combination panel are higher than that of CA 19-9 alone, representing a pronounced improvement over the results for a single marker. In the subgroup analysis, the multi-marker panel showed a higher AUC and sensitivity than CA 19-9 for distinguishing between PDAC and other cancers, suggesting that the multi-marker panel complemented the properties of CA 19-9 when CA 19-9 was expressed nonspecifically (23) and was more effective for distinguishing PDAC from other carcinomas than CA 19-9 alone.

In our previous study, an MS-based targeted proteomic approach was used to select potential PDAC biomarkers for development of an ELISA. In this study, using the same method, the multi-marker panel that we established was designed with the express purpose of identifying PDAC cases at an early stage for clinical use. Our study features the following advantages: (i) the relatively large number of samples (401 patients with PDAC and 558 controls, including healthy controls, patients with benign pancreatic disease, and patients with other cancers), which were obtained from five medical centers, increase the credibility of our results; (ii) the multi-marker panel was tested on two independent validation sets, including one whose samples were obtained from a single medical center; and (iii) the use of a targeted proteomics approach (MRM-MS) to assess biomarker candidates. Although new cancer-related biomarker candidates have been discovered recently, through many biomarker studies, only a very small number of these biomarkers have been used in clinical practice (25). To develop a meaningful candidate PDAC biomarker with the potential for clinical application, we have tested these candidates using a large number of samples from multiple medical centers, and these biomarkers exhibited reproducible predictive performances in various datasets. To verify the diagnostic abilities of the multi-marker panel with confidence, the use of an independent validation process with a different population is recommended (26). The training set that was used to build the multi-marker panel may not accurately reflect all of the characteristics of the underlying populations of interest (26). By subjecting the multi-marker panel to an independent validation set, especially one that features a distinct population from a single medical center, the general applicability of the panel and its robustness can be assessed. If the multi-marker panel performs well in several independent sets, it is more likely to be viable for clinical applications (27).

The targeted proteomics approach (MRM-MS), which produces robust measurements with a high degree of quantitative and analytical reproducibility while increasing throughput compared with conventional antibody-based methods (28), represents an alternative strategy that can overcome the limited multiplex capabilities inherent to methods such as ELISA (29). On the basis of these characteristics, the MRM-MS assay is a viable alternative to conventional assays, and several assays are already being tested in clinical trials (30). One particular MRM-MS assay that measures thyroglobulin, a biomarker for monitoring thyroid carcinoma, is being used in six clinical laboratories throughout North America (31). Another study has attempted to predict overall survival by measuring HER2 expression in patients with breast cancer by MRM-MS assay (32), and a separate clinical trial was conducted to differentiate benign and malignant nodules in the lung by measuring two proteins (33). These impressive cases demonstrate that the MRM-MS assay is applicable to clinical practice.

Challenges in the clinical application of the MRM-MS assay include the establishment of standard operating procedures for sample preparation and the confirmation of the analytic validity of the MRM-MS assay. The latter can be obtained by applying the validation guidelines for bioanalytical methods (e.g., calibration curve, precision, accuracy, reproducibility) of the FDA (34) and European Medicines Agency (EMA; ref. 35). Our group validated the analytic method of the MRM-MS assay according to guidelines in a previous study (36).

Using such large-scale data, with various biological and patient-to-patient variations and different distributions in the raw states can challenge the development of an appropriate predictive model (37). In particular, skewed distributions are indicative of bias, which may be reflected in the developed model. The transformation and normalization of data can reduce the skewness of data distribution, making biomarker candidates more comparable or normally distributed (20). By transforming the scale of each biomarker candidate, we have not only reduced natural biological variations but also impacted the potential performance of the multi-marker panel. This result suggests that the preprocessing of large-scale quantitative data can improve the further development of multi-marker panels.

The 14 biomarkers in our multi-marker panel, CLU, C5, KLKB1, PPBP, IFRD1, IGFBP2, ICAM1, C4BPA, PTPRJ, ECM1, VIM, C4BPB, SERPINA5, and TTR, represent proteins that are significantly differentially expressed between patients with PDAC and healthy control groups or patients with benign pancreatic disease, based on the entire dataset (Supplementary Fig. S4). The capacity of our l4 markers for comparing PDAC and non-PDAC is described in Supplementary Materials and Methods (Supplementary Table S6). These biomarkers were identified as being associated with PDAC in previous studies. TTR, a major circulating thyroxine-binding protein, was reported to be elevated in patients with PDAC compared with healthy controls (38). IGFBP2, which belongs to the family of IGFBP, is also known to be associated with PDAC. IGFBP2 levels have been shown to increase with increasing PDAC stages, suggesting that this protein may serve as a biomarker for disease progression (39). PPBP (also called CXCL7) belongs to the angiogenic CXC chemokine family and functions to promote angiogenesis. In a proteomic study, the level of PPBP was found to significantly decreased in patients with PDAC (40). Complement C5 is produced by cancer cells to promote immunosuppression and allow tumor growth in lung cancer. C5 is also significantly elevated in patients with PDAC, particularly in the presence of jaundice, compared with healthy controls (41). CLU has been associated with tumorigenesis and apoptosis and was overexpressed in PDAC tissue (42). ICMA 1 is known to modulate adhesion (cell-to-cell and cell-to-matrix) and has been well studied in inflammation. The enhanced expression of serum ICAM 1 in PDACs can discriminate patients with PDAC from healthy controls with high sensitivity (43). C4b-binding protein, which is involved in the regulation of the classical complement pathway, has two major isoforms, C4BPA and C4BPB. The serum C4BPA levels in patients with PDAC were significantly increased compared with healthy control and patients with pancreatitis (44). C4BPB has been reported to belong to the PDAC subtype associated with intermediate survival among patients with PDAC (45). PTPRJ plays a regulatory role in angiogenesis, cell proliferation, and migration and may serve as a promising candidate for PDAC gene therapy (46). ECM1 is involved in cell proliferation, angiogenesis, migration, and metastasis and demonstrates significantly increased expression in PDAC tissues, associated with poor prognosis in patients with PDAC (47). VIM is a cytoplasmic intermediate filament protein associated with tumor development and progression. The proteomic analysis of VIM revealed the increased expression of VIM under conditions of high PDAC cell invasion (48). KLKB1 is a glycoprotein associated with the surface-dependent activation of blood coagulation, fibrinolysis, kinin generation, and inflammation. KLKB1 was differentially expressed in PanNET (pancreatic neuroendocrine tumors), accounting for 1%–2% of all pancreatic tumors (49). SERPINA5 is loosely associated with PDAC, but some studies have indicated that this protein is overexpressed in PDAC or subtypes of PDAC (45). To our knowledge, IFRD1 has not been explored in PDAC studies. However, IFRD1 expression in colon cancer is associated with poorer patient prognosis (50).

The limitation of our study was that the control group used for the independent validation set consisted of patients with benign pancreatic disease; therefore, this validation set was unable to evaluate the discriminating performance of the panel to distinguish patients with PDAC from the healthy control group. Validating the panel using an independent dataset that includes a healthy control group will further increase the credibility of the diagnostic power of the panel. For clinical application of the multi-marker panel, its performance must be validated in a separate cohort, because the cohort that was used in this study was homogeneous with regard to ethnicity. A future study on samples that originate from several ethnic and genetic backgrounds will confirm the multi-marker panel as a generalizable model. Also, the multi-marker panel must be validated externally to determine its reproducibility across laboratories. If the MRM-MS assay meets analytical validation standards across facilities, the marker panel can be deemed to be reproducible. We expect other researchers who wish to apply our work using their own data to follow the detailed methods in Supplementary Materials and Methods (Supplementary Table S7).

In conclusion, using a targeted proteomics approach (MRM-MS), this study developed and validated a diagnostic, multi-marker panel for PDAC that significantly outperformed CA 19-9 alone for distinguishing of patients with PDAC from healthy individuals or patients with benign pancreatic disease in several datasets. These results suggested that the developed multi-marker panel has the potential for clinical application, serving a complementary role to CA 19-9 screening tests.

No disclosures were reported by the authors.

Y. Kim: Data curation, investigation, visualization, writing–original draft. I. Yeo: Data curation, investigation, methodology, writing–review and editing. I. Huh: Conceptualization, investigation, methodology, writing–review and editing. J. Kim: Visualization. D. Han: Data curation. J.-Y. Jang: Conceptualization, resources, funding acquisition, writing–review and editing. Y. Kim: Conceptualization, resources, supervision, funding acquisition, writing–review and editing.

This work was supported by the Industrial Strategic Technology Development Program (No. 10079271 and No. 20000134), funded by the Ministry of Trade, Industry and Energy [MOTIE, Republic of Korea (South)] and the Korea Health Industry Development Institute (No. HI14C2640, No. HL19C0020), funded by the Ministry of Health and Welfare, Republic of Korea (South).

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.