Multianalyte Profiling of Serum Antigens and Autoimmune and Infectious Disease Molecules to Identify Biomarkers Dysregulated in Epithelial Ovarian Cancer

There is great interest in the development of a test for the early diagnosis of ovarian cancer—the second most common, yet deadliest, gynecologic cancer in the United States (1). In 2007, an estimated 22,430 new cases of ovarian cancer were detected. When ovarian cancer is diagnosed and treated early, intervention is successful, with 5-year survival rates of 92% (2). Therefore, a diagnostic test for the early detection of ovarian cancer has clinical value. Regrettably, only 19% of ovarian cancers are found early with the majority of cases detected at late-stage disease where the outcome is far less favorable. Indeed, 5-year survival rates are only 30% for patients with distant malignancy and as a result, >15,000 women will die each year from this cancer in the United States (1).

Early symptoms of ovarian cancer, such as fatigue, back pain, abdominal bloating, indigestion, poor appetite, diarrhea, constipation, and menstrual changes, are subtle and typical of many normal and benign conditions. Therefore, diagnosis usually does not occur until a significant amount of abdominal fluid develops, or a pelvic tumor mass is detectable by imaging techniques such as transvaginal sonography, magnetic resonance imaging, or computed tomography. These imaging methods, however, all lack sufficient specificity to distinguish between benign masses and malignant tumors, and abnormal imaging necessitates additional intervention in the form of laparotomy or laparoscopy (3). In a recent study, 93% of abnormal transvaginal sonography findings resolved within 3 to 6 months (4). Therefore, imaging methods are unreliable, especially for early detection, and prone to interobserver variability.

There are no Food and Drug Administration–approved biomarkers for the diagnosis of ovarian cancer. The biomarker cancer antigen 125 (CA-125) is often elevated in ovarian cancer and used by many oncologists as an aid in diagnosis, although it is only Food and Drug Administration–approved for monitoring recurrent disease and therapeutic response (5-7). In studies of women with known or suspected ovarian cancer, the reported sensitivities of CA-125 in detecting stage I and II cancers range widely from 29% to 75% and 67% to 100%, respectively. However, CA-125 is elevated in a wide variety of normal, benign, and malignant conditions including menstruation, endometriosis, pregnancy, pelvic inflammatory disease, heart failure, pancreatitis, lung disease, liver cirrhosis, peritoneal tuberculosis, surgery, nonmalignant gynecologic disease, and abdominal disorders (8-10). It is reported that 86% of abnormal CA-125 tests are not indicative of cancer and resolve in 3 to 6 months (4). Many approaches have been taken to improve the predictive value of CA-125. For example, greater specificity has been reported by serial measurement of CA-125 using a risk of ovarian cancer algorithm but the accuracy is still inadequate for initial diagnosis (11), although it is the subject of an ongoing clinical investigation (12). Although several refinements of the use of CA-125 have been proposed, including the combination of additional markers, a simple and clinically practical ovarian cancer–screening tool remains elusive (13-16). Therefore, there is an unmet medical need for good diagnostic biomarkers.

Ovarian cancer is a collection of diverse entities with >30 subtypes of malignancies, each with a distinctive histology, pathology, and clinical behavior (17). Of these, epithelial tumors—those arising from cells covering the ovaries—account for >90% of cases in the United States. The most common subtypes of epithelial tumor are serous (40%), endometrioid (20%), clear cell (6%), and mucinous (1%) tumors. The diversity and low incidence of ovarian cancer hinder the search for biomarkers. Many individual biomarkers have been proposed as having potential clinical utility, but most were assessed on small, discovery-size sample sets that cannot represent the diversity of the disease (18). Moreover, different studies have used different collection protocols and inclusion and exclusion criteria and tend to focus on late-stage samples that are more readily accessible than early-stage samples. In this setting, both positive and negative findings can be misleading.

As the first step in a two-part study, we have undertaken a broad biomarker survey of the sera of patients with ovarian cancer or benign conditions. We have used a large, well-characterized, multisite collection of sera, drawn under a standard protocol before intervention or knowledge of disease status. We have analyzed the data for evidence of useful single markers of ovarian cancer across disease type, stage, and pathology. Our findings point to the absence of a single diagnostic marker and support the increasing interest in the use of multivariate assays, which form the ongoing focus of our work.

Sera were obtained from the National Cancer Institute–funded Gynecologic Oncology Group (GOG). Samples were from multiple collection sites, collected under a protocol approved by the GOG institutional review boards. Eligible patients were women scheduled for surgery with suspicion of having a gynecologic cancer or were having prophylactic surgery because of increased ovarian cancer risk (first- or second-degree relative with the disease). All samples were collected before any intervention or knowledge of disease status. Serum aliquots forwarded to Correlogic Systems, Inc., had been de-identified and contained a unique coded GOG identifier. Complete clinicopathology reports obtained after surgery, along with the patient age, race, staging, subtype, and coded collection site, accompanied each sample. Samples were selected from the GOG collection to remove as many potential biases as possible, representing a balanced age distribution, number of cases and controls from each collection site, and storage time. The Western Institutional Review Board approved the studies by Correlogic Systems.

Before any intervention, blood samples (5-20 mL) were collected into red top glass Vacutainer tubes. The blood was allowed to stand for 30 to 180 min at 4°C and was then centrifuged at 3,500 × g for 10 min at 4°C. Serum was carefully decanted into prelabeled cryotubes and stored promptly at -80°C. Aliquots were taken from storage and shipped to Correlogic on dry ice where they were immediately stored at -80°C. Samples to be analyzed were removed from storage, warmed gently by hand until almost thawed, completely thawed on ice, vortexed gently, and then stored as 150 μL aliquots at -80°C. Finally, samples were shipped on dry ice to Rules-Based Medicine (RBM), Inc. An accompanying document provided a coded sample identification number and a specific order of analysis. Therefore, the RBM analytic site was completely blinded to the cancer status, pathology, and all other sample details.

A total of 204 analytes representing 104 antigens and 44 autoimmune and 56 infectious disease molecules were measured using a set of proprietary multiplexed immunoassays (Human MAP; Supplementary Table S1) at RBM in their Luminex-based, CLIA-certified laboratory. Each antigen assay was calibrated using 8-point standard curves and done in duplicate, and raw intensity measurements were interpreted into final protein concentrations using RBM's proprietary software. Machine performance was verified using quality control samples at low, medium, and high levels for each analyte in duplicate. All standard and quality control samples were in a complex plasma-based matrix to match the sample background. The autoimmune and infectious disease assays were qualitative and the results obtained for unknown samples were compared with established cutoff values. Because sera were analyzed at a previously optimized dilution, any sample exceeding the maximum concentration of the calibration curve was arbitrarily assigned the concentration of the highest standard, whereas those assayed below the minimum concentration of the calibration curve were assigned the value 0.0. For analysis, samples were ordered in a manner to avoid any sequential bias due to the presence or absence of disease, subtype or stage of disease, patient age, or age of serum sample. Generally, samples alternated between cases and controls.

Descriptive statistics, Receiver Operating Characteristic (ROC) curves, and graphical displays (dot plots) for serum analyte concentrations were done using GraphPad Prism version 5.0a (GraphPad Software, Inc.). Unless indicated, statistical differences were determined using the nonparametric Kruskal-Wallis test (ANOVA) that does not assume Gaussian distribution followed by Dunn's multiple comparison posttest. For all statistical comparisons, P < 0.05 was interpreted as statistically significant. A Spearman correlation matrix was created using the multispectral analysis application SpectraViewer (Correlogic Systems). One-Rule (OneR) classification was done to identify potentially informative single markers (19). The analysis was done using a 5-fold, cross-validation strategy to minimize sampling errors.

Using a bead-based, multianalyte profiling approach, the levels of 204 molecules were measured simultaneously in sera from 147 patients with pathology-confirmed epithelial ovarian cancer and 147 individuals without ovarian cancer. Because early detection and treatment of ovarian cancer increases survival, our study focused on the analysis of early-stage disease with 43% of the cancer sample set representing stage I and II disease (Table 1). Consistent with the literature, the average patient age was higher as disease progressed (Table 1; ref. 20). The subtype distribution was representative of the U.S. population, with a larger proportion of serous carcinoma (43%) over other subtypes (Table 1). Because a key clinical need is to be able to differentiate between a benign gynecologic condition and cancer, the control samples were predominantly from individuals with benign ovarian conditions (71%), including cystadenoma, cystadenofibroma, and fibroma. In addition, the control group had other gynecologic cancers (10%), including endometrial and cervical cancer, whereas the number of normal healthy samples was restricted to just 19%.

The profiled analytes covered a broad range of structures and biological functions including cancer antigens, hormones, clotting factors, tissue modeling factors, lipoprotein constituents, proteases and protease inhibitors, markers of cardiovascular risk, growth factors, cytokine/chemokines, soluble forms of cell signaling receptors, and inflammatory and acute-phase reactants (Supplementary Table S1). Because immune dysfunction and autoimmune responses could play a role in cancer, our profile included a panel of markers associated with autoimmune conditions. Although there has been no credible link between infection or vaccination and ovarian cancer, an estimated 15% of cancers are caused by viruses (21). Therefore, we included a survey of infectious disease markers to complete the broad serum profiling. To our knowledge, this is the broadest profiling of molecules using a large collection of ovarian cancer samples to date.

As one measure of discriminatory ability of markers, ROC curves were generated for each analyte and area under the curve (AUC) values compared with that of an uninformative marker (AUC = 0.5). Seventy-seven analytes had AUC values statistically greater than 0.5. Of these, 32 analytes were up-regulated and 45 analytes were down-regulated in ovarian cancer samples compared with controls (Table 2). The analytes with the greatest AUC values were predominantly up-regulated in ovarian cancer (Table 2; Supplementary Figs. S1 and S2). CA-125 had the highest AUC value of 0.906 followed by C-reactive protein, soluble epidermal growth factor receptor (EGFR), interleukin 10 (IL-10), IL-8, connective tissue growth factor (CTGF), haptoglobin, and tissue inhibitor of metalloproteinase 1 (TIMP-1), with AUC values between 0.756 and 0.701 (Table 2; Supplementary Fig. S1). This group of molecules is predominantly composed of inflammatory markers and acute-phase reactants. The remaining 69 informative proteins had a continuum of AUC values from 0.693 to 0.567. Interestingly, all 26 informative autoimmune and infectious disease markers were down-regulated in ovarian cancer samples with AUC values between 0.568 and 0.627. Anti–cytochrome P450, double-stranded DNA antibody, and collagen type 6 antibody were the most informative autoimmune markers (AUC values 0.627, 0.618, and 0.614, respectively), and Polio virus, Borrelia burgdorferi (Lyme), and varicella zoster (chickenpox) virus were the most informative infectious disease markers (AUC values 0.611, 0.610, and 0.602, respectively).

To examine potential clinical utility, the sensitivity for each analyte with AUC higher than 0.7 was determined for five fixed specificity values (80%, 90%, 95%, 99%, and 99.6%; Supplementary Table S2). In addition, the optimal sensitivity and specificity values (that is, those yielding the greatest combined total) were determined. For any analyte exhibiting more than one solution for the optimal cutoff, the solution yielding the highest specificity was determined. None of the markers, except CA-125, had relatively high sensitivity at a specificity value of 80% or higher. The sensitivity for CA-125 was 82.3%, 72.1%, 64.0%, 17.7%, and 12.9% as specificity values increased from 80%, 90%, 95%, 99%, and 99.6%, respectively. The optimal cutoff for CA-125 was 112 units/mL, giving a sensitivity of 76.9% and specificity of 88.4%.

A more sophisticated approach than ROC analysis is to use multiple cutoff values for a single feature. For example, assume that two different critical values exist for a given analyte. Samples falling below the first cutoff are assigned one state (e.g., cancer). Then, samples falling above the first cutoff but below the second cutoff are classified as noncancer, whereas samples falling above the second cutoff are classified as cancer. The OneR algorithm attempts to classify samples in this manner and, when used with appropriate restrictions on the number of samples in an interval, has been shown to be almost as accurate as state-of-the-art machine learning schemes (19). We therefore used OneR to determine if there were any more informative single biomarkers for ovarian cancer using this strategy. To minimize potential sample set bias and the drive to overfit, we used a 5-fold, cross-validation strategy and modeled over a range of minimal interval sizes ranging from 3 to 15 samples. A minimum interval value of 11 was found to give the optimal fit, creating reasonably large and homogenous intervals. Smaller intervals led to overfitting of the data. When applied to the entire data set, to determine the most informative assay, OneR identified CA-125 with 81.3%, 76.0%, and 78.7% ± 3.21% sensitivity, specificity, and accuracy, respectively. When the analysis was repeated, excluding CA-125 from the data set, C-reactive protein was identified with 65.3%, 73.3%, and 71.3% ± 4.6% sensitivity, specificity, and accuracy, respectively. Dropping both CA-125 and C-reactive protein from the analysis identified several different proteins, all with low-classification strengths, which did not hold up well through cross-validation.

The expression level of analytes did not vary from stage to stage in a statistically significant manner. However, for six analytes, there were clear trends associated with disease progression (Fig. 1). Consistent with previous reports, CA-125 increased markedly from stages I through IV (Fig. 1; ref. 22). Similar trending was evident for C-reactive protein, IL-10, IL-8, and CTGF (Fig. 1). In contrast, EGFR expression seemed to decrease as disease progressed. Only five markers, CA-125, cancer antigen 19-9 (CA 19-9), C-reactive protein, creatine kinase-MB, and EGFR had statistically dysregulated levels in early-stage samples (stages I and II combined, before cancer spread outside of the pelvic area) compared with the controls. In contrast, a total of 40 markers were dysregulated in late-stage disease (stages III and IV combined, disease progressed beyond the pelvic area) compared with the control samples (Supplementary Table S3). These results underscore the requirement of careful sample selection for biomarker discovery studies whereby early-stage samples are used whenever possible.

Several molecules exhibited statistically dissimilar expression levels in different subtypes of ovarian cancer (Fig. 2). CA-125 was clearly overexpressed in serous, endometrioid, and mixed ovarian cancers (P < 0.001) and slightly in clear cell (P < 0.05) compared with the control group but not in mucinous ovarian cancer. C-reactive protein, haptoglobin, and IL-10 were overexpressed in clear cell, endometrioid, and serous but not mucinous or mixed ovarian cancer. IL-8 was overexpressed in endometrioid, mucinous, and serous ovarian cancer, whereas EGFR and insulin-like growth factor (IGF-I) were underexpressed in serous and endometrioid subtypes, and insulin was underexpressed in serous ovarian cancer. These results indicate that the different subtypes of ovarian cancer tumors can have very disparate expression profiles, and, as such, any study must include an adequate number of each tumor subtype to capture the diversity of this disease.

When developing multivariate panels for a disease, it is important to combine markers that classify the data in different ways. We therefore performed covariation analysis on the 77 informative molecules. The covarying molecules were sorted agglomeratively with hierarchical clustering using Spearman correlation coefficients as the distance measure. The pair-wise results were assembled into a 77 × 77 matrix and displayed using a heat map where an intense red color signifies strong positive correlation and blue signifies a negative correlation. Three large groupings of molecules were observed (Fig. 3). Cluster A, the rightmost cluster, was the largest grouping and consisted of all the infectious and autoimmune disease markers as well as infection-related IgM (index values 50-76). All statistically significant covariations in this cluster were positive.

The majority of the antigens were found in the two remaining clusters. In the leftmost cluster (cluster C; index values 0-18), there was high covariation evident for the soluble fragment of the tumor necrosis factor receptor with CD40, β2-macroglobin, and vascular cell adhesion molecule-1 with correlation coefficients of 0.676, 0.827, and 0.652, respectively. High covariation was also observed among apolipoproteins AI, CIII, and H (index values 9, 10, and 11) with correlation values between 0.596 and 0.731. The remaining molecules in this cluster did not seem to be structurally or biologically related and showed minimal covariance. The middle cluster (cluster B; index values 19-42) contained CA-125 (index value 30). Forty-nine molecules covaried with CA-125 with statistical significance, 20 of which had Spearman correlation values higher than 0.250 (Supplementary Table S4). Of the 49 molecules, 19 covaried negatively and 30 elicited a positive covariance. The molecules with the highest positive correlation coefficients with CA-125 were TIMP-1, C-reactive protein, and haptoglobin (0.502, 0.489, and 0.446, respectively), whereas EGFR (-0.359), IGF-I (-0.279), and lymphotactin (-0.229) showed the greatest negative covariance. The middle cluster also contained a small subcluster that contained several interleukins; this subcluster was composed of fatty acid–binding protein, IL-8, TNF-α, IL-10, IL-6, and IL-1ra (cluster D; index values 36-41). In addition, a strong correlation was evident for tenascin C with CTGF (0.820) as well as the functionally related protease inhibitors TIMP-1 and plasminogen activator inhibitor-1 (0.767).

During the preparation of this article, Mor and colleagues described a promising panel of six proteins diagnostic for ovarian cancer (20). Consistent with other literature, the AUC value for CA-125 alone was 0.85 (Table 3; ref. 20). Surprisingly, the five remaining markers had highly significant classification potential as individual markers. Prolactin and total IGF-II outperformed CA-125 with AUC values of 0.97 and 0.94, respectively. The remaining markers—osteopontin, migration inhibitory factor, and leptin—had relatively high AUC values of 0.80, 0.80, and 0.77, respectively. We were interested to see if the most informative markers on this panel were equally informative with our independent sample set. We selected 37 nonovarian cancer and 37 ovarian cancer samples from our collections and tested the individual classification potential of prolactin and total IGF-II. IGF-I, a molecule in the same family as IGF-II, was only slightly discriminatory in our original analysis; however, we tested the free form of this protein. Therefore, we analyzed total IGF-I using these same 74 samples. In addition, we tested several new markers that RBM had since added to their panel. The new markers, along with CA-125 as a baseline comparison, were measured in the 74 samples. The discriminatory potential of CA-125 was comparable with our original study (0.906) and Visintin's study (0.85) with an AUC value of 0.882 (Table 3; ref. 20). In stark contrast to Visintin's data, neither prolactin (AUC = 0.560) nor total IGF-II (0.535) elicited statistically significant classification with our sample set. Furthermore, neither free (AUC = 0.514) nor total (0.619) IGF-I elicited significant discriminatory potential. Of the other new markers tested, complement 3 des Arg and resistin showed the greatest classification potential with AUC values of 0.710 and 0.709, respectively. The difference in the AUC values obtained between our study and that of Visintin may be due to the makeup of the control group. Whereas Visintin's control group was a healthy control cohort, we have focused on a group with benign gynecologic conditions because a key clinical need is to be able to differentiate between cancer and benign conditions. When we omitted the benign samples from ROC analysis, the AUC values were more comparable with Visintin's study. However, this finding has weak statistical support because only six normal samples were available for this analysis and a further study would be required to confirm this was truly significant.

CA-125 is widely used in the clinical assessment for the likely presence of ovarian cancer. However, whereas CA-125 is useful for monitoring treatment and disease recurrence, poor specificity precludes it from Food and Drug Administration approval for screening. Consequently, many groups have sought to identify novel biomarkers that can augment CA-125. Many studies have reported potentially useful markers; however, the small cohorts examined restrict their interpretation. Furthermore, inclusion and exclusion criteria for participant selection, blood processing protocols, and analytic methodologies used differ across the studies. We have addressed these issues by using a large population of patients scheduled for surgery under a common set of inclusion and exclusion criteria. Both cancer and control patients were subjected to the same level of stress caused by the fear of cancer and anticipation of surgery. Cancer samples were supported by a full pathology report and controls were surgically and pathology-confirmed free of ovarian cancer. Samples were collected preintervention, under a well-defined protocol, across multiple institutions with diverse geographic and socioeconomic groupings. Fully qualified, commercial MAPs were done in a CLIA-certified laboratory at the same time on all samples. To our knowledge, this is the largest number of assays panned using a substantial sample size that represents the heterogeneity of ovarian cancers and benign conditions seen in U.S. surgeries. Our approach was to perform an observational analysis. We report trends across a large sample set to determine if there is any single biomarker that is sufficiently informative to warrant a subsequent validation study. This study would not have been possible without the foresight and care of the GOG who established the serum and tissue bank that provided the samples. A particular value of this large collection has been the ability to focus emphasis on early-stage disease, where early detection can benefit patient survival.

We analyzed the discriminatory potential of 204 molecules in almost 300 individuals, with a high proportion of individuals having early-stage cancer. A literature review indicated that 35 of these molecules have been proposed as potentially useful markers for ovarian cancer. Only 12 of these markers—apolipoprotein A1, CA-125, CA 19-9, C-reactive protein, EGFR, haptoglobin, IL-6, IL-8, ferritin, leptin, TNF-α, and vascular endothelial growth factor—were dysregulated in our samples. These results indicate that discovery studies on small sample numbers may not capture the inherent diversity of ovarian cancer and care needs to be taken in interpretation of small-scale discovery studies. Our results show that of all the molecules examined, CA-125 remained the single most dominant biomarker, whether judged using a single or multiple cutoff value assessment. Despite the common concept that CA-125 is not elevated in many early-stage cancers, we found that it had high sensitivity. One reason for this is the use of CA-125 in the work-up to ovarian cancer, which will enrich the patient population with symptomatic women exhibiting elevated CA-125.

Although 26 autoimmune and infectious disease markers elicited some discriminatory potential for ovarian cancer, their clinical use would be very limited with AUC values all below 0.63. Interestingly, the most informative autoimmune marker, anti–cytochrome P450, has been linked with premature ovarian failure and warrants further investigation (23). There is mounting evidence that infectious diseases are involved in cancer—most notably, the Helicobacter pylori bacterium association with gastric cancer, human papilloma virus with cervical cancer, and chronic infection with hepatitis B or C viruses with liver cancer. Indeed, it is estimated that 15% of cancers are a consequence of viral infection (21). However, none of the 56 infectious disease markers tested in this study had high discriminatory potential for ovarian cancer. Interestingly, all infectious and autoimmune markers that had AUC >0.5 were down-regulated in ovarian cancer samples compared with the controls. This implies that the overall immune response may be compromised in individuals with ovarian cancer.

Many of the most discriminatory molecules in our study were inflammatory markers, in particular, acute-phase reactants. There is increasing evidence that inflammation plays a pivotal role in cancer progression, whereby the inflammatory response provides a microenvironment conducive to tumor survival, angiogenesis, and metastasis (24). C-reactive protein, a major acute-phase reactant, was markedly overexpressed and was the second most discriminatory marker. C-reactive protein levels are thought to be elevated primarily due to increases in production of the proinflammatory cytokine IL-6. Interestingly, IL-6 was elevated in the ovarian cancer cohort relative to the control group and there was high correlation between C-reactive protein and IL-6 levels in the sample set as a whole (Spearman correlation of 0.523). C-reactive protein may have a limited predictive value in several cancers, including ovarian, renal, pancreatic, gastrointestinal, melanoma, and lymphoma (18). In addition to C-reactive protein, we also found the positive acute-phase reactants fibrinogen, von Willebrand factor, α1 anti-trypsin, plasminogen activator inhibitor-1, haptoglobin, and EN-RAGE to be up-regulated in the ovarian cancer samples, whereas apolipoprotein AI, a negative acute-phase reactant, was down-regulated. There was high covariance between these seven proteins with absolute Spearman values generally >0.30.

The acute phase is controlled by the balance of proinflammatory cytokines, including IL-6, IL-8, IL-1β, TNF-α, and IFN-γ; anti-inflammatory molecules, including the cytokines IL-4, IL-10, IL-13, and TGF-β; and the inhibitors of proinflammatory cytokines, including soluble TNF-α receptor, soluble IL-1 receptor, and IL-1 receptor antagonist. We found that IL-6, IL-1β, TNF-α, IL-8, IL-10, soluble TNF receptor (TNF RII), and IL-1ra but not IFN-γ, TGF-β, IL-4, and IL-13 were overexpressed in ovarian cancer samples over the control group. These results indicate that there is not an indiscriminate dysregulation of inflammatory markers but rather a subgroup of specific inflammatory and acute-phase pathways are activated, and while inflammation is not specific to any particular disease, the relative ratios or expression patterns of inflammatory markers may be unique to ovarian cancer. It is also important to appreciate that although it is convenient to group many of these proteins as acute-phase reactants, they have multiple biological roles intersecting with many biological pathways. Therefore, although inflammation may be one reason for their elevation, the particular pattern of dysregulation may reflect biological pathways beyond an inflammatory response. For example, C-reactive protein binding to the stimulatory Fcγ receptors increases phagocytosis and the release of inflammatory cytokines, but C-reactive protein can also have an anti-inflammatory activity through its interaction with the FcγRIIb receptor (25).

Several mediators of cell cycle progression, differentiation, and proliferation were dysregulated in the ovarian cancer group. Notably, EGFR, the tyrosine kinase cell surface receptor for EGF and TGF-α, was the third most informative assay in our study. EGFR is a member of the ErbB family of receptors that includes HER2/c-neu, a proto-oncogene associated with poorer prognosis in breast cancer. Situations resulting in EGFR overexpression or overactivity have been associated with breast and other cancers (26). In human cancer cells, EGFR prevents autophagic cell death by maintaining intracellular glucose levels through interaction and stabilization of the sodium/glucose cotransporter 1 (27). In our study and in a study by Baron et al. (28), underexpression of soluble EGFR is evident in ovarian cancer serum (Fig. 1; Table 2). This shed form of the receptor, lacking the transmembrane and intracellular cell signaling domains, behaves as an inhibitor to the intact form of the receptor. Thus, it would be expected that EGFR signaling would be higher in situations where the soluble form is down-regulated (29).

Angiogenesis and extracellular matrix (ECM) remodeling are important steps in tumorigenesis. The onset of new blood vessel formation requires a local imbalance between proangiogenic and antiangiogenic factors. Consistent with this biology, we found that three potent proangiogenic factors—CTGF, IL-8, and VEGF—were overexpressed in ovarian cancer sera compared with controls. IL-8 has also been found to induce neovascularization, a critical step in the path between hyperplasia and neoplasia as well as metastases (30). In addition to angiogenesis, CTGF is implicated in ECM remodeling by controlling the activity of matrix metalloproteinases and their inhibitors (TIMPs; ref. 31). Matrix metalloproteinases and TIMPs regulate ECM synthesis and degradation by controlling the turnover of basement membranes and other ECM components, including collagens, proteoglycans, gelatin, fibronectin, laminin, and elastin. Interestingly, in our study, TIMP-1 seemed to be up-regulated in the ovarian cancer group in a similar fashion to CTGF. Indeed, a strong correlation in levels of CTGF and TIMP-1 in our samples was evident with a Spearman correlation of 0.491. In summary, several important molecules involved in key events of tumor biology seemed to be dysregulated in our ovarian cancer cohort.

Our findings underscore that ovarian cancer is a particularly heterogeneous disease. Poor accuracy precludes the use of any individual marker for the robust detection of ovarian carcinoma, presumably because a single marker cannot capture the diversity of this disease. The well-studied glycoprotein CA-125, several inflammatory markers and acute-phase reactants such as C-reactive protein, cell cycle mediators such as EGFR, angiogenic factors such as VEGF, and ECM regulators such as TIMP-1, are among the most discriminatory markers for ovarian cancer. Currently, we are exploring the use of several multivariate approaches to combine the diagnostic potential of these and other informative markers. These multivariate approaches hold promise in the pursuit of a diagnostic test for ovarian cancer.

No potential conflicts of interest were disclosed.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.