A combined blood based gene expression and plasma protein abundance signature for diagnosis of epithelial ovarian cancer - a study of the OVCAD consortium

Background The immune system is a key player in fighting cancer. Thus, we sought to identify a molecular ‘immune response signature’ indicating the presence of epithelial ovarian cancer (EOC) and to combine this with a serum protein biomarker panel to increase the specificity and sensitivity for earlier detection of EOC. Methods Comparing the expression of 32,000 genes in a leukocytes fraction from 44 EOC patients and 19 controls, three uncorrelated shrunken centroid models were selected, comprised of 7, 14, and 6 genes. A second selection step using RT-qPCR data and significance analysis of microarrays yielded 13 genes (AP2A1, B4GALT1, C1orf63, CCR2, CFP, DIS3, NEAT1, NOXA1, OSM, PAPOLG, PRIC285, ZNF419, and BC037918) which were finally used in 343 samples (90 healthy, six cystadenoma, eight low malignant potential tumor, 19 FIGO I/II, and 220 FIGO III/IV EOC patients). Using new 65 controls and 224 EOC patients (thereof 14 FIGO I/II) the abundances of six plasma proteins (MIF, prolactin, CA125, leptin, osteopondin, and IGF2) was determined and used in combination with the expression values from the 13 genes for diagnosis of EOC. Results Combined diagnostic models using either each five gene expression and plasma protein abundance values or 13 gene expression and six plasma protein abundance values can discriminate controls from patients with EOC with Receiver Operator Characteristics Area Under the Curve values of 0.998 and bootstrap .632+ validated classification errors of 3.1% and 2.8%, respectively. The sensitivities were 97.8% and 95.6%, respectively, at a set specificity of 99.6%. Conclusions The combination of gene expression and plasma protein based blood derived biomarkers in one diagnostic model increases the sensitivity and the specificity significantly. Such a diagnostic test may allow earlier diagnosis of epithelial ovarian cancer.


Background
One of the most deadly malignant diseases in women is ovarian cancer. The high risk of dying is particularly due to late diagnosis, i. e. 67% of patients are diagnosed with advanced disease. The five-year overall survival (OS) rate is only 46% among all stages [1]. Patients with stage I disease have a five-year OS rate of about 90%, whereas patients with advanced disease less than 30% [2]. One reason for the low five-year OS rate is the fact that ovarian cancer presents with few, if any, specific symptoms. Therefore markers for early detection of ovarian cancer could improve OS.
Up to now no screening markers are recommended or routinely used for early detection of ovarian cancer. One of the known serum marker for ovarian cancer is CA-125, described for the first time in 1981 as a murine monoclonal antibody (OC125) reacting against ovarian cancer cell lines and cryopreserved ovarian cancer tissues but not against benign tissues or other carcinomas [3]. CA-125 is a coelomic epithelial antigen produced by mesothelial cells in the peritoneum, pleural cavity and pericardium and in several other epithelia such as the gastrointestinal tract, respiratory tract, and genital tract. Serum CA-125 levels are measurably increased in about 80% of patients with ovarian cancer. An increase is measured to a lesser extent in patients with early stages, resulting in a sensitivity of CA-125 screening of lower than 60% in early stages [4]. Serum concentrations can be elevated by a number of common benign gynecologic conditions, including endometriosis and leiomyomas, as well as by non-gynecologic pathologies such as congestive heart failure and liver cirrhosis. In general, serum concentrations of CA-125 are higher in premenopausal women, compared to postmenopausal women. These facts all together results in an impaired sensitivity and specificity for CA-125 [5]. Nevertheless, there are numerous papers dealing with CA-125 as marker for early detection, diagnosis, response prediction and monitoring, disease recurrence, and for distinguishing malignant from benign pelvic tumors [6].
To increase the sensitivity and specificity of CA-125, this single marker could be expanded to a marker panel. Including other serum markers and building a statistical model, this might result in a more sensitive and specific signature for detection of EOC.
In 2004 Zhang et al. published a four marker panel comprised of CA-125 and three by mass spectroscopy (SELDI) newly identified serum protein peaks, identified as apolipoprotein A1 (down-regulated in malignant tumors), a truncated form of transthyretin (down-regulated), and a cleaved fragment of inter-α-trypsin inhibitor heavy chain H4 (up-regulated) [7]. A multivariate model combining the three biomarkers and CA-125 reached a sensitivity of 74% by a fixed specificity of 97% for detection of early stage EOC. This set of biomarkers was amended by four additional serum protein peaks leading to a commercialized FDA cleared blood test for assessment of the likelihood that an ovarian mass is malignant, called OVA1™ (Quest Diagnostics, Madison, NJ, USA). Recently, in a prospective study, the effectiveness of the OVA1™ test was compared to the malignancy-assessment by physicians. The multivariate index assay demonstrated higher sensitivity and lower specificity compared to the physician assessment together with the CA-125 serum levels [8,9]. Mor et al. described in 2005 four new serum markers, namely Leptin, Prolactin, OPN, and IGF-II, found by a rolling circle amplification (RCA) immunoassay microarray approach. In a combined predictive model including 19% early stage patients, an overall sensitivity and specificity of approx. 95% was reached [10]. Adding CA-125 and MIF to this four-marker-panel, the specificity was increased to 99.4% at a sensitivity of 95.3%. With this marker panel, 11.1% of stage I and II samples (4 of 36) were misclassified [11].
Recently, Yurkovetsky et al. described a four serum marker panel, namely HE4, CEA, VCAM-1, and CA-125, for early detection of ovarian cancer. A model derived from these four serum markers provided a diagnostic power of 86% sensitivity for early stage, and 93% sensitivity for late stage ovarian cancer at a specificity of 98% [12].
Another approach to find prognostic markers for early detection of ovarian cancer is to use peripheral blood cells instead of serum. In 2005 a set of 37 genes was identified whose expression in peripheral blood cells could detect a malignancy in at least 82% of breast cancer patients [13]. Very recently, a set of 738 genes was identified discriminating breast cancer patients from controls with an estimated prediction accuracy of 79.5% (80.6% sensitivity and 78.3% specificity) [14].
The aim of this study was to investigate if combining gene-expression patterns with a serum protein panel results in a more sensitive and more specific signature for the detection of EOC. Primarily, we isolated a leukocytes fraction from epithelial ovarian cancer (EOC) patients, patients with non-malignant gynecological diseases and healthy blood donors (controls). A whole genome transcriptomics approach (Applied Biosystems Human Genome Survey microarrays V2.0) was used to identify gene expression patterns discriminating between ovarian cancer patients and healthy controls or patients with non-malignant diseases. In the second place we determined a six-protein panel [11] from the plasma samples. Taken together predictive models were built from a large cohort of patients and controls using either RT-qPCR derived expression values or protein abundance values alone or in combination. Validation was performed by means of the bootstrap .632+ cross-validation method.

Patients and controls
In total, blood from 239 epithelial ovarian cancer (EOC) patients (19 FIGO I/II and 220 FIGO III/IV) and 169  Table 1 Overall statistics for EOC patients, patients with benign or low malignant potential (LMP) tumors, and healthy persons and patients with benign diseases as controls (A), clinicopathologic characteristics of FIGO I/II and FIGO III/IV patients (B) and diagnosis of patients with benign diseases (C)

Isolation of the leukocytes fraction and total RNA preparation
A leukocytes fraction depleted from epithelial cells was isolated from EDTA-blood by a density gradient centrifugation protocol, largely according to Brandt and Griwatz [15]. Total RNA was isolated using the RNeasy Mini kit (QIAGEN, Venlo, Netherlands) and quality-checked with the Agilent 2100 Bioanalyzer (Agilent Technologies, Santa Clara, Ca, USA). The RNA-quantity was measured spectrophotometrically.

Microarray analysis and pre-selection
Whole genome expression analysis was performed on single channel Applied Biosystems Human Genome Survey microarrays V2.0 (Applied Biosystems, Foster City, Ca, USA) containing 32,878 probes representing 29,098 genes. Two μg total RNA from 44 ovarian cancer patients and 19 age-matched controls (13 completely healthy controls and 6 patients with benign ovarian cysts (mean 60.8 ± 13.7 years and 61.7 ± 12.9 years, respectively) were labeled with the NanoAmp RT-IVT Labeling Kit and hybridized to the microarrays for 16 hours at 55°C. After washing and visualization of bound digoxigenin-labeled cRNAs with the Chemiluminescence Detection Kit according to the manufacturer's instructions (Applied Biosystems), images were read with the 1700 Chemiluminescent Microarray Analyzer (Applied Biosystems). Raw expression data, signal-to-noise ratios and quality-flags delivered from the Applied Biosystems Expression System software were further processed using Bioconductor's ABarray package (www. bioconductor.org). In brief, raw expression values were log 2 transformed and measurements with quality indicator flag values greater than 5000 were set missing. For interarray comparability, data were quantile-normalized and missing values imputed with 10-nearest neighbors imputation. Several pre-filtering steps of probes were performed. Firstly, 13,520 probeIDs which exhibited a signal-to-noise ratio less than 2 in at least 50% of the two pooled groups (patients with malignant disease and non-malignant controls) were excluded (19,358 probeIDs were remaining). Secondly, 10,125 probeIDs assumed to be potentially affected by batch-effects were excluded, resulting in remaining 9,233 probeIDs. Finally, 205 probeIDs with foldchanges > 3 between both groups were selected. Three further genes were eliminated due to non-available TaqMan W Assay-on-Demand probes and primer sets (Applied Biosystems). From the remaining 202 probeIDs three consecutive predictive models were built using the uncorrelated shrunken centroids (USC) [16] approach with default parameters, implemented in the MultiExperiment Viewer (MeV) [17]. This methods selects uncorrelated genes which best discriminate the two groups in internal cross-validation. Since the method picks only one gene from a group of several highly correlated genes, and this selection may be arbitrarily affected by small-sample variation, we repeated the method twice each time excluding the genes found in the previous step. This iterative approach leads to a richer set of candidate genes for further analyses. Microarray data are accessible on the Gene Expression Omnibus (GEO) under GEO accession: GSE31682.

Evaluation of microarray results by RT-qPCR
The microarray gene expression measurements of the selected genes were validated by real time RT-qPCR. cDNA was synthesized from 1 μg total RNA using the M-MLV reverse transcriptase (Promega, Madison, WI, USA) and a random nonamer primer. For normalization three stably expressed genes were selected from all 63 microarrays and all genes with signal-to-noise ratios greater than 3 in all samples (8,318     Area under the receiver operating characteristic (ROC) curves (AUCs) for all six models built from blood based expression values and/or plasma based protein abundances as derived from cohort 2 (for key metrics see Figure 2 and Table 6). calculated for each sample and this normalizing samplespecific constant was subtracted from each measurement of sample to obtain normalized (delta-CT) values. Delta-CT values were finally multiplied by −1 to be interpretable as log 2 -expression values.

Determination of the six-protein panel
The abundances of the six proteins (MIF, prolactin, CA125, leptin, osteopondin, and IGF2) from the cancer biomarker panel [11] were determined from the plasma samples according to the MILLIPLEX MAP Kit

Statistical analysis and model building
Differences in mean age between the five clinically defined groups (Table 1) were assessed by analysis of variance (ANOVA), followed by Tukey's post hoc tests. Significant up-or down-regulation of the expression of the 13 genes (AP2A1, B4GALT1, C1orf63, CCR2, CFP, DIS3, NEAT1, NOXA1, OSM, PAPOLG, PRIC285, ZNF419, and BC037918) and the 6 proteins between healthy controls and patients with malignant disease For selection the log 2 expression values from 20 genes were compared between samples from healthy patients and patients with malignant tumors by the significance analysis of microarrays (SAM) procedure, employing the t statistic and using R's samr package [18]. 13 Genes with q-values less than 0.15 were finally selected for model building with data from cohort 1. To this end the expression of these genes were determined by RT-qPCR in all 239 malignant (including the 44 ovarian cancer patients from the microarray experiment), 90 healthy (including 13 of the 19 controls from the microarray experiment), and 14 low-malignant potential or benign samples. Gene expression values were normalized as described above, and an L 1 penalized logistic regression model, also known as LASSO, which retained all 13 genes was estimated to obtain a model discriminating between the healthy and diseased groups [19].
Unfortunately, the plasma samples from the original 90 healthy controls were not available and therefore a further cohort of 65 controls (30 healthy blood donors and 35 patients with benign gynecological diseases) was enrolled in the study (cohort 2). The expressions of the 13 genes and the abundances of the six proteins were determined as described above. Using these two groups, one comprised of 224 EOC patients (for the remaining 15 EOC samples, no plasma samples were available) and one comprised of 65 controls (cohort 2), models using either gene expression values or protein abundance values alone or both in combination were built by means of L1 and L2 penalized logistic regressions, also known as LASSO and ridge regression, respectively (cf. Figure 1 for ROCs). Both models impose a penalty on the regression coefficients such that the sum of their absolute values (L1) or the sum of their squared values (L2) does not exceed a threshold value λ. The optimal value of the tuning parameter λ is found by maximizing the leave-one-out cross-validated likelihood. While L1 penalized models may set some regression coefficients exactly to zero, thus selecting a subset of the variables as predictors, L2 models always include all variables. The glmpath R package was used for computing the L1 and L2 models. To assess the differences of the obtained discriminatory models, likelihood ratio tests were performed.

Bootstrap validation
The misclassification error rate and the cross-validated receiver operating characteristic curve were estimated using the bootstrap .632+ cross-validation procedure [20].

Results
Gene expression based biomarkers Figure 2 outlines the gene selection and model building procedure for the mRNA-expression based genes. Starting from 202 genes preselected as described above, three consecutive uncorrelated shrunken centroid (USC) models were built, comprised of 7, 14, and 6 genes, respectively. Expressions of these 27 genes were validated in 63 samples using RT-qPCR with corresponding Assay-on-Demand TaqMan W probes ( Table 2) and a set of three stably expressed genes as normalizers, selected also from the microarray data. Seven of these 27 failed the validation step, because these genes showed no expressions in the 63   samples, indicating microarray artifacts or problems with the Assay-on-Demand TaqMan W probes (Table 2). A further selection step by Significance Analysis of Microarrays (SAM) selected 13 of the remaining 20 genes with qvalues ≤ 0.15 (Table 2). Normalized RT-qPCR expression values of these 13 genes were determined from all 343 samples of cohort 1. Regulation levels for each FIGO group, FIGO I/II and FIGO III/ IV, are shown in Table 3A. Five genes were significantly down-regulated in the leukocytes fraction of FIGO I/II and FIGO III/IV EOC patients compared to 90 healthy blood donors, AP2A1, B4GALT1, CFP, OSM, and PRIC285. One further gene was significantly down-regulated only in FIGO III/IV EOC patients, NOXA1. In addition, two genes were significantly up-regulated in FIGO III/IV EOC patients but not in FIGO I/II EOC patients, namely CCR2 and DIS3.
The expression of five genes was associated with higher probability of EOC ( Figure 3A), two of them nonsignificantly (DIS3 and ZNF419), and eight genes were negatively correlated with the probability of EOC. Using L1 penalized logistic regression, a predictive model was built to discriminate between healthy blood donors as controls and the 239 EOC patients. The model selected all 13 genes including the genes which were not significantly different in the univariate analyses (Table 2). CFP was the only gene whose predictive value changed from its negative direction in the univariate analysis to a positive contribution in the L1 penalized multivariable logistic model.
Since the healthy donors were significantly younger than the EOC patients (Table 1), we investigated whether the risk score from the L1 penalized logistic regression model (i. e., the sum of each subject's gene expressions weighted by the L1 model coefficients) was correlated to age. This was not the case, as confirmed by irrelevant correlation coefficients of the risk score with age of 0.083 (p = 0.449) in healthy donors and 0.104 (p = 0.111) in EOC patients, which indicates clearly the independence of our models from the impact of age on diagnosis of EOC.
The same model discriminated FIGO I + II patients from controls with a sensitivity of 74% at a specificity set at 99% ( Figure 3D, AUC = 0.905, CI 95% 0.781-1.000, Table 4). However, our model could not discriminate well between healthy controls and patients with benign or LMP tumors (AUC = 0.658, p = 0.058). Nevertheless, malignant tumors were distinguished from benign or LMP tumors with a sensitivity of 87% at a specificity fixed at 95% (AUC = 0.939, CI 95% 0.902-0.976) ( Figure 3E, Table 4) and even FIGO I + II EOC tumors were different from benign or LMP tumors with an AUC of 0.853 (CI 95% 0.719-0.987) ( Figure 3F, Table 4). Substantial differences for histological types or grades for all tumors and FIGO I + II stage tumors were not obvious, taking into account the small number of observations in some groups.

Combination with plasma protein abundance-based biomarkers
To combine the information of the 13 expression based biomarkers with plasma protein biomarkers, the abundances of six proteins from a known cancer biomarker panel were determined from 224 EOC-plasma samples and from 65 controls (cohort 2) using a commercially available Luminexbased multiplex assay (Figures 2 and 4). In Table 5 the coefficients of the L1 and L2 penalized models, in Figure 2 the corresponding AUC-values, and in Figure 1 the ROC-curves are shown. In Table 6 the characteristics of the two regression models (L1 and the L2 penalized)-are tabularized using the combination of both types of biomarkers. The discriminatory models built from the 13 expression based biomarkers combined with the plasma protein biomarkers proved to be significantly better than the models built from the plasma protein biomarkers alone (p < 0.0001, likelihood ratio test).

Bootstrap validation
The ability of the two combined models to discriminate cancer patients from healthy controls (ROC analysis), and their classification errors were estimated using bootstrap .632+ validation, simulating external validation by resampling. This corrects for the over optimism that would result from an internal validation of our results ( Table 6). The L1 model, comprised of five gene expression and five protein abundance based values (excluding osteopontin), proved to be slightly more sensitive (97.8% compared to 95.6% at a given specificity of 99.6%). The L2 model, using all 13 gene expression and all six protein abundance values, resulted in less misclassification (bootstrap .632+ crossvalidated classification error of 2.8% vs. 3.1%).

Discussion
In this study, the combination of gene expression values with a serum protein biomarker panel significantly increased the capacity to distinguish between EOC patients and controls. Serum proteins used for serum based tests are thought to be derived from the tumor-microenvironment and are therefore directly correlated with the amount of tumor mass. We speculate that among others, differences in leukocytes expressions, representing the systemic status of the immune system, are also driven by the malignant processes. Therefore, discrimination between benign and malignant tumors could probably be easier using leukocyte expression patterns than with only serum protein patterns, especially to detect patients with early EOC stages.
Applying a whole genome transcriptomics approach, we identified gene expression patterns of 7 or 13 genes in a leukocytes fraction from peripheral blood, discriminating healthy controls and patients with benign diseases from EOC patients. We are the first to show that this discrimination can be performed with an AUC of 0.984 (CI 95% 0.972-0.996) and 0.987 (CI 95% 0.976-0.997), respectively. We reached a sensitivity of 88% at a specificity set at 99%. A limitation of this study is that our models were not tested with other cancer entities and thus not enough evidence for cancer-type specificity can be provided. Furthermore, patients with other diseases, i.e. diseases which are inflammatory active like arthritis, should be included in a furtherlargercontrol cohort. Therefore, the term 'specificity' is only related to the statistical differentiation between the controls and the ovarian cancer patients of this study. The diagnostic power of this gene expression patterns is similar or even stronger to marker panels derived from serum proteins [7,[10][11][12]. Furthermore, our gene expression model can distinguish benign or LMP tumors from malignant tumors with a rather high sensitivity and specificity (87% and 95%, respectively). Only Zhang at al. [7] had tested their multi-marker serum panel for the discriminatory potential between benign or low malignant  potential tumors and malignant tumors, with sensitivities and specificities in the range of 33% to 50% and 33% to 45%, respectively. Combining the expressions of the 13 genes that we have identified with the protein abundance values from a commercially available serum protein biomarker panel significantly increases our predictive model. A model comprised of five gene expression values and five protein abundance values showed a sensitivity of 97.8 at a specificity of 99.6%. The high sensitivity and specificity reached by our models highlight a possible applicability of our combined model as a diagnostic test in high-risk individuals or as second test in combination with a CA-125/transvaginal sonography-based screening approach. The bootstrap .632+ validated classification error for this model was 3.1%. Our models were not tested with other cancer entities and thus our study does not provide enough evidence for cancer-type specificity. Hence, our use of the term 'specificity' relates to the statistical differentiation between the controls and the ovarian cancer patients of this study only. Nevertheless, we think that the combination of a sensitive blood gene expression test (even if it is not cancer typeor evencancer specific) with a cancer type specific protein test provides in combination both, a high sensitivity and a high specificity.
The functional specifications of the 13 genes that we identified are widespread among the pool of functional clusters and pathways, which is not a big surprise given the model building approaches used for generating the discriminative models, i.e. methods which exclude correlated genes from the final model explicitly (Table 3). Nevertheless, four genes are involved directly in inflammatory response and the immune system (B4GALT1, CCR2, CFP, and OSM), and two of them in the JAK/STAT pathway (CCR2 and OSM), known to be a common signaling pathway used by many cytokines [21]. Two genes seem to be non-coding RNAs (NEAT1 and transcript BC037918), presumably involved in regulation of transcription. The other protein functions are completely incoherent: one is a zinc-finger protein (ZNF419) of unknown function, one a poly(A) polymerase (PAPOLG), one a coactivator for several nuclear receptors like PPARA, PPARG, TR-beta-1, ER-alpha, and RXR-alpha (PRIC285), one a activator of catabolic NADPH oxidases (NOXA1), one is an RNase enzyme and can be part of the exosome complex (DIS3), and one is involved in the assembly of clathrin coated vesicles (AP2A1). From one transcript, C1orf63, no homologue protein is known.
The next step will be to validate the combined gene expression-and protein abundance-based predictive model using an independent large cohort of various controls and patients including patients with a systematic inflammatory status and including a larger sample of patients with FIGO I/II stages, which is an apparent shortcoming of this study.

Conclusion
The combination of two different types of biomarker signatures, one derived from blood plasma and one derived from the peripheral immune system, improved the discriminative power between control persons and ovarian cancer patients significantly, compared to the two single signatures alone. The concept of combining different types of biomarker (signatures) for one diagnostic or prognostic test opens new avenues, particularly by expanding this concept to further types of blood based biomarker, e.g. derived from circulating tumor cells or cell-free nucleic acids and involving genetic, epigenetic, or microRNA associated biomarker (signatures).