Cancer Research Meeting Calendar Sign up for Cancer Research eTOC's
HOME HELP FEEDBACK SUBSCRIPTIONS ARCHIVE SEARCH TABLE OF CONTENTS
Cancer Research Clinical Cancer Research
Cancer Epidemiology Biomarkers & Prevention Molecular Cancer Therapeutics
Molecular Cancer Research Cancer Prevention Research
Cancer Prevention Journals Portal Cancer Reviews Online
Annual Meeting Education Book Meeting Abstracts Online
This Article
Right arrow Abstract Freely available
Right arrow Full Text (PDF)
Right arrow Alert me when this article is cited
Right arrow Alert me if a correction is posted
Services
Right arrow Email this article to a friend
Right arrow Similar articles in this journal
Right arrow Similar articles in PubMed
Right arrow Alert me to new issues of the journal
Right arrow Download to citation manager
Right arrow reprints & permissions
Citing Articles
Right arrow Citing Articles via HighWire
Right arrow Citing Articles via Google Scholar
Google Scholar
Right arrow Articles by Honda, K.
Right arrow Articles by Yamada, T.
Right arrow Search for Related Content
PubMed
Right arrow PubMed Citation
Right arrow Articles by Honda, K.
Right arrow Articles by Yamada, T.
[Cancer Research 65, 10613-10622, November 15, 2005]
© 2005 American Association for Cancer Research

Clinical Research

Possible Detection of Pancreatic Cancer by Plasma Protein Profiling

Kazufumi Honda1, Yasuharu Hayashida1,2, Tomoko Umaki1, Takuji Okusaka4, Tomoo Kosuge5, Satoru Kikuchi1,2, Mitsufumi Endo2, Akihiko Tsuchida2, Tatsuya Aoki2, Takao Itoi3, Fuminori Moriyasu3, Setsuo Hirohashi1 and Tesshi Yamada1

1 Chemotherapy Division and Cancer Proteomics Project, National Cancer Center Research Institute; 2 Third Department of Surgery and 3 Fourth Department of Internal Medicine, Tokyo Medical University; and 4 Hepatobiliary and Pancreatic Oncology Division and 5 Hepatobiliary and Pancreatic Surgery Division, National Cancer Center Hospital, Tokyo, Japan

Requests for reprints: Tesshi Yamada, Chemotherapy Division, National Cancer Center Research Institute, 5-1-1 Tsukiji Chuoh-ku, Tokyo 104-0045, Japan. Phone: 81-3-3547-5201, ext. 4270; Fax: 81-3-3547-6045; E-mail: tyamada{at}ncc.go.jp.


   Abstract
Top
 Abstract
Introduction
 Materials and Methods
 Results
 Discussion
 References
 
The survival rate of pancreatic cancer patients is the lowest among those with common solid tumors, and early detection is one of the most feasible means of improving outcomes. We compared plasma proteomes between pancreatic cancer patients and sex- and age-matched healthy controls using surface-enhanced laser desorption/ionization coupled with hybrid quadrupole time-of-flight mass spectrometry. Proteomic spectra were generated from a total of 245 plasma samples obtained from two institutes. A discriminating proteomic pattern was extracted from a training cohort (71 pancreatic cancer patients and 71 healthy controls) using a support vector machine learning algorithm and was applied to two validation cohorts. We recognized a set of four mass peaks at 8,766, 17,272, 28,080, and 14,779 m/z, whose mean intensities differed significantly (Mann-Whitney U test, P < 0.01), as most accurately discriminating cancer patients from healthy controls in the training cohort [sensitivity of 97.2% (69 of 71), specificity of 94.4% (67 of 71), and area under the curve value of 0.978]. This set discriminated cancer patients in the first validation cohort with a sensitivity of 90.9% (30 of 33) and a specificity of 91.1% (41 of 45), and its discriminating capacity was further validated in an independent cohort at a second institution. When combined with CA19-9, 100% (29 of 29 patients) of pancreatic cancers, including early-stage (stages I and II) tumors, were detected. Although a multi-institutional large-scale study will be necessary to confirm clinical significance, the biomarker set identified in this study may be applicable to using plasma samples to diagnose pancreatic cancer.


   Introduction
Top
 Abstract
 Introduction
Materials and Methods
 Results
 Discussion
 References
 
The 5-year survival rate of pancreatic cancer sufferers is the lowest among patients with common solid tumors. Pancreatic cancer is the fifth leading cause of cancer-related mortality in Japan and the fourth in the United States, with >19,000 estimated annual deaths in Japan and >28,000 in the United States (1–3). Pancreatic cancer is characterized by massive local invasion and early metastasis to the liver and regional lymph nodes. Because surgical resection is the only reliable curative treatment, early detection is essential to improve the outcomes of pancreatic cancer patients. However, the clinical symptoms of pancreatic cancer, except for obstructive jaundice, are often unremarkable until the advanced stages of the disease, and the anatomic location of the pancreas deep in the abdomen makes physical and ultrasonic detection of pancreatic cancer difficult. As a result, only 20% to 40% of pancreatic cancer patients undergo surgical resection (1, 4). Mass screening by computed tomography (CT), magnetic resonance imaging (MRI), or positron emission tomography (PET) may not be cost-effective because of the relatively low incidence of pancreatic cancer, and the long-term safety of these modalities has not been established (5). Thus, new diagnostic modalities allowing early detection of pancreatic cancer in a safe/noninvasive and cost-effective way are needed.

Recently, mass spectrometry (MS)–based proteomic approaches have gained considerable attention as effective modalities for identifying new biomarkers of various diseases because of their high sensitivity, but proteomic analysis of blood samples has been hampered by the marked dominance of a handful of particularly abundant proteins, including albumin, immunoglobulins, and transferrins (6). Surface-enhanced laser desorption/ionization (SELDI)-MS was developed to resolve these problems and is considered to be among the most useful tools available for the analysis of serum and plasma (7–9). Proteins are captured, concentrated, and purified on the small chemical surface of a SELDI chip, and the molecular weight (m/z) and relative intensity of each protein captured on the chip are measured with sensitive time-of-flight (TOF)-MS. As a result, a comprehensive proteomic profile can be created from as little as 20 µL serum/plasma samples. Combined with multivariate bioinformatical analysis, serum proteomics by SELDI-TOF-MS has been reported be successfully applied to the diagnosis of ovarian and prostate cancers (10–13).

The ProteinChip system is a sophisticated commercial platform designed for SELDI-TOF-MS. This system has been widely used because of its high-throughput automated measurements. However, relatively low resolution and poor mass accuracy have been recognized as drawbacks of the TOF-MS instrument of this system, and the reproducibility of SELDI-MS data has been controversial (14–16). Multivariate discrimination is dependent on stacks of small differences between cases and controls. Recently, Petricoin and Liotta reported the use of high-resolution performance hybrid quadrupole TOF-MS (QqTOF-MS) instruments to significantly improve the resolution and mass accuracy of SELDI-MS compared with results obtained with low-resolution instruments (17, 18).

Koopmann et al. (19) identified a set of biomarkers for pancreatic adenocarcinoma using the ProteinChip system. They increased the number of detectable peaks using stepwise anion-exchange chromatography, but only two of the six fractions were used for subsequent analyses. The two protein peaks that most effectively discriminated between pancreatic cancer patients and healthy controls reportedly achieved a sensitivity of 78% and a specificity of 97%, but this sensitivity was below the level necessary for clinical application. More importantly, diagnostic performance was not validated in an independent cohort. We reviewed and refined various aspects of SELDI-MS. In this study, we first compared the results obtained using low-resolution TOF-MS and high-resolution QqTOF-MS instruments and confirmed the high reproducibility of data obtained using the latter. Computerized machine learning may identify even a perfect multivariate classifier within a closed sample set in a nonbiological/mathematical way (16). Erroneous identification by machine learning must be eliminated by validation experiments using an independent sample set. Herein, we report the identification and validation of a set of biomarkers that can detect pancreatic cancer with high accuracy.


   Materials and Methods
Top
 Abstract
 Introduction
 Materials and Methods
Results
 Discussion
 References
 
Patients and plasma samples. Plasma samples (n = 245) were obtained from two institutes, the National Cancer Center Hospital (NCCH; Tokyo, Japan) between August 2002 and October 2003 and the Tokyo Medical University Hospital (TMUH; Tokyo, Japan) between February 2004 and February 2005. The 220 NCCH cases included untreated pancreatic ductal adenocarcinoma patients (n = 104) and healthy controls (n = 116), whereas the 25 TMUH cases included untreated pancreatic ductal adenocarcinoma patients (n = 9), individuals with pancreatic tumors and/or cysts (n = 6), chronic pancreatitis patients (n = 5), and healthy controls (n = 5). The pancreatic tumor and/or cyst category included two pathologically unproven mucinous cystic tumors, two pathologically unproven serous papillary tumors, and two clinically diagnosed nonmalignant mass lesions in the pancreas. These cases are currently being followed, and a final diagnosis has not been obtained to date. The patients in the chronic pancreatitis category had no detectable mass lesions in the pancreas. Written informed consent was obtained from all of the subjects. Blood samples were collected in EDTA glass tubes. The supernatant was separated by centrifugation and cryopreserved at –80°C until analysis. All samples were processed in the same manner. The study was reviewed and approved by the ethics committees of the National Cancer Center (Tokyo, Japan; authorization nos. 16-36 and 16-71) and Tokyo Medical University (Tokyo, Japan; authorization no. 341).

The clinical characteristics of the patients are summarized in Table 1. Patients were classified as having clinical disease stage I, II, III, or IV according to the Fifth Edition of the General Rules for the Study of Pancreatic Cancer (Japanese Pancreas Society; ref. 20).


View this table:
[in this window]
[in a new window]

  		Table 1. Clinicopathologic characteristics of the 220 cases seen at NCCH

 
Surface-enhanced laser desorption/ionization. Ninety microliters of U9 buffer [9 mol/L urea, 2% 3-[(3-cholamidopropyl)dimethylammonio]-1-propanesulfonic acid, and 50 mmol/L Tris-HCl (pH 9)] were added to 10 µL of each plasma sample and vortexed for 20 minutes. Parts of the denatured plasma samples were fractionated using stepwise anion-exchange chromatography (pH 9 plus flow trough, pH 7, pH 5, pH 4, pH 3, and organic wash) with QHyper DF resin (Ciphergen Biosystems, Inc., Fremont, CA) using a Biomek 2000 Laboratory Automation Robot (Beckman Coulter, Fullerton, CA) according to a previously described method (12, 21).

Each sample was randomly assigned, with a 96-spot format, to 12 ProteinChip arrays (8 spots per array; Ciphergen) in duplicate using the Biomek 2000 Robot. Three types of ProteinChip arrays with different surface chemistries [i.e., immobilized metal affinity capture coupled with copper (IMAC-Cu2+), weak hydrophobic (H50), or cationic (CM10) arrays] were used (21). The CM10 arrays were used under either low-stringent (pH 4) or high-stringent (pH 7) conditions as instructed by the supplier. The arrays were air-dried and applied to the matrix (50% sinapinic acid in 50% acetonitrile/0.1% trifluoroacetic acid).

Time-of-flight mass spectrometry. TOF-MS analysis was done using two types of mass spectrometers, a low-resolution TOF-MS (PBS IIc, Ciphergen) and a high-resolution QqTOF-MS [Q-star XL (Applied Biosystems, Framingham, CA) equipped with a PCI 1000 (Ciphergen)]. Peak detection for the low-resolution instrument was done using CiphergenExpress software version 2.1 (Ciphergen). All of the spectra were compiled and normalized to the total ion currents, and the baselines were subtracted. Peaks between 3,000 and 30,000 m/z were autodetected using a signal-to-noise ratio of >3, and the peaks were clustered using second-pass peak selection with a signal-to-noise ratio of >2 and 0.3% mass windows. The permissible range of m/z drift between samples was set at 0.3% (21).

The high-resolution instrument was set to measure the range between 2,000 and 40,000 m/z. The laser intensity, laser frequency, and accumulation time were set to 60%, 25 Hz, and 90 seconds, respectively. The mass data obtained using the high-resolution instrument were converted to text files consisting of m/z and intensity after mass calibration by Analyst QS (Applied Biosystems) and were processed using newly developed in-house peak detection, normalization, and quantification software (22).

The peak data were visualized using Mass Navigator software (Mitsui Knowledge Industry, Tokyo, Japan). Mass accuracy was calibrated externally on the day of the measurements using an all-in-one-peptide molecular mass standard (Ciphergen).

Statistical analysis. Statistically significant differences were detected using the Fisher exact probability test, the Student's t test, and the Mann-Whitney U test. Receiver operator characteristics (ROC) curves were generated and the area under the curve (AUC) values were calculated using StatFlex software version 5.0 (Artech, Osaka, Japan; ref. 23).

We compiled the multivariate intensity data of the mass peaks into the distance from a support vector machine (SVM) hyperplane using the following formula (details in Supplementary Data; ref. 24):

where yi is label (1 or –1), k(xj,xi) is Gaussian kernel function, and {lambda}i is a value that maximizes [1] target function under [2] constrained conditions, where is the [1] target function, 0 ≤ {lambda}i ≤ C are the [2] constrained conditions, and {alpha} and C are constants 0.25 and 10, respectively.

Immunoradiometric assay of CA19-9. Plasma (100 µL) was analyzed using a commercially available immunoradiometric assay kit (Fujirebio Diagnostic, Inc., Malvern, PA) according to the manufacturer's recommendations.


   Results
Top
 Abstract
 Introduction
 Materials and Methods
 Results
Discussion
 References
 
Comparison between low-resolution and high-resolution instruments. The reproducibility of data obtained using the low-resolution TOF-MS instrument of the ProteinChip system has been a concern. We compared the number of detectable peaks and the reproducibility of data obtained using low-resolution TOF-MS and high-resolution QqTOF-MS instruments. From unfractionated plasma samples (24 pancreatic cancer patients and 24 healthy controls), a total of 226 unique peaks were detected using the low-resolution instrument and 637 unique peaks were detected using the high-resolution instrument (Table 2). This difference seems to be attributable to the mass resolutions of the instruments (Fig. 1A). In addition, we noticed significant mass drifts (<0.3%) in the data obtained with the low-resolution instrument (Fig. 1B). In contrast, the mass deviation was <0.05% for the high-resolution instrument (Fig. 1B). As a result, the correlation coefficients for three independent measurements of a pooled plasma sample done every other day with the high-resolution instrument reached 0.97 to 0.99 (data not shown).


View this table:
[in this window]
[in a new window]

  		Table 2. Comparison of low-resolution and high-resolution instruments

 


View larger version (25K):
[in this window]
[in a new window]
[Download PPT slide]
  		Figure 1. Comparison of low-resolution and high-resolution instruments. A, representative spectra of an unfractionated plasma sample in the range of 8,000 to 10,000 m/z obtained using a low-resolution TOF instrument (top) and a high-resolution QqTOF instrument (bottom). B, spectra of an unfractionated plasma sample in the range of 8,500 to 9,500 m/z obtained thrice every other day using a low-resolution TOF instrument (top) and a high-resolution QqTOF instrument (bottom). The spectra (green, blue, and red lines) were superimposed to allow visualization of the day-to-day variations. Note that only the green line is visible in the bottom because of the high reproducibility of results obtained with the QqTOF instrument.

 
Chromatographic fractionation reduced the reproducibility of measurements. Fractionation via stepwise anion-exchange chromatography has been widely done to increase the number of detectable peaks obtained with low-resolution instruments. Actually, the total number of detectable peaks increased from 226 to 872 with fractionation of the same plasma samples (Table 2). However, the fractionation procedure seemed to compromise the reproducibility of the measurements. Forty-eight plasma samples (24 pancreatic cancer patients and 24 healthy controls) were analyzed in duplicate, and the mean correlation coefficient of all the peaks calculated between the duplicates was 0.87 to 0.96 for the unfractionated samples and 0.61 to 0.76 for the fractionated samples (Table 2). Fig. 2A (unfractionated) and Fig. 2B (fractionated) show the results of duplicate assays of a representative plasma sample.



View larger version (17K):
[in this window]
[in a new window]
[Download PPT slide]
  		Figure 2. Reproducibility of data from the low-resolution and high-resolution instruments. Two-dimensional plot analyses of the mass intensities corresponding to the duplicated peaks that appeared in the H50 (blue diamonds), IMAC-Cu2+ (red squares), CM10 pH 4 (yellow triangles), and CM10 pH 7 (light blue crosses) arrays. Unfractionated (A and C) or fractionated (B) samples of the same plasma were measured using a low-resolution TOF instrument (A and B) and a high-resolution QqTOF instrument (C).

 
Based on these quality-control experiments, we decided to measure unfractionated plasma samples using the high-resolution QqTOF-MS instrument. More than 90% of the duplicate protein peaks measured with the QqTOF-MS instrument were plotted within a 2-fold difference (Fig. 2C), and the mean correlation coefficient between duplicate assays was at least 0.95 (Table 2).

Identification of a candidate classifier in the training cohort by machine learning. From the total of 220 samples obtained at the NCCH, we selected 71 pancreatic cancer patients and 71 healthy controls with no statistically significant differences in age or sex distribution as a training cohort (Table 1). The remaining 78 cases served as a validation cohort. The clinicopathologic characteristics of these pancreatic cancer patients in the training and validation cohorts are summarized in Table 1.

The acquired MS peak information was stored in a large-capacity server computer, and the data set that most accurately discriminated pancreatic cancer patients from healthy controls was extracted using a rbf SVM learning algorithm (24). The set, or classifier, was composed of four protein peaks at 17,272 m/z (CM10 pH 4), 8,766 m/z (CM10 pH 4), 28,080 m/z (CM10 pH 4), and 14,779 m/z (H50). The selection of these four peaks was evaluated by leave-one-out (LOO) cross-validation. Representative spectra profiles and pseudo-gel images of the four peaks are shown in Fig. 3. Akaike information criterion procedure (25) selected another peak at 11,516 m/z (H50; indicated by a red arrowhead in Fig. 3). Although the 11,516 m/z peak was only detected in 1 of the 71 (1.4%) healthy controls, it was not included in the above discriminating data set generated by machine learning because of its low-positive rate in pancreatic cancer patients [19.7% (14 of 71)].



View larger version (72K):
[in this window]
[in a new window]
[Download PPT slide]
  		Figure 3. Representative mass spectra [a healthy control (c067) and a pancreatic cancer patient (p048)] and converted gel-like images [23 healthy controls (c023-c091) and 22 pancreatic cancer patients (p040-p124)] showing the peaks at 8,766, 17,272, 28,080 (CM10 pH 4), and 14,779 (H50) m/z. Red arrowhead, peak at 11,516 m/z, which was extracted using the Akaike information criterion (25).

 
Statistical differences in all four peaks were recognized between the pancreatic cancer patients and the healthy controls (Mann-Whitney U test, P < 0.0022; Table 3). The ROC and AUC values of each peak and their combination in the 142 cases of the training cohort are shown in Fig. 4.


View this table:
[in this window]
[in a new window]

  		Table 3. Intensities of the 17,272, 8,766, 14,779, and 28,080 m/z peaks

 


View larger version (28K):
[in this window]
[in a new window]
[Download PPT slide]
  		Figure 4. ROC curves and AUC values showing the discriminating capacities of the 17,272, 8,766, 28,080 (CM10 pH 4), and 14,779 (H50) m/z peaks individually and in combination.

 
The intensity data of the four peaks obtained in each individual were compiled into a single value, the distance from a fixed SVM hyperplane, using the formula described in Materials and Methods and Supplementary Data. When the distance was positive, the individual was classified as having pancreatic cancer and vice versa. This classifier correctly diagnosed 97.2% (69 of 71) of the cancer patients and 94.4% (67 of 71) of the healthy controls in the training cohort (Fig. 5A).



View larger version (24K):
[in this window]
[in a new window]
[Download PPT slide]
  		Figure 5. Calculated SVM distances of healthy controls (black columns) and pancreatic cancer patients (gray columns) in the training (A) and first validation (B) cohorts. Cases separated into the positive direction from the SVM hyperplane were classified as having "cancer" and those separated into the negative direction were classified as being "healthy."

 
Confirmation of the classifier in the first validation cohort. We next validated the discriminating performance of the classifier in a blinded manner using an independent cohort consisting of 78 individuals (NCCH) who had not been included in the training cohort (Table 1). Again, statistically significant differences in the mean intensities of every peak were observed between the 33 pancreatic cancer patients and the 45 healthy controls (Mann-Whitney U test, P < 0.0078; Table 3).

The SVM hyperplane determined in the training cohort was applied to the diagnosis of the 78 cases in the validation set. The same SVM hyperplane separated 90.9% (30 of 33) of the pancreatic cancer patients into the positive direction group and 91.1% (41 of 45) of the healthy controls into the negative direction group (Fig. 5B). The overall accuracy of the classification was 91.0% (71 of 78) in the validation cohort.

Combination of the surface-enhanced laser desorption/ionization classifier and CA19-9. Overall, the classifier was able to detect 95.2% (99 of 104) of the pancreatic cancer patients in the training and validation cohorts (Table 4). Although the number of cases was small, 83.3% (10 of 12) of stage I and II cases were detected (training and first validation cohorts). No statistically significant differences in detection rates were seen among cases with different tumor locations or different clinical stages (Table 4). To improve the detection rate, we measured plasma CA19-9 levels in all individuals whose residual samples were sufficient (29 pancreatic cancer patients and 39 healthy controls; Table 5). The sensitivity of CA19-9 (cutoff value of 37 units/mL) was 86.2% (25 of 29) and specificity was 94.9% (37 of 39). The SELDI classifier and the CA19-9 level were complementary. Combining CA19-9 and the SELDI classifier detected 100% (29 of 29) of cancer patients, but this combination yielded six false-positive cases [15.4% (6 of 39); Table 5].


View this table:
[in this window]
[in a new window]

  		Table 4. Diagnostic accuracy of the SELDI classifier

 

View this table:
[in this window]
[in a new window]

  		Table 5. Detection rates with CA19-9, the SELDI classifier, and their combination

 
Confirmation of the classifier in a second validation cohort obtained at a different institution. Finally, we did a second confirmatory experiment using samples collected prospectively at another institution. In total, 25 plasma samples from pancreatic cancer patients, individuals with other pancreatic diseases, and healthy volunteers were obtained from TMUH and analyzed in a blinded manner. Although the discovery of biomarkers useful for the differential diagnosis of pancreatic diseases was not the primary goal of this study, the classifier was able to discriminate pancreatic cancer patients and individuals with pancreatic tumors/cysts from healthy controls and pancreatitis patients (Table 4; Fig. 6). Four of the six patients with pathologically unproven pancreatic tumors/cysts were classified into the positive direction group. A close follow-up of these patients has been undertaken, because they may have premalignant or preclinical conditions. The SELDI classifier correctly identified 88.9% (8 of 9) of the pancreatic cancer patients and 80% (4 of 5) of the healthy controls, whereas the CA19-9 level correctly identified 66.7% (6 of 9) of the pancreatic cancer patients and 100% (5 of 5) of the healthy controls (Fig. 6). Again, in all the pancreatic cancer patients (9 of 9), the SELDI classifier and the CA19-9 level provided complementary results, even in this second validation cohort.



View larger version (12K):
[in this window]
[in a new window]
[Download PPT slide]
  		Figure 6. Confirmation in a second cohort treated at a different institution. A, calculated SVM distances of nine pancreatic cancer patients, six individuals with pancreatic tumors and/or cysts, five chronic pancreatitis patients, and five healthy controls seen at TMUH. B, plasma CA19-9 levels in nine pancreatic cancer patients, six individuals with pancreatic tumors and/or cysts, five chronic pancreatitis patients, and five healthy controls seen at TMUH. The cutoff value was set at 37 units/mL.

 

   Discussion
Top
 Abstract
 Introduction
 Materials and Methods
 Results
 Discussion
References
 
Comparative proteomic profiling coupled with a computerized machine learning approach may revolutionize medical practice and cancer diagnosis. We compared the plasma protein profiles of a large number of pancreatic cancer patients and healthy controls with identical age and gender distributions (Table 1) to identify a biomarker for detecting pancreatic cancer patients in a large population composed mainly of healthy individuals. The reproducibility of data obtained using the low-resolution instrument of the ProteinChip system has been a concern, but employing a high-resolution QqTOF instrument was found to significantly improve mass accuracy and minimize day-to-day variations (Fig. 1B). The high reproducibility of measurements was confirmed not by using a few high-intensity peaks selected intentionally but rather by using all the peaks detectable in the entire range (intensity and m/z) of mass spectra (Fig. 2). We also eliminated fractionation procedures, which increased the number of detectable peaks but significantly decreased reproducibility (Table 2; Fig. 2). A minimal set of four low-molecular weight proteins (Fig. 3) was found to be sufficient for discriminating pancreatic cancer patients with a sensitivity of 97.2% (69 of 71) and a specificity of 94.4% (67 of 71; Fig. 5A). This high discriminating capacity was confirmed by LOO cross-validation and ROC analysis (Fig. 4). We confirmed the discriminating capacity of our classifier in two independent validation cohorts (Figs. 5B and 6) to eliminate accidental identification of nonbiological/mathematical multivariate classifiers within a closed cohort by overfitting.

We noticed that a peak at 11,516 m/z (H50) was detected in 19.4% of the pancreatic cancer patients in the training cohort but in only 1.4% of the healthy controls (the peaks are indicated by a red arrowhead in Fig. 3). Tolson et al. (26) reported that an 11.5-kDa protein was detected in 32% of renal cell carcinoma patients but in none of the normal controls. Howard et al. (27) identified 11,682 m/z proteins in the sera of lung cancer patients as a diagnostic biomarker using matrix-assisted laser desorption/ionization (MALDI)-TOF-MS. Both groups identified the proteins as fragments of serum amyloid A. Serum amyloid A is an acute-phase reactant and a biomarker for inflammatory disease. The serum amyloid A level is elevated up to 1,000-fold during tissue damage and inflammation and is also increased in patients with various solid tumors and hematopoietic malignancies. However, serum amyloid A has not been recognized as a tumor marker because of its low positive rate (28, 29). Consistently, the 11,516 m/z peak was not incorporated into our classifier. The discovery of a single biomarker differing markedly between cancer patients and controls as well as having a high positive rate in cancer patients would be ideal but is perhaps not realistic. Since the discovery of CA19-9 in 1982 (30), no single tumor marker applicable to the clinical diagnosis of pancreatic cancer has been identified. The carcinogenesis of pancreatic cancer is probably mediated via a variety of molecular pathways (2, 31), and multimarker analysis of proteins with different specificities is a realistic alternative to a conventional single biomarker assay.

There are pros and cons to SELDI-MS with high-resolution instruments. Although the primary goal of our study was the development of a bioassay applicable to the detection of pancreatic cancer, attempts to purify proteins from these four low-intensity peaks without contamination by neighboring high-intensity peaks have not been successful to date. However, the high reproducibility of QqTOF-MS warrants direct clinical application of its measurements and does not necessitate the actual protein identification of these peaks. Zhang et al. (12) reported that a set of three peaks, at 3,272, 12,828, and 28,043 m/z, could be used to detect early-stage ovarian cancer. The 28,043 m/z peak was down-regulated in ovarian cancer patients and was found to be derived from apolipoprotein A1. The relatively abundant 28,080 m/z protein identified as one of the peaks down-regulated in pancreatic cancer patients in this study (Table 3; Fig. 3) may be related to apolipoprotein A1. The mass deviation of 0.3% seen in the low-resolution TOF-MS may represent a drift in this region as large as 84 m/z (28,080 x 0.003 = 84). At least four peaks were detected between 28,000 and 28,100 m/z using the high-resolution QqTOF-MS instrument (Fig. 3). These peaks merged and were detected as a single peak with the low-resolution instrument (data not shown). The intensities of the 8,766, 17,272, and 14,779 m/z peaks were one magnitude smaller than that of the 28,080 m/z peak (Table 3) and were apparently below the sensitivity of tandem MS. So-called top-down proteomics using Fourier transform (FT)-MS (32) may be necessary to identify the proteins indicated by the low-intensity peaks of our classifier. However, an interface to the SELDI arrays is currently not available for FT-MS.

No significant differences in the detection rates for our classifier were observed among different stages of pancreatic cancer (Table 4). Koomen et al. (33) did plasma protein profiling of pancreatic cancer patients using MALDI-MS and identified a set of eight peaks distinguishing pancreatic cancer patients from controls with a sensitivity of 88% and a specificity of 75%. Protein identification revealed these peaks to be derived mainly from host response proteins. Many low molecular weight proteins detected by SELDI-MS in serum or plasma samples have also been reported to be metabolic products, proteolytic fragments, or peptide hormones. These proteins may not always be attributable to direct secretion or production by cancer cells, instead being the results of host responses in the microenvironment of the tumor (7, 18, 34), such as stromal desmoplastic reactions, inflammation, and angiogenesis. Two of eight pancreatic cancer patients who were classified as having "cancer," but none of normal controls in the TMUH validation cohort, had diabetes (data not shown). This raises the possibility that diabetic conditions, which are often associated with pancreatic cancer patients, also may influence the classifier.

All the pancreatic cancers were detected by complementary use of CA19-9 and/or the SELDI classifier (Table 5). CA19-9 is a tumor marker widely used for the evaluation of therapeutic effects and the detection of pancreatic cancer recurrence but is not considered to be applicable to mass screening (35–38). Ten percent to 15% of humans do not secrete CA19-9 because of their genetic Lewis antigen status (39). The CA19-9 level is often within reference range when pancreatic cancer is still at an early stage and is often elevated in benign biliary and pancreatic diseases. When the cutoff value for CA19-9 was set at 37 units/mL, which is widely used for clinical purposes, the false-positive rate of the combined CA19-9 and SELDI strategy reached 15.4% (Table 5). To increase diagnostic accuracy, the CA19-9 cutoff value may need to be adjusted and the selection of SELDI peaks may need to be further refined.

Early detection seems to be essential for improving the outcomes of pancreatic cancer patients. The SELDI classifier identified in this study has high potential for detecting pancreatic cancers (Tables 4 and 5), but one of the five pancreatitis patients in the TMUH validation cohort was classified into the pancreatic cancer category (Fig. 6). This pancreatitis patient may have a premalignant or preclinical condition and is currently being followed. Alternatively, because inflammatory conditions were not used in training, it is also possible that the classifier may not be entirely specific for the cancer phenotype. Machine learning was done with the training cohort, in which there were no cases with benign pancreatic diseases, because the discovery of biomarkers useful for pancreatic cancer screening in a large population made up mostly of healthy individuals was a primary goal of this study. The final diagnosis of pancreatic cancer is not made solely based on plasma protein profiling. CT, MRI, PET, ultrasound, and endoscopic and/or surgical approaches are employed as well. To evaluate the clinical significance of the biomarkers identified in this study and to refine the selection of biomarkers using a large number of subjects, including patients with pancreatic cancer and other pancreatic diseases, we need to undertake a prospective multi-institutional study.


   Acknowledgments
 
Grant support: "Third Term Comprehensive Control Research for Cancer" from the Ministry of Health, Labor and Welfare; "Program for Promotion of Fundamental Studies in Health Sciences" of the National Institute of Biomedical Innovation of Japan; and Foundation for the Promotion of Cancer Research resident fellowship to Y. Hayashida (patent pending in Japan, no. 2005-070512).

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.

We thank Drs. T. Kakizoe, N. Moriyama, and T. Yoshida (National Cancer Center) for helpful discussions and encouragement, Y. Ishiyama for her secretarial assistance, and Dr. K. Aoshima, H. Kuwabara, T. Isobe, and H. Matsuzuki (Mitsui Knowledge Industry) for the statistical analyses.


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

Received 5/27/05. Revised 8/ 5/05. Accepted 9/ 9/05.


   References
Top
 Abstract
 Introduction
 Materials and Methods
 Results
 Discussion
 References
 

  1. Yamamoto M, Ohashi O, Saitoh Y. Japan Pancreatic Cancer Registry: current status. Pancreas 1998;16:238–42.[Medline]
  2. Lowenfels AB, Maisonneuve P. Epidemiology and prevention of pancreatic cancer. Jpn J Clin Oncol 2004;34:238–44.[Abstract/Free Full Text]
  3. Landis SH, Murray T, Bolden S, Wingo PA. Cancer statistics, 1999. CA Cancer J Clin 1999;49:8–31.[Abstract/Free Full Text]
  4. Shimamura T, Sakamoto M, Ino Y, et al. Dysadherin overexpression in pancreatic ductal adenocarcinoma reflects tumor aggressiveness: relationship to e-cadherin expression. J Clin Oncol 2003;21:659–67.[Abstract/Free Full Text]
  5. Berrington de Gonzalez A, Darby S. Risk of cancer from diagnostic X-rays: estimates for the UK and 14 other countries. Lancet 2004;363:345–51.[CrossRef][Medline]
  6. Anderson NL, Anderson NG. The human plasma proteome: history, character, and diagnostic prospects. Mol Cell Proteomics 2002;1:845–67.[Abstract/Free Full Text]
  7. Issaq HJ, Veenstra TD, Conrads TP, Felschow D. The SELDI-TOF MS approach to proteomics: protein profiling and biomarker identification. Biochem Biophys Res Commun 2002;292:587–92.[CrossRef][Medline]
  8. Chapman K. The ProteinChip Biomarker System from Ciphergen Biosystems: a novel proteomics platform for rapid biomarker discovery and validation. Biochem Soc Trans 2002;30:82–7.[CrossRef][Medline]
  9. von Eggeling F, Junker K, Fiedle W, et al. Mass spectrometry meets chip technology: a new proteomic tool in cancer research? Electrophoresis 2001;22:2898–902.[CrossRef][Medline]
  10. Petricoin EF, Ardekani AM, Hitt BA, et al. Use of proteomic patterns in serum to identify ovarian cancer. Lancet 2002;359:572–7.[CrossRef][Medline]
  11. Adam BL, Qu Y, Davis JW, et al. Serum protein fingerprinting coupled with a pattern-matching algorithm distinguishes prostate cancer from benign prostate hyperplasia and healthy men. Cancer Res 2002;62:3609–14.[Abstract/Free Full Text]
  12. Zhang Z, Bast RC, Jr., Yu Y, et al. Three biomarkers identified from serum proteomic analysis for the detection of early stage ovarian cancer. Cancer Res 2004;64:5882–90.[Abstract/Free Full Text]
  13. Banez LL, Prasanna P, Sun L, et al. Diagnostic potential of serum proteomic patterns in prostate cancer. J Urol 2003;170:442–6.[CrossRef][Medline]
  14. Baggerly KA, Morris JS, Coombes KR. Reproducibility of SELDI-TOF protein patterns in serum: comparing datasets from different experiments. Bioinformatics 2004;20:777–85.[Abstract/Free Full Text]
  15. 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.[CrossRef][Medline]
  16. Ransohoff DF. Lessons from controversy: ovarian cancer screening and serum proteomics. J Natl Cancer Inst 2005;97:315–9.[Abstract/Free Full Text]
  17. Conrads TP, Fusaro VA, Ross S, et al. High-resolution serum proteomic features for ovarian cancer detection. Endocr Relat Cancer 2004;11:163–78.[Abstract]
  18. Petricoin EF, Liotta LA. SELDI-TOF-based serum proteomic pattern diagnostics for early detection of cancer. Curr Opin Biotechnol 2004;15:24–30.[CrossRef][Medline]
  19. 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.[Abstract/Free Full Text]
  20. Japanese Pancreas Society. General rule for the study of pancreatic cancer. Tokyo: Kanehara Shuppan; 2002.
  21. Hara T, Honda K, Ono M, Naito K, Hirohashi S, Yamada T. Identification of two serum biomarkers of renal cell carcinoma by surface-enhanced laser desorption/ionization mass spectrometry. J Urol 2005;174:1213–7.[CrossRef][Medline]
  22. Hayashida Y, Honda K, Osaka Y, et al. Possible prediction of chemoradiosensitivity of esophageal cancer by serum protein profiling. Clin Cancer Res. In press.
  23. Metz CE. Basic principles of ROC analysis. Semin Nucl Med 1978;8:283–98.[Medline]
  24. Byvatov E, Schneider G. Support vector machine applications in bioinformatics. Appl Bioinformatics 2003;2:67–77.[Medline]
  25. Kadota K, Nishimura S, Bono H, et al. Detection of genes with tissue-specific expression patterns using Akaike's information criterion procedure. Physiol Genomics 2003;12:251–9.[Abstract/Free Full Text]
  26. Tolson J, Bogumil R, Brunst E, et al. Serum protein profiling by SELDI mass spectrometry: detection of multiple variants of serum amyloid {alpha} in renal cancer patients. Lab Invest 2004;84:845–56.[CrossRef][Medline]
  27. Howard BA, Wang MZ, Campa MJ, Corro C, Fitzgerald MC, Patz EF, Jr. Identification and validation of a potential lung cancer serum biomarker detected by matrix-assisted laser desorption/ionization-time of flight spectra analysis. Proteomics 2003;3:1720–4.[CrossRef][Medline]
  28. Weinstein PS, Skinner M, Sipe JD, Lokich JJ, Zamcheck N, Cohen AS. Acute-phase proteins or tumour markers: the role of SAA, SAP, CRP and CEA as indicators of metastasis in a broad spectrum of neoplastic diseases. Scand J Immunol 1984;19:193–8.[Medline]
  29. Khan N, Cromer CJ, Campa M, Patz EF, Jr. Clinical utility of serum amyloid A and macrophage migration inhibitory factor as serum biomarkers for the detection of nonsmall cell lung carcinoma. Cancer 2004;101:379–84.[CrossRef][Medline]
  30. Herlyn M, Sears HF, Steplewski Z, Koprowski H. Monoclonal antibody detection of a circulating tumor-associated antigen. I. Presence of antigen in sera of patients with colorectal, gastric, and pancreatic carcinoma. J Clin Immunol 1982;2:135–40.[CrossRef][Medline]
  31. Gansauge S, Gansauge F, Beger HG. Molecular oncology in pancreatic cancer. J Mol Med 1996;74:313–20.[CrossRef][Medline]
  32. Gershon D. Proteomics technologies: probing the proteome. Nature 2003;424:581–7.[Medline]
  33. Koomen JM, Shih LN, Coombes KR, et al. Plasma protein profiling for diagnosis of pancreatic cancer reveals the presence of host response proteins. Clin Cancer Res 2005;11:1110–8.[Abstract/Free Full Text]
  34. Liotta LA, Ferrari M, Petricoin E. Clinical proteomics: written in blood. Nature 2003;425:905.[CrossRef][Medline]
  35. Goggins M, Canto M, Hruban R. Can we screen high-risk individuals to detect early pancreatic carcinoma? J Surg Oncol 2000;74:243–8.[CrossRef][Medline]
  36. Abrams RA, Grochow LB, Chakravarthy A, et al. Intensified adjuvant therapy for pancreatic and periampullary adenocarcinoma: survival results and observations regarding patterns of failure, radiotherapy dose and CA19-9 levels. Int J Radiat Oncol Biol Phys 1999;44:1039–46.[CrossRef][Medline]
  37. Ritts RE, Pitt HA. CA 19-9 in pancreatic cancer. Surg Oncol Clin N Am 1998;7:93–101.[Medline]
  38. Safi F, Schlosser W, Kolb G, Beger HG. Diagnostic value of CA 19-9 in patients with pancreatic cancer and nonspecific gastrointestinal symptoms. J Gastrointest Surg 1997;1:106–12.[CrossRef][Medline]
  39. Narimatsu H, Iwasaki H, Nakayama F, et al. Lewis and secretor gene dosages affect CA19-9 and DU-PAN-2 serum levels in normal individuals and colorectal cancer patients. Cancer Res 1998;58:512–8.[Abstract/Free Full Text]



This article has been cited by other articles:


Home page
 J. Biol. Chem.Home page
M. Ono, J. Matsubara, K. Honda, T. Sakuma, T. Hashiguchi, H. Nose, S. Nakamori, T. Okusaka, T. Kosuge, N. Sata, et al.
Prolyl 4-Hydroxylation of {alpha}-Fibrinogen: A NOVEL PROTEIN MODIFICATION REVEALED BY PLASMA PROTEOMICS
J. Biol. Chem., October 16, 2009; 284(42): 29041 - 29049.
[Abstract] [Full Text] [PDF]

Home page
 JCOHome page
J. Matsubara, M. Ono, A. Negishi, H. Ueno, T. Okusaka, J. Furuse, K. Furuta, E. Sugiyama, Y. Saito, N. Kaniwa, et al.
Identification of a Predictive Biomarker for Hematologic Toxicities of Gemcitabine
J. Clin. Oncol., May 1, 2009; 27(13): 2261 - 2268.
[Abstract] [Full Text] [PDF]

Home page
 J. Clin. Pathol.Home page
T Qi, J Han, Y Cui, M Zong, X Liu, and B Zhu
Comparative proteomic analysis for the detection of biomarkers in pancreatic ductal adenocarcinomas
J. Clin. Pathol., January 1, 2008; 61(1): 49 - 58.
[Abstract] [Full Text] [PDF]

Home page
 Clin. Cancer Res.Home page
D. V. Gold, Z. Karanjawala, D. E. Modrak, D. M. Goldenberg, and R. H. Hruban
PAM4-Reactive MUC1 Is a Biomarker for Early Pancreatic Adenocarcinoma
Clin. Cancer Res., December 15, 2007; 13(24): 7380 - 7387.
[Abstract] [Full Text] [PDF]

Home page
 Ann. Thorac. Surg.Home page
Z. T. Hammoud, L. Dobrolecki, K. A. Kesler, E. Rahmani, K. Rieger, L. H. Malkas, and R. J. Hickey
Diagnosis of Esophageal Adenocarcinoma by Serum Proteomic Pattern
Ann. Thorac. Surg., August 1, 2007; 84(2): 384 - 392.
[Abstract] [Full Text] [PDF]

Home page
 JCOHome page
W. Wang, S. Chen, K. A. Brune, R. H. Hruban, G. Parmigiani, and A. P. Klein
PancPRO: Risk Assessment for Individuals With a Family History of Pancreatic Cancer
J. Clin. Oncol., April 10, 2007; 25(11): 1417 - 1422.
[Abstract] [Full Text] [PDF]

Home page
 Clin. Chem.Home page
C. Melle, G. Ernst, N. Escher, D. Hartmann, B. Schimmel, A. Bleul, H. Thieme, R. Kaufmann, K. Felix, H. M. Friess, et al.
Protein Profiling of Microdissected Pancreas Carcinoma and Identification of HSP27 as a Potential Serum Marker
Clin. Chem., April 1, 2007; 53(4): 629 - 635.
[Abstract] [Full Text] [PDF]

Home page
 Mol. Cell. ProteomicsHome page
M. Ono, M. Shitashige, K. Honda, T. Isobe, H. Kuwabara, H. Matsuzuki, S. Hirohashi, and T. Yamada
Label-free Quantitative Proteomics Using Large Peptide Data Sets Generated by Nanoflow Liquid Chromatography and Mass Spectrometry
Mol. Cell. Proteomics, July 1, 2006; 5(7): 1338 - 1347.
[Abstract] [Full Text] [PDF]

Home page
 Clin. Chem.Home page
G. L. Hortin
The MALDI-TOF Mass Spectrometric View of the Plasma Proteome and Peptidome
Clin. Chem., July 1, 2006; 52(7): 1223 - 1237.
[Abstract] [Full Text] [PDF]

Home page
 Cancer Res.Home page
E. P. Diamandis
Serum Proteomic Profiling by Matrix-Assisted Laser Desorption-Ionization Time-of-Flight Mass Spectrometry for Cancer Diagnosis: Next Steps
Cancer Res., June 1, 2006; 66(11): 5540 - 5541.
[Full Text] [PDF]

Home page
 Molecular Cancer TherapeuticsHome page
A. Jimeno and M. Hidalgo
Molecular biomarkers: their increasing role in the diagnosis, characterization, and therapy guidance in pancreatic cancer.
Mol. Cancer Ther., April 1, 2006; 5(4): 787 - 796.
[Abstract] [Full Text] [PDF]

This Article
Right arrow Abstract Freely available
Right arrow Full Text (PDF)
Right arrow Alert me when this article is cited
Right arrow Alert me if a correction is posted
Services
Right arrow Email this article to a friend
Right arrow Similar articles in this journal
Right arrow Similar articles in PubMed
Right arrow Alert me to new issues of the journal
Right arrow Download to citation manager
Right arrow reprints & permissions
Citing Articles
Right arrow Citing Articles via HighWire
Right arrow Citing Articles via Google Scholar
Google Scholar
Right arrow Articles by Honda, K.
Right arrow Articles by Yamada, T.
Right arrow Search for Related Content
PubMed
Right arrow PubMed Citation
Right arrow Articles by Honda, K.
Right arrow Articles by Yamada, T.

HOME HELP FEEDBACK SUBSCRIPTIONS ARCHIVE SEARCH TABLE OF CONTENTS
Cancer Research Clinical Cancer Research
Cancer Epidemiology Biomarkers & Prevention Molecular Cancer Therapeutics
Molecular Cancer Research Cancer Prevention Research
Cancer Prevention Journals Portal Cancer Reviews Online
Annual Meeting Education Book Meeting Abstracts Online