The lack of a standard approach to study the cancer transcriptome as well as in analysis of gene expression is a major impediment to formulating hypotheses about the etiology and effective treatment of cancer (for review, see Gillet et al. (14
)). The variety of gene expression profiling platforms and normalization processes that have been utilized by various researchers renders an integrative computational and analytical approach extremely challenging (16
). A second drawback linked to the technology is insufficient specificity and sensitivity of the assay utilized (14
). This is a major problem, especially in studies investigating multidrug resistance mechanisms, as many of the mediators belong to highly homologous gene superfamilies. We and others recently showed that Taqman low-density arrays (TLDAs) provide the highest sensitivity and specificity in measuring ABC transporter gene-expression patterns, a superfamily of 48 highly homologous members, initially studied in the NCI-60 cancer cell line panel (8
). Therefore, we chose to configure such a platform to study the expression profiles of multidrug resistance-associated genes, selected through the literature published over the last thirty years, in clinical specimens and to assess their utility in survival prediction.
We found that the expression signature of a group of 11 genes statistically improves the prognostic power of four clinical covariates (age, stage, residual tumor status after cytoreductive surgery and CA-125 level) for overall survival. The predictive ability of the 11-gene signature with the four covariates is estimated using a leave-one-out cross-validation scheme. Note that the cross-validation is only used as a method to estimate the predictive ability. The final model (which would potentially be used on other samples or as a clinical diagnostic) is created using all the samples, without any samples left out. This final model is the one that contains the 11 genes.
To date, three clinical covariates, age, stage, and residual tumor status after cytoreductive surgery, have been validated as prognostic variables for ovarian cancer. Another variable, the presence of elevated serum levels of a protein known as CA-125 (the product of the MUC16
gene) is being used here as another prognostic factor. Although CA-125 has proven useful in the clinic to predict response to treatment and to detect the recurrence of ovarian cancer, the independent prognostic power of this antigen is more controversial (13
). The specificity of the test for CA-125 is poor, as a number of benign and malignant conditions may result in falsely elevated CA-125 values, and while it is found to be elevated in the serum of approximately 80% of the patients with advanced stage ovarian cancer, it is found in only 50% of the patients with early stage disease. When CA-125 is removed from the statistical analysis, the addition of gene expression data does not result in any improvement in the prediction of overall survival (see supplementary Fig. 2
), indicating that this marker adds prognostic power to our cohort, which was composed exclusively of advanced-stage (III and IV) ovarian cancer samples.
Clinically, 75 – 80% of the patients that present with ovarian cancer are stage III or IV. Some of these patients are much more responsive to treatment than others. If clinicians could have a better handle on the molecular profile of patients that will or will not respond to standard chemotherapy (if they know upfront which patients are likely not to respond to the standard chemotherapeutic regimens), they could provide them with an alternative initial treatment. For that reason, we attempted to better categorize the high and low risk patient groups predicted by the covariates only into more specific risk groups (i.e., high-high, high-low, low-high, low-low) using the expression levels of the 11 genes. We observed that the highest and lowest risk patient groups are confirmed by adding the gene expression signature. Remarkably, we also found that patients considered as high risk by clinical covariates have a better prognosis than the low risk patient group if the expression of the 11 genes is low. Similarly, the low risk patient group identified by clinical covariates has a worse prognosis if these patients highly express those 11 genes. Although it will be necessary to repeat this analysis in an independent set of samples from ovarian cancer patients, the statistical approach used in this work argues strongly that any appropriate sample size of patients at a similar disease stage, treated similarly, should yield a similar signature. In a much smaller group of 23 ovarian cancer patients from the Norwegian Radium Hospital, we were unable to confirm the 11-gene signature as a predictor of poor response to chemotherapy, but this is likely due to the small sample size and the differences in diagnosis and treatment between the United States and Norway (data not shown). The data also indicate that our gene expression profile alone is insufficient to provide significant prognostic or predictive information. Others have successfully correlated gene expression profiles with either overall or progression-free survival (18
). However, these gene signatures, identified through whole genome microarrays, exhibit very little overlap. The lack of similarity does not necessarily preclude the value or robustness of these signatures, and can be explained in part by the technical limitations of microarray analysis previously mentioned as well as by the heterogeneity among the various cohorts analyzed. For example, Birrer and colleagues uncovered two prognostic signatures using two different strategies (19
). One of the signatures was found to be relevant for suboptimally debulked patients based on the analysis of 185 untreated late-stage serous ovarian cancer patients (19
). Later, another prognostic signature was discovered from the study of 53 laser-captured microdissected samples from untreated late-stage serous ovarian cancer patients, further validated in 64 additional samples (20
). Jazaeri et al
. investigated 45 samples of serous ovarian cancer composed of 21 chemosensitive samples (defined as those with a complete response to chemotherapy and a platinum-free interval of ≥13 months) and 24 chemoresistant samples including patients who had either surgical debulking following chemotherapy (i.e., neoadjuvant chemotherapy) or residual tumor at the time of a second-look procedure (21
). Berchuck et al
. highlighted a predictive signature studying 101 samples composed of early stage, late stage and borderline cases of serous ovarian cancer (18
), while Spentzos and colleagues studied a cohort of 68 samples mostly composed of stage III and IV serous ovarian cancer (22
). To our knowledge, none of these predictors, some of which were poorly characterized open reading frames (ORFs), have yet been validated in prospective studies.
The 11-gene signature identified in the current study includes some genes for which a role as markers of poor prognosis in ovarian cancer had previously been suggested, but is still unclear. The EGFR and MAPK3 (ERK1) genes may affect numerous cell processes as mediators in cell signalling. A recent meta-analysis of 15 studies addressing the role of EGFR in ovarian cancer showed considerable publication bias and concluded that this receptor is unlikely to be useful as a prognostic marker in clinical practice (23
). However, two other studies showed that the use of EGFR inhibitors stabilized disease in 11–44% of patients and produced objective regression in 4–6% (24
). It is known that MAPK3 (ERK1) is downstream from EGFR and activated mainly through Ras. Therefore, MAPK3 is a relevant target for patients in whom inhibitors against both EGFR and Ras are ineffective due to mutations in those proteins. Matrix Metallopeptidase 9 (MMP9), which is involved in cell invasion/adhesion, was found to be expressed in surgical specimens of ovarian cancer (26
). Three genes are linked to the apoptotic pathway; BAG3
is an anti-apoptotic mediator, FASL and BNIP3
are pro-apoptotic markers. BNIP3 is related to the BH3-only family, which induces both cell death and autophagy (27
). This gene was found to be strongly correlated with poor prognosis in non-small cell lung cancer (28
). Further analysis showed BNIP sequestration in the nucleus, which was confirmed in primary glioblastoma multiforme tumors (29
). Another study reported the inhibition of BNIP3 cell death activity by growth factors such as EGF and IGF in epithelial cells (30
), which is consistent with the presence of EGFR in our 11-gene signature. This mechanism of inhibition may be explained by the conflicting signals promoting cell death and the survival mediators that may allow the selection of tumor cells that survive chronic BNIP3 overexpression. Fas ligand (FASL) induces apoptosis through the binding of its receptor, FAS. Recent studies have shown that Fas loss is associated with bad prognosis in patients with ovarian cancer (31
). The phase II metabolism enzyme Glutathione peroxidase 3 (GPX3) was found to be highly expressed in clear cell adenocarcinoma cell lines. Its down-regulation in these cells increased cisplatin sensitivity 3- to 4-fold (33
). TAP is a heterodimeric protein complex consisting of TAP1 and 2 (ABCB2-B3) subunits, which transports antigens into the ER lumen for subsequent loading onto major histocompatibility complex (MHC) class I molecules (34
). A recent study reported the upregulation of these genes in approximately 70% of the specimens profiled from 150 patients with invasive epithelial ovarian cancers (35
). Defects in the antigen processing machinery (APM) may allow tumor cells to escape immune recognition. The ITGAE gene included in our signature was also in a gene signature that predicted platinum resistance in a set of 72 ovarian tumors (36
). The current study confirms that expression of the genes mentioned above predicts poor outcome in patients with ovarian cancer, providing independent verification of their importance in this type of cancer.
Some of the eleven genes in our signature had not previously been linked to ovarian cancer. To our knowledge, there have been no previous reports on the role of either S100A10
in ovarian cancer. The tumor suppressor adenomatosis polyposis coli (APC) has been confirmed to participate in the Wnt signaling pathway by downregulating beta-catenin and thereby controlling gene transcription and cell proliferation. Mutations inactivating the APC
gene are found in approximately 80% of all human colon tumors (37
). S100 proteins regulate cellular differentiation, energy metabolism, cytoskeletal membrane interactions and cell-cycle progression (38
). They are commonly upregulated in tumors of epithelial origin. For example, S100A4
upregulation was found to be associated with an unfavourable outcome in early breast cancer (39
) and its nuclear expression in advanced-stage ovarian carcinoma was associated with poor response to chemotherapy and with worse overall survival (40
). Another example is S100A7
, which was found to be upregulated in high-grade ductal carcinoma in situ (DCIS) and some invasive breast carcinomas (41
). Although one study reported the down-regulation of APC
in both primary and metastatic serous ovarian carcinomas (43
), the role of both APC
as markers in ovarian cancer has yet to be elucidated. These genes, whose prognostic significance was previously unreported in regards to ovarian cancer, provide potential new targets for improved therapy.
We also sought to understand why the signature did not predict progression-free survival, the clinical surrogate for drug-resistant cancer. Although a high percentage of our patients (57%) failed to respond to carboplatin, indicating the advanced state of disease in these banked tumor specimens, we must acknowledge that a relatively small number of samples analyzed in our study were from patients presenting with progressive disease (22.5%), while more than half (56%) of the samples came from those who exhibited a complete response. This significant imbalance may explain the failure to find a statistically significant predictive gene signature. We must also recognize the dramatic effect of chemotherapy on gene expression profiling, which could not be assessed due to the lack of availability of samples taken after recurrence of the disease. Finally, we can exclude neither the potential role of any of the 380 genes in intrinsic drug resistance in individual primary ovarian serous carcinomas, nor the possibility of an undetected acquired drug resistance signature. The inter-individual specificity in gene expression, as well as the relevance of the 380 genes to platinum and taxane agents was confirmed experimentally in a study investigating 32 unpaired ovarian serous carcinoma effusion samples obtained at diagnosis or at disease recurrence following chemotherapy. Our analyses demonstrated that gene expression alone can effectively predict the survival outcome of women with ovarian serous carcinoma (OS: log-rank p=0.0000 and PFS: log-rank p=0.002) (10
In summary, our study reveals that the expression level of 11 genes associated with MDR define an intrinsic drug resistance signature that significantly improves prediction of overall survival in patients with ovarian cancer compared to predictions based on clinical covariates only. The potential role of these genes in drug resistance or alternatively, how they support the malignant phenotype and/or growth rate of the tumor remains to be determined ().
Genes included in the gene signature, their functions and association with ovarian cancer