MICROARRAYS AND OTHER NEW TECHNOLOGIES. Diagnosis of Gastric Cancer by Serum Proteomic Fingerprinting

Transcription

1 GASTROENTEROLOGY 2006;130: MICROARRAYS AND OTHER NEW TECHNOLOGIES Diagnosis of Gastric Cancer by Serum Proteomic Fingerprinting TERENCE C. W. POON, JOSEPH J. Y. SUNG, SHUK M. CHOW, ENDERS K. W. NG, ALEX C. W. YU, EAGLE S. H. CHU, ANGELA M. Y. HUI, and WAI K. LEUNG Institute of Digestive Diseases of the Chinese University of Hong Kong and Prince of Wales Hospital, Shatin, New Territories, Hong Kong, Special Administrative Region, The People s Republic of China See editorial on page Background & Aims: Accurate serum biomarkers for gastric cancer currently are lacking. We attempted to identify potential diagnostic serum markers for gastric cancer with the use of the surface-enhanced laser desorption/ionization ProteinChip technology. Methods: The study was divided into 3 phases: (1) discovery of potential diagnostic markers using sera of gastric cancer patients and controls, (2) development of a diagnostic model, and (3) independent validation of the diagnostic model using a different cohort of gastric cancer and control patients. The serum proteins/peptides were analyzed with 2 types of Protein- Chip arrays, IMAC30 arrays loaded with copper (II) ion and CM10 (weak cation exchange) arrays. Results: In the discovery set, peak intensities of 31 surface-enhanced laser desorption/ionization proteomic features were significantly higher in gastric cancer patients. The tumor-specific nature of 6 proteomic features with the mass/charge (m/z) values of 5098, 8592, 8610, 11,468, 11,804, and 50,140 was verified by their lower peak intensities in postoperative sera. After excluding the sodium adduct peak (8610 m/z) of the 8592 m/z protein, the peak intensities of the tumor-specific proteomic features were used to develop a linear regression model for calculating a diagnostic index. The area under the receiver operating characteristic curve of the corresponding diagnostic index was 0.92 (95% confidence interval, ) in the independent validation set. At a specificity of 95%, the sensitivity for gastric cancer detection was 83%. Conclusions: A unique serum proteomic fingerprint can be detected in the sera of gastric cancer patients, which may be useful in the noninvasive diagnosis of gastric cancer. Gastric cancer is the second leading cause of cancerrelated deaths worldwide. One of the main reasons for the high cancer-related mortality is the delay in diagnosis because early cancers typically are asymptomatic. To date, the invasive endoscopic examination remains the most reliable method for the accurate diagnosis of gastric cancers. Although endoscopic screening for gastric cancer is practiced widely in Japan, the feasibility and cost effectiveness of this aggressive approach in most other countries remains questionable. 1 In this regard, the lack of a simple and reliable blood test for the diagnosis of gastric cancer has hindered the development of a noninvasive screening test for this fatal cancer. Specifically, there is no reliable serum biomarker for gastric cancer. 2 5 Surface-enhanced laser desorption/ionization (SELDI) ProteinChip technology is a proteomic tool that has been applied in the discovery of diagnostic proteomic fingerprints for various diseases including cancer and infectious diseases. 6 8 In this technology, proteins are bound on a solid-phase chromatographic surface (ie, the ProteinChip array [Ciphergen Biosystems, Inc., Fremont, CA]), ionized, and detected by time-of-flight mass spectrometry. The normalized peak intensity is directly proportional to the serum level. Because most of the proteins are singlecharged during the time-of-flight mass spectrometry analysis, each peak in the mass spectrum usually corresponds to a single peptide/protein with a molecular weight equivalent to the m/z value. By using different chromatographic surfaces, different proteins can be enriched and detected selectively by the system in a quantitative manner. Advantages of this technology are its high-throughput capability and the requirement of small amounts of sera. In the present study, we aimed to identify serum proteomic fingerprints specifically associated with gastric cancer with the SELDI ProteinChip technology and to develop a proteomic-fingerprinting model for gastric cancer. Abbreviations used in this paper: m/z, mass/charge; ROC, receiver operating characteristic; SAA, serum amylase A; SELDI, surface-enhanced laser desorption/ionization by the American Gastroenterological Association Institute /06/$32.00 doi: /j.gastro

2 May 2006 SERUM PROTEOMIC FINGERPRINT OF GASTRIC CANCER 1859 Figure 1. Study design and patient flow. The study was divided into 3 parts: discovery phase, model construction phase, and validation phase. The diseased and control patients involved in the validation phase were different from the patients involved in the proteomic feature selection and model training. Materials and Methods Study Design This study comprised 3 parts: (1) biomarker discovery, (2) diagnostic model construction, and (3) independent validation (Figure 1). In addition, we used a 2-stage approach in the biomarker discovery part. In the first stage, the serum proteomic profiles of the pretreatment samples of cancer patients were compared with controls to identify differential proteomic features. In the second stage, the differential proteomic features were verified in the postoperative sera of cancer patients. Patients All blood samples were collected from gastric cancer patients or controls, who presented to the Prince of Wales Hospital in Hong Kong. The part I study (biomarker discovery set) included 34 pretherapeutic and 24 postoperative serum samples from gastric cancer patients. The mean age of the cancer patients was 57.3 years (range years). Gastric cancers were staged according to the 5th edition of the TNM staging system. 9 Thirty-one patients had complete TNM staging: 5 stage I, 5 stage II, 15 stage III, and 6 stage IV. Control sera were collected from 29 age-matched patients who underwent upper endoscopy for dyspepsia and were found to have normal endoscopy results. In the third part of the study (independent validation set), pretherapeutic sera were collected from a different cohort of 40 gastric cancer patients and postoperative sera were available from 14 patients. The mean age of the cancer patients was 69.4 years (range, years). There were 35 patients with complete TNM staging: 5 stage I, 5 stage II, 6 stage III, and 19 stage IV. Again, control sera from another 20 age-matched healthy patients with normal upper-endoscopy results were included. All blood samples were collected in a fasting state in the early morning. Sera were separated immediately and then were stored at 80 C until processing. All patients and controls gave informed consent for drawing blood and the study procedure was approved by the Institutional Review Board of the Chinese University of Hong Kong. Serum Proteomic Profiling Serum samples were subjected to the SELDI Protein- Chip analysis to obtain a quantitative proteomic profile with a molecular mass ranging from 0.9 to 250 kilodaltons as previously described. 8,10 All serum samples were processed and assayed in a random order. The investigator was blinded to the diagnosis. The SELDl ProteinChip analysis was performed with CM10 ProteinChip arrays at a ph of 4, and IMAC30 ProteinChip arrays were loaded with copper (II) ion (Cu 2 ion). Two microliters of the serum sample were denatured by adding 4 L of U9 solution (9 mol/l urea, 2% 3-[(3-Cholamidopropyl) dimethylammonio]-1-propanesulfonate, 50 mmol/l Tris-HCl, ph 9), and diluted with 34 L oft4 binding buffer (50 mmol/l sodium acetate, 0.1% Triton X-100, ph 4.0) for CM10 ProteinChip array experiments or metal affinity-binding buffer (0.1% n-octyl glucoside, 1 mol/l NaCl, 200 mmol/l Na phosphate, ph 7.0) for copper-loaded IMAC30 ProteinChip array to result in a final dilution of 20-fold. Five microliters of the diluted sample were applied to the pre-equilibrated ProteinChip array in duplicate, and incubated with shaking at room temperature for 90 minutes. After the incubation, each array was washed 5 times with the binding buffer and rinsed twice with deionized water. After air drying, sinapinic acid matrix in 50% acetonitrile and 0.5% trifluoroacetic acid was added to each array. The ProteinChip arrays were read on a ProteinChip PBS II reader of the ProteinChip Biomarker System to measure the masses and intensities of the protein peaks (Ciphergen Biosystems, Inc.).

3 1860 POON ET AL GASTROENTEROLOGY Vol. 130, No. 6 The spectra were smoothed, baseline subtracted, and externally calibrated. The common peaks among the SELDI mass spectra were identified and quantified using Biomarker Wizard software (Ciphergen Biosystems). The peak intensities were normalized with the total ion current, and subsequently with the total peak intensities. Before data mining, the normalized peak intensities of the duplicate measurements were averaged, followed by log 2 transformation. The intra-assay coefficient of variations of the normalized peak intensities of various peaks were less than 15%. Statistical Analyses The Significance Analysis of Microarray (Stanford University, Stanford, CA) was used to identify proteomic features with peak intensities that were significantly different between cancer patients and controls as described previously. 8,10,11 In the Significance Analysis of Microarray, 2 classed but unpaired data were selected as the data type, and 5000 permutations were performed. The false discovery rate of significant differential features was set to 0 to avoid the identification of false significant proteomic features caused by multiple comparisons. The potential diagnostic proteomic features were used to develop a linear regression model for discriminating gastric cancer patients from controls. The training dataset was composed of the samples used in the discovery phase (34 pretreatment gastric cancer samples, 29 control samples). The normalized peak intensities of the diagnostic proteomic features were used as the input variables (S-PLUS; MathSoft, Inc., Cambridge, MA). The output was the diagnostic index in the range of 0 1. During model training, the diagnostic index of a healthy patient was defined as 0; the diagnostic index of a gastric cancer case was defined as 1. The constructed model then was evaluated on the independent validation sample set (40 pretreatment gastric cancer samples, 20 control samples). The investigator was blinded to the diagnoses of the validation sample set while performing the validation analysis. The Mann Whitney U test was performed to compare the levels of a biomarker between gastric cancer patients and controls (SPSS; SPSS Inc., Chicago, IL). Correlation between biomarkers was analyzed by the Spearman rank-order correlation test. Receiver operating characteristic (ROC) curves were constructed by calculating the sensitivities and specificities of the biomarkers at different cut-off points for differentiating gastric cancer patients from control subjects. Result Identification of Serum Proteomic Features Associated With Gastric Cancer After matching the peaks ( ,000 m/z) across the mass spectra of the serum samples of cancer patients and controls in the discovery set, 716 peaks were obtained (446 by using the IMAC30 ProteinChip arrays loaded with copper (II) ion, and 270 by using the CM10 ProteinChip array). To search for proteins/polypeptides that were associated significantly with gastric cancer, we performed a Significance Analysis of Microarrays to identify for differential proteomic features between the 34 pretreatment sera of gastric cancer patients and the 29 control sera (biomarker discovery sets). At a median false-discovery rate of 0, we identified 32 peaks that were significantly higher and 3 peaks that were significantly lower in the cancer group. To confirm the specificity of these 32 up-regulated peaks in gastric cancer patients, the normalized intensities of these peaks in the 24 postoperative samples were determined. When compared with the pretreatment samples, 6 protein peaks were found to be significantly lower in the postoperative samples (Significance Analysis of Microarrays, false-discovery rate 0). The mean m/z values (minimum maximum) of the 6 peaks were 5098 ( ), 8592 ( ), 8610 ( ), 11,468 (11,452 11,484), 11,804 (11,794 11,815), and 50,140 (50,050 50,240) (Figure 2). To evaluate the diagnostic values of these 6 proteomic features, the normalized peak intensity of each feature was subjected to ROC curve analysis. The areas under the ROC curves of these proteomic features were between 0.64 and 0.87 in the discovery set (P.05, Table 1). The capability of these proteomic features in classifying gastric cancer was verified in the independent validation set, and the areas under the ROC curves of these proteomic features in identifying the gastric cancer cases ranged from 0.71 to 0.84 (P.01, Table 1). On further examination of the SELDI mass spectra, the m/z 8610 peak was found to be the secondary peak of the m/z 8592 peak because there was a high correlation between the normalized intensities of the m/z 8592 peak and the m/z 8610 peak (correlation factor 0.762, P.0005, Spearman correlation test). In most of the mass spectra, the m/z values of these 2 peaks differed by about 22 daltons, suggesting that these 2 peaks resulted from the signals of the same protein. The former corresponded to the protonated protein (MH ) whereas the latter corresponded to the sodium adduct (MNa ). Hence, only the former was used in the construction of the diagnostic model. Construction of Linear Regression Model by Serum Proteomic Fingerprinting Because the gastric cancer specific serum proteomic fingerprint comprised 5 proteomic features only (5098, 8592, 11,468, 11,804, and 50,140 m/z), a linear regression algorithm was used to construct a classification model for the diagnosis of gastric cancer.

4 May 2006 SERUM PROTEOMIC FINGERPRINT OF GASTRIC CANCER 1861 Figure 2. Representative mass spectra of a gastric cancer associated proteomic feature identified by SELDI ProteinChip technology. The intensities of the protein peak of 8592 m/z were significantly higher in the preoperative serum samples of the gastric cancer patients than in the normal controls. The cancer-specific nature of the corresponding protein was suggested by its lower levels in the postoperative serum samples. After training the model with the discovery dataset excluding the postoperative serum samples, the multiple R 2 values of the resulting model was 0.54 (P.0005). The area under the ROC curve in the discovery dataset was 0.96 (95% confidence interval, ; P.0005). To validate this model, the performance was examined with the independent validation set, which was not involved in the proteomic feature selection or model training. The mean diagnostic index ( SD) for the gastric cancer and control groups was (0.292) and (0.196), respectively (P.0005, Mann Whitney U test). The area under the ROC curve in identifying gastric cancer cases was 0.92 (95% confidence interval, ; P.0005; Figure 3). At 95% specificity, the sensitivity and accuracy of this diagnostic model was 83% and 87%, respectively. In the 14 postoperative serum samples (weeks after surgery: median, 17.3; range, 13 56), the mean diagnostic index ( SD) was (0.214), which was significantly lower than the preoperative serum samples (P.0005, Mann Whitney U test) and was not Table 1. Summary of the Gastric Cancer Specific Proteomic Features Identified by SELDI ProteinChip Technology Intensity of proteomic feature (%), mean SD (relative to total sum of proteomic features) Mean m/z (minimum maximum) ROC curve area (95% confidence interval) P value Normal control Gastric cancer Biomarker discovery set 5098 ( ).71 (.58.84) [.005] ( ).87 (.78.96) [.0005] ( ).82 (.71.93) [.0005] ,468 (11,452 11,484).64 (.51.78) [.05] ,804 (11,794 11,815).74 (.61.86) [.001] ,140 (50,050 50,240).86 (.75.96) [.0005] Independent validation set 5098 ( ).79 (.68.91) [.0005] ( ).81 (.70.92) [.0005] ( ).83 (.73.94) [.0005] ,468 (11,452 11,484).71 (.59.84) [.01] ,804 (11,794 11,815).75 (.63.87) [.005] ,140 (50,050 50,240).84 (.74.94) [.0005]

5 1862 POON ET AL GASTROENTEROLOGY Vol. 130, No. 6 Table 2. Effect of TNM Staging on the Sensitivity of the Linear Regression Model on the Detection of Gastric Cancer Patients in the Validation Set Stage I (n 5) Stage II (n 5) Stage III (n 6) Stage IV (n 19) Sensitivity 20% 80% 83% 95% correlation between the diagnostic model and tumor staging was evaluated. Among the 35 patients with available tumor staging in the independent validation set, a significant correlation between the diagnostic index and tumor staging was shown (correlation coefficient, r 0.393, P.02, Spearman rank-order correlation). The sensitivity was lower in stage I cancer (20%) and progressively increased from stage II (80%) to stage IV tumors (95%) (Table 2). Figure 3. ROC curve analysis of the diagnostic index. The diagnostic index was generated by the linear regression model, which could discriminate between the gastric cancer and control samples in the independent validation set. different from that of normal controls (P.847, Mann Whitney U test) (Figure 4). Tumor Stage and Performance of the Diagnostic Model Although the earlier-described validation was performed without considering the tumor staging, the Figure 4. Scatter plot and box plot of the diagnostic index of serum samples in the independent validation groups. The diagnostic index of each sample in the independent validation set were shown including sera samples from normal controls, preoperative samples, and postoperative samples from gastric cancer patients. The box represents the interquartile range whereas the line across the box indicates the median value. Discussion In this study, we explored the potential of a noninvasive diagnosis of gastric cancer by using the latest serum proteomic technology. With the use of very stringent criteria, we have identified 5 proteomics features that are expressed differentially in the sera of gastric cancer patients. The individual ROC curve area of the 5 proteomic features ranged from 0.64 to 0.87 in the discovery set and from 0.71 to 0.84 in the validation set. When the 5 markers were combined to generate a diagnostic index, the best sensitivity and specificity of this panel of proteomic markers was 83% and 95%, respectively. Notably, we found that the sensitivity of serum proteomic fingerprinting is dependent on the staging of cancer. In stage I cancer when the tumor is confined to mucosa or submucosa, the sensitivity of serum proteomic fingerprinting is 20%. The sensitivity improved to 80% in stage II, 83% in stage III, and 95% in stage IV tumors. These results may implicate that invasion through the submucosa plays an important role in the leaking of these cancer-specific proteins into the serum. Despite major advances in our understanding of gastric carcinogenesis, there is no reliable serum biomarker for this malignancy. Although the use of Helicobacter pylori serology and serum pepsinogen levels helps to identify a subgroup of individuals at higher risk for developing cancer, 12,13 it fails to differentiate tumor patients from nontumor controls. 14 The sensitivities of the conventional tumor markers including carcinoembryonic antigen, cancer antigen 19-9, and cancer antigen 72-4 remain low (20% 30%). 2 5 Apart from protein markers, the detection of altered DNA in blood, particularly the detection of methylated DNA, appears to be a potential

6 May 2006 SERUM PROTEOMIC FINGERPRINT OF GASTRIC CANCER 1863 cancer screening test. In our previous study, 15 aberrant DNA methylation was detected in the serum of 55% of patients with gastric cancer. It thus appears that the current approach of serum proteomic fingerprinting has a much higher sensitivity than all previous efforts of serologic diagnoses of gastric cancer. The findings of SELDI ProteinChip technology could be biased by artifacts related to the nature of the clinical samples, the sample storage conditions, the experimental details, the mass spectrometric instruments, and/or bioinformatic analyses. 17,18 In the present study we used various methods to reduce the chance of false-positive results. First, the cases for diagnostic feature selection and model training were different from the cases for validation of the resulting diagnostic model. Second, the investigators were blinded to the diagnosis while performing the SELDI profiling experiments and while validating the diagnostic models. Third, we identified the differential proteomic features by setting the false-discovery rate of the significant proteomic features at 0. Fourth, the present study adopted a unique experimental design in which the serum differential proteomic features identified in the discovery set were considered as potential diagnostic markers only if their levels were reversed in the postoperative sera (Figure 1). This is owing to the fact that a significantly larger amount of a proteomic feature in the gastric cancer group could be caused by other systematic biases from preanalytic values other than the presence of gastric cancer. Among the 31 differential proteomic features identified in the discovery set, only 6 features were reversed in the postoperative samples. Therefore, more than 85% of these features might not be associated specifically with tumors and further verification with postoperative serum samples appears to be a promising approach to increase the specificity of this assay. Finally, we used simple regression algorithms, such as linear regression, for diagnostic model construction to reduce the chance of overfitting. To confirm the practical value of our diagnostic model further, in the future it should be tested on independent samples run on a SELDI instrument in a different laboratory. Recently, Ebert et al 16 used the SELDI technology in identifying patients with gastric cancer with very high sensitivity and specificity. Instead of using a simple linear regression algorithm, the results were obtained by the combined use of 50 decision trees and 28 proteomic features. A proteomic feature of 11,537 m/z, which was similar to the cancer-specific feature of 11,468 m/z found in the present study, was observed in 3 of the 50 decision trees. Among various differentially expressed proteomic features, the m/z 3949 feature was involved in 28 (56%) of the 50 decision trees and was regarded as the most informative biomarker in their study. Although we also found that the peak intensity of 3953 m/z was significantly higher in gastric cancer patients by using either the cation exchange ProteinChip arrays or the copperloaded IMAC30 ProteinChip arrays, the intensities of this feature were not reduced significantly in postoperative patients. Because the 28 diagnostic features from the previous study were identified by direct comparison of gastric cancer patients and controls, their specificities await further investigation. The SELDI ProteinChip assays could be performed in a 96-well microplate format in a high-throughput manner. Based on our findings, it is possible to develop a rapid SELDI ProteinChip assay for the detection of gastric cancer. Because each protein has a unique m/z value and physicochemical properties, they can be detected unambiguously and quantified in patients sera using specific ProteinChip arrays, even without knowing their protein identities. However, further characterization of the 5 gastric cancer specific proteomic features (5098, 8592, 11,468, 11,804, and 50,140 m/z) may help us to understand the disease pathology, which would enable us to develop a simple immunoassay to measure these potential markers. Experiments currently are ongoing in our laboratory to identify the gastric cancer specific proteomic features. Among the 5 gastric cancer proteomic features, the serum proteins of 11,468 and 11,804 m/z identified on the weak cation exchange ProteinChip arrays (CM10 [formerly WCX2]) at ph 4.0 had been identified previously to be serum amyloid A isoforms (SAA). 18,19 SAA isoforms have been shown to be upregulated in various cancers, including renal cancer, 18 nasopharyngeal carcinoma, 19 and prostate cancer. 20 In keeping with our findings, SAA levels decreased significantly in the serum of patients with nasopharyngeal carcinoma in remission. 19 Because SAA is an acute-phase response protein that is increased in various nontumor conditions, 21 the use of SAA alone may not be an ideal candidate for the specific diagnosis of gastric cancer. Hence, the combination of the levels of SAA and other serum proteomic markers that form a disease-specific fingerprint for diagnosis appears to be a more rational approach for the diagnosis of gastric cancer. In conclusion, we have shown the presence of unique proteomic features in the serum of gastric cancer patients. Serum proteomic fingerprinting by SELDI ProteinChip technology has enormous potential on the noninvasive diagnosis of gastric cancer.

7 1864 POON ET AL GASTROENTEROLOGY Vol. 130, No. 6 References 1. Genta RM. Screening for gastric cancer: does it make sense? Aliment Pharmacol Ther 2004;20: Marrelli D, Roviello F, De Stefano A, Farnetani M, Garosi L, Messano A, Pinto E. Prognostic significance of CEA, CA 19-9 and CA72-4 preoperative serum levels in gastric carcinoma. Oncology 1999;57: Marrelli D, Pinto E, De Stefano A, Farnetani M, Garosi L, Roviello F. Clinical utility of CEA, CA 19-9, and CA 72-4 in the follow-up of patients with resectable gastric cancer. Am J Surg 2001;181: Ishigami S, Natsugoe S, Hokita S, Che X, Tokuda K, Nakajo A, Iwashige H, Tokushige M, Watanabe T, Takao S, Aikou T. Clinical importance of preoperative carcinoembryonic antigen and carbohydrate antigen 19-9 levels in gastric cancer. J Clin Gastroenterol 2001;32: Gaspar MJ, Arribas I, Coca MC, Diez-Alonso M. Prognostic value of carcinoembryonic antigen, CA 19-9 and CA 72-4 in gastric carcinoma. Tumour Biol 2001;22: Petricoin EF III, Ardekani AM, Hitt AB, Levine PJ, Fusaro VA, Steinberg SM, Mills GB, Simone C, Fishman DA, Kohn EC, Liotta LA. Use of proteomic patterns in serum to identify ovarian cancer. Lancet 2002;359: Poon TCW, Yip TT, Chan ATC, Yip C, Yip V, Mok TSK, Lee CCY, Leung TWT, Ho SKW, Johnson PJ. Comprehensive proteomic profiling identifies serum proteomic signatures for detection of hepatocellular carcinoma and its subtypes. Clin Chem 2003;49: Poon TCW, Chan KCA, Ng PC, Chiu RW, Ang IL, Tong YK, Ng EKO, Cheng FWT, Li AM, Hon EKL, Fok TF, Lo YMD. Serial analysis of plasma proteomic signatures in pediatric patients with severe acute respiratory syndrome and correlation with viral load. Clin Chem 2004;50: Sobin LH, Wittekind CH,TNM classification of malignant tumors. 5th ed. New York: John Wiley & Sons, 1997: Poon TCW, Hui AY, Chan HLY, Ang IL, Chow SM, Wong N, Sung JJY. Prediction of liver fibrosis and cirrhosis in chronic hepatitis B infection by serum proteomic fingerprinting: a pilot study. Clin Chem 2005;51: Tusher VG, Tibshirani R, Chu G. Significance analysis of microarrays applied to the ionizing radiation response. Proc Natl Acad Sci U S A 2001;98: Varis K, Sipponen P, Laxen F, Samloff IM, Huttunen JK, Taylor PR, Heinonen OP, Albanes D, Sande N, Virtamo J, Harkonen M, the Helsinki Gastritis Study Group. Implications of serum pepsinogen I in early endoscopic diagnosis of gastric cancer and dysplasia. Scand J Gastroenterol 2000;9: Bodger K, Wyatt JI, Heatley RV. Variation in serum pepsinogens with severity and topography of Helicobacter pylori-associated chronic gastritis in dyspeptic patients referred for endoscopy. Helicobacter 2001;6: Miki K, Morita M, Sasajima M, Hoshina R, Kanda E, Urita Y. Usefulness of gastric cancer screening using the serum pepsinogen test method. Am J Gastroenterol 2003;98: Lee TL, Leung WK, Chan MWY, Ng EKW, Tong JHM, Lo KW, Chung SCS, Sung JJY, To KF. Detection of gene promoter hypermethylation in the tumor and serum of patients with gastric carcinoma. Clin Cancer Res 2002;8: Ebert MP, Meuer J, Wiemer JC, Schulz HU, Reymond MA, Traugott U, Malfertheiner P, Rocken C. Identification of gastric cancer patients by serum protein profiling. J Proteome Res 2004;3: Diamandis EP. Analysis of serum proteomic patterns for early cancer diagnosis: drawing attention to potential problems. J Natl Cancer Inst 2004;96: Tolson J, Bogumil R, Brunst E, Beck H, Elsner R, Humeny A, Kratzin H, Deeg M, Kuczyk M, Mueller GA, Mueller CA, Flad T. Serum protein profiling by SELDI mass spectrometry: detection of multiple variants of serum amyloid alpha in renal cancer patients. Lab Invest 2004;84: Cho WC, Yip TT, Yip C, Yip V, Thulasiraman V, Ngan RK, Yip TT, Lau WH, Au JS, Law SC, Cheng WW, Ma VW, Lim CK. Identification of serum amyloid a protein as a potentially useful biomarker to monitor relapse of nasopharyngeal cancer by serum proteomic profiling. Clin Cancer Res 2004;10: Le L, Chi K, Tyldesley S, Flibotte S, Diamond DL, Kuzyk MA, Sadar MD. Identification of serum amyloid A as a biomarker to distinguish prostate cancer patients with bone lesions. Clin Chem 2005;51: Lannergard A, Larsson A, Kragsbjerg P, Friman G. Correlations between serum amyloid A protein and C-reactive protein in infectious diseases. Scand J Clin Lab Invest 2003;63: Received September 10, Accepted January 25, Address requests for reprints to: Wai K. Leung, MD, Department of Medicine and Therapeutics, Prince of Wales Hospital, Ngan Shing Street, Shatin, New Territories, Hong Kong, China. wkleung@cuhk.edu.hk; fax: (852)