Large-Scale Serial Analysis of Gene Expression Reveals Genes Differentially Expressed in Ovarian Cancer

Cell Culture and Tissue Samples.

Ovarian cancer cell lines OV1063, ES2, and MDAH 2774 were obtained from the American Type Culture Collection (Manassas, VA). Cell lines A222, AD10, UCI101, and UCI107 were obtained from Dr. Michael Birrer (NCI, Rockville, MD). Cell line A2780 was obtained from Dr. Vilhelm Bohr (National Institute on Aging, Baltimore, MD). The SV40-immortalized cell lines IOSE29 (14) and ML10 (15) were kindly provided by Drs. Nelly Auersperg (University of British Columbia, British Columbia, Canada) and Louis Dubeau (University of Southern California, Los Angeles, CA), respectively. Except for IOSE29, ML-10, and HOSE-4, all cell lines were cultured in McCoy’s 5A growth medium (Life Technologies, Inc, Gaithersburg, MD) supplemented with 10% FBS and antibiotics (100 units/ml penicillin and 100 μg/ml streptomycin). IOSE29 was cultivated in Medium 199 (Life Technologies, Inc, Gaithersburg, MD) supplemented with 5% newborn calf serum. ML10 was cultivated in MEM (Life Technologies, Inc, Gaithersburg, MD) supplemented with 10% FBS and antibiotics as above.

Three high-grade serous ovarian cancer specimens OVT6, OVT7, and OVT8, composed of at least 80% tumor cells as determined by histopathology, were chosen for SAGE. The ovarian tumor samples were frozen immediately after surgical resection and were obtained form the Johns Hopkins gynecological tumor bank in accordance with institutional guidelines on the use of human tissue. Normal human ovarian surface epithelial (HOSE-4) cells were cultured from the right ovary of a patient undergoing hysterectomy and bilateral salpingo-oophorectomy for benign disease. The OSE cells were obtained by gently scraping the surface of the ovary with a cytobrush and were grown for two passages in RPMI 1640 supplemented with 10% FBS and 10 μg/ml epidermal growth factor.

SAGE.

Total RNA was obtained from guanidinium isothiocyanate cell lysates by centrifugation on CsCl. Polyadenylated mRNA was purified from total RNA using the Messagemaker kit (Life Technologies), and the cDNA was generated using the cDNA Synthesis System (Life Technologies). For the “POOL” library, 100 μg of total RNA from each of 10 ovarian cancer cell lines (A222, A2780, AD10, BG-1, ES-2, MDAH 2774, OVCA432, OV1063, UCI101, and UCI107) were combined, and the mRNA was purified. SAGE was performed essentially as described (12) for all of the libraries except HOSE. To create the HOSE library, MicroSAGE, a modified SAGE technique developed for limited sample sizes (16) , was used. Approximately 1 × 10⁶ OSE cells in short-term culture were lysed, and the mRNA was purified directly using Oligo (dT)₂₅ Dynabeads (Dynal, Oslo, Norway). As part of the CGAP SAGE consortium, the SAGE libraries were arrayed at the Lawrence Livermore National Laboratories and sequenced at the Washington University Human Genome Center or the National Institute Sequencing Center (NIH, Bethesda, MD). The data have been posted on the CGAP website 4 as part of the SAGEmap database (17) .

Sequence data from each library were analyzed by the SAGE software (12) to quantify tags and identify their corresponding transcripts. The data for the colon libraries NC1, NC2, Tu98, Tu102, HCT116, and SW837 were obtained from the SAGEmap database and analyzed in the same way. Because the different libraries contained various numbers of total tags, normalization (to 100,000 tags) was performed to allow meaningful comparisons. The 10,000 most highly expressed genes in the 16 SAGE libraries of interest were formatted in a Microsoft Excel spreadsheet, and Pearson correlation coefficients were calculated for each pair-wise comparison using normalized tag values for each library. The value for the Pearson correlation coefficient (r) represents the degree of similarity (the strength of the relationship) between two libraries and is calculated using the following equation:

where, x_i is the number of tags per 100,000 for tag i in the first library and y_i is the number of tags per 100,000 for tag i in the second library. For our purposes, n equals 10,000 because 10,000 tags are compared. A dendrogram representing the hierarchical relationships between samples was then generated using hierarchical cluster analysis as described (18) . In addition, the identification of differentially expressed genes was also done using this subset of the SAGE data.

IHC.

Deparafinized 5-μm sections of formalin-fixed ovarian cancer specimens were submitted to heat-induced antigen retrieval and processed using the LSAB2 system (DAKO, Carpinteria, CA) with 3,3′-diaminobenzidine as the chromatogen and a hematoxylin counterstain. Monoclonal antibody against ApoJ/clusterin (clone CLI-9) was obtained from Alexis Corporation (San Diego, CA) and used at a 1:500 dilution. Monoclonal antibody against Ep-CAM (clone 323/A3) from NeoMarkers (Fremont, CA) was used at a 1:500 dilution. Polyclonal antibodies against claudin-3 and -4 were a generous gift from Drs. M. Furuse and S. Tsukita (Kyoto University, Kyoto, Japan) and were used at a dilution of 1:1000.

Ovarian SAGE Library Construction and Analysis.

Gene expression alterations that arise during malignant transformation can be identified a number of ways. We chose the unbiased, comprehensive method SAGE to create global gene expression profiles from 10 different ovarian sources. The expression patterns are generated by sequencing thousands of short sequence tags that contain sufficient information to uniquely identify the corresponding transcripts (12) . Ten different SAGE libraries were constructed and sequenced for this study (Table 1) ⇓ . Our libraries included two derived from OSE cells (IOSE29 and HOSE-4), one derived from immortalized cystadenoma cells (ML-10), three primary tumors (OVT6, -7, and -8), and four libraries derived from ovarian cancer cell lines (OV1063, ES-2, A2780, and a pool of cell lines). Almost 20,000 sequencing reactions were performed, yielding a total of 384,497 tags of which 82,533 were unique. Accounting for a SAGE tag error rate of 6.8% (attributable to sequencing errors; see Ref. 13 ), we estimate that we have identified a total of 56,387 genes expressed in ovarian tissues. Except for the A2780 cell line and the pooled lines (POOL) sample, a minimum of 12,000 genes were obtained from every library. Typically, for each library, 10% of the genes were expressed at levels of ≥0.01%, and collectively, these genes accounted for >50% of all of the tags sequenced. Among the tags that appeared more than once, up to 95% matched to known sequences in the current GenBank database. For example, of the 6637 tags that appeared more than once in ML10, only 311 had no matches in the current database, excluding the EST databases. The complete expression profiles for the 10 libraries described here are available from the authors or from the CGAP website. 4

Comparisons of Global Gene Expression between Ovarian Tissue Samples.

Although progression to malignancy requires a number of gene expression changes, the transcript levels from the vast majority of genes remain unaltered (13 , 19) . Similarities between the global expression profiles of two given samples can be readily visualized using scatterplots and quantitated through the calculation of Pearson correlation coefficients (Fig. 1A) ⇓ . As expected, the immortalized IOSE29 and ovarian cystadenoma strain ML10 are much more similar to ovarian tumors than to colon tumors (average correlation coefficients of 0.70 versus 0.51, respectively; Fig. 1B ⇓ ). In addition, IOSE29 and ML10 are very similar to each other, with a correlation coefficient of 0.82. The primary culture of OSE cells (HOSE-4) exhibited higher similarities to the ovarian tumors than to the colon tumors, although the similarity levels were much lower than those observed for IOSE29. Interestingly, HOSE-4 and IOSE29 appear to be much more distantly related than expected considering the fact that they were both derived from “normal” OSE cells. The differences in gene expression between these cells may be attributable to a number of factors. The age of the patient, the pathological state of the ovaries, the presence of nonepithelial cells in the culture, and the fact that IOSE29 is SV40-immortalized may all contribute to the differences in gene expression observed. However, it is unlikely that the main differences are attributable to SV40-immortalization because IOSE29 is much more similar to normal colon (a non-SV40-immortalized epithelium) than HOSE-4. It is, of course, possible that the lower degree of similarity between HOSE-4 and the ovarian tumors compared with IOSE29 and ML-10 reflects the fact that HOSE-4 represents a better approximation of the normal in vivo OSE cell. A comparative immunochemical study of carefully chosen gene products in normal ovarian sections and IOSE29, HOSE-4, and ML-10 cells should allow the determination of the optimal model for normal OSE.

• Download figure
• Open in new tab
• Download powerpoint

Fig. 1.

Global gene expression analysis. A, scatterplots of IOSE versus ML10, OVT6 (ovarian), and Tu98 (colon) generated using the Spotfire Pro 4.0 software (Cambridge, MA). The tag frequency for each tag (per 100,000) is plotted on a logarithmic scale for the indicated libraries. B, Pearson correlation coefficients were calculated for each pairwise comparison of the 16 ovarian and colon SAGE libraries. C, dendrograms were created from hierarchical cluster analysis (18) of all colon and ovarian SAGE libraries (left), ovarian samples only (middle), and nonmalignant ovarian and colon epithelia as well as ovarian and colon primary tumors (right). Normalized values for the 10,000 most highly expressed genes in each library were used for all manipulations.

The Cluster software (18) was used to generate the dendograms shown in Fig. 1C ⇓ . When all of the samples from Fig. 1B ⇓ were included in the hierarchical clustering analysis, the primary colon tumors clustered with the normal colon epithelium, but colon cell lines clustered with the ovarian specimens. Clearly, the tissue clustering that was readily apparent when comparing primary tissues or immortalized lines was lost when including carcinoma cell lines. For example, A2780, a widely used ovarian cancer cell line was just as similar to colon cancer cell lines as it was to ovarian cancer cell lines. This observation supports the idea that in the process of establishment, cell lines may lose many of the gene expression characteristics of their tissue of origin, although tissue-specific expression is clearly not completely lost in cancer cell lines (20) . In addition, the origin of a particular cell line can be difficult to determine with certainty. The origins of widely used ovarian cancer cell lines SW626 and OV1063 have recently been questioned. American Type Culture Collection reported the presence of a Y chromosome in OV1063, and SW626 displays several characteristics of colon cancer, including an adenomatous polyposis coli mutation (21) . For these reasons, cancer cell lines may not represent the best model for tissue-specific gene expression studies.

It is widely believed that epithelial ovarian cancer and benign ovarian cysts, although not necessarily part of a progression sequence toward malignancy, are both derived from the OSE (22) . OSE cells themselves are mesodermal in origin and are believed to undergo metaplasia before progressing to neoplasia (22 , 23) . On the other hand, it has also been argued that ovarian cancers are not derived from OSE but rather from the secondary Mullerian system, structures lined by Mullerian epithelium but located outside the uterus, cervix, and fallopian tubes (3) . This hypothesis would explain some of the shortcomings of the OSE model, such as the requirement for metaplasia and the lack of well-defined precursors in the ovary. In any event, our results are consistent with the widely accepted dogma of the OSE origin of ovarian cancer. Indeed, IOSE29 showed high degrees of similarity to the ovarian tumors, and both IOSE29 and HOSE were much more closely related to ovarian than colon primary cancers.

E-Cadherin expression has been proposed to be a major determinant in the formation of metaplastic OSE (14 , 23) . Consistent with this hypothesis, E-cadherin was absent in IOSE29, HOSE, and ML10 but was expressed in all three ovarian tumors (Table 2) ⇓ . Other cadherins are also shown for comparison. Interestingly, VE-cadherin is absent in most libraries except in two of the preneoplastic ovarian samples, again suggesting metaplasia. As expected, LI-cadherin was expressed exclusively in the colon-derived libraries. Interestingly, vimentin, a mesenchymal marker, was present in essentially all of the ovarian libraries but was very low in the colon specimens. Although the specificity of vimentin as a mesenchymal marker has been questioned, this suggests that OSE cells may retain some of their mesenchymal characteristics, even after the expression of E-cadherin is turned on.

The CKs and CEA have been used to differentiate between colon cancer and ovarian cancer (24 , 25) . Typically, colon cancer expresses CK20 and CEA, whereas ovarian cancer expresses CK7. The expression patterns in our libraries were consistent with previously reported observations: CK20 and CEA were found in normal colon and colon tumors but absent from all of our ovarian samples (Table 2) ⇓ . Conversely, CK7 was expressed in all three primary ovarian tumors and, although not absent, was much lower in the colon samples. Examination of the differential expression patterns of a variety of established ovarian cancer markers thus provided validation of the SAGE database and cluster analysis.

Differential Gene Expression.

The ultimate goal of comparing SAGE libraries is to identify differentially expressed genes. Criteria for differential expression can be determined for each comparison and transcripts within the determined range selected for study. We found a large number of genes that were up-regulated in only one or two of the three tumors on which SAGE was performed. For example, a total of 444 genes were up-regulated> 10-fold in at least one of the three ovarian primary cancers compared with IOSE29. However, only 45 genes were overexpressed >10-fold in all three ovarian tumors analyzed compared with IOSE29. This tumor heterogeneity is not unexpected but emphasizes the importance of analyzing multiple specimens for gene expression studies. Our analysis of three different primary ovarian cancers allowed us to reduce the number of candidates by looking for consistency between samples.

To identify genes that are very likely to be frequently up-regulated during ovarian tumorigenesis, we set the following conservative criteria for our analysis. First, the fold induction was calculated by adding the number of normalized tags from the three primary tumors and dividing this number by the total normalized tags in the three nonmalignant specimens. Cell lines were not included here for reasons described above. In addition, although HOSE-4 appeared more distantly related to the other nontransformed specimens, we believe that the inclusion of HOSE-4, although possibly eliminating real candidates, makes our analysis more conservative and more likely to identify truly overexpressed genes in ovarian cancer. Second, all three primary tumors were required to consistently show elevated levels (>12 tags/100,000) of the gene in question. This eliminated genes that may be very highly overexpressed in one tumor but not in others. Finally, the candidate genes were required to be expressed in at least one ovarian cell line at a level greater than 3 tags/100,000. This last criterion was used to reduce the possibility of identifying genes because of their high level of expression in inflammatory cells or in the stroma of the primary tumors. Using these criteria, we identified the genes that exhibited more than 10-fold overexpression (Table 3) ⇓ .

Interestingly, two members of the claudin family of tight junction proteins, claudin-3 and -4, were found among the top six differentially expressed genes and likely represent transmembrane receptors (26) . Another claudin family member, claudin-7, was recently found to be overexpressed in breast cancer, using SAGE (27) , but the functions of the claudin proteins in cancer remain unclear. In addition, ApoJ and ApoE were both overexpressed in ovarian cancer. These two lipophilic proteins are involved in lipid homeostasis and have been implicated in Alzheimer’s disease (28) . It will be interesting to elucidate the function of these proteins in ovarian cancer. Of the 27 overexpressed genes shown in Table 3 ⇓ , 14 were relatively specific for the ovary (HLA-DR, two different ESTs, GA733-1, ceruloplasmin, glutathione peroxidase-3, the secretory leukocyte protease inhibitor, ApoJ, ApoE, complement 1r, BCAM, Mal, HLA-DPB1, and mesothelin), whereas the others were also expressed in colon tissues. In any event, it is significant that MUC1, HE4, Ep-CAM, and mesothelin, four genes already known to be up-regulated in epithelial ovarian cancer (11 , 29, 30, 31) , were identified in this study. This fact validates our approach as well as our set of criteria used to determine the genes differentially expressed.

Similarly, stringent criteria were used to identify genes down-regulated in ovarian tumors compared with IOSE29, HOSE-4 and ML10. Again, the fold difference was calculated by adding tag frequency for all three “normal” specimens and dividing by the total number of tags in the three ovarian tumors. A candidate was required to be expressed at a level of 12 tags/100,000 or greater in all three normal samples. The genes found elevated >10-fold in normal tissue compared with tumors are shown in Table 3 ⇓ . These proteins may be important for ovarian tissue homeostasis. Indeed, several of these proteins, (i.e., PAI-1, EMP-3, galectin-1, lysyl oxidase-like 2 and vinexin-β) have been implicated in apoptosis, proliferation, adhesion, or tissue maintenance. Interestingly, a tumor suppressor role has previously been suggested for lysyl oxidase because it is decreased in H-ras-transformed cells and up-regulated in spontaneous revertants of H-ras-transformed fibroblasts (32) . Lysyl oxidase-like 2 itself has been implicated in cellular adhesion and senescence (33) .

Validation of SAGE Data by IHC of Selected Candidates.

To validate the candidates identified by SAGE, we performed immunohistochemical analysis of 13 cases of serous cancer of the ovary, using antibodies against four of the genes identified as up-regulated in ovarian cancer (Table 3) ⇓ . This was particularly important because the SAGE analysis was initially performed from primary ovarian cancers, which contain a mixture of cell types. Ep-CAM exhibited diffuse strong staining of all 13 tumors without blood cell or stromal staining (Fig. 2) ⇓ . Importantly, only one of six samples of the OSE present in the cases showed weak focal staining, and the rest were negative. The strong immunoreactivity of all 13 ovarian tumors confirms the validity of our approach to identify genes highly and consistently up-regulated in ovarian cancer. Similarly, ApoJ was found to be expressed in ovarian cancer cells and absent from the surface epithelium (Fig. 2) ⇓ . Although some expression was detected in non-tumor stroma and inflammatory cells, most of the immunoreactivity was in tumor cells, and a majority (9 of 13) of the cases showed staining. This represents the first report of ApoJ expression in ovarian cancer and may represent a novel target for diagnosis or therapy. Claudin-3 and -4 also exhibited staining limited to the tumor component of the specimens. Most tumor cells showed strong membrane staining with weak cytoplasmic reactivity (Fig. 2) ⇓ . Some tumor specimens showed decreased membrane staining with strong cytoplasmic reactivity. Interestingly, it has been shown that deregulation of the mitogen-activated protein kinase pathway can lead to mislocalization of tight junction proteins, including claudin-1 (34) . The normal surface epithelial component (or mesothelial cells) examined did not stain or stained only weakly with the claudin-4 antibody, whereas the determination of claudin-3 levels in normal epithelium was complicated by a low background reactivity with this antibody.

• Download figure
• Open in new tab
• Download powerpoint

Fig. 2.

Immunohistochemical staining of serous carcinomas. A–F, two representative cases are illustrated. A–C, case 1. Sections are stained as follows: A, H&E; B, Ep-CAM; and C, ApoJ. Tumor cells show strong, diffuse membrane staining for Ep-CAM and diffuse cytoplasmic staining for ApoJ. D–F, case 2. Sections as stained as follows: D, H&E; E, Ep-CAM; and F, ApoJ. Tumor implant on ovarian surface shows strong, diffuse membrane staining for Ep-CAM in tumor cells, whereas adjacent OSE is negative. ApoJ expression is more focal than in case 1 and localizes to the luminal aspect of tumor cells. G–I, staining patterns of claudin-3 and -4. Claudin-4 shows strong membrane staining on tumor cells. Adjacent detached normal epithelium (G) is negative. Claudin-3 shows predominantly membrane staining in the tumor illustrated in H and shows both cytoplasmic and membrane staining in the tumor in I. Overlying mesothelium in I is negative.

Human tumors are highly heterogeneous, even within a tissue type. This heterogeneity likely arises from the increased genomic instability of tumor cells. This instability is responsible for driving the cells through the tumorigenesis process. Although the defects induced in unstable cells are random, the changes found in cancers may not be completely random, as the changes favorable to the tumorigenesis process are selected for. For these reasons, although we expect a high level of heterogeneity in ovarian cancer, it is our hypothesis that there will be consistent changes that reflect the inherent physiological and pathological conditions of ovarian cancer progression. To eliminate changes that may be attributable to the random nature of the tumorigenesis process, we used strict and conservative criteria to identify genes that were both consistently and highly differentially regulated in ovarian cancer. Consequently, although we may be missing genes that are relevant to ovarian cancer, we believe that the genes in Table 2 ⇓ represent a list of genes relevant to the tumorigenesis process. Three of the candidates reported here and validated by IHC, ApoJ, claudin-3 and claudin-4, have never been previously linked to ovarian cancer. At present, we are evaluating our numerous candidates, including novel genes, using a variety of techniques such as quantitative reverse transcription-PCR and IHC. It will be interesting to identify novel genes that are specifically expressed in ovarian cancer and tissues. The complete expression profiles for all of the samples reported here are publicly available through SAGEmap (17) and will likely stimulate and facilitate further research in the field.