Purpose: Mass spectrometry–based serum peptidome profiling is a promising tool to identify novel disease-associated biomarkers, but is limited by preanalytic factors and the intricacies of complex data processing. Therefore, we investigated whether standardized sample protocols and new bioinformatic tools combined with external data validation improve the validity of peptidome profiling for the discovery of pancreatic cancer–associated serum markers.

Experimental Design: For the discovery study, two sets of sera from patients with pancreatic cancer (n = 40) and healthy controls (n = 40) were obtained from two different clinical centers. For external data validation, we collected an independent set of samples from patients (n = 20) and healthy controls (n = 20). Magnetic beads with different surface functionalities were used for peptidome fractionation followed by matrix-assisted laser desorption/ionization time-of-flight (MALDI-TOF) mass spectrometry (MS). Data evaluation was carried out by comparing two different bioinformatic strategies. Following proteome database search, the matching candidate peptide was verified by MALDI-TOF MS after specific antibody-based immunoaffinity chromatography and independently confirmed by an ELISA assay.

Results: Two significant peaks (m/z 3884; 5959) achieved a sensitivity of 86.3% and a specificity of 97.6% for the discrimination of patients and healthy controls in the external validation set. Adding peak m/z 3884 to conventional clinical tumor markers (CA 19-9 and CEA) improved sensitivity and specificity, as shown by receiver operator characteristics curve analysis (AUROCcombined = 1.00). Mass spectrometry–based m/z 3884 peak identification and following immunologic quantitation revealed platelet factor 4 as the corresponding peptide.

Conclusions: MALDI-TOF MS-based serum peptidome profiling allowed the discovery and validation of platelet factor 4 as a new discriminating marker in pancreatic cancer.

Translational Relevance

Based on matrix-assisted laser desorption/ionization time-of-flight peptidome profiling, the present study identifies platelet factor 4 as a discriminating serum marker in patients with pancreatic cancer. In conjunction with the conventional tumor markers CA 19-9 and CEA, platelet factor 4 strongly improves the diagnostic power of tumor marker testing. This might be of special relevance in the differential diagnosis of pancreatic cancer and pancreatitis.

Pancreatic cancer is the fourth leading cause of cancer death in the United States. Most patients diagnosed with pancreatic cancer develop clinical symptoms late in the course of the disease (1). Therefore, only 20% will be amenable to potentially curative therapy and only 3% to 5% of patients survive 5 years or more (2). Earlier diagnosis of the disease and early relapse monitoring are probably the best available options to improve patient survival (3). Currently, no single clinical chemical marker meets the sensitivity and specificity criteria required for screening or stratification purposes (4). Established serum markers such as carbohydrate antigen CA 19-9 or carcinoembryonic antigen (CEA) are useful to monitor the course of disease on and off treatment, but they lack the prerequisites for screening and to estimate the prognosis of a patient (2, 5).

Peptidome-based studies using high-throughput spectrometric methods such as matrix-assisted laser desorption/ionization time-of-flight (MALDI-TOF) mass spectrometry (MS) promise to be valuable for the identification of new “disease signatures” and cancer-associated biomarkers, especially combined with hitherto known biomarker patterns in a multivariate approach (613). However, there is a controversy regarding the diagnostic potential and reliability of the clinical proteomics approach (1418). It was perceived that standardization of preanalytic and analytic factors, as well as improvements in bioinformatic tools are important preconditions for translating serum peptidomics from the bench to the bedside (16, 1921). For instance, it was shown that many preanalytic factors have a major effect on the results of biomarker discovery and limit the use of pre-existing sample banks (2226). Recently, we developed standardized sample protocols and new bioinformatic tools for spectral data preprocessing and peak selection to enhance the sensitivity of data analysis (proteomics.net; refs. 22, 27).6

The present clinical study investigated the effect of our standardized sample protocol and new bioinformatic tools combined with external data validation on the efficiency of MALDI-TOF–based peptidome profiling for the discovery and clinical replication of novel serum markers of pancreatic cancer.

Patients and samples

A total of 120 patients with pancreatic cancer and controls were recruited for this study. For the discovery study, sera were obtained from two different clinical centers [University Hospital Leipzig (UHL, set A) and University Hospital Heidelberg (UHH, set B)]. Consequently, we obtained two sets from patients with pancreatic cancer [Ap (n = 20), Bp (n = 20)] and two sets of healthy controls [Ac (n = 20), Bc (n = 20)]. Following finalization of the discovery study, additional blood samples from patients with pancreatic cancer and from healthy controls were collected at the University Hospital Leipzig for independent external validation [Cp (n = 20) and Cc (n = 20)]. Subjects were adjusted according to age and gender (Table 1). Additionally, serum samples of 26 patients suffering from acute pancreatitis were collected as the inflammatory control group.

Table 1.

Clinical characteristics of patients with pancreatic cancer and controls

Study complexSample setsVariablesPancreatic cancerControls
Discovery study     
    Set A UHL n 20 20 
  Male/female 10/10 10/10 
  Mean age (y) 59.3 50.1 
  Age range (y) 46-71 37-71 
    Set B UHH n 20 20 
  Male/female 10/10 15/5 
  Mean age (y) 59.3 56.9 
  Age range (y) 47-70 41-85 
External validation     
    Set C UHL n 20 20 
  Male/female 10/10 10/10 
  Mean age (y) 63.8 52.2 
  Age range (y) 33-72 32-70 
Study complexSample setsVariablesPancreatic cancerControls
Discovery study     
    Set A UHL n 20 20 
  Male/female 10/10 10/10 
  Mean age (y) 59.3 50.1 
  Age range (y) 46-71 37-71 
    Set B UHH n 20 20 
  Male/female 10/10 15/5 
  Mean age (y) 59.3 56.9 
  Age range (y) 47-70 41-85 
External validation     
    Set C UHL n 20 20 
  Male/female 10/10 10/10 
  Mean age (y) 63.8 52.2 
  Age range (y) 33-72 32-70 

Abbreviations: UHL, University Hospital Leipzig; UHH, University Hospital Heidelberg.

Blood sampling from patients was done before the initiation of specific therapy. Diagnosis of pancreatic cancer was confirmed by histologic examination in all cases. Healthy controls showed no evidence of actual disease proven by physical examination and routine laboratory testing [differentials, C-reactive protein, creatinine, transaminases, alkaline phosphatase, γ-glutamyl transferase, bilirubin, and tumor markers (CA 19-9, CEA)]. Serum samples were collected and stored (−80°C) using standardized techniques and protocols (22).

The study was approved by the local ethics committees and fulfills the requirements of the Helsinki Declaration. All subjects gave informed consent to participate in the study.

Chemicals, standards, and consumables

Gradient grade acetonitrile, ethanol, and high-performance liquid chromatography water were obtained from J.T. Baker; p.a. trifluoroacetic acid and acetone were purchased from Sigma-Aldrich. The peptide and protein MALDI-TOF calibration standards I and α-cyano-4-hydroxycinnamic acid were purchased from Bruker Daltonics. Automated magnetic bead preparations were done using 96-well plates, TubePlates from Biozym, polypropylene tubes (low profile) from Abgene, and modular reservoir quarter modules from Beckman. For sample storage, 450 μL CryoTubes were purchased from Sarstedt. Multifly needle sets and polypropylene serum monovettes with clotting activators were also obtained from Sarstedt.

Peptidome separation

All serum samples of the discovery set were processed once and analyzed simultaneously to avoid procedure-dependent errors. The external validation set was prepared, processed, and analyzed separately.

Peptidome separation of the samples was done using the CLINPROT profiling purification kits from Bruker Daltonics. Magnetic particles with defined surface functionalities [magnetic bead–immobilized metal ion affinity chromatography (MB-IMAC Cu), magnetic bead–hydrophobic interaction (MB-HIC C8), and weak cation exchange (MB-WCX)] were processed by the CLINPROT liquid handling robot according to the manufacturer's protocol (Bruker Daltonics). Serum specimens were thawed on ice for 30 min and immediately processed according to our standardized protocol for serum peptidomics (22).

Mass spectrometry

A linear MALDI-TOF mass spectrometer (Autoflex I; Bruker Daltonics) was used for the peptidome profiling. Daily mass calibration was done using the standard calibration mixture of peptides and proteins in a mass range of 1 to 10 kDa. Mass spectra were recorded and processed using the AutoXecute tool of the flexControl acquisition software (ver. 2.0; Bruker Daltonics). For candidate biomarker discovery using MS/MS analysis, a MALDI-TOF/TOF mass spectrometer (Ultraflex III; Bruker Daltonics) was used.

Mass spectrometry data analysis

Bioinformatic processing. For data analysis, two bioinformatic tools were applied, Bruker Daltonics ClinPro Tools (CPT) 2.0.365 Software and proteomics.net, a novel statistically driven preprocessing, peak-finding, and analyzing pipeline (27). Mass spectra were generated equally for CPT and proteomics.net. Device-dependent raw data served as the database for CPT. For proteomics.net, data were transformed into device-independent raw data using an ASCII converter (American Standard Code for Information Interchange) provided by Bruker Daltonics.

The CPT workflow starts by spectra loading of two selected classes (e.g., pancreatic cancer and healthy controls of one set). The used CPT software package includes an automated raw data pretreatment workflow, comprising baseline subtraction with 80% baseline flatness (Convex Hull), normalization of spectra according to the total ion count, an alignment of peaks with a signal-to-noise ratio (S/N) >3 to prominent peaks with S/N > 100 and a peak-picking procedure resulting in peaks defined as dynamic m/z ranges. Savitzky-Golay smoothing was deactivated to avoid the blurring of peaks (CPT manual version 2.0; Bruker Daltonics). Peak statistics were done using Welch's t test without multiple testing correction. Finally, the software provides a list of peaks sorted along the statistical difference between two classes, which was used for further data analysis (28, 29).

Furthermore, we introduced proteomics.net as a novel web-based bioinformatic platform for peptidome data analysis in cooperation with the Department for Mathematics of the Free University Berlin and Microsoft Research (Cambridge, United Kingdom; ref. 27). The fundamental difference with former approaches is based on the omission of postanalysis spectra alignment and of S/N cutoffs. This results in an up to 10-fold improvement of detection sensitivity (27). Following single spectra processing, proteomics.net software creates a master-peak list by detecting clusters using statistical distribution analysis. Single spectra are matched to the master-peak list and the peak features are derived. For candidate creation, we selected peaks with significant feature (via e.g., Jensen-Shannon divergence) in patient and control groups. P values were calculated for every significant peak feature using an extreme values distribution method. Terminal classification of spectra was done by established classification algorithms (e.g., Support Vector Machine [SVM] from WEKA). An internal validation (20% leave-out internal five cycle cross-validation) was realized to avoid model-specific overfitting.

Following data analysis and peak calculation, we obtained peak lists according to statistical significance for every study set and both software tools (Fig. 1). Because CPT performs no correction of the multiple testing problem, we finally focused on the 50 best-discriminating peaks.

Fig. 1.

Schematic overview of the entangled peak selection procedure including reciprocal cross-validation and external confirmation. Discovery sets (A) and (B), consisting of patients (p) and controls (c) differentiating peak lists (boxes) are linearly and reciprocally derived. Comparison of these peak lists reveals the common candidate peak (D), which is applied as a discriminator on the external validation set. This discriminator was subsequently analyzed with respect to its discriminatory power and structure.

Fig. 1.

Schematic overview of the entangled peak selection procedure including reciprocal cross-validation and external confirmation. Discovery sets (A) and (B), consisting of patients (p) and controls (c) differentiating peak lists (boxes) are linearly and reciprocally derived. Comparison of these peak lists reveals the common candidate peak (D), which is applied as a discriminator on the external validation set. This discriminator was subsequently analyzed with respect to its discriminatory power and structure.

Close modal

Data analysis

In the discovery study, each disease group was cross-tested against both control groups to create lists of candidate markers applying both bioinformatic approaches (Fig. 1). Subsequently, we filtered discriminating candidate peaks present in all comparisons. In the following external validation study, we applied patterns of these candidate peaks alone and in combination with the tumor markers CA 19-9 and CEA to classify the independent samples of the external validation set (Table 1, Set C). In addition, we did receiver operator characteristics curves (ROC) analyses using SPSS Software (version 16, SPSS, Inc.) to detect the gain in sensitivity and specificity introduced by the additional candidate peaks.

In silico search of candidate biomarkers

An in silico search against SwissProt database was done using TagIdent tool from ExPASy Proteomics Server with the following criteria: m/z 7767.6, error tolerance 350 ppm, taxonomy: Homo sapiens.

Antibody-based confirmation of MB-MALDI-TOF MS–based marker

Antibody capture beads magnetic bead–based immunoaffinity chromatography on immobilized protein G (MB-IAC Prot G; Bruker Daltonics) were used for the antibody-based confirmation of the specific candidate peaks. Selected sample aliquots were prepared according to the manufacturer's instructions. Briefly, 15 μL of the MB-IAC Prot G beads were incubated for 60 min at room temperature with 10 μL of anti–platelet factor 4 (PF4) antibodies (concentration, 1 μg/μL). Thereafter, 5 μL of a MB-IMAC Cu prepared serum and a PF4 standard solution were added and incubated for 120 min at room temperature. Samples were eluted using 10 μL elution buffer and incubated for 20 min at room temperature. MALDI target preparation was done as mentioned above.

Immunoassay for quantification of PF4

PF4 levels were measured in the serum samples using an Asserachrom PF4 antibody-based ELISA (Roche). The serum samples were diluted 1:2,100 v/v with a dilution buffer and analyzed according to the manufacturer's instructions on a Tecan MT plate reader (Tecan) at a wavelength of 450 nm.

Immunoassays for quantification of CA 19-9 and CEA

CA 19-9 and CEA were measured in serum samples by an electrochemiluminescence immunoassay (Roche) on Modular analytics E 170 analyzer (Roche) according to the manufacturer's instructions.

Detection of discriminating peaks

In the discovery phase, 80 samples [subset A: 20 cancer (Ap) and 20 controls (Ac); subset B: 20 cancer (Bp) and 20 controls (Bc)] were included for peptidome profiling (Table 1). We obtained ∼750 signals for each single serum specimen using MB-IMAC Cu, MB-WCX, and MB-HIC C8, respectively. Each sample was 4-fold processed to improve the data reproducibility (22). Therefore, 960 peptidome profiles in the mass range of 1,000 to 10,000 daltons were generated. In total, ∼60,000 mass signals were available for further bioinformatic analysis.

Data processing using CPT. The cancer subsets Ap and Bp were cross-tested against the healthy control subsets Ac and Bc for significant differences in the peptide profile using Welch's t test on the basis of peak areas (Fig. 1). As a result, four candidate lists of the 50 most significant peaks were generated. This was separately done for each of the three bead functionalities. In total, 150 promising candidates were selected for each group comparison. In summary, as a result of the discovery phase, 600 candidates were selected for further validation. In the next step, all candidates, which could not all be found in the four candidate lists, were excluded from further evaluation. As a result, only seven significant peaks, which fulfilled this criterion, remained. In the following verification procedure, unfortunately, all candidates had to be excluded as discriminators due to their inconstant rectification as shown for m/z 1945 in Fig. 2. Consequently, further data processing by genetic algorithm and support vector machine using CPT was omitted.

Fig. 2.

Peaks identified by CPT as significant were contrarily oriented; example peak m/z 1945, which is higher in the control subset Ac (A), but lower in the control subset Bc (B), compared with the corresponding patients.

Fig. 2.

Peaks identified by CPT as significant were contrarily oriented; example peak m/z 1945, which is higher in the control subset Ac (A), but lower in the control subset Bc (B), compared with the corresponding patients.

Close modal

Proteomics.net. Based on the same design, the cancer subsets Ap and Bp were cross-tested against the healthy control subsets Ac and Bc for significant differences in the peptide profile using protemics.net. This bioinformatic tool calculates P values using an extreme values distribution method. In the four candidate lists of each functionality, 8 to 19 significant different peaks could be identified (in total 57 peaks), of which six peaks (m/z 1003, 1021, 3194, 3884, 4055, and 5959) were significant in at least two comparisons (Supplementary Table S1). Signals up to an m/z 1500 were excluded from further data analysis due to ion suppression effects of the MALDI matrix resulting in disturbing high background noise. Therefore, only four peaks (m/z 3194, 3884, 4055, and 5959) could be confirmed as discriminating peaks in the following verification procedure. These four peaks were applied for linear SVM classification analysis. The peak pattern m/z 3884 and 5959 (Fig. 3) showed the best discriminating power with a sensitivity of 86.3% and specificity of 97.6%, adding a further third or fourth peak did not significantly improve the sensitivity and specificity. Therefore, we decided to limit our further analyses to the peak couple m/z 3884 and m/z 5959.

Fig. 3.

Peak m/z 3884 in the controls (A) and the pancreatic carcinoma patients (B) and peak m/z 5959 in the controls (C) and pancreatic carcinoma patients (D) of the external validation set C, as revealed by proteomics.net.

Fig. 3.

Peak m/z 3884 in the controls (A) and the pancreatic carcinoma patients (B) and peak m/z 5959 in the controls (C) and pancreatic carcinoma patients (D) of the external validation set C, as revealed by proteomics.net.

Close modal

Models for AUROC analysis

The SVM data were not applicable for AUROC analysis. Therefore, we did all AUROC calculations and model generations on the quantitative data of the peak heights and serum tumor marker levels using the SPSS 16 statistical package (SPSS).

The four model variables of interest were the established tumor markers CA 19-9 and CEA as well as the aforementioned peaks m/z 3884 and m/z 5959 as nondiscrete quantitative variables for a given patient P as follows:

  • W(P) = numerical value of the serum concentration of CEA in ng/mL for patient (P)

  • X(P) = numerical value of the serum concentration of CA 19-9 in ng/mL for patient (P)

  • Y(P) = numerical value of the peak height of m/z 3884 for patient (P)

  • Z(P) = numerical value of the peak height of m/z 5959 for patient (P)

We built simple factorial models out of these variables, which resulted in composite scores:

  • Model for tumor markers alone: score 1 = W(P) × X(P)

  • Model for peaks alone: score 2 = Y(P) × Z(P)

  • Model for tumor markers and peaks combined: score 3 = [Y(P) × Z(P)] / [W(P) × X(P)]

  • Model for tumor markers and peak m/z 5959: score 3a = Z(P) / [W(P) × X(P)]

  • Model for tumor markers and peak m/z 3884: score 3b = Y(P) / [W(P) × X(P)]

These models were equally applied to the discovery and validation sets.

AUROCs in the discovery set

Based on these quantitative data of the discovery set, the AUROC of the serum tumor marker concentrations alone (score 1) was 0.925 [95% confidence interval, (95% CI), 0.856-0.994], the AUROC consisting of peak heights of peaks m/z 3884 and m/z 5959 (score 2) was 0.734 (95% CI, 0.620-0.848), and the AUROC of the combination of peak heights and serum tumor marker concentrations (score 3) was 0.960 (95% CI, 0.922-0.997).

AUROCS in the external validation set

The aim of the external replication was to validate the discriminatory power of candidate m/z 3884 and 5959 in an independently collected sample subset of 20 patients with pancreatic cancer (Cp) and 20 healthy controls (Cc) alone and in combination with the established tumor markers CA 19-9 and CEA. Computing ROCs, CA 19-9 and CEA (score 1) revealed an area under the curve (AUROC) of 0.868. Adding peak m/z 5959 to CA 19-9 and CEA (score 3a) increased the sensitivity, but decreased the specificity, resulting in an unimproved AUROC of 0.868. However, adding peak m/z 3884 to CA 19-9 and CEA (score 3b) led to an increase in sensitivity and specificity with an AUROC of 1.0 (Fig. 4).

Fig. 4.

ROC curves of the external validation set. A, sensitivity and specificity for the combination of the two tumor markers CA 19-9 and CEA, resulting in an AUROC of 0.868; B, the first model plus marker m/z 5959, also resulting in an AUROC of 0.868; C, the first model plus marker m/z 3884, resulting in an AUROC of 1.00.

Fig. 4.

ROC curves of the external validation set. A, sensitivity and specificity for the combination of the two tumor markers CA 19-9 and CEA, resulting in an AUROC of 0.868; B, the first model plus marker m/z 5959, also resulting in an AUROC of 0.868; C, the first model plus marker m/z 3884, resulting in an AUROC of 1.00.

Close modal

Identification of biomarkers

The MALDI-TOF/TOF MS analysis of the purified MB-IMAC CU eluate revealed that besides the signal m/z 3884, a second prominent peak at m/z 7767, which suggests that both peaks are differentially charged ions from the same molecule. However, no sufficient fragment spectra could be obtained for structural identification and database search. High-resolution MS proved m/z 3884.3 as the double-charged ion of the signal at m/z 7767.6. Database search using TagIdent tool from SwissProt database revealed PF4 as the potential underlying peptide. Next, we used anti-PF4 MB-IAC Prot G particle–based MALDI-TOF analysis to confirm m/z 3884 as a double-charged ion of PF4. As shown in Fig. 5, the signals at m/z 7767 as well as m/z 3884 could be unambiguously identified as single-charged and double-charged ions of PF4.

Fig. 5.

Identification of m/z 3884 and m/z 7767, respectively. A, IMAC Cu eluate of a characteristic serum sample after standard preparation; B, IMAC Cu eluate of a PF4 standard after incubation with anti-PF4 MB-IAC Prot G particles; C, IMAC Cu eluate of a serum sample after incubation with anti-PF4 MB-IAC Prot G particles.

Fig. 5.

Identification of m/z 3884 and m/z 7767, respectively. A, IMAC Cu eluate of a characteristic serum sample after standard preparation; B, IMAC Cu eluate of a PF4 standard after incubation with anti-PF4 MB-IAC Prot G particles; C, IMAC Cu eluate of a serum sample after incubation with anti-PF4 MB-IAC Prot G particles.

Close modal

Validation of PF4 by ELISA techniques

For the direct validation and quantification of the identified PF4 in all serum samples of the present study, the ELISA technique was used. The PF4 concentrations of healthy controls (median/2.5th/97.5th percentile: 7.3/3.3/13.8 kU/mL) and patients with pancreatic cancer (median/2.5th/97.5th percentile: 5.6/0.8/12.3 kU/mL) differed significantly (P = 0.001) and confirmed the MB-MALDI-TOF MS results. In addition, 26 serum samples from patients with acute pancreatitis were analyzed as inflammatory controls. The PF4 concentrations of patients with acute pancreatitis (median/2.5th/97.5th percentile: 8.7/4.6/15.4 kU/mL) were slightly elevated compared with healthy controls and were significantly (P < 0.001) higher compared to patients with pancreatic cancer (Supplementary Fig. S1). The AUROCs of immunologically determined PF4 concentrations for the discrimination between healthy controls and patients with pancreatic cancer was 0.833 (95% CI, 0.725-0.941), and for the discrimination between patients with pancreatic cancer and patients with acute pancreatitis, it was 0.829 (95% CI, 0.720-0.938).

In this study, we identified and confirmed PF4 as a potential marker peptide for pancreatic cancer using MALDI-TOF MS–based clinical serum peptidome profiling with special consideration for the preanalytic preconditions and bioinformatic intricacies. The additional application of PF4 strongly improves the diagnostic power of conventional serum tumor marker panels consisting of CA 19-9 and CEA.

Clinical proteomics and peptidomics have rapidly grown over the past years, especially in the discovery of potential biomarkers for cancer diagnosis (613). However, missing standardization of preanalytic factors, methodologic shortcomings, and bioinformatic artifacts led to controversies regarding the applicability of these techniques in clinical settings (1418). To achieve the objective of extracting true positive marker peptides from a haystack of interference-based candidates, well-designed, bias-free, and prospective investigations are demanded, in which sample collection and storage are highly standardized and appropriate bioinformatic tools for analysis of putative informative peptides are applied (10, 30). Therefore, we collected, stored, and processed the serum samples of the study according to a feasible and highly standardized preanalytic protocol to minimize any sampling-related disturbances (22). Following MALDI-TOF peptidome analysis, we applied a proprietary bioinformatic software (ClinPro Tools) and a recently developed bioinformatic approach (protemics.net) for spectra analysis, which promised to improve the detection of candidate peaks (31). Primarily, we used the proprietary CPT software with data analysis based on an averaged mass spectrum generation for each sample set. In the discovery phase, we could initially select 600 candidates. However, following cross-validation and verification, no single candidate marker remained. Secondarily, we applied protemics.net, which was specially developed for the demands of large-scale peptidome profiling studies (27). This bioinformatic tool supports the processing of each single peptidome profile. Therefore, it allows data analysis with respect to the variance and the statistical distribution of each single peak even below the common noise level and avoids artifactual peak findings as well as data overfitting (27). Using the proteomics.net software, in total, 57 candidate peaks were obtained in the discovery phase. Following cross-testing and verification, four reproducible candidate peptides remained, of which peaks m/z 3884 and 5959 showed the best discriminating power with a sensitivity of 86.3% and a specificity of 97.6%.

To assure the reproducibility of the results, independent validation studies are necessary (17), but often are not practical due to the limited access to comparable patient subsets of similar source and possible storage time–dependent degradation of the samples (21). Hence, following discovery study, we collected an independent sample set for external validation of our first results adhering to the same preanalytic, analytic, and bioinformatic protocols. Complementing the suggestions proposed by Diamandis (20) and Pepe et al. (32), this procedure provided the possibility to immediately sort out irreproducible peaks and confirm the true positives. As a result of this external validation, the two peaks m/z 3884 and 5959 could be confirmed as potential candidate markers. Our data also support the requirement of a multicenter study design to detect reproducible candidate markers and to rule out center-specific influences on the peptidome profiles (33).

The diagnostic power of the single peaks m/z 3884 and 5959, as well as their pattern, were proven alone and in combination with the conventional tumor markers CA 19-9 and CEA by performing ROC analysis. The selectivity of the conventional tumor marker model (AUROC 0.868) was comparable to previous data (34). Introducing peak m/z 5959 sensitized the model, but lowered the specificity, resulting in unimproved discriminatory power (AUROC 0.868). However, the addition of peak m/z 3884 enabled the correct assignment of the whole evaluation set and increased the selectivity by 13.2% (AUROC 1.000). This information surplus might be attributable to disease-associated alterations not covered by conventional tumor markers.

Interestingly, the mass signal m/z 3884 was also found as a potential discriminating peak in patients suffering from pancreatic cancer in a study by Koopmann et al. using SELDI-TOF MS analysis (35), but the underlying peptide was not identified. Using MALDI-TOF/TOF MS analysis, we could prove m/z 3884 as a double-charged ion of m/z 7767 by high-resolution MS. A subsequent database search revealed PF4 as the most likely candidate peptide. Using a specific G protein–coupled antibody-based MALDI-TOF approach, we could identify the underlying peptide of candidate m/z 3884 as PF4. The subsequent direct immunologic quantification of PF4 in all study samples corroborated our mass spectrometric findings and proved that PF4 levels significantly decreased (P = 0.001) in patients suffering from pancreatic cancer. To exclude an inflammatory response effect, we analyzed the PF4 concentration in serum samples of patients suffering from acute pancreatitis. The PF4 concentrations of patients with pancreatitis were significantly higher (P < 0.001) compared with patients with cancer and even slightly elevated compared with healthy controls. This finding resembles results in previous studies of inflammatory bowel disease (36). Thus, it is unlikely that concomitant inflammation in pancreatic cancer is causing the decreased PF4 levels. In patients suffering from prostate cancer, PF4 was also significantly decreased compared with controls (37). Therefore, the PF4 decrease itself does not seem to be a pancreatic cancer–specific effect, but it adds information to the conventional serum tumor marker panels consisting of CA 19-9 and CEA, and thereby strongly improves the sensitivity and specificity of the laboratory tumor marker testing in patients suffering from pancreatic cancer. The application of a monoclonal PF4 ELISA specific for m/z 3884 (and m/z 7767) might further enhance these results.

PF4 is a member of the C-X-C chemokine family (CXCL4) and is present in α granules of all mammalian platelets as well as in the granules of mast cells (38, 39). The implication of PF4 in tumor growth and vascularization is still in discussion, and possible mechanisms of action are only partially elucidated (40). Recent evidence suggests that PF4 might pleiotropically both mark and mediate the expansion of pancreatic malignancies (41, 42). The cancer-associated reduction of PF4 serum concentration in patients with pancreatic cancer might be explained by recent observations from Villanueva et al. (43, 44). They could show that differential exoprotease activities might contribute to cancer type–specific serum peptidome degradation. In pancreatic cancer, several matrix metalloproteinases are up-regulated and partially secreted into the blood (45, 46). This has recently been shown for matrix metalloproteinase-9 (47), which is also capable of degrading PF4 (45).

In conclusion, we identified and replicated PF4 as an additional discriminating marker in pancreatic cancer, which improves, in combination with the conventional markers CA 19-9 and CEA, the diagnostic power of tumor marker testing. This might be of additional relevance for the differential diagnosis of pancreatic cancer and pancreatitis. Further investigations are necessary to enlighten the complexity of PF4's action in pancreatic cancer and to evaluate the effect of PF4 as a novel and additional tumor marker in prospective clinical studies.

No potential conflicts of interest were disclosed.

Grant support: BMBF/SAB and by a grant from Microsoft Research, Cambridge (J. Thiery), and by a “formel-1 grant” of the Medical Faculty of the University Leipzig (S. Baumann and A.B. Leichtle).

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.

Note: Supplementary data for this article are available at Clinical Cancer Research Online (http://clincancerres.aacrjournals.org/).

G.M. Fiedler, A.B. Leichtle, and J. Kase contributed equally to this study.

We thank Drs. Markus Kostrzewa and Thomas Elssner from Bruker Daltonics for skillful technical support regarding our mass-spectrometric platform. We also thank Dr. André Hagehülsmann from Microsoft Research for his outstanding support in planning and implementation of the proteome database.

1
Lowenfels AB, Maisonneuve P. Epidemiology and risk factors for pancreatic cancer.
Best Pract Res Clin Gastroenterol
2006
;
20
:
197
–209.
2
Michl P, Pauls S, Gress TM. Evidence-based diagnosis and staging of pancreatic cancer.
Best Pract Res Clin Gastroenterol
2006
;
20
:
227
–51.
3
Chen R, Yi EC, Donohoe S, et al. Pancreatic cancer proteome: the proteins that underlie invasion, metastasis, and immunologic escape.
Gastroenterology
2005
;
129
:
1187
–97.
4
Misek DE, Patwa TH, Lubman DM, Simeone DM. Early detection and biomarkers in pancreatic cancer.
J Natl Compr Canc Netw
2007
;
5
:
1034
–41.
5
Hess V, Glimelius B, Grawe P, et al. CA 19-9 tumour-marker response to chemotherapy in patients with advanced pancreatic cancer enrolled in a randomised controlled trial.
Lancet Oncol
2008
;
9
:
132
–8.
6
Aebersold R, Mann M. Mass spectrometry-based proteomics.
Nature
2003
;
422
:
198
–207.
7
Liotta LA, Ferrari M, Petricoin E. Clinical proteomics: written in blood.
Nature
2003
;
425
:
905
.
8
Rai AJ, Chan DW. Cancer proteomics: serum diagnostics for tumor marker discovery.
Ann N Y Acad Sci
2004
;
1022
:
286
–94.
9
Coombes KR, Morris JS, Hu J, Edmonson SR, Baggerly KA. Serum proteomics profiling—a young technology begins to mature.
Nat Biotechnol
2005
;
23
:
291
–2.
10
Diamandis EP, van der Merwe DE. Plasma protein profiling by mass spectrometry for cancer diagnosis: opportunities and limitations.
Clin Cancer Res
2005
;
11
:
963
–5.
11
Hortin GL, Jortani SA, Ritchie JC, Jr., Valdes R, Jr., Chan DW. Proteomics: a new diagnostic frontier.
Clin Chem
2006
;
52
:
1218
–22.
12
Beretta L. Proteomics from the clinical perspective: many hopes and much debate.
Nat Methods
2007
;
4
:
785
–6.
13
Grantzdorffer I, Carl-McGrath S, Ebert MP, Rocken C. Proteomics of pancreatic cancer.
Pancreas
2008
;
36
:
329
–36.
14
Check E. Proteomics and cancer: running before we can walk?
Nature
2004
;
429
:
496
–7.
15
Diamandis EP. Analysis of serum proteomic patterns for early cancer diagnosis: drawing attention to potential problems.
J Natl Cancer Inst
2004
;
96
:
353
–6.
16
Hortin GL. Can mass spectrometric protein profiling meet desired standards of clinical laboratory practice?
Clin Chem
2005
;
51
:
3
–5.
17
Diamandis EP. Oncopeptidomics: a useful approach for cancer diagnosis?
Clin Chem
2007
;
53
:
1004
–6.
18
Liotta LA, Petricoin EF. Putting the “bio” back into biomarkers: orienting proteomic discovery toward biology and away from the measurement platform.
Clin Chem
2008
;
54
:
3
–5.
19
Banks RE, Stanley AJ, Cairns DA, et al. Influences of blood sample processing on low-molecular-weight proteome identified by surface-enhanced laser desorption/ionization mass spectrometry.
Clin Chem
2005
;
51
:
1637
–49.
20
Diamandis EP. Peptidomics for cancer diagnosis: present and future.
J Proteome Res
2006
;
5
:
2079
–82.
21
Banks RE. Preanalytical influences in clinical proteomic studies: raising awareness of fundamental issues in sample banking.
Clin Chem
2008
;
54
:
6
–7.
22
Baumann S, Ceglarek U, Fiedler GM, Lembcke J, Leichtle A, Thiery J. Standardized approach to proteome profiling of human serum based on magnetic bead separation and matrix-assisted laser desorption/ionization time-of-flight mass spectrometry.
Clin Chem
2005
;
51
:
973
–80.
23
Findeisen P, Sismanidis D, Riedl M, Costina V, Neumaier M. Preanalytical impact of sample handling on proteome profiling experiments with matrix-assisted laser desorption/ionization time-of-flight mass spectrometry.
Clin Chem
2005
;
51
:
2409
–11.
24
Rai AJ, Vitzthum F. Effects of preanalytical variables on peptide and protein measurements in human serum and plasma: implications for clinical proteomics.
Expert Rev Proteomics
2006
;
3
:
409
–26.
25
Fiedler GM, Baumann S, Leichtle A, et al. Standardized peptidome profiling of human urine by magnetic bead separation and matrix-assisted laser desorption/ionization time-of-flight mass spectrometry.
Clin Chem
2007
;
53
:
421
–8.
26
Bruegel M, Planert M, Baumann S, et al. Standardized peptidome profiling of human cerebrospinal fluid by magnetic bead separation and matrix-assisted laser desorption/ionization time-of-flight mass spectrometry.
J Proteomics
2009
;
72
:
608
–15.
27
Conrad TOF, Leichtle A, Hagehülsmann A, et al. Beating the noise: new statistical methods for detecting signals in MALDI-TOF spectra below noise level. In: M.R. Berthold R.G., Fischer I., editors. Springer-Verlag Berlin Heidelberg; 2006, p. 119–28.
28
Ketterlinus R, Hsieh SY, Teng SH, Lee H, Pusch W. Fishing for biomarkers: analyzing mass spectrometry data with the new ClinProTools software.
Biotechniques
2005
;
Suppl
:
37
–40.
29
Bruker. Clinprotools 2.0 user manual: Bruker Daltonik GmbH; 2005.
30
McLerran D, Grizzle WE, Feng Z, et al. Analytical validation of serum proteomic profiling for diagnosis of prostate cancer: sources of sample bias.
Clin Chem
2008
;
54
:
44
–52.
31
Diamandis EP. Mass spectrometry as a diagnostic and a cancer biomarker discovery tool: opportunities and potential limitations.
Mol Cell Proteomics
2004
;
3
:
367
–78.
32
Pepe MS, Etzioni R, Feng Z, et al. Phases of biomarker development for early detection of cancer.
J Natl Cancer Inst
2001
;
93
:
1054
–61.
33
van der Merwe DE, Oikonomopoulou K, Marshall J, Diamandis EP. Mass spectrometry: uncovering the cancer proteome for diagnostics.
Adv Cancer Res
2007
;
96
:
23
–50.
34
Chen R, Pan S, Brentnall TA, Aebersold R. Proteomic profiling of pancreatic cancer for biomarker discovery.
Mol Cell Proteomics
2005
;
4
:
523
–33.
35
Koopmann J, Zhang Z, White N, et al. Serum diagnosis of pancreatic adenocarcinoma using surface-enhanced laser desorption and ionization mass spectrometry.
Clin Cancer Res
2004
;
10
:
860
–8.
36
Meuwis MA, Fillet M, Geurts P, et al. Biomarker discovery for inflammatory bowel disease, using proteomic serum profiling.
Biochem Pharmacol
2007
;
73
:
1422
–33.
37
Lam YW, Mobley JA, Evans JE, Carmody JF, Ho SM. Mass profiling-directed isolation and identification of a stage-specific serologic protein biomarker of advanced prostate cancer.
Proteomics
2005
;
5
:
2927
–38.
38
Maurer AM, Zhou B, Han ZC. Roles of platelet factor 4 in hematopoiesis and angiogenesis.
Growth Factors
2006
;
24
:
242
–52.
39
McLaren KM, Pepper DS. The immunoelectronmicroscopic localization of human platelet factor 4 in tissue mast cells.
Histochem J
1983
;
15
:
795
–800.
40
Bikfalvi A. Platelet factor 4: an inhibitor of angiogenesis.
Semin Thromb Hemost
2004
;
30
:
379
–85.
41
Soucek L, Lawlor ER, Soto D, Shchors K, Swigart LB, Evan GI. Mast cells are required for angiogenesis and macroscopic expansion of Myc-induced pancreatic islet tumors.
Nat Med
2007
;
13
:
1211
–8.
42
Theoharides TC. Mast cells and pancreatic cancer.
N Engl J Med
2008
;
358
:
1860
–1.
43
Villanueva J, Shaffer DR, Philip J, et al. Differential exoprotease activities confer tumor-specific serum peptidome patterns.
J Clin Invest
2006
;
116
:
271
–84.
44
Villanueva J, Martorella AJ, Lawlor K, et al. Serum peptidome patterns that distinguish metastatic thyroid carcinoma from cancer-free controls are unbiased by gender and age.
Mol Cell Proteomics
2006
;
5
:
1840
–52.
45
Van den Steen PE, Proost P, Wuyts A, Van Damme J, Opdenakker G. Neutrophil gelatinase B potentiates interleukin-8 tenfold by aminoterminal processing, whereas it degrades CTAP-III, PF-4, and GRO-α and leaves RANTES and MCP-2 intact.
Blood
2000
;
96
:
2673
–81.
46
Patten LC, Berger DH. Role of proteases in pancreatic carcinoma.
World J Surg
2005
;
29
:
258
–63.
47
Tian M, Cui YZ, Song GH, et al. Proteomic analysis identifies MMP-9, DJ-1 and A1BG as overexpressed proteins in pancreatic juice from pancreatic ductal adenocarcinoma patients.
BMC Cancer
2008
;
8
:
241
.

Supplementary data