Analysis of Gene Expression in Early-Stage Ovarian Cancer

Purpose: Gene expression profile was analyzed in 68 stage I and 15 borderline ovarian cancers to determine if different clinical features of stage I ovarian cancer such as histotype, grade, and survival are related to differential gene expression.

Experimental Design: Tumors were obtained directly at surgery and immediately frozen in liquid nitrogen until analysis. Glass arrays containing 16,000 genes were used in a dual-color assay labeling protocol.

Results: Unsupervised analysis identified eight major patient partitions, one of which was statistically associated to overall survival, grading, and histotype and another with grading and histotype. Supervised analysis allowed detection of gene profiles clearly associated to histotype or to degree of differentiation. No difference was found between borderline and grade 1 tumors. As to recurrence, a subset of genes able to differentiate relapsers from nonrelapsers was identified. Among these, cyclin E and minichromosome maintenance protein 5 were found particularly relevant, as their expression was inversely correlated to progression-free survival (P = 0.00033 and 0.017, respectively).

Conclusions: Specific molecular signatures define different histotypes and prognosis of stage I ovarian cancer. Mucinous and clear cells histotypes can be distinguished from the others regardless of tumor grade. Cyclin E and minichromosome maintenance protein 5, whose expression was found previously to be related to a bad prognosis of advanced ovarian cancer, appear to be potential prognostic markers in stage I ovarian cancer too, independent of other pathologic and clinical variables.

In spite of much recent preclinical and clinical efforts to identify novel markers and investigate new therapies, epithelial ovarian cancer (EOC) remains the fourth most common cause of cancer-related death in women in the western world, with ∼70% of patients dying within 5 years of diagnosis (1–4). This high death rate results from late diagnosis, when the disease has spread beyond the pelvis (stage III and IV). In contrast, when EOC is diagnosed at stage I, when it is still organ confined, the 5-year survival rate exceeds 90%. Unfortunately, only ∼10% of all ovarian cancers are diagnosed at this early stage (5–7). Fewer than 20% to 25% of stage I patients relapse within 5 years from primary surgical-medical therapy. A similar percentage of stage I patients present with a more aggressive disease, which is difficult to predict based on currently known clinicopathologic features. These facts highlight the need for better tools for screening and staging ovarian cancer.

The exploitation of the knowledge of EOC genetic changes to benefit early disease detection or specific therapy has been limited (8, 9). Predictions based on changes in single genes or sets of genes have yet to be used in the clinic, and most of the genetic insights have been gained in small numbers of patients that do not allow firm conclusions (10, 11). The situation is complicated by the heterogeneity among ovarian cancer subtypes. Based on morphologic criteria, EOC is classified into four major subtypes: serous, mucinous, endometrioid, and clear cell. Each class is further divided into benign, malignant, and borderline with different degrees of differentiation (2). Recent studies have shown that molecular defects appear to differ among the most common types of EOC, supporting the notion that they represent histopathologically, genetically, and biologically distinct, albeit related, diseases (8, 12). The histologic grade of the tumor is probably the variable that is most significantly associated with the prognosis of stage I ovarian carcinoma, but its predictive value is far from being satisfactory (13).

Many groups have applied cDNA microarray technology to the study of advanced EOC or cell lines derived from EOC (14–18), but very little is known about stage I and borderline carcinomas, which, although relatively rare, are potentially more useful sources of information concerning early events associated with tumor pathogenesis. This lack of information prompted us to test the hypothesis that cDNA microarray analysis can detect properties of stage I and borderline tumors that might not be discovered by the commonly used clinical or histopathologic analyses at diagnosis. Using a single center tumor bank of >1,200 ovarian carcinoma biopsies frozen immediately after surgical removal, 68 cases of stage I and 15 borderline tumors were selected to assess the expression of 16,000 genes. The aim of the present study was to determine if different clinical features of stage I ovarian cancer such as histotype, grade, and survival are related to differential gene expression.

A total of 83 patients were recruited among the 1,200 samples stored into the frozen tissue bank between September 1992 and March 2003. Sixty-eight were diagnosed as having stage I ovarian cancer and 15 as borderline. The histology, grade, and stage of each tumor are listed in Table 1. Tumors were staged according to the International Federation of Gynecology and Obstetrics criteria; 63% of the stage I cases (43 of 68) were treated after surgery with platinum-based therapy (carboplatin, cisplatin, or CDBCA + paclitaxel), whereas the remaining 37% (25 of 68) did not receive adjuvant therapy. One of 15 borderline tumors (serous, substage c) received carboplatin therapy, whereas the remaining were untreated (Table 1). Ovarian tumor samples were collected in the operating theater and frozen within 15 min in liquid nitrogen after surgery and stored at -80°C. The tumor content of the specimens was assessed by H&E stain at San Gerardo anatomopathologic unit. Only specimens containing >70% of tumor were used for these experiments. The collection and use of tumor samples was approved by the local scientific ethical committee and written consent was obtained from the patients.

Frozen primary tumor tissues (30 mg) were homogenized in a Ultra-turrax at 4°C and total RNA was purified using SVTotal RNA isolation system according to the manufacturer's instructions (Promega). RNA was measured by spectrophotometer and quality assessed by a Bioanalyzer (Agilent). Only samples with a RNA quality score (RIN) larger than 5 were further processed and aliquots were stored at −80°C until use.

Real-time quantitative reverse transcription-PCR (qRT-PCR) was used to validate the differential expression of selected genes in RNA samples from primary EOC. Automatic liquid handling station (epMotion 5075 LH; Eppendorf) and real-time PCR (ABI-7900; Applied Biosystems) were used as described (19, 20). The comparative threshold cycle (C_t) method was used for the calculation of amplification fold as specified by the manufacturer. Differences among histologic types in qRT-PCR expression data were analyzed by Welch t test. Arrays and qRT-PCR data correlation were assessed by Spearman's ρ and test. Primers pair sequences were designed as described (20) and their sequences were reported in Supplementary Table S1. Actin B and cyclophilin A were used as internal control.

Gene expression was analyzed using Agilent slides onto which 16,000 cDNAs were spotted (Agilent). Human Atlas control RNA from a pool of human cell lines was used as a reference for the whole experimental protocol (BD Biosciences). The T7 amplification was done according to Message Amp aRNA kit (Ambion) and cRNA was reverse transcribed in Cy3/Cy5 cDNA (GEH) using the amino allyl cDNA labeling kit (Ambion). For each slide, Cy3-labeled samples were compared with their Cy5-labeled controls and labels were swapped to increase the robustness of our data.

Arrays were scanned using a laser scanner Axon 4000-B (Axon Instruments) using 5 μm resolution and with photomultiplier variables that defined the quanta intensity density distribution of the array to give a Cy3/Cy5 ratio close to 1.0. Separate 16-bit images were acquired for Cy3 and Cy5. Two similar but not equivalent procedures were adopted for image analysis and data processing. Details of the two procedures are attached as Supplementary Data. Briefly, the first method consisted in analyzing images with Spot and processing data with R (Procedure 1), whereas the second one used GenePixPro (Axon Instruments) for image analysis and the Rosetta Resolver SE software (Rosetta Biosoftware) for data processing and analysis (Procedure 2). Raw and processed data are available in the Gene Expression Omnibus repository (GSE8841 and GSE8842).⁴

Available clinical and pathologic variables were tested for association to gene expression profiles in a supervised approach (a priori defining classes to compare) and in an unsupervised one (clustering based on expression data only; ref. 21).

Unsupervised analysis: cluster reproducibility and statistical correlation to clinical variables. Self-organizing maps analysis (as described in Supplementary Data) and hierarchical clustering (with average linkage and Euclidean distance) were applied to cluster samples. The reproducibility of clustering results was assessed using BRB-Array Tools version 3.3.0.⁵

Measures of reproducibility (R) and discrepancy (D) were used to evaluate the robustness of the clusters as explained in the program manual (22).

The unsupervised analysis was focused on the detection of clusters, their correlation with known prognostic factors, and the prognostic role of each cluster in terms of overall survival (OS; primary endpoint) and progression-free survival (PFS) benefit. Correlations between each cluster and known prognostic variables (substage, grading, and histology) were described with relative and absolute frequencies and analyzed with the χ² test for association or trend, as appropriate. Cumulative OS and PFS curves were constructed as time-to-event plots by the Kaplan-Meier method. Comparison of the curves was done using log-rank statistics. Hazard ratio was calculated for each cluster by using the Cox model. Results are reported as hazard ratio estimates with 95% confidence intervals and P value at Wald's test. A hazard ratio of 1 denotes the absence of a difference between the two arms (or the two categories of a compared covariate), whereas a hazard ratio of >1 or <1 denotes an increase or decrease in the risk in a given patient group with respect to the reference. The proportional hazards assumption implied by Cox's model was checked by analysis of Schoenfeld residuals. All the computations were made using SAS software.

Supervised analysis. Supervised analysis, which directly links gene expression and clinical variables, was accomplished using BRB-Array Tools, PAM (23), Significance Analysis of Microarrays (24), and the Rosetta Resolver SE software.

(a) Class comparison. With BRB-Array Tools, genes that were differentially expressed in two or more classes were identified by using a multivariate permutation test (25) to provide 95% confidence interval that the false discovery rate was ≤10%. We also applied the random variance model (26) and performed the global test of significance (with default algorithm settings) described in the program manual. One-way error-weighted ANOVA with Benjamini and Hochberg false discovery rate method for multiple test correction implemented on the Resolver System was applied to extract differentially expressed genes between two or more classes. This algorithm was applied to expression data obtained with the second analysis method. Significance Analysis of Microarrays⁶

was also applied on patient Z scores (analysis method 1) and log ratios (analysis method 2) to extract differentially expressed genes. For the latter, the k-nn input algorithm available within R was adopted to complete missing values in the gene per patient matrix before submission to Significance Analysis of Microarrays. When there were two classes, the “two-class unpaired data” option was applied, and when there were more than two, the “multiclass” option was chosen.

(b) Correlation with survival outcomes. Using BRB-Array Tools, we searched for genes whose expression was significantly related to survival. Algorithm settings were the same as in the class comparison case. OS and PFS analysis was also done by using the “survival” option in the SAM package using the ratio experiment matrix as input and specifying each patient's survival time and status. Association between gene expression and survival was described using Kaplan-Meier curves and quantified using hazard ratio computed by the Cox model as summary measure.

(c) Class prediction. Models to predict the class of samples with unknown class, using gene expression profiles, were developed. These were based on different algorithms (27, 28) and incorporated genes differentially expressed at different significance levels as assessed by the random variance t test or F test. The level of significance (among 0.01, 0.005, 0.001, and 0.0005) is automatically selected in a fully cross-validated way for each algorithm as a tuning variable value. We estimated the prediction error of each model using two methods: cross-validation (leave-one-out and 10-fold coefficient of variance) and bootstrap (0.632 bootstrap) as described in refs. 29, 30. For each coefficient of variance (and bootstrap) training set, we repeated the entire model building process, including gene selection. We also evaluated whether the cross-validated error rate estimate for a model was significantly less than one would expect from random prediction as reported in BRB-Array Tools manual.

(d) Gene ontology and pathways analysis. Gene ontology (GO) groups of genes whose expression was differentially regulated among classes were identified using the “Gene set comparison” tool in BRB-Array Tools with default settings. We considered a GO category significantly differentially regulated if significance levels of both the Fisher and the Kolmogorov-Smirnov statistics were <0.005 (31).⁷

Pathway analysis is similar to the GO analysis, except that genes are grouped by Kegg or BioCarta pathways. Functional annotation of gene lists was done using DAVID.⁸

Comparison of the gene lists and correlation. Gene lists obtained by applying supervised procedures on the expression data from analysis methods 1 and 2 were tested for correlation if their intersections contained at least two genes. The rankings of differential statistics in the original lists were taken as input for the algorithm. A nonparametric correlation test was done in R⁹

by the cor test function in the stats package and using both Spearman's ρ and Kendall's τ to find correlations statistically different from 0 (32).

Agilent human arrays with 16,000 cDNAs were employed to identify global gene expression changes in 68 stage I EOC biopsies and in 15 borderline tumors. Table 1 shows histopathologic features and chemotherapeutic regimens of the cohort of patients enrolled in this study. Cox proportional hazards model shows that our series appears to be representative of stage I ovarian cancer, as PFS and OS are similar to those reported in literature according to substage, grading, and histotype (Supplementary Table S2A and S2B; ref. 2). Patients enrolled in the study described here did not harbor any mutation in the p53 gene (20).

The use of two different preprocessing procedures, which gave Z scores and log ratios, respectively, did not markedly affect unsupervised or supervised analysis outputs (see Supplementary Data). Reported here are results that showed high consistency within the two procedures.

Molecular profiling of ovarian cancer subtypes. Hierarchical cluster analysis on a subset of 1,783 cDNAs obtained by feature filtering based on percentage of missing values and SD (Fig. 1) shows that the 83 samples are overall molecularly homogeneous. The selected genes can be roughly subdivided into three distinct subsets: slightly differentially expressed compared with the universal reference (light green and light red spots) at the top of the figure, strongly down-regulated (green spots) in the middle, and strongly up-regulated at the bottom. Notably, the clear cell and mucinous histotypes are readily distinguished from the serous, endometrioid, and undifferentiated ones, which appear to be intermingled among themselves regardless of tumor grade. In particular, 10 of 16 clear cells, grade 3, show an overlapping signature. Borderline samples are scattered around and cocluster with stage I tumors regardless of their grade.

To unravel the complexity of the molecular fingerprint of the 68 stage I tumors, self-organizing maps analysis was applied to a gene expression matrix of 1,116 rows (similar filtering criteria as above), which yielded 34 patient partitions showing statistically significant associations to clinical variables (P = 0.001-0.0001; see Supplementary Table S3).

We focused our attention on two partitions characterized by genes up-regulated in the clear cell grade 3 subgroup, named partition 1 (Fig. 2A), or in the mucinous subtypes, named partition 2 (Fig. 2B). To support the robustness of our observations, we validated partition 1 signature by qRT-PCR in an independent set of total RNA purified from the same cohort of patients. Table 2 shows a representative panel of observed differences in 6 genes between the clear cells and the other histotypes analyzed using microarray or qRT-PCR. With a P value threshold of 0.05 and a ρ value ranging from 0.4 to 0.8, in 75% of the analyzed samples, the results obtained by qRT-PCR mirror those gained by microarray in both qualitative and quantitative terms. This suggests that most array probe sets are likely to accurately measure single transcript levels within a complex mixture of transcripts.

Partition 2 is characterized by a subset of 202 genes, the expression of which was found consistently modified in 6 mucinous tumors. To ascertain whether particular functional categories could be overrepresented, GO enrichment analysis was applied to the differentially regulated probe sets identified for mucinous histotypes. This provided a list of GO categories having more genes differentially expressed between mucinous and nonmucinous histotypes than expected by chance. DAVID functional annotation, according to GO molecular functions level 4, yielded an enrichment of 6 functions (P < 0.05; Fig. 2C). Attention was focused on genes involved in “metabolism of xenobiotics by cytochrome P450” and “signal transduction through the interleukin-1 receptor pathway.” In 70% of the selected genes, differences in gene expression between mucinous and nonmucinous subtypes were confirmed by qRT-PCR in the same cohort of patients (Table 3). The mucinous histotype was characterized by an increase in the expression of different families of genes involved in the metabolisms of xenobiotics as the cytochrome P450 family (CYP2S1 and CYP4F12), glutathione metabolism isozymes (MGST2), nuclear orphan receptor (NR1I2), and NADH dehydrogenase quinone 1 (NQO1). These differences were also confirmed in an independent cohort of stage I patients (“test” set) selected from our tumor bank. Of these, only 9 of 21 tumors were diagnosed as mucinous (Supplementary Table S4). Data obtained in this test set (Table 3, right column) confirmed the trend in mean differences obtained previously in the original cohort of patients, although the P value did not reach significance. Nevertheless, the tumor necrosis factor gene, which belongs to the interleukin-1 receptor signal transduction pathway, was confirmed to be strongly down-regulated in the mucinous histotype compared with the nonmucinous one in both data sets.

Relationship between gene expression patterns and histopathology. Supervised analysis was done to define molecular features associated to histopathologic variables. Analyses were carried out on the complete set of cDNAs without any kind of feature filtering. This allowed detection of cDNAs with small variance among samples but homogeneous behavior within predefined sample categories. Figure 3 shows a multidimensional scaling (MDS) three-dimensional scatter plot for 67 stage I samples, all but one with undifferentiated histotype, built on 692 genes obtained by one-way ANOVA on histotypes. In particular, it can be noted that serous (red) and endometrioid samples (blue) have a much lower distance between each other than clear cells (cyan) or mucinous samples (green).

Further supervised analysis was done by class comparison between grade 1 and grade 3 tumors. Hierarchical clustering applied to 128 genes shared between results from two different class comparison algorithms is shown in Fig. 4A. Grade 1 (red) are separated from grade 3 (cyan) stage I tumors. Only 1 of 13 grade 1 samples coclusters with grade 3 tumors. Functional annotation done on these 128 genes revealed that 6 of 10 biological processes with significant enrichment score are related to cell cycle regulation, supporting the notion that changes in expression profiles of genes involved in cellular proliferation account for different histologic grade within stage I (Fig. 4B).

To refine and improve the quality of the comparison among histologic grades, grade 2 tumors were included into the analysis and the borderlines were compared with the other grades. The MDS three-dimensional scatter plot built on 139 genes obtained from one-way ANOVA on grades 1 to 3 (Fig. 4C) shows that grade 2 tumors (blue) are spread between grade 1 (green) and grade 3 (red), which are distributed at the two edges of the plot. Interestingly, the 15 borderlines (cyan) overlap grade 1 samples and are separated from the more undifferentiated tumors. Statistical analysis confirmed the MDS scatter plot results.

Gene expression pattern in relapsed patients. Class comparison was applied to the 68 stage I patients to discriminate the recurrence status of the samples. One hundred ninety-one genes distinguish relapsed samples from those who did not relapse. The MDS scatter plot built on these genes is shown in Fig. 5A, whereas Fig. 5B reports the GO biological processes with significant DAVID score (P < 0.05) obtained using Homo sapiens as background. The complete set of genes subdivided for the seven GO biological processes is reported in Supplementary Table 5. Attention was focused on genes involved in cell cycle control and chromosome organization, particularly on CCNE1 (cyclin E) and minichromosome maintenance protein 5 (MCM5), a member of the minichromosome maintenance complex. The log₂ ratio and −ΔΔC_t values for the group mean differences were 1.12 and 1.16, respectively, for cyclin E and 0.76 and 0.44 for MCM5. Both qRT-PCR and array data confirmed that CCNE1 and MCM5 were significantly (P < 0.05) up-regulated in relapsing patients compared with nonrelapsing ones (Table 4). Absence of relapsing patients in our 21 test sample set confounded confirmation of this observation in an independent cohort of patients.

Survival analysis. Survival analysis was done in 68 patients. Recurrence occurred in 18 patients, 13 (19%) of which died for progression. We correlated the expression levels of CCNE1 and MCM5 with OS and PFS, plotted with the Kaplan-Meier method (Fig. 6). Even if the survival curves for CCNE1 and MCM5 did not show a significant difference of OS (P = 0.203 and 0.295, respectively), patients with expression levels of CCNE1 higher than 0.124 had a worst prognosis than those with levels lower than 0.124. The power test (1 - β = 80%) suggested that a cohort of at least 2,675 patients should be analyzed to reach significance of 95%. However, when expression levels of CCNE1 and MCM5 were correlated with PFS, patients with expression levels higher than 0.34 and 0.58, respectively, had a shorter PFS than patients with lower levels, and such differences reached statistically significance (P = 0.0003 and 0.018 for CCNE1 and MCM5, respectively). Patients were stratified according to chemotherapy and results reported in Supplementary Fig. S1 confirm similar differences in both chemotherapy-treated and untreated group. To assess the effect of each variable on PFS and OS, a Cox's model was done using univariate analysis. Variables of interest were gene expression (CCNE1 and MCM5), grading, histotype, and chemotherapy. In addition, a further Cox's model was done according to the level of expression of CCNE1 and MCM5 gene, adjusted for chemotherapy. Grading was found associated with both OS and PFS (P = 0.0098 and 0.046 for PFS and OS, respectively), whereas gene expression of CCNE1 was found significantly associated with PFS only (P = 0.0069). Chemotherapy alone did not affect OS and PFS in the whole sample (P = 0.09 and 0.066 for OS and PFS, respectively) or in stratified patients according to CCNE1 or MCM5 levels (P = 0.059 and 0.08 for PFS and P = 0.52 and 0.56 for OS, respectively; Supplementary Table S6, sections a and b). A multivariate Cox's regression model for PFS was built using variables that were found significant on univariate analysis. In this case, both CCNE1 and grading were not statistically significant (Supplementary Table S6, section c).

The data presented above represent to our knowledge the first gene profiling study focused on stage I ovarian cancer. Our results show that gene expression profiles are related to both tumor pathologic features and patients' survival. As to tumor morphology, it should be noted that the histologic origin of the different ovarian cancer subtypes is still a matter of debate. According to the most recent studies, it appears that clear cell, endometrioid, and mucinous EOCs arise probably from the metaplastic Mullerian epithelium rather than directly from ovarian surface celomatic sheets, which is more likely in the case of serous carcinomas (12). The strong correlation of stage IC and clear cell histotype with inferior prognosis in our series is consistent with previous reports indicating that the gene profiling approach adopted here was done on a series adequately representing stage I ovarian cancer (2). Morphologic studies suggest there are at least two distinct ovarian cancer subtypes, defined as weakly and highly malignant with different pathogenesis. Consistent with this definition, the study presented here indicates that changes between the different histologic subtypes are associated with specific gene expression patterns regardless of the substage and grade of the disease.

Unsupervised and supervised analysis of these homogeneous samples identified a sizeable number of genes associated with grading and histotype. Regardless of the tumor grade, mucinous and clear cells are largely separable and distinct from the other histotypes. Partition 1, which strongly identifies patients with a poor prognosis, is also associated with the clear cell histotype. Partition 2 strongly associates with the mucinous type. The nature and function of the specific genes associated with these clusters needs thorough analysis to identify those most likely to predict whether a patient belongs to one group or another. The identification of several clear cell or mucinous specific markers constitutes the basis for future studies in which some of these biomarkers can be tested for their suitability to serve in diagnosis. For example, the overexpression of genes coding for insulin-like growth factor binding proteins and its intracellular adaptor protein GRB14 in clear cell tumors may explain the poor prognosis associated with this histotype compared with other types of EOC. The insulin-like growth factor binding protein family of proteins is often associated with increased tumor aggressiveness (33, 34) and has been tentatively linked to resistance against cisplatin treatment (35).

Previous studies done in advanced ovarian cancer have shown that the gene expression profile in mucinous histotype cancer differs from that in other histotypes (36). The present study confirms this finding and shows that clear-cut differences between histotypes other than the mucinous are based on the expression of oxidative stress response genes (CYP2S1, CYP4F12, MGST2, NQO1, and NR1I2). The fact that the tumor necrosis factor gene was down-regulated in the mucinous histotype may be clinically significant. The contribution of tumor necrosis factor to human ovarian cancer progression has long been recognized (37, 38). Models predict that constitutive tumor necrosis factor-α production by tumor cells may generate and sustain a tumor-promoting cytokine network in the ovarian cancer microenvironment (39). The present study suggests that this cytokine is less expressed in mucinous tumors, a finding possibly relevant for the selection of patients undergoing therapeutic approaches against this target.

As observed previously for late-stage EOC (40) also within stage I, grade 1 tumors profile was similar to that of borderline tumors, both being different from grade 2 or 3 tumors. This is a novel observation that might justify similar clinical management of borderline and grade 1 stage I ovarian cancer perhaps obviating postsurgical chemotherapy.

As far as the relationship between gene profile and prognosis, class comparison assessing differential expression between relapsers and nonrelapsers in conjunction with GO analysis showed a different expression of genes involved in cell cycle regulation (those associated with cytoskeleton organization and biogenesis, regulation, and progression through the cell cycle and regulation of nucleotide metabolism). One of the most interesting finding of our analysis is the inverse relationship between the expression of cyclin E and MCM5 on one side and patients' disease-free survival on the other. These differences were found in both patients treated with chemotherapy and untreated patients, supporting the idea that, in our cohort of patients, the prognostic value of MCM5 and CCNE1 is not affected by treatment.

The importance of cyclin E as a predictor of tumor progression and poor prognosis of ovarian cancer patients has been reported previously (41–43). Cyclin E levels were assessed by immunohistochemistry in archived paraffin-embedded tissue blocks from a relative large series of patients, the majority of whom were at an advanced stage. The present study confirms this observation in stage I ovarian cancer and corroborates the idea that cyclin E could be an important prognostic marker in the management of ovarian cancer (41, 44). Less information is available in the literature on MCM5, a protein recently found to be a good proliferation marker related to the prognosis in different tumors (45–48). Very recently, a study done in ovarian cancer indicated a strong correlation between the expression of MCM2 and MCM5 and Ki-67 LI and p53 levels, presumably associated to mutations of this protein, as well as with patients' survival (49). Because all our cases expressed wild-type p53, it appears that the relevance of MCM5 is independent of p53 status and might be of prognostic value also in patients affected by stage I ovarian cancer. Studies on a larger cohort of patients are mandatory to confirm our observation.

In conclusion, our study shows the genetic diversity of ovarian cancer according to histotype and degree of differentiation as reflected by their profile signature. Evaluation of tumor gene profiling can provide immensely useful information germane to prognosis in early-stage ovarian cancer. Cyclin E and MCM5 expression levels appear to be promising as predictors of disease-free survival in stage I ovarian cancer.

No potential conflicts of interest were disclosed.

We thank the Nerina and Mario Mattioli Foundation, Italian Association for Cancer Research, and Luigi and Francesca Lucchetti, for the generous contribution and Prof. Andreas Gescher for insightful comments, suggestions, and critical review of the article.