- Email: [email protected]
Using proteomic approaches to identify new biomarkers for detection and monitoring of ovarian cancer
Gynecologic Oncology 100 (2006) 247 – 253 www.elsevier.com/locate/ygyno
Using proteomic approaches to identify new biomarkers for detection and monitoring of ovarian cancer Fanbing Kong a,1, C. Nicole White b,1, Xueyuan Xiao c,1, Youji Feng a, Congjian Xu a, Dacheng He c, Zhen Zhang b, Yinhua Yu d,* a
d
Gynecology and Obstetrics Hospital of Fudan University, Shanghai, China b Johns Hopkins Medical Institutions, Baltimore, MD 21205, USA c Beijing Normal University, Beijing, China U.T.M.D. Anderson Cancer Center, 1515 Holcombe Blvd., Box 354, Houston, TX 77030, USA Received 31 March 2005 Available online 17 October 2005
Abstract Objectives. Early detection and monitoring the treatment remain the most important factors in improving long-term survival of ovarian cancer patients. New biomarkers that individually or in combination improve the diagnostic performance of existing tumor markers are critically needed. This study uses proteomic approaches to identify new biomarkers for detection and monitoring of ovarian cancer. Methods. We analyzed protein profiles of three sets of sera using surface enhanced laser desorption and ionization time-of-light mass spectroscope (SELDI-TOF-MS) on IMAC ProteinChip arrays and ProPeak software for bioinformatics data analysis. The first set of patients included 21 ovarian cancers, 18 benign diseases, and 20 normal patients. The second set included 32 ovarian cancers, 30 benign ovarian diseases, and 30 age-matched healthy controls. The third set included samples collected before and after chemotherapy from 18 ovarian cancer patients. All samples were collected at the Gynecology and Obstetrics Hospital of Fudan University in Shanghai, China. The datasets from low-intensity and high-intensity spectra were analyzed separately. Results. Seven peaks were selected for their contribution to the separation of ovarian cancers from controls using the first and second set of samples. The same dysregulation trends were confirmed for six of the seven peaks in independent validation using the third set of samples. Conclusions. Using SELDI-TOF analysis of 195 unique specimens, we discovered with preliminary validation six distinct peaks that may potentially be useful in the detection and monitoring of ovarian cancer. Additional studies are required to determine the protein identities of these peaks and to further validate their performance as biomarkers. D 2005 Elsevier Inc. All rights reserved. Keywords: Proteomics; Biomarker; SELDI; Ovarian cancer
Introduction Ovarian cancer is the sixth most common cancer among women and the fifth most common cause of cancer death among women, in the United States, with approximately 25,580 new cases and 16,090 deaths anticipated in 2004 [1]. Ovarian cancer causes more deaths than any other cancer of the female reproductive system. Ovarian cancer occurs most frequently in women aged 50 – 79; over 70% of the cancers * Corresponding author. Fax: +1 713 745 2107. E-mail address: [email protected] (Y. Yu). 1 Contributed equally to this study. 0090-8258/$ - see front matter D 2005 Elsevier Inc. All rights reserved. doi:10.1016/j.ygyno.2005.08.051
occur after age 50. Ninety percent of women with ovarian cancer have no family history of the disease [2]. After diagnosis and standard treatment with surgery and chemotherapy, nearly 80% of women are expected to survive 1 year and 52% to survive 5 years. With early detection and prompt treatment, the overall survival rate climbs to 95%. However, early-stage diagnosis is rare. Currently, only 25% of all ovarian cancers are found at an early stage [3], suggesting the pressing need to improve the early detection of ovarian cancer. Monitoring treatment is another important factor to improve the outcome of ovarian cancer. At present, the optimal followup/monitoring strategy for the asymptomatic patient with advanced ovarian cancer after initial treatment is undefined.
248
F. Kong et al. / Gynecologic Oncology 100 (2006) 247 – 253
Two screening and monitoring tests for ovarian cancer are currently in use: the CA125 serum tumor marker and imaging using transvaginal sonography (TVS). CA125 is a highmolecular-weight glycoprotein, detectable in serum that has been used for many years to detect ovarian cancer recurrence and to monitor treatment response. CA125 has been studied more recently for use in early detection. Less then 1% of healthy post-menopausal women have a CA125 > 35 U/ml. Over 85% of all advanced ovarian cancer cases have CA125 levels exceeding this threshold. Some 50– 60% of stage I cases with disease confined to the ovary exceed this threshold [4]. CA125 may be useful as an indicator of risk, as it is over 30 U/ml in 50% of ovarian cancer patients more than 18 months prior to clinical detection and in 23% of patients more than 5 years before diagnosis [5]. However, no single screening test exists that can detect all malignant transformations of the ovary. In a pre-menopausal population, the specificity of CA125 is not adequate for early detection. Nearly 6% of pre-menopausal women without cancer have levels of CA125 exceeding 35 U/ml [6]. Many conditions, including pregnancy, uterine fibroids, inflammation or infection within the abdomen (such as colitis, acute pancreatitis, or hepatitis), or other malignancies, can cause elevations of the CA125 level. The Gynecologic Cancer Intergroup (GCIG) consists of representatives from the major gynecologic cancer trial groups around the world. After considerable debate – and after several versions – the GCIG has proposed a precise but simple definition for response according to CA125, based on a 50% reduction in CA125 levels that is confirmed and maintained for at least 28 days. This definition, with examples demonstrating its implementation, is posted on the GCIG Web site [7]. It is important to recognize the limitations of CA125 when monitoring the course of disease. Levels can be altered dramatically by abdominal surgery or peritonitis. There is also the possibility of laboratory error and considerable variation in results among laboratories. Torizuka et al. tested the value of whole-body positron emission tomography (PET) using 2-[fluorine-18]-fluoro-2-deoxy-d-glucose (FDG) for the diagnosis of recurrent ovarian cancer. They demonstrated that FDG-PET may be accurate and useful for the detection of tumor recurrence when conventional imaging is inconclusive or negative, especially in patients with abnormal CA125 levels [8]. In order to identify more effective biomarkers to detect early-stage ovarian cancer patients and monitor biological responses to therapy, proteomic-based approaches have been extended to medical application and diagnosis [9,10]. Recently, several reports demonstrated that proteomic approaches could improve our ability to detect and monitor ovarian cancer [11]. Proteomic expression patterns can be potentially evaluated as a new way to track biological responses to therapy, even become the starting point for individualized therapy [12]. Our recent work has also identified three new biomarkers for ovarian cancer using a proteomics approach [13]. In the early stage of ovarian cancer, some low-molecularweight serum protein profiling might reflect the pathologic state of patient. Screening these proteins may be very useful for distinguishing early-stage ovarian cancer and monitoring
biological responses to therapy. Matrix-assisted laser desorption and ionization time-of-light (MALD-TOF) and surface enhanced laser desorption and ionization time-of-flight (SELDTOF) mass spectroscope can profile proteins in this range [14,15]. With the help of bioinformatics analytical tools, we will be able to identify significant changes in complex mass spectra in a high-throughput fashion. SELDI-TOF technology has been evaluated as a new means to track biological responses to therapy. The measurement of a protein profile at different stages of disease progression or before and after treatment, combined with protein microarray technologies, constitutes a new paradigm for detecting disease and monitoring disease response to therapy [12]. Ultimately, proteomics and genomics will become integrated into cancer patient management through the design and tracking of individualized therapy. In this study, we use SELDI-TOF to analyze three sets of sera. First, we generated comparative protein profiles from biopsy-confirmed ovarian cancer, benign ovarian tumors, and healthy controls. Then we use the protein profiles from the first two sets to compare the changes before and after chemotherapy in the third set of cancer patients’ samples. This study may provide important information in detecting and monitoring ovarian cancer. Materials and methods Patients and clinical samples Serum samples were collected at the Gynecology and Obstetrics Hospital, Fudan University in Shanghai, China, from October 2002 to November 2003. Informed consents were obtained from all patients and healthy controls. According to a pre-defined protocol, sera were collected in the morning. Samples of patients were collected before operation or 1 to 2 months (average 32 days) after chemotherapy. Sera were kept at room temperature for 1 to 3 h for clotting before being centrifuged at 1000 rpm and again at 12000 rpm at 4-C. Samples were immediately aliquoted and stored at 70-C. All sera were allowed to thaw once only. All samples had CA125 value by CA125 EIA test (Canag Company, Sweden). All ovarian cancer and benign tumor patients had received ultrasound examination. A total of 195 unique specimens were included in this study. Diagnoses were pathologically confirmed. A total of three sets of spectra were analyzed. While there is some overlap between the sample sets, new spectra from overlapping samples were collected with each dataset (see Table 1). A total of 58 ovarian cancer patients, 48 ovarian benign tumor patients, and 45 healthy controls are included in this research study. All normal controls were without evidence of gynecological diseases at least 3 years before sera collection and were healthy more then 2 years after collection. In the first dataset, 20 healthy women between the ages of 21 and 56 years (mean T SD: 39 T 11) were included in the normal control group. 18 women between the ages of 18 and 72 years (44 T 14) were included with benign disease. 21 women with ovarian cancer between the ages of 18 and 79 years (54 T 16) were included as well. The normal and benign groups were younger then the cancer group in the first set. We had therefore also collected a second serum set of agematched samples. The second set included 30 health women aged 27 to 75 years (43 T 10), 30 women with benign ovarian tumor aged from 21 to 73 years (age 45 T 9), and finally 32 women aged 35 – 72 years (50 T 10) with malignant ovarian cancer. All benign tumor patients and cancer patients received open surgical or laparoscopic treatment. There are a number of samples in the second set overlap with those in the first set. The second set had in total 57 new samples (27 cancer patients, 15 benign patients, and 15 healthy controls).
F. Kong et al. / Gynecologic Oncology 100 (2006) 247 – 253
249
Table 1 Sample size of three datasets Disease status
Cancer Benign Normal
Dataset 1
Dataset 2
Dataset 3
Overlapping samples
Unique observations
Unique observations
Unique observations
Overlap between datasets 1 and 2
Overlap between datasets 1 and 3
Overlap between datasets 2 and 3
21 18 20
27 15 15
10 15 10
5 15 15
0 0 0
8 0 0
Unique and overlapping observations presented. In the third dataset, 18 patients aged 41 – 73 (54 T 9) with ovarian cancer were enrolled in ovarian cancer group. All these 18 cancer patients had completed surgery, remained tumor size < 2 cm. Sera samples were collected before operation and after regular chemotherapy (DDP + ADM + CTX). 3 patients had one cycle of chemotherapy, 5 patients had two cycles, 9 had three cycles, and 1 patient had four cycles. A total of 10 cancer patients were unique to this dataset. The other 8 cancer patients’ pre-operation samples overlapped with the second set. In addition, 15 benign patients and 10 normal controls were also included in this set. Table 2 summarizes the clinical stages of cancer group in these three datasets. To assess inter- and intra-assay reproducibility, a pooled serum sample (from 5 normal sera) was processed multiple times during experiments on the second and third sample sets. The order in which samples were processed and the spotting allocation of samples in chips and bioprocessors were randomized using an in-home experiment design software.
Protein chip array analysis 7.5 Al of 9 M urea, 1% CHAPS in PBS, pH 7.4, was added to 5 Al of each serum sample. The mixture was vortexed at 4-C for 30 min and diluted 1:40 in PBS. Immobilized metal affinity capture arrays (IMAC3) were activated with 100 mM CuSO4 according to manufacturer’s instructions (Ciphergen Biosystems, CA). 50 Al of diluted samples was applied to each spot on the ProteinChip Array in a 96-well bioprocessor (Ciphergen Biosystems, CA). After binding at room temperature for 60 min on a platform shaker, the array was washed three times with 200 Al of washing buffer (100 mM Na3PO4 + 0.5 M NaCl PH7) for 5 min followed by two quick rinses with 200 Al of 1 mM HEPES (PH7.0). After air drying, 0.5 Al of saturated sinapinic acid (SPA) prepared in 50% acetonitrile, 0.5% trifluoroacetic acid was applied twice to each spot. Proteins bound to the chelated metal (through histidine, tryptophan, cysteine, or phosphorylated amino acids) were detected with the ProteinChip Reader model PBSII. Data were collected by averaging 50 laser shots at an intensity of 200 (low intensity) or 220 (high intensity) and a detector sensitivity of 7. External calibration of the instrument was performed using a set of molecular weight standards less than 200 kDa.
Bioinformatics and statistics All spectra were compiled and qualified mass peaks (S/N > 5) with M/Z between 2 K and 150 K were auto-detected. Peak clusters were completed using second pass peak selection (S/N > 2, within 0.3% mass window) and estimated peaks added. The peak intensities were normalized to the total ion current of M/Z between 2 K and 150 K. All of these were performed using
Table 2 The clinical stages of ovarian cancer patients in three datasets Cancer
Dataset 1
Dataset 2
Dataset 3
Stage 1 Stage 2 Stage 3 Stage 4 Unknown Total
7 0 9 2 3 21
7 3 12 10 0 32
10 2 5 1 0 18
ProteinChip Software 3.2 (Ciphergen Biosystems, CA). Peak amplitudes were exported and log transformed before analysis. First, datasets 1 and 2 were merged and analyzed. The low-intensity and high-intensity data were analyzed separately using the ProPeak software package (3 Z Informatics). ProPeak calculates and ranks the contribution of each individual peak toward the separation of the two diagnostic groups. This software package implements the linear version of Unified Maximum Separability Analysis (UMSA) algorithm. The UMSA algorithm uses data distribution information to identify a direction along which two pre-defined sets of data achieve maximal separation. Two ProPeak analysis modules were used for this study. In the first module, component analysis, UMSA calculates the optimal separation, ranks the peaks’ contribution toward this separation, and plots the data in a three-dimensional display. The separability is assessed visually. In the second module, Bootstrap selection, a fixed percentage of samples are left out, and the peak ranks from multiple runs are collected. In this analysis, 30 iterations were completed where 30% of the samples were left out per iteration. The relative contribution of each peak toward the separation of patient groups is reported as mean, median, and standard deviation of the rank. The top ranking peaks were selected and input back into the component analysis module. Using only the topranking peaks, the separability was visually assessed to determine if separation was retained. Once a set of peaks was selected from the first two datasets, the third dataset was used to verify potential utility of those peaks. The same peaks were collected from the third dataset. The peaks’ trend was compared between the two analyses to ensure that, for instance, any peak up-regulated in datasets 1 and 2 was up-regulated in dataset 3. We compared pre-surgery cancer with normal healthy controls and benign patients. Then we compared the pre-surgery cancer and post-chemotherapy observations. Receiver operating characteristic (ROC) curves are constructed using peaks that maintained the same trend to determine the accuracy of these peaks.
Results Datasets 1 and 2 were used to discover candidate biomarkers. We were able to achieve a reasonable level of separation between the cancer and non-cancer serum samples (benign and normal samples) using 7 protein peaks from the low-intensity spectra (M/Z value of 7 peaks: Peak 1 M7676_21, Peak 2 M11463_8, Peak 3 M11545_9, Peak 4 M11681_2, Peak 5 M11706_6, Peak 6 M13790_8, Peak 7 M15908_3). Fig. 1 showed the gel view of peaks produced by mass spectrometry. All seven peaks were statistically different as tested by two-sample t test. The separation achieved using ProPeak is presented in Fig. 2 and Table 3. Log transformation reduces the range of intensity data. As a result, the variance of the transformed peak intensity tends to be less volatile across the spectrum. The 7 peaks identified from datasets 1 and 2 were then collected from dataset 3. The same separation was observed in dataset 3, and three peaks were statistically different when cancer and non-cancer samples were compared. Peaks
250
F. Kong et al. / Gynecologic Oncology 100 (2006) 247 – 253
Fig. 1. The gel view of peaks produced by mass spectrometry. Each peak included in this paper is highlighted in the gel view. In order to fit all three datasets in one screen shot, the first 10 subjects in each of the study groups (normal, benign, and cancer) are included. The remaining subjects were deleted from view.
6 and 7 only showed clear differences between cancer and non-cancer in datasets 1 and 2, but not in dataset 3. This may due to the smaller sample size (Table 3). In addition to non-cancer samples, the third dataset included 18 ovarian cancer patients who underwent surgical removal of ovarian cancer. A serum sample was collected before and after surgery for each of these women. The 7 peaks identified from datasets 1 and 2 were also collected in these two groups (pre-surgery and post-surgery). Comparing the pre-surgery spectra with the post-surgery spectra, we
found that the separation observed was similar to the separation observed between cancer and non-cancer samples in datasets 1 and 2 (see Fig. 3). Six of the seven peaks maintained the same trend, although no statistical difference was observed between pre- and post-surgery samples. Only Peak 7 showed a different trend; it is down-regulated in the cancer group in datasets 1 and 2 but is up-regulated in the pre-surgery sample in dataset 3. Table 4 compares the peak means and standard deviations calculated from analysis of these two groups.
Fig. 2. Separation observed using 7 peaks collected using the low-intensity spectra. Red: cancer; green: non-cancer. Data were from combined set 1 and set 2.
F. Kong et al. / Gynecologic Oncology 100 (2006) 247 – 253
251
Table 3 Comparison of 7 peaks identified in datasets 1 and 2 and 3
Table 4 Comparison of 7 peaks collected from pre and post-surgery cancer patients
Candidate Analysis of datasets 1 and 2 peaks Cancer Non-cancer
Candidate peaks
Mean Peak Peak Peak Peak Peak Peak Peak
1 2 3 4 5 6 7
0.229 0.520 0.505 0.435 0.216 0.525 0.475
std 0.220 0.623 0.624 0.565 0.553 0.272 0.445
Mean 0.094 0.851 0.825 0.826 0.546 0.384 0.158
std 0.240 0.390 0.376 0.406 0.237 0.224 0.541
Analysis of dataset 3 Pre-surgery Cancer
Non-cancer
Mean
Mean
0.251 0.330 0.432 0.077 0.153 0.283 0.167
std 0.132 0.494 0.491 0.517 0.231 0.137 0.352
Analysis of dataset 3 Pre-surgery Mean
0.158 0.522 0.667 0.403 0.004 0.288 0.166
std 0.256 0.167 0.233 0.226 0.140 0.186 0.539
Means highlighted in bold are statistically different as tested by two-sample t test. Log-transformed peak amplitudes presented.
We assumed that patients undergoing chemotherapy after surgery were correctly classified as cancer free; CA125 values were significantly reduced in 14 out of 15 patients in dataset 3. ROC analysis was determined by comparing presurgery observations and the first time-point post-surgery because all patients had at least one follow-up sera collection (Fig. 4). All but one of the selected peaks had an ROC analysis area-under-curve (AUC) greater than 0.5 (corresponding to an ROC curve that is a diagonal straight line indicating no discriminatory power), indicating some level of discriminatory power. However, in contrast to the ROC result of CA125 (Fig. 5), when these peaks were analyzed individually, the differences between their AUCs and 0.5 were not statistically significant. This could be partially explained by the limited number of samples used for the analysis. Furthermore, progressively fewer patients’ samples were collected at the later time points. Therefore, we could not address classification by proteomics analysis for the remaining samples. Future studies are planned where
Peak Peak Peak Peak Peak Peak Peak
1 2 3 4 5 6 7
0.251 0.330 0.432 0.077 0.153 0.283 0.167
Post-surgery std 0.132 0.494 0.491 0.517 0.231 0.137 0.352
Mean 0.226 0.409 0.494 0.166 0.117 0.196 0.352
std 0.210 0.342 0.432 0.331 0.119 0.133 0.381
Log-transformed peak amplitudes presented.
all individuals are observed for the same number of followup visits. Discussion A fairly large number of studies have been reported on the application of SELDI technology for clinical proteomics and biomarker discovery [10,11,13,16]. However, the seemingly very high performances of some of the reports have been questioned as partially due to experiment design and other reason [17]. Care must be taken when designing and carrying out the study to ensure quality SELDI spectra are collected. Recently, multiple serum sets from a five-center casecontrol study were analyzed using SELDI, and three biomarkers were identified [13]. In independent validation to detect early-stage invasive epithelial ovarian cancer from healthy controls, the sensitivity of a multivariate model combining the three biomarkers and CA125 was higher than that of CA125 alone at a matched specificity of 97%. When compared at a fixed sensitivity of 83%, the specificity of the model was significantly better than that of CA125 alone. These results
Fig. 3. Separation observed using 7 peaks collected from dataset 3. Red: pre-surgery cancer; green: post-surgery cancer after first chemotherapy cycle.
252
F. Kong et al. / Gynecologic Oncology 100 (2006) 247 – 253
Fig. 4. ROC curve of Peaks 1 through 5 in dataset 3. Peak 1 (M7676_21), gray line, AUC = 0.577. Peak 2 (M11463_8), gray short line and dot, AUC = 0.537. Peak 3 (M11545_9), black, short line, AUC = 0.590. Peak 4 (M11545_9), black dot, AUC = 0.522. Peak 5 (M11706_6), black line, AUC = 0.454. Pre-surgery cancer and post-surgery compared.
confirmed that SELDI technique is a useful tool to identify new biomarkers for ovarian cancer. In this study, using SELDI technique and bioinformatics analysis, we were able to achieve a reasonable level of separation between the cancer samples and the non-cancer samples by selecting seven peaks from patients in the first and second dataset. Those seven peaks were then evaluated using the third dataset by comparing cancer and non-cancer patients
and cancer patients before and after surgery. Six of the seven peaks maintained the same trend. These six peaks are the candidates of potential biomarkers. Since the sample size is not big enough to address the question of efficacy of the test for screening or early detection. A large scale of sample collection and longer term of patients’ following up are necessary. Additional studies are going on to further identify and validate these biomarkers.
Fig. 5. ROC curve of CA125 in dataset 3. Pre-surgery cancer and post-surgery cancer samples compared.
F. Kong et al. / Gynecologic Oncology 100 (2006) 247 – 253
Acknowledgment This work was supported by ‘‘985’’ funding of Fudan University, China. References [1] Cannistra SA. Cancer of the ovary. N Engl J Med 2004;351:2519 – 29 [Review]. [2] Schildkraut JM, Thompson WD. Familial ovarian cancer: a population based case-control study. Am J Epidemiol 1988;128:456 – 66. [3] Ries LAG, Kosary CL, Hankey BF, Miller BA, Clegg L, Edwards BK. SEER Cancer Statistics Review, 1973B1996. Bethesda’ National Cancer Institute; 1999. [4] Bast Jr RC, Xu FJ, Yu YH, Barnhill S, Zhang Z, Mills GB. CA 125: the past and the future. Int J Biol Markers 1998;13:179 – 87. [5] Jacobs IJ, Skates S, Davies AP, Woolas RP, Jeyerajah A, Weidemann P, et al. Risk of diagnosis of ovarian cancer after raised serum CA 125 concentration: a prospective cohort study. BMJ 1996;313:1355 – 8. [6] Jacobs I, Bast Jr RC. The CA-125 tumor associated antigen: a review of the literature. Hum Reprod 1989;4:1 – 12. [7] Rustin GJ. Can we now agree to use the same definition to measure response according to CA-125? J Clin Oncol 2004;22:4035 – 6. [8] Torizuka T, Nobezawa S, Kanno T, Futatsubashi M, Yoshikawa E, Okada H, et al. Ovarian cancer recurrence: role of whole-body positron emission tomography using 2-[fluorine-18]-fluoro-2-deoxy-d-glucose. Eur J Nucl Med Mol Imaging 2002;29:797 – 803.
253
[9] Jungblut PR, Zimny-Arndt U, Zeindl-Eberhart E, Stulik J, Koupilova K, Pleissner KP, et al. Proteomics in human disease: cancer, heart and infectious diseases. Electrophoresis 1999;20:2100 – 10 [Review]. [10] Banks RE, Dunn MJ, Hochstrasser DF, Sanchez JC, Blackstock W, Pappin DJ, et al. Proteomics: new perspective, new biomedical opportunities. Lancet 2000;356:1749 – 56. [11] Petricoin EF, Ardekani AM, Hitt BA, Levine PJ, Fusaro VA, Steinberg SM, et al. Use of proteomic patterns in serum to identify ovarian cancer. Lancet 2002;359:572 – 7. [12] Espina V, Dettloff KA, Cowherd S, Petricoin III EF, Liotta LA. Use of proteomic analysis to monitor responses to biological therapies. Expert Opin Biol Ther 2004;4:83 – 93. [13] Zhang Z, Bast Jr RC, Yu Y, Li J, Sokoll LJ, Rai AJ, et al. Three biomarkers identified from serum proteomic analysis for the detection of early stage ovarian cancer. Cancer Res 2004;64:5882 – 90. [14] Ardekani AM, Liotta LA, Petricoin III EF. Clinical potential of proteomics in the diagnosis of ovarian cancer. Expert Rev Mol Diagn 2002;2:312 – 20 [Review]. [15] Banerjee H, Hawkins Z, Williams J, Blackshear M, Sawyer C, Cezares L, et al. Search for a novel biomarker for the brain cancer astrocytoma by using surface enhanced laser desorption/ionisation (SELDI) technique. Cell Mol Biol (Noisy-le-grand) 2004;50:733 – 6. [16] Ebert MP, Meuer J, Wiemer JC, Schulz HU, Reymond MA, Traugott U, et al. Identification of gastric cancer patients by serum protein profiling. J Proteome Res 2004;3:1261 – 6. [17] Baggerly KA, Morris JS, Coombes KR. Reproducibility of SELDI-TOF protein patterns in serum: comparing datasets from different experiments. Bioinformatic 2004;20:777 – 85.
Using proteomic approaches to identify new biomarkers for detection and monitoring of ovarian cancer Fanbing Kong a,1, C. Nicole White b,1, Xueyuan Xiao c,1, Youji Feng a, Congjian Xu a, Dacheng He c, Zhen Zhang b, Yinhua Yu d,* a
d
Gynecology and Obstetrics Hospital of Fudan University, Shanghai, China b Johns Hopkins Medical Institutions, Baltimore, MD 21205, USA c Beijing Normal University, Beijing, China U.T.M.D. Anderson Cancer Center, 1515 Holcombe Blvd., Box 354, Houston, TX 77030, USA Received 31 March 2005 Available online 17 October 2005
Abstract Objectives. Early detection and monitoring the treatment remain the most important factors in improving long-term survival of ovarian cancer patients. New biomarkers that individually or in combination improve the diagnostic performance of existing tumor markers are critically needed. This study uses proteomic approaches to identify new biomarkers for detection and monitoring of ovarian cancer. Methods. We analyzed protein profiles of three sets of sera using surface enhanced laser desorption and ionization time-of-light mass spectroscope (SELDI-TOF-MS) on IMAC ProteinChip arrays and ProPeak software for bioinformatics data analysis. The first set of patients included 21 ovarian cancers, 18 benign diseases, and 20 normal patients. The second set included 32 ovarian cancers, 30 benign ovarian diseases, and 30 age-matched healthy controls. The third set included samples collected before and after chemotherapy from 18 ovarian cancer patients. All samples were collected at the Gynecology and Obstetrics Hospital of Fudan University in Shanghai, China. The datasets from low-intensity and high-intensity spectra were analyzed separately. Results. Seven peaks were selected for their contribution to the separation of ovarian cancers from controls using the first and second set of samples. The same dysregulation trends were confirmed for six of the seven peaks in independent validation using the third set of samples. Conclusions. Using SELDI-TOF analysis of 195 unique specimens, we discovered with preliminary validation six distinct peaks that may potentially be useful in the detection and monitoring of ovarian cancer. Additional studies are required to determine the protein identities of these peaks and to further validate their performance as biomarkers. D 2005 Elsevier Inc. All rights reserved. Keywords: Proteomics; Biomarker; SELDI; Ovarian cancer
Introduction Ovarian cancer is the sixth most common cancer among women and the fifth most common cause of cancer death among women, in the United States, with approximately 25,580 new cases and 16,090 deaths anticipated in 2004 [1]. Ovarian cancer causes more deaths than any other cancer of the female reproductive system. Ovarian cancer occurs most frequently in women aged 50 – 79; over 70% of the cancers * Corresponding author. Fax: +1 713 745 2107. E-mail address: [email protected] (Y. Yu). 1 Contributed equally to this study. 0090-8258/$ - see front matter D 2005 Elsevier Inc. All rights reserved. doi:10.1016/j.ygyno.2005.08.051
occur after age 50. Ninety percent of women with ovarian cancer have no family history of the disease [2]. After diagnosis and standard treatment with surgery and chemotherapy, nearly 80% of women are expected to survive 1 year and 52% to survive 5 years. With early detection and prompt treatment, the overall survival rate climbs to 95%. However, early-stage diagnosis is rare. Currently, only 25% of all ovarian cancers are found at an early stage [3], suggesting the pressing need to improve the early detection of ovarian cancer. Monitoring treatment is another important factor to improve the outcome of ovarian cancer. At present, the optimal followup/monitoring strategy for the asymptomatic patient with advanced ovarian cancer after initial treatment is undefined.
248
F. Kong et al. / Gynecologic Oncology 100 (2006) 247 – 253
Two screening and monitoring tests for ovarian cancer are currently in use: the CA125 serum tumor marker and imaging using transvaginal sonography (TVS). CA125 is a highmolecular-weight glycoprotein, detectable in serum that has been used for many years to detect ovarian cancer recurrence and to monitor treatment response. CA125 has been studied more recently for use in early detection. Less then 1% of healthy post-menopausal women have a CA125 > 35 U/ml. Over 85% of all advanced ovarian cancer cases have CA125 levels exceeding this threshold. Some 50– 60% of stage I cases with disease confined to the ovary exceed this threshold [4]. CA125 may be useful as an indicator of risk, as it is over 30 U/ml in 50% of ovarian cancer patients more than 18 months prior to clinical detection and in 23% of patients more than 5 years before diagnosis [5]. However, no single screening test exists that can detect all malignant transformations of the ovary. In a pre-menopausal population, the specificity of CA125 is not adequate for early detection. Nearly 6% of pre-menopausal women without cancer have levels of CA125 exceeding 35 U/ml [6]. Many conditions, including pregnancy, uterine fibroids, inflammation or infection within the abdomen (such as colitis, acute pancreatitis, or hepatitis), or other malignancies, can cause elevations of the CA125 level. The Gynecologic Cancer Intergroup (GCIG) consists of representatives from the major gynecologic cancer trial groups around the world. After considerable debate – and after several versions – the GCIG has proposed a precise but simple definition for response according to CA125, based on a 50% reduction in CA125 levels that is confirmed and maintained for at least 28 days. This definition, with examples demonstrating its implementation, is posted on the GCIG Web site [7]. It is important to recognize the limitations of CA125 when monitoring the course of disease. Levels can be altered dramatically by abdominal surgery or peritonitis. There is also the possibility of laboratory error and considerable variation in results among laboratories. Torizuka et al. tested the value of whole-body positron emission tomography (PET) using 2-[fluorine-18]-fluoro-2-deoxy-d-glucose (FDG) for the diagnosis of recurrent ovarian cancer. They demonstrated that FDG-PET may be accurate and useful for the detection of tumor recurrence when conventional imaging is inconclusive or negative, especially in patients with abnormal CA125 levels [8]. In order to identify more effective biomarkers to detect early-stage ovarian cancer patients and monitor biological responses to therapy, proteomic-based approaches have been extended to medical application and diagnosis [9,10]. Recently, several reports demonstrated that proteomic approaches could improve our ability to detect and monitor ovarian cancer [11]. Proteomic expression patterns can be potentially evaluated as a new way to track biological responses to therapy, even become the starting point for individualized therapy [12]. Our recent work has also identified three new biomarkers for ovarian cancer using a proteomics approach [13]. In the early stage of ovarian cancer, some low-molecularweight serum protein profiling might reflect the pathologic state of patient. Screening these proteins may be very useful for distinguishing early-stage ovarian cancer and monitoring
biological responses to therapy. Matrix-assisted laser desorption and ionization time-of-light (MALD-TOF) and surface enhanced laser desorption and ionization time-of-flight (SELDTOF) mass spectroscope can profile proteins in this range [14,15]. With the help of bioinformatics analytical tools, we will be able to identify significant changes in complex mass spectra in a high-throughput fashion. SELDI-TOF technology has been evaluated as a new means to track biological responses to therapy. The measurement of a protein profile at different stages of disease progression or before and after treatment, combined with protein microarray technologies, constitutes a new paradigm for detecting disease and monitoring disease response to therapy [12]. Ultimately, proteomics and genomics will become integrated into cancer patient management through the design and tracking of individualized therapy. In this study, we use SELDI-TOF to analyze three sets of sera. First, we generated comparative protein profiles from biopsy-confirmed ovarian cancer, benign ovarian tumors, and healthy controls. Then we use the protein profiles from the first two sets to compare the changes before and after chemotherapy in the third set of cancer patients’ samples. This study may provide important information in detecting and monitoring ovarian cancer. Materials and methods Patients and clinical samples Serum samples were collected at the Gynecology and Obstetrics Hospital, Fudan University in Shanghai, China, from October 2002 to November 2003. Informed consents were obtained from all patients and healthy controls. According to a pre-defined protocol, sera were collected in the morning. Samples of patients were collected before operation or 1 to 2 months (average 32 days) after chemotherapy. Sera were kept at room temperature for 1 to 3 h for clotting before being centrifuged at 1000 rpm and again at 12000 rpm at 4-C. Samples were immediately aliquoted and stored at 70-C. All sera were allowed to thaw once only. All samples had CA125 value by CA125 EIA test (Canag Company, Sweden). All ovarian cancer and benign tumor patients had received ultrasound examination. A total of 195 unique specimens were included in this study. Diagnoses were pathologically confirmed. A total of three sets of spectra were analyzed. While there is some overlap between the sample sets, new spectra from overlapping samples were collected with each dataset (see Table 1). A total of 58 ovarian cancer patients, 48 ovarian benign tumor patients, and 45 healthy controls are included in this research study. All normal controls were without evidence of gynecological diseases at least 3 years before sera collection and were healthy more then 2 years after collection. In the first dataset, 20 healthy women between the ages of 21 and 56 years (mean T SD: 39 T 11) were included in the normal control group. 18 women between the ages of 18 and 72 years (44 T 14) were included with benign disease. 21 women with ovarian cancer between the ages of 18 and 79 years (54 T 16) were included as well. The normal and benign groups were younger then the cancer group in the first set. We had therefore also collected a second serum set of agematched samples. The second set included 30 health women aged 27 to 75 years (43 T 10), 30 women with benign ovarian tumor aged from 21 to 73 years (age 45 T 9), and finally 32 women aged 35 – 72 years (50 T 10) with malignant ovarian cancer. All benign tumor patients and cancer patients received open surgical or laparoscopic treatment. There are a number of samples in the second set overlap with those in the first set. The second set had in total 57 new samples (27 cancer patients, 15 benign patients, and 15 healthy controls).
F. Kong et al. / Gynecologic Oncology 100 (2006) 247 – 253
249
Table 1 Sample size of three datasets Disease status
Cancer Benign Normal
Dataset 1
Dataset 2
Dataset 3
Overlapping samples
Unique observations
Unique observations
Unique observations
Overlap between datasets 1 and 2
Overlap between datasets 1 and 3
Overlap between datasets 2 and 3
21 18 20
27 15 15
10 15 10
5 15 15
0 0 0
8 0 0
Unique and overlapping observations presented. In the third dataset, 18 patients aged 41 – 73 (54 T 9) with ovarian cancer were enrolled in ovarian cancer group. All these 18 cancer patients had completed surgery, remained tumor size < 2 cm. Sera samples were collected before operation and after regular chemotherapy (DDP + ADM + CTX). 3 patients had one cycle of chemotherapy, 5 patients had two cycles, 9 had three cycles, and 1 patient had four cycles. A total of 10 cancer patients were unique to this dataset. The other 8 cancer patients’ pre-operation samples overlapped with the second set. In addition, 15 benign patients and 10 normal controls were also included in this set. Table 2 summarizes the clinical stages of cancer group in these three datasets. To assess inter- and intra-assay reproducibility, a pooled serum sample (from 5 normal sera) was processed multiple times during experiments on the second and third sample sets. The order in which samples were processed and the spotting allocation of samples in chips and bioprocessors were randomized using an in-home experiment design software.
Protein chip array analysis 7.5 Al of 9 M urea, 1% CHAPS in PBS, pH 7.4, was added to 5 Al of each serum sample. The mixture was vortexed at 4-C for 30 min and diluted 1:40 in PBS. Immobilized metal affinity capture arrays (IMAC3) were activated with 100 mM CuSO4 according to manufacturer’s instructions (Ciphergen Biosystems, CA). 50 Al of diluted samples was applied to each spot on the ProteinChip Array in a 96-well bioprocessor (Ciphergen Biosystems, CA). After binding at room temperature for 60 min on a platform shaker, the array was washed three times with 200 Al of washing buffer (100 mM Na3PO4 + 0.5 M NaCl PH7) for 5 min followed by two quick rinses with 200 Al of 1 mM HEPES (PH7.0). After air drying, 0.5 Al of saturated sinapinic acid (SPA) prepared in 50% acetonitrile, 0.5% trifluoroacetic acid was applied twice to each spot. Proteins bound to the chelated metal (through histidine, tryptophan, cysteine, or phosphorylated amino acids) were detected with the ProteinChip Reader model PBSII. Data were collected by averaging 50 laser shots at an intensity of 200 (low intensity) or 220 (high intensity) and a detector sensitivity of 7. External calibration of the instrument was performed using a set of molecular weight standards less than 200 kDa.
Bioinformatics and statistics All spectra were compiled and qualified mass peaks (S/N > 5) with M/Z between 2 K and 150 K were auto-detected. Peak clusters were completed using second pass peak selection (S/N > 2, within 0.3% mass window) and estimated peaks added. The peak intensities were normalized to the total ion current of M/Z between 2 K and 150 K. All of these were performed using
Table 2 The clinical stages of ovarian cancer patients in three datasets Cancer
Dataset 1
Dataset 2
Dataset 3
Stage 1 Stage 2 Stage 3 Stage 4 Unknown Total
7 0 9 2 3 21
7 3 12 10 0 32
10 2 5 1 0 18
ProteinChip Software 3.2 (Ciphergen Biosystems, CA). Peak amplitudes were exported and log transformed before analysis. First, datasets 1 and 2 were merged and analyzed. The low-intensity and high-intensity data were analyzed separately using the ProPeak software package (3 Z Informatics). ProPeak calculates and ranks the contribution of each individual peak toward the separation of the two diagnostic groups. This software package implements the linear version of Unified Maximum Separability Analysis (UMSA) algorithm. The UMSA algorithm uses data distribution information to identify a direction along which two pre-defined sets of data achieve maximal separation. Two ProPeak analysis modules were used for this study. In the first module, component analysis, UMSA calculates the optimal separation, ranks the peaks’ contribution toward this separation, and plots the data in a three-dimensional display. The separability is assessed visually. In the second module, Bootstrap selection, a fixed percentage of samples are left out, and the peak ranks from multiple runs are collected. In this analysis, 30 iterations were completed where 30% of the samples were left out per iteration. The relative contribution of each peak toward the separation of patient groups is reported as mean, median, and standard deviation of the rank. The top ranking peaks were selected and input back into the component analysis module. Using only the topranking peaks, the separability was visually assessed to determine if separation was retained. Once a set of peaks was selected from the first two datasets, the third dataset was used to verify potential utility of those peaks. The same peaks were collected from the third dataset. The peaks’ trend was compared between the two analyses to ensure that, for instance, any peak up-regulated in datasets 1 and 2 was up-regulated in dataset 3. We compared pre-surgery cancer with normal healthy controls and benign patients. Then we compared the pre-surgery cancer and post-chemotherapy observations. Receiver operating characteristic (ROC) curves are constructed using peaks that maintained the same trend to determine the accuracy of these peaks.
Results Datasets 1 and 2 were used to discover candidate biomarkers. We were able to achieve a reasonable level of separation between the cancer and non-cancer serum samples (benign and normal samples) using 7 protein peaks from the low-intensity spectra (M/Z value of 7 peaks: Peak 1 M7676_21, Peak 2 M11463_8, Peak 3 M11545_9, Peak 4 M11681_2, Peak 5 M11706_6, Peak 6 M13790_8, Peak 7 M15908_3). Fig. 1 showed the gel view of peaks produced by mass spectrometry. All seven peaks were statistically different as tested by two-sample t test. The separation achieved using ProPeak is presented in Fig. 2 and Table 3. Log transformation reduces the range of intensity data. As a result, the variance of the transformed peak intensity tends to be less volatile across the spectrum. The 7 peaks identified from datasets 1 and 2 were then collected from dataset 3. The same separation was observed in dataset 3, and three peaks were statistically different when cancer and non-cancer samples were compared. Peaks
250
F. Kong et al. / Gynecologic Oncology 100 (2006) 247 – 253
Fig. 1. The gel view of peaks produced by mass spectrometry. Each peak included in this paper is highlighted in the gel view. In order to fit all three datasets in one screen shot, the first 10 subjects in each of the study groups (normal, benign, and cancer) are included. The remaining subjects were deleted from view.
6 and 7 only showed clear differences between cancer and non-cancer in datasets 1 and 2, but not in dataset 3. This may due to the smaller sample size (Table 3). In addition to non-cancer samples, the third dataset included 18 ovarian cancer patients who underwent surgical removal of ovarian cancer. A serum sample was collected before and after surgery for each of these women. The 7 peaks identified from datasets 1 and 2 were also collected in these two groups (pre-surgery and post-surgery). Comparing the pre-surgery spectra with the post-surgery spectra, we
found that the separation observed was similar to the separation observed between cancer and non-cancer samples in datasets 1 and 2 (see Fig. 3). Six of the seven peaks maintained the same trend, although no statistical difference was observed between pre- and post-surgery samples. Only Peak 7 showed a different trend; it is down-regulated in the cancer group in datasets 1 and 2 but is up-regulated in the pre-surgery sample in dataset 3. Table 4 compares the peak means and standard deviations calculated from analysis of these two groups.
Fig. 2. Separation observed using 7 peaks collected using the low-intensity spectra. Red: cancer; green: non-cancer. Data were from combined set 1 and set 2.
F. Kong et al. / Gynecologic Oncology 100 (2006) 247 – 253
251
Table 3 Comparison of 7 peaks identified in datasets 1 and 2 and 3
Table 4 Comparison of 7 peaks collected from pre and post-surgery cancer patients
Candidate Analysis of datasets 1 and 2 peaks Cancer Non-cancer
Candidate peaks
Mean Peak Peak Peak Peak Peak Peak Peak
1 2 3 4 5 6 7
0.229 0.520 0.505 0.435 0.216 0.525 0.475
std 0.220 0.623 0.624 0.565 0.553 0.272 0.445
Mean 0.094 0.851 0.825 0.826 0.546 0.384 0.158
std 0.240 0.390 0.376 0.406 0.237 0.224 0.541
Analysis of dataset 3 Pre-surgery Cancer
Non-cancer
Mean
Mean
0.251 0.330 0.432 0.077 0.153 0.283 0.167
std 0.132 0.494 0.491 0.517 0.231 0.137 0.352
Analysis of dataset 3 Pre-surgery Mean
0.158 0.522 0.667 0.403 0.004 0.288 0.166
std 0.256 0.167 0.233 0.226 0.140 0.186 0.539
Means highlighted in bold are statistically different as tested by two-sample t test. Log-transformed peak amplitudes presented.
We assumed that patients undergoing chemotherapy after surgery were correctly classified as cancer free; CA125 values were significantly reduced in 14 out of 15 patients in dataset 3. ROC analysis was determined by comparing presurgery observations and the first time-point post-surgery because all patients had at least one follow-up sera collection (Fig. 4). All but one of the selected peaks had an ROC analysis area-under-curve (AUC) greater than 0.5 (corresponding to an ROC curve that is a diagonal straight line indicating no discriminatory power), indicating some level of discriminatory power. However, in contrast to the ROC result of CA125 (Fig. 5), when these peaks were analyzed individually, the differences between their AUCs and 0.5 were not statistically significant. This could be partially explained by the limited number of samples used for the analysis. Furthermore, progressively fewer patients’ samples were collected at the later time points. Therefore, we could not address classification by proteomics analysis for the remaining samples. Future studies are planned where
Peak Peak Peak Peak Peak Peak Peak
1 2 3 4 5 6 7
0.251 0.330 0.432 0.077 0.153 0.283 0.167
Post-surgery std 0.132 0.494 0.491 0.517 0.231 0.137 0.352
Mean 0.226 0.409 0.494 0.166 0.117 0.196 0.352
std 0.210 0.342 0.432 0.331 0.119 0.133 0.381
Log-transformed peak amplitudes presented.
all individuals are observed for the same number of followup visits. Discussion A fairly large number of studies have been reported on the application of SELDI technology for clinical proteomics and biomarker discovery [10,11,13,16]. However, the seemingly very high performances of some of the reports have been questioned as partially due to experiment design and other reason [17]. Care must be taken when designing and carrying out the study to ensure quality SELDI spectra are collected. Recently, multiple serum sets from a five-center casecontrol study were analyzed using SELDI, and three biomarkers were identified [13]. In independent validation to detect early-stage invasive epithelial ovarian cancer from healthy controls, the sensitivity of a multivariate model combining the three biomarkers and CA125 was higher than that of CA125 alone at a matched specificity of 97%. When compared at a fixed sensitivity of 83%, the specificity of the model was significantly better than that of CA125 alone. These results
Fig. 3. Separation observed using 7 peaks collected from dataset 3. Red: pre-surgery cancer; green: post-surgery cancer after first chemotherapy cycle.
252
F. Kong et al. / Gynecologic Oncology 100 (2006) 247 – 253
Fig. 4. ROC curve of Peaks 1 through 5 in dataset 3. Peak 1 (M7676_21), gray line, AUC = 0.577. Peak 2 (M11463_8), gray short line and dot, AUC = 0.537. Peak 3 (M11545_9), black, short line, AUC = 0.590. Peak 4 (M11545_9), black dot, AUC = 0.522. Peak 5 (M11706_6), black line, AUC = 0.454. Pre-surgery cancer and post-surgery compared.
confirmed that SELDI technique is a useful tool to identify new biomarkers for ovarian cancer. In this study, using SELDI technique and bioinformatics analysis, we were able to achieve a reasonable level of separation between the cancer samples and the non-cancer samples by selecting seven peaks from patients in the first and second dataset. Those seven peaks were then evaluated using the third dataset by comparing cancer and non-cancer patients
and cancer patients before and after surgery. Six of the seven peaks maintained the same trend. These six peaks are the candidates of potential biomarkers. Since the sample size is not big enough to address the question of efficacy of the test for screening or early detection. A large scale of sample collection and longer term of patients’ following up are necessary. Additional studies are going on to further identify and validate these biomarkers.
Fig. 5. ROC curve of CA125 in dataset 3. Pre-surgery cancer and post-surgery cancer samples compared.
F. Kong et al. / Gynecologic Oncology 100 (2006) 247 – 253
Acknowledgment This work was supported by ‘‘985’’ funding of Fudan University, China. References [1] Cannistra SA. Cancer of the ovary. N Engl J Med 2004;351:2519 – 29 [Review]. [2] Schildkraut JM, Thompson WD. Familial ovarian cancer: a population based case-control study. Am J Epidemiol 1988;128:456 – 66. [3] Ries LAG, Kosary CL, Hankey BF, Miller BA, Clegg L, Edwards BK. SEER Cancer Statistics Review, 1973B1996. Bethesda’ National Cancer Institute; 1999. [4] Bast Jr RC, Xu FJ, Yu YH, Barnhill S, Zhang Z, Mills GB. CA 125: the past and the future. Int J Biol Markers 1998;13:179 – 87. [5] Jacobs IJ, Skates S, Davies AP, Woolas RP, Jeyerajah A, Weidemann P, et al. Risk of diagnosis of ovarian cancer after raised serum CA 125 concentration: a prospective cohort study. BMJ 1996;313:1355 – 8. [6] Jacobs I, Bast Jr RC. The CA-125 tumor associated antigen: a review of the literature. Hum Reprod 1989;4:1 – 12. [7] Rustin GJ. Can we now agree to use the same definition to measure response according to CA-125? J Clin Oncol 2004;22:4035 – 6. [8] Torizuka T, Nobezawa S, Kanno T, Futatsubashi M, Yoshikawa E, Okada H, et al. Ovarian cancer recurrence: role of whole-body positron emission tomography using 2-[fluorine-18]-fluoro-2-deoxy-d-glucose. Eur J Nucl Med Mol Imaging 2002;29:797 – 803.
253
[9] Jungblut PR, Zimny-Arndt U, Zeindl-Eberhart E, Stulik J, Koupilova K, Pleissner KP, et al. Proteomics in human disease: cancer, heart and infectious diseases. Electrophoresis 1999;20:2100 – 10 [Review]. [10] Banks RE, Dunn MJ, Hochstrasser DF, Sanchez JC, Blackstock W, Pappin DJ, et al. Proteomics: new perspective, new biomedical opportunities. Lancet 2000;356:1749 – 56. [11] Petricoin EF, Ardekani AM, Hitt BA, Levine PJ, Fusaro VA, Steinberg SM, et al. Use of proteomic patterns in serum to identify ovarian cancer. Lancet 2002;359:572 – 7. [12] Espina V, Dettloff KA, Cowherd S, Petricoin III EF, Liotta LA. Use of proteomic analysis to monitor responses to biological therapies. Expert Opin Biol Ther 2004;4:83 – 93. [13] Zhang Z, Bast Jr RC, Yu Y, Li J, Sokoll LJ, Rai AJ, et al. Three biomarkers identified from serum proteomic analysis for the detection of early stage ovarian cancer. Cancer Res 2004;64:5882 – 90. [14] Ardekani AM, Liotta LA, Petricoin III EF. Clinical potential of proteomics in the diagnosis of ovarian cancer. Expert Rev Mol Diagn 2002;2:312 – 20 [Review]. [15] Banerjee H, Hawkins Z, Williams J, Blackshear M, Sawyer C, Cezares L, et al. Search for a novel biomarker for the brain cancer astrocytoma by using surface enhanced laser desorption/ionisation (SELDI) technique. Cell Mol Biol (Noisy-le-grand) 2004;50:733 – 6. [16] Ebert MP, Meuer J, Wiemer JC, Schulz HU, Reymond MA, Traugott U, et al. Identification of gastric cancer patients by serum protein profiling. J Proteome Res 2004;3:1261 – 6. [17] Baggerly KA, Morris JS, Coombes KR. Reproducibility of SELDI-TOF protein patterns in serum: comparing datasets from different experiments. Bioinformatic 2004;20:777 – 85.