Editorial Consortium. Proteomic Approaches to Tumor Marker Discovery. Identification of Biomarkers for Ovarian Cancer

Transcription

1 Editorial Consortium Proteomic Approaches to Tumor Marker Discovery Identification of Biomarkers for Ovarian Cancer Alex J. Rai, PhD; Zhen Zhang, PhD; Jason Rosenzweig, BS; Ie-ming Shih, MD, PhD; Thang Pham, PhD; Eric T. Fung, MD, PhD; Lori J. Sokoll, PhD; Daniel W. Chan, PhD Context. Current tumor markers for ovarian cancer still lack adequate sensitivity and specificity to be applicable in large populations. High-throughput proteomic profiling and bioinformatics tools allow for the rapid screening of a large number of potential biomarkers in serum, plasma, or other body fluids. Objective. To determine whether protein profiles of plasma can be used to identify potential biomarkers that improve the detection of ovarian cancer. Design. We analyzed plasma samples that had been collected between 1998 and 2001 from patients with sporadic ovarian serous neoplasms before tumor resection at various International Federation of Gynecology and Obstetrics stages (stage I [n 11], stage II [n 3], and stage III [n 29]) and from women without known neoplastic disease (n 38) using proteomic profiling and bioinformatics. We compared results between the patients with and without cancer and evaluated their discriminatory performance against that of the cancer antigen 125 (CA125) tumor marker. Results. We selected 7 biomarkers based on their collective contribution to the separation of the 2 patient groups. Among them, we further purified and subsequently identified 3 biomarkers. Individually, the biomarkers did not perform better than CA125. However, a combination of 4 of the biomarkers significantly improved performance (P.001). The new biomarkers were complementary to CA125. At a fixed specificity of 94%, an index combining 2 of the biomarkers and CA125 achieves a sensitivity of 94% (95% confidence interval, 85% 100.0%) in contrast to a sensitivity of 81% (95% confidence interval, 68% 95%) for CA125 alone. Conclusions. The combined use of bioinformatics tools and proteomic profiling provides an effective approach to screen for potential tumor markers. Comparison of plasma profiles from patients with and without known ovarian cancer uncovered a panel of potential biomarkers for detection of ovarian cancer with discriminatory power complementary to that of CA125. Additional studies are required to further validate these biomarkers. (Arch Pathol Lab Med. 2002;126: ) The identification of tumor markers suitable for the early detection and diagnosis of cancer holds great promise to improve the clinical outcome of patients. It is especially important for patients presenting with vague or no symptoms or with tumors that are relatively inaccessible to physical examination. Ovarian carcinoma represents one of such insidious and aggressive cancers. It is the most lethal gynecologic malignancy in women, 1,2 with new cases and deaths expected in Despite considerable effort directed at early detection, no cost-effective screening tests have been developed, 4 and women generally present with disseminated disease at diagnosis. 5 Cancer antigen 125 (CA125) is the best-characterized serologic tumor marker for advanced epithelial ovarian cancers. However, its use as a population-based screening tool for early detection and diagnosis of ovarian cancer is hin- Accepted for publication July 17, From the Department of Pathology, the Johns Hopkins University School of Medicine, Baltimore, Md (Drs Rai, Zhang, Shih, Sokoll, and Chan, and Mr Rosenzweig), and Ciphergen Biosystems, Fremont, Calif (Drs Pham and Fung). Reprints: Daniel W. Chan, PhD, Department of Pathology, Division of Clinical Chemistry, the Johns Hopkins University School of Medicine, 600 N Wolfe St, Meyer B121, Baltimore, MD ( dchan@jhmi.edu). dered by its low sensitivity and specificity. 6 8 Although pelvic and more recently vaginal sonography has been used to screen high-risk patients, neither technique has the sufficient sensitivity and specificity to be applied to the general population. 6 Recent efforts in using CA125 in combination with additional tumor markers, 9 12 in a longitudinal risk of cancer model, 13 and in tandem with ultrasound as a second-line test 14,15 have shown promising results in improving overall test specificity, which is critical for diseases such as ovarian cancer that have a relatively low prevalence. However, it is still well recognized that there is a critical need for new serologic tumor markers that individually or in combination with other markers or diagnostic modalities deliver the required sensitivity and specificity for early detection of ovarian cancer. 16 Recently, there has been an explosion of interest in exploring the genome and proteome for biomarkers that might provide a better understanding of the molecular basis of cancer and serve as targets for further development in molecular diagnostics and therapeutics. This newly spawned interest and the increasing availability of highthroughput instruments for genomic and proteomic profiling will greatly accelerate the discovery of potential tumor markers in body fluids, compared with the progress over the last 40 years The ProteinChip Biomarker Sys Arch Pathol Lab Med Vol 126, December 2002 Biomarkers for Ovarian Cancer Rai et al

2 tem is one such high-throughput instrument for protein expression profiling of clinical specimens such as serum or plasma. Briefly, this system uses chromatographic ProteinChip Arrays to assay the samples using SELDI (surface enhanced laser desorption/ionization). Proteins bound to the arrays are read in a ProteinChip Reader, a time-of-flight mass spectrometer. This system has been successfully used to profile specimens from prostate cancer, transitional cell carcinoma of the bladder, and cervical cancer A more thorough review of this technology and clinical applications can be found in the existing literature The identification of a small subset of differentially expressed genes or proteins from a large volume of profiling data is a challenging task. Sophisticated bioinformatics tools are required to evaluate the complex associations between genes or proteins and the phenotypes of specimens. In research to discover tumor markers, the phenotypes of specimens are often known beforehand. In such cases, supervised analytic methods that take into consideration disease classification of specimens tend to be more efficient in information use and relevant to the desired research end point of interest. A basic approach would be to first construct linear or nonlinear classification functions that best separate the predefined groups of specimens. The relative importance of individual markers would then be estimated and ranked according to the way in which they contributed to the classification functions. In this study, we used the ProteinChip Biomarker System to generate comparative protein profiles from patients diagnosed with ovarian carcinoma and patients without known neoplastic diseases. A subset of biomarkers was selected based on collaborative results from 2 supervised analytic methods. We were able to purify and subsequently identify 3 of these biomarkers. The selected biomarkers, together with the tumor marker CA125, were evaluated individually and in combination through multivariate logistic regression. We show that high-throughput protein profiling combined with effective use of bioinformatics tools offers a viable approach to screening for tumor markers. The new biomarkers as a panel had significant power for discriminating between patients with and without ovarian cancer in this study and are complementary to CA125. Additional study will be needed to further validate these markers for their potential in the detection of ovarian cancer. MATERIALS AND METHODS Samples A total of 81 specimens were collected for this study. Blood samples were collected at the Johns Hopkins Hospital between 1998 and 2001 from 43 patients with sporadic ovarian carcinomas before tumor resection. These patients included 11 patients with International Federation of Gynecology and Obstetrics (FIGO) stage I disease, 3 patients with FIGO stage II disease, and 29 patients with FIGO stage III disease. The median age of these patients was 53 years (range, years). We used specimens from 38 women with nongynecologic diseases but without known neoplastic disease as controls. The median age of the control patients was 57 years (range, years). Specimens were collected in Vacutainer tubes containing EDTA and were centrifuged at 1500g for 20 minutes; plasma samples were harvested to avoid leukocyte contamination. Specimens obtained before 2000 were analyzed for CA125II using Centocor CA125II assays (Fujirebio Diagnostics, Malvern, Pa). For the remaining specimens, CA125 levels were measured in either serum or EDTAanticoagulated plasma using the Tosoh AIA-PACK CA125 assay on the 600 II analyzer (Tosoh Medics, South San Francisco, Calif). The Centocor CA125II assay is equivalent to the Tosoh CA125 assay (D.W.C., oral communication, October 2001). The Tosoh CA125 assay is approved for use in serum; however, the assay was validated for plasma in house, and the results for serum and plasma were determined to be equivalent. Results were available for 68 of the 80 total specimens used in this study (1 sample from stage III disease was not repeated due to insufficient quantity). The median, mean, and SD values for CA125 concentration for the cancer group (n 32) were 58, 174.8, and U/mL, respectively, and for the control group (n 36) were 7.6, 7.8, and 8.9 U/mL, respectively. Among the total 80 plasma samples, we initially analyzed samples from a group of 67 patients (29 with ovarian cancer and 38 control cases) for biomarker selection and identification. We then repeated the analysis on the entire collection of 80 specimens to include more patients with early-stage cancer. We performed statistical analysis of biomarker performance based on the entire 80 patients. ProteinChip Analysis Fifteen microliters of each plasma sample was diluted into 25 L of 9M urea, 2% CHAPS (3-[(3-cholamidopropyl)dimethylammonio]- 1-propane-sulfonate), and 50mM Tris-HCl ph 9.0. Each sample was then diluted 1:40 in phosphate-buffered saline (PBS) ph 7.4 for use with immobilized metal affinity capture type 3 (IMAC3) 8-spot arrays. The IMAC3 ProteinChips were pretreated with nickel sulfate per the manufacturer s instructions. Each array was then prewashed in PBS ph 7.4 on a bioprocessor. Fifty microliters of each sample was applied to each spot and incubated on a shaker for 40 minutes at room temperature. Samples were washed using 100 L ofpbs ph 7.4, repeated twice, followed by 2 quick rinses in distilled water. After the spots were air dried, sinapinic acid, prepared per the manufacturer s instructions, was applied to each spot. The arrays were analyzed on a PBS-II mass reader (Ciphergen Biosystems, Fremont, Calif) using SELDI 2.1b software (Ciphergen). Data were collected by averaging 60 laser shots with an intensity of 240 and a detector sensitivity of 8. Bioinformatics and Statistics We used the Ciphergen ProteinChip software system to identify qualified peaks from the raw spectrum data by applying a threshold to peak intensities that had been normalized against total ion current. Since more-sophisticated procedures were used for the final peak selection, we set the initial threshold to capture the largest number of candidate peaks. We applied logarithmic transformation to the data when needed to reduce peak intensity ranges. The final result was an m (peaks) by n (specimens) matrix, where an entry at row i, column j presented the normalized relative abundance of proteins at mass weight corresponding to peak i in specimen j. We used 2 supervised pattern classification methods the classification and regression tree (CART), 29 implemented in Biomarker Patterns Software (BPS) version 4.0 (Ciphergen), and the unified maximum separability analysis (UMSA) procedure, 30 implemented in ProPeak (3Z Informatics, Mt Pleasant, SC) individually and in cross-comparison to screen for peaks that most contributed to discrimination between the patients with ovarian cancer and the cancer-free controls. The CART procedure constructs a binary decision tree that recursively partitions a given data set into blocks of predicted positive and negative samples. The procedure minimizes a cost function that balances prediction errors and the total number of markers used. The relative importance of a peak is measured by the order in which it was selected in the decision tree and the number of correct predictions it is credited for. Support vector machine (SVM) 31 has been applied to a number of applications for processing biological expression data. 32 The UMSA procedure modifies the SVM learning algorithm to allow for the incorporation of data distribution information. For data sets with a small sample size relative to the number of variables, Arch Pathol Lab Med Vol 126, December 2002 Biomarkers for Ovarian Cancer Rai et al 1519

3 Figure 1. Representative spectrum obtained from surface enhanced laser desorption/ionization analysis (relative intensity vs m/z). Plasma sample was run on an immobilized nickel (IMAC-Ni) ProteinChip array. Upper panel shows a portion of the protein profile in spectrum view. Lower panel is same profile shown in pseudogel view. UMSA tends to be less sensitive than the typical SVM to possible labeling errors in data, such as those resulting from specimen contamination or misdiagnosis. Currently, ProPeak offers 2 analytic modules. The first is a UMSA component analysis module, which projects the original specimen as individual points into a 3-dimensional component space. The components (axes) are linear combinations of the original spectrum peaks determined so that 2 prespecified groups of data achieve maximum separability. The results can then be viewed in an interactive 3-dimensional display. The second module in ProPeak uses a backward stepwise process to compute a significance score to rank individual markers according to their collective contribution to the separation of the 2 groups of specimens under UMSA. We used logistic regression analysis to evaluate the peaks selected by BPS and UMSA analysis, both individually and in combinations of multiple peaks, for their diagnostic performance. We assessed diagnostic performance by estimating sensitivity and specificity, and by determining the area under receiver-operating characteristic (ROC) curves. For specimens with available CA125 values, we also compared the results with the diagnostic performance of CA125. Biomarker Identification Based on the relative expression levels of the candidate biomarkers of interest within the plasma samples, we chose a subset of samples to be used in protein purification. Plasma samples (27 L each) were first buffer exchanged into 20mM Tris-HCl ph 9.0 buffer using K-30 size-selection spin columns (Ciphergen) equilibrated with the same buffer. Proteins were then fractionated on anion exchange spin columns based on their isoelectric point. Each sample was applied to a spin microcolumn containing 100 L of Q HyperD anion exchanger resin (BioSepra, Fremont, Calif), equilibrated in 20mM Tris-HCl ph 9.0 buffer. After binding occurred, the flow-through fraction was collected. Subsequent fractions were collected using 100 L of ph 9.0 buffer, a series of 20mM Tris-HCl buffers with decreasing ph 8, and 20mM phosphate/citrate combination buffers of ph 7.0, 6.0, 5.0, 4.0, and 3.0. Finally, columns were washed in an organic buffer containing 16.7% isopropanol, 33.3% acetonitrile, and 0.1% trifluoroacetic acid, to remove the remaining proteins. Fractionation was monitored on both NP (normal phase) and IMAC-Ni (immobilized nickel) arrays. An aliquot of 1 L (of 120 L total) of each fraction was applied to each spot on the NP array, and 2 L was used for each spot on the IMAC-Ni array. We used the ProteinChip reader (PBS II, Ciphergen Biosystems) to detect proteins in each spot of the array through automatic data acquisition mode at fixed laser intensity. The mass spectrometric profiles (intensity vs m/z) of all plasma samples were compared to identify fractions containing the biomarkers of interest, as well as the purity of each biomarker. After we identified the fractions of interest, we separated the samples by sodium dodecyl sulfate polyacrylamide gel electrophoresis. A 16% acrylamide Tris-glycine gel (Invitrogen/Novex, Carlsbad, Calif) was used to isolate the 7- to 12- kd proteins; a 4% to 20% acrylamide Tris-glycine gel was used for the 15- to 50-kd proteins; and a 6% acrylamide Tris-glycine gel was used for the 52- to 80-kd proteins. Gels were stained with colloidal blue (Invitrogen/Novex) and destained with deionized water. By correlating the mass spectra and Coomassie blue stained protein bands for high- and low-abundance proteins, we were able to identify the particular protein bands of interest. We subsequently isolated these bands using a disposable Pasteur pipette. The gel slices were destained, and then the purified proteins in the gel slices were digested with 10 L of 0.02 g/ L(1 M) modified trypsin in 25mM ammonium bicarbonate ph 8.0 buffer. Peptides generated by in-gel tryptic digestion were profiled using NP and H4 (hydrophobic) arrays. An aliquot (1 2 L) of each digest was applied to each spot on the array, and the proteins were allowed to concentrate to dryness before 0.5 L of 20% saturated cyano-4-hydroxycinnamic acid in 50% acetonitrile, 0.5% trifluroacetic acid solution was applied to each spot. After the arrays were completely dry, peptides were mapped with the ProteinChip reader (PBS II). Peptide standards were used to internally calibrate the mass spectrometry spectra for accurate peptide mass determination, and those obtained from control samples (trypsin incubated with blank gel plugs) were subtracted from the peptide maps. Subsequently, peptide masses were used for database searching and protein identification using ProFound (Rockefeller University, New York, NY) and MASCOT (MatrixScience, We further confirmed protein identity by sequencing selected peptides from the tryptic digest using a ProteinChip interface PCI-1000 (Ciphergen) coupled to a Q-TOF II MS/MS (MicroMass, Manchester, United Kingdom). RESULTS We obtained mass spectra for the initial group of 67 patients (29 with cancer, 38 without cancer) from SELDI analysis using IMAC-Ni ProteinChips. Figure 1 shows a representative view of the spectra showing proteins retained on the chip, in both spectrum and pseudogel view. We analyzed spectra of the 67 samples using 2 bioinformatics software packages, BPS and ProPeak. We compared the results to select a subset of peaks that possessed the maximum discriminatory power. Using the UMSA component analysis module in ProPeak, we were able to project the patient data onto a 3-dimensional space in which the patients with and without cancer were best separated (Figure 2, A). Subsequently, using the backward stepwise peak selection module, we selected 7 peaks (8.6, 9.2, 19.8, 39.8, 54, 60, and 79 kd), for further analysis. Among them, peaks at 9.2, 19.8, and 60 kd showed higher expression levels on average among the specimens from 1520 Arch Pathol Lab Med Vol 126, December 2002 Biomarkers for Ovarian Cancer Rai et al

4 Figure 2. ProPeak analysis of 67 samples. The unified maximum separability analysis component analysis module of ProPeak was used to project 67 samples onto a 3-dimensional space (control: green; cancer: red). A, Projection using all peaks. B, Projection using only 7 selected peaks. Figure 3. Biomarker Patterns Software analysis of 67 samples. A, Tree diagram shows that the 2 peaks can be used to separate the patient data into control and cancer groups. Green squares indicate decision nodes, whereas terminal nodes are in shades of blue (control) and red (cancer), indicating classification into the 2 groups. B, Sample composition of terminal nodes (blue: control; green: cancer); nodes are left to right, as numbered in the tree diagram. C, The cost value in relation to the number of terminal nodes. Arch Pathol Lab Med Vol 126, December 2002 Biomarkers for Ovarian Cancer Rai et al 1521

5 Figure 4. Pseudogel view of surface enhanced laser desorption/ionization analysis of 67 plasma samples showing relative abundance of all markers in 3 panels: 6 to 10, 15 to 45, and 50 to 90 kd. Asterisks indicate markers of interest. Samples from patients without cancer (n 38) are shown above the solid line; samples from patients with cancer (n 29) are shown below the solid line. Figure 5. Schematic diagram of the protein purification protocol. SELDI indicates surface enhanced laser desorption/ionization; SDS- PAGE, sodium dodecyl sulfate polyacrylamide gel electrophoresis; and Q-TOF, quadrupole time of flight. the patients with cancer compared with the specimens from controls, whereas the remaining peaks had the inverse expression pattern. We then reapplied the UMSA component analysis using only these 7 peaks to test whether they retained most of the discriminatory power of the original full spectrum (Figure 2, B). Using BPS, we identified the peaks at 79 and 9.2 kd as the top choices that provided the optimal classification rate for the data set (Figure 3). Compared with the protein peaks identified through ProPeak analysis, these 2 peaks were ranked numbers 1 and 6, respectively. The pseudogel view of the 7 selected protein peaks is given in Figure 4. We were able to purify and identify only 3 proteins at peaks 9.2, 54, and 79 kd. The flow diagrams describe the steps in protein purification (Figure 5) and identification using tandem mass spectrometry (Figure 6). We determined that the 79-kd protein corresponded to transferrin and the 9.2-kd protein was a fragment of the haptoglobin precursor protein. We identified the third, 54- kd protein as immunoglobulin heavy chain. We used 4 peaks (9.2, 54, 60, and 79 kd) in the final statistical evaluation of diagnostic performance. We selected these peaks for their relative high scores in UMSA analysis. We compared the performance of individual peaks with that from the logistic regression functions of all 4 peaks and of 2 of the peaks (60 and 79 kd) using ROC analysis (Figure 7). In the scatter plot (Figure 8), the y axis represents the combination of the 60- and 79-kd 1522 Arch Pathol Lab Med Vol 126, December 2002 Biomarkers for Ovarian Cancer Rai et al

6 Figure 6. Protein identification. Molecular weights of peptide fragments were measured by tandem mass spectrometry using quadrupole time of flight. Data from the 9.2-kd candidate marker are shown. Selected peaks were further analyzed by tandem mass spectrometry fragmentation, as shown in the inset. Figure 7. Receiver-operating characteristic curve analysis based on all 80 patients to compare the diagnostic performance of 4 biomarkers (9.2-, 54-, 60-, and 79-kd peaks) individually and in combinations through logistic regression. Curves are shown for the 9.2- kd peak (1); 54-kd peak (2); 60-kd peak (3); 79-kd peak (4); logistic regression of the 60- and 79-kd peaks (a); and logistic regression of the 9.2-, 54-, 60-, and 79-kd peaks (b). Areas under the curve are (1), (2), (3), (4), (a), and (b). Two-sided P values were.02 (1 4 vs a),.001 (1 4 vs b), and.01 (a vs b). peaks through a logistic regression function. The x axis is the CA125 value in logarithmic scale, with the recommended cutoff value at 35 U/mL marked as a vertical line. The dashed line shows that by combining the 2 biomarkers with CA125 level, the separation between the 2 groups of patients is much improved compared with that obtained using CA125 level alone. Based on this observation, we performed ROC analysis for 68 patients with available CA125 values to compare the diagnostic performance of the combination of 60- and 79-kd peaks, CA125 level alone, and the combination of all 3 markers (Figure 9). The addition of the 2 biomarkers improved on the overall performance obtained with the CA125 level. Table 1 compares the estimated sensitivities and speci- Arch Pathol Lab Med Vol 126, December 2002 Biomarkers for Ovarian Cancer Rai et al 1523

7 Figure 8. Scatter plot showing that the combination of the 60- and 79-kd biomarkers complements cancer antigen 125 (CA125) in separating patients with ovarian cancer from control patients. The dashed line indicates the decision boundary of a possible linear classification function. The vertical line at a CA125 concentration of 35 U/mL indicates the recommended cutoff value for CA125 level. Figure 9. Receiver-operating characteristic curve analysis based on 68 patients with available cancer antigen 125 (CA125) values to compare the diagnostic performance of the combination of biomarkers 60- and 79-kd peaks, CA125 level alone, and a diagnostic index combining the 2 biomarkers and CA125 level. Curves are shown for the logistic regression of 60- and 79-kd peaks (a), CA125 level alone (b), and a combination of logistic regression of the 60- and 79-kd peaks and CA125 level (c). Areas under the curve were (a), (b), and (c). Twosided P values were.17 (a vs b),.03 (a vs c), and.31 (b vs c). Biomarker(s) Table 1. Sensitivity and Specificity of Various Combinations of Biomarkers* No. of Cancer Patients Sensitivity (95% Cl), % No. of Control Patients Specificity (95% Cl), % CA125 level with a cutoff of 35 U/mL ( ) ( ) CA125 level with a cutoff of 18.5 U/mL ( ) ( ) Logistic regression of 60 and 79-kd peaks ( ) ( ) Combination of logistic regression 60- and 79-kd peaks and CA 125 level ( ) ( ) * Differences in numbers of patients reflect missing CA125 results for 10 patients with cancer and 2 control patients. CA125 indicates cancer antigen 125; CI, confidence interval. ficities of (1) CA125 alone at 2 different cutoff values, (2) logistic regression of the 60- and 79-kd peaks, and (3) a diagnostic index that is the linear combination of (1) and (2). In the table, the first cutoff value of CA125 was the recommended value of 35 U/mL. The second value at 18.5 U/mL was selected so that CA125 level achieved maximum efficiency based on ROC analysis. This resulted in a specificity of 94.4%. The remaining comparison, performed using this set specificity, indicated that the diagnostic index from the combination of the 2 biomarkers and 1524 Arch Pathol Lab Med Vol 126, December 2002 Biomarkers for Ovarian Cancer Rai et al

8 Table 2. Sensitivities of Various Combinations of Biomarkers Calculated Separately According to Cancer Stage* Sensitivity, % (No. Positive/No. True Positive) Biomarker(s) Stages I/II Stage III CA125 level with a cutoff of 35 U/mL 44.4 (4/9) 73.9 (17/23) CA125 level with a cutoff of 18.5 U/mL 88.9 (8/9) 91.3 (21/23) Logistic regression of 60- and 79-kd peaks 71.4 (10/14) 53.6 (15/28) Combination of logistic regression of 60- and 79-kd peaks and CA125 level (9/9) 91.3 (21/23) * Due to the small sample size, confidence intervals were not computed. CA125 indicates cancer antigen 125. CA125 improves the sensitivity from 81.3% to 93.8%. Finally, we calculated test sensitivities separately according to early and late disease stages (Table 2). The result shows that the diagnostic index from combining the 2 biomarkers and CA125 retained a high level of sensitivity for the patients with early-stage cancer. Of the early-stage samples misclassified (4/14), 1 was serous and 3 were mucinous. The mean and SD of the diagnostic index in the cancer group were and 0.037, respectively, and those in the control group were and 0.062, respectively. The difference was statistically significant (P.001). COMMENT Numerous studies have documented that genomic expression analysis does not necessarily reflect protein levels or posttranslational modifications. By globally cataloging protein contents, proteomics complements genomics in the discovery and identification of biomarkers that are useful for the detection of human cancer. In this study, we show that the plasma protein profile in women with ovarian cancer can be reliably distinguished from that of women without cancer using SELDI analysis. This technique provides a simple yet sensitive approach to diagnosing ovarian cancer using plasma samples. Two important features of this study are (1) the use of sophisticated bioinformatics tools to process and select a small number of peaks with the greatest discriminatory power for further analysis and (2) the complementary behavior of the selected peaks when used with CA125 level, as shown in results from the study patient data. The poor prognosis of ovarian cancer diagnosed at late stages, the cost and risk associated with confirmatory diagnostic procedures, and the relatively low prevalence of this cancer in the general population together pose extremely stringent requirements on the sensitivity and specificity for a test to be used in screening for ovarian cancer in the general population. Despite more than a decade of effort in this direction, there is still not a costeffective screening test that satisfies these requirements. For example, results of tests for the best-characterized tumor marker, CA125, are negative in approximately 30% to 40% of women with stage I ovarian carcinoma, and levels of this marker are elevated in a variety of benign diseases Statistical analysis of data from this study shows that the diagnostic index combining 2 of the biomarkers (the 60- and 79-kd peaks) and CA125 level improves sensitivity by more than 10% over that of CA125 level alone when the specificity is fixed at 94%. The overlap of 95% confidence intervals of the estimated sensitivities is most likely due to the small sample size, since the diagnostic index values in the 2 patient groups are well separated (P.001). Although the selected biomarkers performed well on the small number of samples from women with early-stage ovarian cancer, the early-stage tumors included in this report were relatively large ( 6 cm), and the patients had markedly high CA125 levels (mean, 45.6 U/mL; median, 29.0 U/mL). This is due to the fact that most sporadic ovarian tumors are relatively large in size at diagnosis because of the absence of effective screening tools. 5 Since early diagnosis is the key for reducing mortality in ovarian cancer, it will be important to determine what is the smallest ovarian tumor that can be detected using this panel of biomarkers. One way that this can be accomplished is to screen high-risk patients with a familial history of ovarian cancer and mutations in the BRCA1 and BRCA2 genes. 36 The control group in this study consisted of patients with nongynecologic diseases. The false-positive rate of CA125 levels in this group at the cutoff of 35 U/mL was about 3%, which is comparable to those found in studies involving healthy donors. It is not clear from this study population whether the selected biomarkers also complement CA125 in reducing the false-positive rate in women with a high CA125 level due to noncancerous conditions. The usefulness of an ovarian cancer detection test will ultimately have to be determined by its positive predictive value in finding clinically asymptomatic cancer through prospective cohort studies. Before such studies can take place, however, the test will have to be evaluated with blinded data from retrospective studies using specimens that are predominantly from patients with early-stage disease and a large number of healthy donors. The identification of these representative candidate markers is promising, as these molecules have been shown to have altered expression in malignant states. 37 More specifically, plasma levels of transferrin are decreased in patients with cancer, whereas haptoglobin levels are increased. Furthermore, based on serial analysis of gene expression data, 38 haptoglobin mrna is up-regulated in ovarian cancer tissue as compared with normal ovarian surface epithelium. These changes are consistent with the results obtained from this study. Traditional multivariate statistical methods typically involve the use of the covariance matrix. It is difficult or even impossible to obtain a stable estimate of covariance matrix from data with a very small sample size and a relative large number of variables to be analyzed. This difficulty has forced a number of current approaches to expression data analysis to evaluate the significance of individual variables (genes, proteins, or biomarkers in general) independent of other variables. 39 The 2 supervised analytic methods used in this study for peak selection do not explicitly require the estimation of the covariance matrix. Both methods evaluate the significance of an individ- Arch Pathol Lab Med Vol 126, December 2002 Biomarkers for Ovarian Cancer Rai et al 1525

9 ual biomarker based on its role in a multivariate effort to separate the different groups of patients. With such methods, it is possible to select biomarkers that by themselves may not provide superior performance. However, their collective expression patterns as a whole are strongly associated with the presence or absence of a given disease condition. This is clearly illustrated in the profound differences between the ROC analysis results for individual biomarkers and for combinations of multiple biomarkers. The UMSA algorithm as implemented in ProPeak is a linear classifier, whereas the CART algorithm in BPS is a binary decision tree type nonlinear classifier. In general, the ranking and selection of peaks based on linear classification tend to be more robust, especially with the inherent variances and noise in the raw spectrum data. On the other hand, a nonlinear classifier might give a better classification result even though extra caution needs to be exercised to avoid overfitting data with superfluous biomarkers. The apparent consistency between results from these 2 approaches on our data provides additional confidence that the selected peaks reflect pathophysiologic changes rather than artifactual differences. For simplicity and because of the relatively small sample size of available data, we combined the multiple biomarkers in this study using only simple forms of logistic regression. Once these biomarkers are further validated through additional studies, it will be possible to construct better predictive models using more-advanced statistical and computational tools such as artificial neural networks. The CART algorithm in BPS and the nonlinear version of UMSA can also be used to construct such models. However, in the initial stage of biomarker discovery, the data sets are typically small. The direct application of overly flexible nonlinear models using a large number of variables from raw profile data could result in models that rely heavily on non disease-related artifacts to produce misleading results on the training data. In conclusion, the combined use of bioinformatics tools and proteomic profiling provides an effective approach to screen for potential tumor markers. Comparison of plasma profiles from patients with and without known ovarian cancer uncovered a limited panel of potential biomarkers. These biomarkers, in combination with CA125, provide significant discriminatory power for the detection of ovarian cancer. A clinical study using a large population will be beneficial for further validation of these promising biomarkers. We thank the members of the Division of Clinical Chemistry and Center for Biomarker Discovery at the Johns Hopkins University School of Medicine for helpful discussions. We also acknowledge Debra Bruzek and Phaedre Mohr for technical assistance with CA125 analysis. References 1. Banks E, Beral V, Reeves G. The epidemiology of epithelial ovarian cancer: a review. Int J Gynecol Cancer. 1997;7: Parkin DM, Muir CS, Whelan SF. Cancer Incidence in Five Continents. Lyon, France: IARC Scientific; Greenlee RT, Hill-Harmon MB, Murray T, Thun M. Cancer statistics, CA Cancer J Clin. 2001;51: Paley PJ. Ovarian cancer screening: are we making any progress? Curr Opin Oncol. 2001;13: Ozols RF, Rubin SC, Thomas GB, Robboy SJ. Epithelial ovarian cancer. In: Hoskins WJ, Perez CA, Young RC, eds. Principles and Practice of Gynecologic Oncology. 3rd ed. Philadelphia, Pa: Lippincott, Williams and Wilkins; 2000: MacDonald ND, Jacobs IJ. Is there a place for screening in ovarian cancer? Eur J Obstet Gynecol Reprod Biol. 1999;82: Jacobs I, Bast RC Jr. The CA 125 tumour-associated antigen: a review of the literature. Hum Reprod. 1989;4: Shih I-M, Sokoll LJ, Chan DW. Tumor markers in ovarian cancer. In: Diamandis EP, Fritsche H, Lilja H, Chan DW, Schwartz M, eds. Tumor Markers Physiology, Pathobiology, Technology and Clinical Applications. Philadelphia, Pa: AACC Press; 2002: Woolas RP XF, Jacobs IJ, Yu YH, et al. Elevation of multiple serum markers in patients with stage I ovarian cancer. J Natl Cancer Inst. 1993;85: Woolas RP, Conaway MR, Xu F, et al. Combinations of multiple serum markers are superior to individual assays for discriminating malignant from benign pelvic masses. Gynecol Oncol. 1995;59: Zhang Z, Barnhill SD, Zhang H, et al. Combination of multiple serum markers using an artificial neural network to improve specificity in discriminating malignant from benign pelvic masses. Gynecol Oncol. 1999;73: Zhang Z, Xu F, Yu Y, Berchuck A, et al. Use of multiple markers to detect stage I epithelial ovarian cancers: neural network analysis improves performance. Presented at: Annual Meeting of the American Society of Clinical Oncology; May ; San Francisco, Calif. Abstract Skates SJ, Xu FJ, Yu YH, et al. Toward an optimal algorithm for ovarian cancer screening with longitudinal tumor markers. Cancer. 1995;76(suppl): Jacobs I, Davies AP, Bridges J, Stabile I, et al. Prevalence screening for ovarian cancer in postmenopausal women by CA 125 measurement and ultrasonography. BMJ. 1993;306: Menon U, Talaat A, Rosenthal AN, Macdonald ND, et al. Performance of ultrasound as a second line test to serum CA125 in ovarian cancer screening. Br J Obstet Gynaecol. 2000;107: Bast RC, Jr, Urban N, Shrindhar V, et al. Early detection of ovarian cancer: promise and reality. Cancer Treat Res. 2002;107: Bichsel VE, Liotta LA, Petricoin EF 3rd. Cancer proteomics: from biomarker discovery to signal pathway profiling. Cancer J. 2001;7: Velculescu VE, Vogelstein B, Kinzler KW. Analysing uncharted transcriptomes with SAGE. Trends Genet. 2000;16: Schaefer C, Grouse L, Buetow K, Strausberg RL. A new cancer genome anatomy project web resource for the community. Cancer J. 2001;7: Vlahou A, Schellhammer PF, Mendrinos S, et al. Development of a novel proteomic approach for the detection of transitional cell carcinoma of the bladder in urine. Am J Pathol. 2001;158: Banks RE, Dunn MJ, Forbes MA, et al. The potential use of laser capture microdissection to selectively obtain distinct populations of cells for proteomic analysis preliminary findings. Electrophoresis. 1999;20: Paweletz CP, Gillespie JW, Ornstein DK, et al. Rapid protein display profiling of cancer progression directly from human tissue using a protein biochip. Drug Dev Res. 2000;49: von Eggeling F, Junker K, Fiedler W, et al. Mass spectrometry meets chip technology: a new proteomic tool in cancer research? Electrophoresis. 2001;22: Merchant M, Weinberger SR. Recent advancements in surface-enhanced laser desorption/ionization time-of-flight mass spectrometry. Electrophoresis. 2000;21: Weinberger SR, Morris TS, Pawlak M. Recent trends in protein biochip technology. Pharmacogenomics. 2000;1: Palmer-Toy D, Kuzdzal S, Chan, DW. Proteomic approaches to tumor marker discovery. In: Diamandis EP, Fritsche H, Lilja H, Chan DW, Schwartz M, eds. Tumor Markers Physiology, Pathobiology, Technology and Clinical Applications. Philadelphia, Pa: AACC Press; 2002: Fung ET, Thulasiraman V, Weinberger SR, Dalmasso EA. Protein biochips for differential profiling. Curr Opin Biotechnol. 2001;12: Fung ET, Wright GL Jr, Dalmasso EA. Proteomic strategies for biomarker identification: progress and challenges. Curr Opin Mol Ther. 2000;2: Breiman L, Friedman JH, Olshen RA, Stone CJ. Classification and Regression Trees. Monterey, Calif: Wadsworths & Brooks; Zhang Z, Page G, Zhang H. Applying classification separability analysis to microarray data. In: Lin SM, Johnson KF, eds. Methods of Microarray Data Analysis: Papers From CAMDA 00. Boston, Mass: Kluwer Academic Publishers; Vapnik VN. Statistical Learning Theory. New York, NY: John Wiley & Sons; Brown M, Grundy WN, Lin D, et al. Knowledge-based analysis of microarray gene expression data by using support vector machines. Proc Natl Acad Sci USA.2000;97: Meyer T, Rustin GJ. Role of tumour markers in monitoring epithelial ovarian cancer. Br J Cancer. 2000;82: Buamah P. Benign conditions associated with raised serum CA-125 concentration. J Surg Oncol. 2000;75: Tuxen MK, Soletormos G, Dombernowsky P. Tumor markers in the management of patients with ovarian cancer. Cancer Treat Rev. 1995;21: Werness BA, Eltabbakh GH. Familial ovarian cancer and early ovarian cancer: biologic, pathologic, and clinical features. Int J Gynecol Pathol. 2001;20: Burtis CA, Ashwood ER. Tietz Textbook of Clinical Chemistry. 3rd ed. Philadelphia, Pa: WB Saunders Co; 1999: SAGE (Serial Analysis of Gene Expression) home page. Available at: Accessed August 6, Tusher VG TRaCG. Significance analysis of microarrays applied to the ionizing radiation response. Proc Natl Acad Sci U S A. 2001;98: Arch Pathol Lab Med Vol 126, December 2002 Biomarkers for Ovarian Cancer Rai et al