Multiple, dissimilar genetic defects in cancers of the same origin contribute to heterogeneity in tumor phenotypes and therapeutic responses of patients, yet the associated molecular mechanisms remain elusive. Here, we show at the systems level that serous ovarian carcinoma is marked by the activation of interconnected modules associated with a specific gene set that was derived from three independent tumor-specific gene expression data sets. Network prediction algorithms combined with preestablished protein interaction networks and known functionalities affirmed the importance of genes associated with ovarian cancer as predictive biomarkers, besides “discovering” novel ones purely on the basis of interconnectivity, whose precise involvement remains to be investigated. Copy number alterations and aberrant epigenetic regulation were identified and validated as significant influences on gene expression. More importantly, three functional modules centering on c-Myc activation, altered retinoblastoma signaling, and p53/cell cycle/DNA damage repair pathways have been identified for their involvement in transformation-associated events. Further studies will assign significance to and aid the design of a panel of specific markers predictive of individual- and tumor-specific pathways. In the parlance of this emerging field, such networks of gene-hub interactions may define personalized therapeutic decisions. Cancer Res; 70(12); 4809–19. ©2010 AACR.

Therapeutic decisions in oncology are based on correlations between tumor characteristics and possibility of disease relapse (1). Limitations of such approaches, however, are now leading to the development of therapies considering individual-specific genetic defects. The resolution of breast cancer into four molecular gene expression–based classes represents a successful outcome of such approaches. Besides suggesting distinct cell origins, these classes correlate well with histologic grading and clinical characteristics (2). Over the last decade, ovarian cancer is realized to represent a group of histologically distinct diseases that correlate with different origins (3) and, hence, require appropriation with individual “molecular signatures” (4). Gene expression profiling–based identification of prognostic and/or predictive biomarkers toward improving disease and therapy risk assessment along with a mechanistic understanding of gene interactions in pathways, networks, and/or complexes is desirable to unravel the biological behavior of tumors. Unfortunately, most studies are restricted by limited sample size and a minimal commonality of genes between different analyses. We present here an alternative analysis of gene expression data to extract a “signature” of commonly modulated genes derived from three independent data sets of serous ovarian carcinoma. Further, resolution of expression-based interaction networks of these genes and known protein-protein interactions (PPI) confirmed existing markers with predictive/prognostic value and “discovered” others. More importantly, the predicted network-based interactions and known functionality of the identified gene set provide novel insights toward a mechanistic understanding of cellular transformation processes.

Cells

We have earlier reported the establishment of a comprehensive in vitro panel of 19 isogenic cell lines from a patient with grade IV serous adenocarcinoma (5). Nontumorigenic A4 and transformed A4 (A4-T) cultures were cultured as described earlier (5).

RNA extraction and A4 microarray hybridization

Total RNA was isolated from cells using the Qiagen RNeasy minikit (Qiagen, Inc.) according to the manufacturer's protocol. Microarray hybridization was carried out at Agilent Genotypic Technologies. Briefly, samples were labeled using the Agilent Low Input RNA amplification kit and were hybridized to Agilent Human Whole Genome 4 × 44 k Array (AMADID: 14850) using the Agilent In situ Hybridiztion kit. Dye swapping was done for one of the replicates to minimize the noise introduced due to dye intensity. Data normalization was done in GeneSpring-GX using the recommended Per Spot and Per Chip: intensity-dependent (Lowess) normalization. After normalization, log ratios of each gene from the three replicates were averaged, and Student's t test was performed. Genes with a P value of <0.05 were extracted for further studies. Gene expression data were further collapsed so as to retain only a single probe per gene having significant P value. Genes with a (log2)fold change of >1 (transformed A4T versus nontransformed A4) were considered upregulated, whereas those with a (log2)fold change of <(−1) were considered downregulated. The remaining genes were categorized as stable genes.

Database description and data analyses

We compiled gene expression data from the following databases:

  1. TCGA, The Cancer Genome Atlas (6)

  2. IST, In Silico Transcriptomics (7).

TCGA is a comprehensive collaborative project that helps the science community to better understand the genomic changes associated with various cancers and is publicly accessible. TCGA stores only high-quality and completely annotated data. The main advantage of using the TCGA gene expression data set is its multidimensional data, i.e., it holds gene expressions for each single sample derived from different array origin, which is normalized, annotated, and validated for the expression variation relevance with the type of tissue rather than with type of array generation. Thus, the analysis of data from different array origins increases its robustness.

We extracted and analyzed level 2 ovarian serous cystadenocarcinoma gene expression data from two platforms, Affymetrix (169 tumor and 10 normal samples) and Agilent (216 tumor and 6 normal samples; February 2009 to June 2009). Level 2 data are obtained by processing raw expression data using different normalization algorithms, which involves equalization transformation (Quantile normalization), probe level, and gene level normalization specific for different platforms (e.g., Affymetrix-Robust Multiarray Average method, Agilent-Lowess Normalization). This processing helps in enabling a comparative analysis of data obtained from different experiments. These data are available in log2-transformed state. Differential gene expression (fold change) is calculated by comparing the expression value for each probe with average expression value in normal cells of respective array platforms. The genes thus identified were further classified in terms of upregulated and downregulated genes as described above.

The IST database archives normalized, quality checked, and annotated data, which reduces array generation–based variations in expression levels, while retaining different tissue type–based dependencies of differential gene expression. The IST data have been developed from a collection of publicly available gene expression data of over 9,000 human samples from >150 normal and diseased states including 124 to 141 ovarian adenocarcinomas. This data mining helps in the comprehensive analysis of the functional, clinical, and therapeutic roles of genes in different tissue and tumor types. The normalization methods included probe-level preprocessing (which removes ambiguous probes), equalization transformation (Quantile normalization), and array generation–based gene centering (which avoids bias introduced due to different array platform generation). These steps produced data comparable across the major array platforms by reducing the array generation–based variations in expression levels while retaining the different tissue type–based dependencies of the differential gene expression.

To evaluate the differential expression pattern of genes obtained from the A4 cell system (GSE18054), the corresponding gene expression profile across different tissue samples for tumor state and normal cell state, as well as correlation plots were retrieved from the IST database. Depending on the expression state of these genes (average fold change value of >2 across two data sets: tumor and normal) in the IST database, these genes are classified in terms of upregulated and downregulated genes.

A comparative analysis of gene expression obtained from multiple sources helps in reducing the dependency of expression variation originating from the noise of array expression analysis and improves robustness of the data. Overlaps of upregulated and downregulated genes extracted from the three databases, i.e., A4, IST, and TCGA gene expression data, provided us with a small pool of 30 commonly modulated genes, which had a consistent pattern of expression in all the data sets studied (Fig. 1). This was termed by us as the SeOvCa gene signature. The analytic pipeline of analyses to derive this signature is represented in Fig. 1, and gene lists of each data set are compiled in Supplementary Table S1.

Figure 1.

Derivation of the SeOvCa gene signature and their expression profiles in the three data sets.

Figure 1.

Derivation of the SeOvCa gene signature and their expression profiles in the three data sets.

Close modal

To further validate the significance of the common 30 genes in SeOvCa, the comparison of gene expression dynamics with clinical data is required. We retrieved raw expression data for the Gene Expression Omnibus entry GSE3149 that contains a total of 153 ovarian tumor samples in four tumor stages (8). The raw data for tumor samples were initially grouped based on the tumor stages and were normalized separately by Robust Multiarray Average normalization tools in the Expression Profiler available at the EBI server Probe. Annotations were retrieved from Gene Expression Omnibus, and each probe was mapped to an Entrez Gene Symbol by querying the accompanied public identifier in the UniGene database. Because corresponding expression profiles for normal tissue type were not available in the data set, relative gene expression level in terms of fold change in gene expression could not be determined. Hence, an analysis of comparative gene expression variation across different tumor stages was carried out in which gene expression values were normalized by array mean normalization. Further, the expression data for SeOvCa genes was analyzed as relative gene expression between the tumor grades that were assessed for a positive (increasing) gradient across the four grades for upregulated genes and a negative (decreasing) gradient for downregulated genes. Genes with average fold change difference of 0.8 to 1.3 across stage I to IV samples were considered as stably expressed genes.

Algorithm for the reconstruction of Accurate Cellular Networks (ARACNe) was used to predict the interaction network of SeOvCa genes and their potential interactors toward the delineation of complex regulatory networks (9). This utility was accessed through the open source platform, “geWorkbench” version 1.5.1. Level 2 TCGA data from both Affymetrix and Agilent platforms were uploaded and normalized using the “array-centered normalization option” available on geWorkbench. Each of the 30 SeOvCa genes were considered as “nodes,” whereas interacting partners (not a signature gene) were called “interactors.”

ARACNe is purely based on the expression data with no influence of gene ontology or other factors. The relationships are calculated using mutual information (MI), which is a measure of the statistical dependency between two variables, implying correlation between the two. Genes that are indirectly correlated through an intermediary interactor are also represented in these networks. False-positive interactions are removed after comparing MIs using the data processing inequality that effectively infers the most likely path (direct interaction) by eliminating indirect interactors by comparing their respective MIs. In the present study, the MI was fixed at 0.20, and the data processing inequality was set at 0.10. The interaction network was converted from probe based to gene based using the appropriate microarray platform annotations. The new network was represented in “Cytoscape” using the “Force directed Layout.”

Protein Interaction Network Analysis (PINA) was used to predict PPIs. PINA represents an integrated platform for protein interaction network construction, filtering, analyses, visualization, and management that integrates PPI data from public curated databases that were mined to generate the PPI networks (10).

Fluorescence in situ hybridization

Fluorescence in situ hybridization (FISH) analysis was carried out by a commercial cyotogenetic laboratory (Sahyadri Medical Genetics and Tissue Engineering Facility). Commercially available probes specific for the 6q21, 8q24, and 20q11 locus were used for the analyses. Five hundred cells were observed for every chromosomal locus.

ChIP-on-chip

Chromatin immunoprecipitation (ChIP) combined with microarray analysis was performed as described earlier (11) using the Agilent Human Promoter CoC 244 k (AMADID:19469) oligoarrays. The genes listed were screened for enrichment with a cutoff of 2-fold change in normalized log ratio. For further analysis, only those genes that had probe enrichment in at least two replicates of each experiment were retained. From this primary list of genes for each type of histone modification, a set of common genes having at least two different types of histone marks was identified. ChIPs and PCRs were performed as per standard protocols using anti-K4, anti-K9, and anti-K27 antibodies from Millipore.

Methyl DNA immunoprecipitation

Genome-wide promoter methylation analysis by methyl DNA immunoprecipitation was performed as described earlier (12) using the above oligoarrays. An enrichment criterion for probes that have at least 2-fold changes in normalized methylation intensity levels compared with the control sample was considered.

Immunoblotting

Immunobloting was performed as described (11). Specification of antibodies used is available on request.

Pathway analyses were carried out using the gene expression analysis tool of Protein Analysis through Evolutionary Relationships (13).

Statistical analysis

All experiments were carried out in triplicate; data are expressed as mean ± SEM of at least three independent experiments. The significance of difference in the mean values was determined using two-tailed Student's t test; P < 0.05 was considered significant.

Derivation and validation of a serous ovarian adenocarcinoma gene signature

Genes likely to be important in serous ovarian carcinoma were identified using three differential gene expression data sets: (a) A4 in vitro ovarian carcinoma progression cell model established earlier from a patient presenting with advanced grade IV serous adenocarcinoma (5, 11, 1416) that presents a pliable cell system for validating top-down data-driven analyses, (b) TCGA gene expression database for ovarian serous cystadenocarcinomas (TCGA), and (c) individual gene profiling in the IST serous adenocarcinoma database. Overlaps of differentially expressed genes from these data sets (Supplementary Table S1) led to the derivation of a prioritized list of 30 commonly regulated genes termed the SeOvCa signature as these genes expressed a consistent pattern (Fig. 1).

SeOvCa also validated in an independent expression data set (Gene Expression Omnibus accession no. GSE3149; ref. 8) associated with different grades of ovarian tumors. This identified similar expression gradients of 11 upregulated and 5 downregulated SeOvCa genes with increasing tumor stage, expression levels of eight other genes that were stable irrespective of tumor stage and could be either predictive of early alterations during transformation, or outliers (Supplementary Fig. S1); probes for the remaining six genes could not be identified.

The significance of SeOvCa genes was determined by curating published data from research literature (until November 30, 2009). This revealed the association of 11 SeOvCa genes with ovarian cancer (MAL, MCM2, MMP9, RRM2, SOX17, SYNCRIP, DAB2, FBN1, HNMT, KLF2, and SMARCA2), of 12 genes with cellular transformation events in other cancers (ATAD2, BCAT1, CDCA4, EXO1, LAMA5, MEST, SLC39A4, EFEMP1, LHFP, SGK1, PAPSS2, and PTGIS), and of 7 genes with no preidentified association with cancer (discovery group: TM7SF2, TNNT1, DIXDC1, GNB5, LRRC17, PROS1, and RNASE4).

System network prediction of SeOvCa interactions from gene expression data

Associations between SeOvCa genes were identified using a previously described algorithm (ARACNe; ref. 9) to the TCGA microarray data sets. ARACNe identifies statistically significant gene-gene coregulation by MI, an information-theoretical measure of relatedness that integrates data processing inequality to eliminate indirect relationships. The final reconstructed network effectively removes bias from partially known functional similarities between genes to reveal relationships with the highest probability of direct interactions. This allows the independent prediction of regulatory phenomena within a defined context even for genes without preestablished functionalities. Each SeOvCa gene was considered as an independent node to identify interactions at three levels:

  1. Node-node interactions. Strikingly, these were restricted to downregulated nodes (Fig. 2A), indicating repression to be more coregulated than activation.

  2. Node-linker-node interactions involving 23 nodes and 99 connecting linker genes that generate a network (Supplementary Table S2; Fig. 2B). Of these, DAB2 and HNMT interact exclusively with downregulated nodes, whereas the SOX17 upregulated hub distanced itself from the main network and interacted with two downregulated nodes. Node-linker networks can identify coregulatory modules; an example is the FBN1-LHFP node-node interaction supported by a large number of linkers, which is thus predicted as being strongly coregulated than with either PAPSS2 or LRRC17 (Fig. 2A).

  3. Node-interactor networks that encompass all possible interactions in the gene expression data. The resulting network has three types of nodes (Supplementary Data Set S1; Fig. 2C): (a) isolated nodes (MAL, GNB5, and TM7SF2) that do not associate with any gene in the data set, (b) stand-alone nodes (SLC39A4, MEST, LAMA5, and MMP9) that although distanced from the main network, maintain their influence on specific genes to form isolated or “stand-alone” hubs, and (c) social nodes, which are mostly SeOvCa genes that interact with other nodes, linkers, and exclusive interacting genes (interactors) to generate a complex network in which upregulated and downregulated nodes segregate distinctly. Edge analysis within this network revealed mutual exclusivity with positively correlating node-partner gene interactions being more frequent than negative correlations. A few specific negative correlations identified at the interface between 20 upregulated and downregulated SeOvCa nodes, and 13 linker genes create an interface network, within which linkers correlate positively with downregulated and negatively with upregulated nodes (Supplementary Fig. S2). Such interface networks may represent a central regulatory module, targeting which could be critical for effective therapy, as subtle perturbations within such inclusive hubs may have a pronounced effect on network stability.

Figure 2.

ARACNe generated SeOvCa interaction networks for (A) node-node, (B) node-linker-node, and (C) node-interacting gene interactions. Red and green circles, upregulated and downregulated nodes, respectively; dark gray nodes, interacting partners predicted by ARACNe. Red and green edges, positively and negatively correlating interacting partners; light blue edges, interacting partners that maintain a stable expression.

Figure 2.

ARACNe generated SeOvCa interaction networks for (A) node-node, (B) node-linker-node, and (C) node-interacting gene interactions. Red and green circles, upregulated and downregulated nodes, respectively; dark gray nodes, interacting partners predicted by ARACNe. Red and green edges, positively and negatively correlating interacting partners; light blue edges, interacting partners that maintain a stable expression.

Close modal

Copy number change prediction and identification

Chromosomal distribution of all SeOvCa genes and interactors considering localization to the same cytoband led to the identification of region-specific enriched/repressed hubs (Supplementary Table S3). To further increase prediction stringencies, only those hubs were considered wherein at least half the interactors were located in the same cytoband as the node and/or the interactors generated secondary networks within the same location (Supplementary Data Set S2; Fig. 3A).

Figure 3.

Prediction and validation of copy number changes. A, secondary ARACNe networks at 6q12-6q16.1, 6q21, 8q24.3, and 20q11.2-13.33. Red circles, nodes; the 6q21 hub is a subnetwork of SYNCRIP at 6q12-6q16.1. B, gene amplification frequencies at each region. C, representative FISH analyses indicating 6q21(rhodamine) and 8q24(FITC; top), and 20q11 (rhodamine) amplification.

Figure 3.

Prediction and validation of copy number changes. A, secondary ARACNe networks at 6q12-6q16.1, 6q21, 8q24.3, and 20q11.2-13.33. Red circles, nodes; the 6q21 hub is a subnetwork of SYNCRIP at 6q12-6q16.1. B, gene amplification frequencies at each region. C, representative FISH analyses indicating 6q21(rhodamine) and 8q24(FITC; top), and 20q11 (rhodamine) amplification.

Close modal

The 6q12-q21 associated gene cluster involves SYNCRIP with 19 of its 23 interactors around two loci at 6q12-q16.1 (node+17 interactors) and 6q21 (4 interactors). The secondary network at 6q12-16.1 contained 14 additional region-specific genes, whereas that at 6q21 had 12 region-specific genes. The 8q22.1-q24.3 associated gene cluster involves two loci: (a) ATAD2 with 16 of its 33 interactors (8q22.1-24.13) and (b) the entire SLC39A4 hub (8q24.3). Although the secondary network generated by ATAD2 did not include region-specific genes, the SLC39A4 secondary network enlisted 51 additional genes of which 31 were cluster specific. The 20q11.2-q13.33 associated gene cluster involves two upregulated nodes, i.e., matrix metalloproteinase 9 (MMP9) that does not generate a region-specific network and LAMA5 with its primary interactors ARFGAP1 and SS18L1 that generates a secondary network of 41 interactors of which 18 are region specific. Overall, the involvement of primary and secondary node components ranged from 17.5% to 41.94% of total genes in the specific chromosomal regions (Fig. 3B).

We validated the suggested amplifications in the A4 cell model using FISH (Fig. 3C). All three amplifications seem to be early events during transformation; however, regulatory mechanisms seemed to be in place in untransformed cells because the node gene expression was significantly lower than in transformed cells. 6q, 8q, and 20q copy number changes are reported in ovarian cancer (17, 18); 6q21-25 amplification in cisplatin-resistant ovarian cancer (19) suggests SYNCRIP to be a useful predictive biomarker. LAMA5 has been identified as a predictor of amplification and biomarker for cervical cancer (20). Mutations, single nucleotide polymorphisms, copy number changes, and rearrangements in 8q24.3 that affect Wnt signaling and target Myc are frequent in several cancers (2123). RAD21, NUDCD1, DCC1, YWHAZ, LY6K, HSF1, and BOP1 identified in the 8q24.3 cluster in our study are reportedly associated with transformation and tumor progression events (2426).

Epigenetic regulation of SeOvCa genes

Global hypomethylation and specific promoter hypermethylation in tumor-suppressor genes mediate aberrant expression in human malignancies (27). Gene silencing is often supplemented by repressive histone modifications including the methylation of lysine-9 on histone 3 (K9) or the methylation of lysine-27 on histone 3(K27), whereas activation is supported by the methylation of lysine-4 on histone 3 (K4; ref. 28). Bivalent and trivalent combinations of histone modifications are also reported in cancer (29). In an ongoing study, we have profiled the epigenetic status of genes in the A4 cell model on a genome-wide scale through Me-DIP for DNA methylation and ChIP-on-chip for histone methylation (K4, K9, and K27) from which we extracted data relating to SeOvCa genes MAL, MEST, PTGIS, PAPSS2, EFEMP1, and FBN1 because these are reported to be epigenetically regulated in human malignancies (18, 3036). The transformed state was strikingly associated with hypomethylated promoters of the upregulated genes (MAL and MEST); this was further supported by two activation K4 marks upstream of the MAL transcription start site (TSS) and an enriched K4-K9 bivalent mark upstream of the MEST-TSS (Fig. 4A). EFEMP1 promoter methylation may be an early event in A4 transformation; contrarily, PTGIS, PAPSS2, and FBN1 seemed to be demethylated (data not shown). PAPSS2 and PTGIS harbored bivalent repressive marks; EFEMP1 had two trivalent repressive marks; whereas FBN1 had monovalent K27 and trivalent repressive marks in their promoters; all of which validated through specific ChIP-PCRs (Fig. 4B).

Figure 4.

Epigenetic regulation. A, schematic of enriched Histone-DNA methylation probes with respect to transcription start site (TSS) of each gene. A1 marks the amplicon analyzed by ChIP-PCR; A2 is another enriched amplicon. B, ChIP-PCR detection of H3K4, H3K9, or H3K27 enrichment. C, methylation status of the six genes in TCGA methylation database (75 tumor and 10 normal samples;*, P < 0.05).

Figure 4.

Epigenetic regulation. A, schematic of enriched Histone-DNA methylation probes with respect to transcription start site (TSS) of each gene. A1 marks the amplicon analyzed by ChIP-PCR; A2 is another enriched amplicon. B, ChIP-PCR detection of H3K4, H3K9, or H3K27 enrichment. C, methylation status of the six genes in TCGA methylation database (75 tumor and 10 normal samples;*, P < 0.05).

Close modal

Retrieval of data from the TCGA DNA methylation database for these six genes supported MEST promoter hypomethylation (Fig. 4C). Promoter hypomethylation of MAL is reported to be a promising predictive marker associated with short patient survival and aggressive ovarian cancer (29). Our data identify a role for K4 methylation in additionally regulating MAL expression. Loss of imprinting and promotor switching in >MEST is linked with its reduced expression. Together, demethylation and K4 modification seem likely mechanisms for the increased expression of MAL and MEST. Aberrant promoter methylation of EFEMP1 in sporadic breast cancer, FBN1 in prostate, and PTGIS in colorectal and lung cancers are also known. The present study is the first report of probable involvement of epigenetic regulation of these genes in ovarian cancer.

SeOvCa protein interactions

We scanned the Human Protein Atlas (HPR, 38) for SeOvCa protein expression and identified positive correlation in ovarian cancer of some upregulated (ATAD2, MCM2, SYNCRIP, and TNNT1), downregulated (FBN1, PROS1, PTGIS, and RNASE4), and marginal correlations (MEST, DAB2, and SGK1), whereas MMP9 was an outlier (Supplementary Table S4). Literature search additionally supported the involvement of MAL, HNMT, and SMARCA2 proteins in ovarian cancer (28). However, most SeOvCa proteins remain to be evaluated in HPR, and their association with ovarian cancer was established. Further exploration of known PINA networks led to the delineation of three types of PPIs between SeOvCa proteins:

  1. Node-node PPIs between MCM2 and GNB5 (Fig. 5A).

  2. Node-linker-node PPIs: 12 nodes and 15 connecting linkers generate a network (Supplementary Table S5; Fig. 5B) analogous to the ARACNe-generated interface network.

  3. Node-interactor networks (Supplementary Table S6; Fig. 5C) involve all possible interactions within SeOvCa to identify (a) independent proteins with no known interactions (MAL, SLC39A4, EFEMP1, HNMT, LHFP, LRRC17, PTGIS, and RNASE4), (b) stand-alone protein network hubs (BCAT1, CDCA4, MEST, SOX17, TM7SF2, KLF2, PROS1, and PAPSS2) that maintain their influence on specific genes to form isolated hubs, and (c) social nodes that are SeOvCa proteins that have exclusive PPIs with other nodes or indirect ones through linkers to generate a complex network.

Figure 5.

SeOvCa PPI networks generated in PINA for (A) node-node, (B) node-linker-node, and (C) node-interacting protein interactions.

Figure 5.

SeOvCa PPI networks generated in PINA for (A) node-node, (B) node-linker-node, and (C) node-interacting protein interactions.

Close modal

Integration of interaction networks suggests c-Myc transformation supported by altered p53 and retinoblastoma signaling

Differences between PINA-generated networks and the contemporary ones in ARACNe exist due to comparisons at different levels of gene regulation, i.e., experimentally established PPIs versus gene expression profile–based interactions. Thus, molecules common to both networks (six nodes and nine interactors) would be strongly associated with ovarian cancer (Fig. 6A). The MCM2 hub (CDC7, MCM3, MCM4, and MCM7) involved early in DNA replication, cell cycle progression, and p53 inactivation interacts with the RRM2 hub to link DNA mismatch repair and replication (37, 38). Significantly, Cdc7 kinase is a predictive marker in ovarian cancer (39), maintains cell viability during replication stress, and is required for loading the MCM2-MCM7 complex onto chromatin. FBN1(DCN) and DAB2(TGFBR2) together with another SeOvCa gene SGK1 are linked to transforming growth factor β signaling and are critical in controlling its apoptotic effects, microsatellite instability, and DNA mismatch repair (40). All these effects are highly probable in the current situation wherein we identified extended networks and functional pathways of each hub (Supplementary Figs. S3 and S4). The information along with published SeOvCa associations (Supplementary Table S7) led us to derive a regulatory network involving altered retinoblastoma (Rb) signaling, c-Myc activation, and p53/cell cycle/DNA damage repair pathways (Fig. 6B).

Figure 6.

Derivation of functional modules. A, overlapping nodes and interactors in the ARACNe and PINA networks. B, schematic representation of the three functional modules derived to be significant in serous ovarian carcinoma. C, Western blots validating some of the proteins and pathways implicated in the predicted functional modules (Un, untransformed; T, transformed A4 cells).

Figure 6.

Derivation of functional modules. A, overlapping nodes and interactors in the ARACNe and PINA networks. B, schematic representation of the three functional modules derived to be significant in serous ovarian carcinoma. C, Western blots validating some of the proteins and pathways implicated in the predicted functional modules (Un, untransformed; T, transformed A4 cells).

Close modal

Rb pathway alterations primarily involve inactivation of its function in senescence that depends on the transient recruitment of SMARCA2 into RB/HDAC1 megacomplexes. SMARCA2 downregulation abrogates G1-S growth arrest through the modulation of E2F, Cdk4/6-cyclin D, Cdk2-cyclin E, and Cip/Kip/Ink4a, which are downstream effectors of Rb. Such evasion of stasis/senescence barriers may be considered as a first step toward immortalization, a recognized hallmark of cancer (41). ATAD2 is a physiologic target of pRB/E2F and functions as a coactivator for the transcription factors ERα, AR, and c-myc by recruiting cAMP response element binding protein (CREB) to target E2F1, Cyclin D1, c-MYC, and BIRC5 (survivin) through a positive feedback loop. SGK1 (AR target) downregulation suggests the irrelevance of AR signaling in ovarian cancer. 8q chromosomal region amplifications of ATAD2 and c-Myc (Fig. 3C), together with BCAT1 (c-myc target) upregulation, contribute to tumor development and progression. SYNCRIP associates with insulin-like growth factor II mRNA binding protein 1 (IGF2BP1) to limit the transfer translation–coupled decay of c-myc RNA that enhances its stability. The phosphoinositide 3-kinase–mediated interaction of c-myc with the prereplicative minichromosome maintenance complex MCM2-MCM7 leads to its localization to early sites of DNA synthesis and replication initiation. MYCN activation by E2F1 enhances the transcription of MCM2-MCM7 members and downregulates p27 that together with BMI1 further mediates stem cell self-renewal.

Myc deregulation generates DNA damage, replication stress, and genomic instability through the inactivation of the p53-mediated DNA damage response involving ATM-ATR-CHK1-CHK2 checkpoints. CDCA4, an E2F1 target, regulates E2F and p53 transcription, cellular proliferation, and cell fate determination. ATR phosphorylates MCM2, resulting in the aberrant loading of the MCM complex onto chromatin and cell cycle progression. EXO1 mediates DNA mismatch repair, suppresses replication fork instability, and enhances resistance to DNA-damaging agents. RRM2, another downstream target of the p53-ATM-ATR-CHK1 axis, mediates DNA repair and cooperates with MCM2 toward cell proliferation. RRM2 also enhances invasion through NF-κB–dependent MMP9 activation, and angiogenesis through decreased thrombspondin-1 and increased vascular endothelial growth factor production. Derepression of CXCR4 through KLF2 downregulation further supports migrating cancer cells. Loss of LRRC17 leads to enhanced interactions between RANKL and its ligand NF-κB, whereas EFEMP1 downregulation signifies the loss of antiangiogenesis activity in transformed cells. LAMA5-associated 20q13.3 amplifications may involve the upregulation of CAS and ZNF217 (a putative oncogene), and correlate with Cyclin D1, Rb, and p53 alterations (42). A part of our predicted model (p53-Rb inactivation) in mouse OSE has been shown to lead to the formation of neoplasms comparable with high-grade human serous ovarian carcinomas (43). Finally, because model systems are essential to validate any set of interactions and regulations, we used the A4 cell system that has wild-type p53 (Supplementary Fig. S5) to identify some of the partners implicated in predicted pathway dysregulation (Fig. 6C).

Tumor screening necessitates the identification of a basic, minimum number of markers representative of disease heterogeneity. Gene expression analysis in the maximum number of samples across different public databases specific for a cancer type are important in specific, sensitive detection through removal of background noise. Thus, the derivation of an unbiased, prioritized list of serous ovarian cancer–specific genes that validates in multiple databases holds the promise of improved data robustness. SeOvCa genes include some earlier identified biomarkers; this conformance of a data-driven approach correlating genome-wide gene expression analyses, predicted interactive networks, and transformation-associated molecular events is very encouraging.

A challenge inherent to such analysis is elucidating the biological connection between signatorial components and phenotypic effects. Toward this end, system network–based prediction of gene-protein interactions provides a global understanding of the significance of SeOvCa gene hubs in transformation, without imposing a reductionism approach. Further screening of hub interactions at the cellular and tumor levels will affirm their diagnostic applicability. Some promising genes include the upregulated RRM2-EXO1-MCM2 cluster (DNA licensing, mismatch repair, and aneuploidy), DAB2 (DOC2, differentially expressed in ovarian cancer), KLF2 (cell cycle inhibitor), and components of interface networks, etc. Some genes earlier known to be involved in cancers other than ovarian cancer implies a commonality in cell transformation mechanisms. Another set of “discovery” genes identified without any known transformation-associated functionalities include TNNT1, TM2SF2, RNASE4, PROS1, LRRC17, GNB5, and DIXDC1, whose precise involvement remains to be investigated.

Biological complexity of tumors arises from the fact that cellular transformation arises from a myriad of interactions that are difficult to predict with reductionism. The development of cancer systems biology–based networks that predict dynamic interactions within—between cells and the tumor niche—represents a fresh insight into serous ovarian carcinoma that ultimately will lead to an increased understanding of regulatory networks in tumors. Further validation of predicted outcomes based on copy number changes and epigenetic regulation presents new viewpoints and assigns prediction credibility in coordination of theory and experiments.

Although correlation between transcript and protein levels is generally poor, the present study generated several tangible leads in ovarian cancer including the identification of three functional modules, centering on c-Myc activation linked to altered Rb signaling and p53/cell cycle/DNA damage repair; together, these define distinct pathways. Dysregulated Rb signaling leads to the bypass of cell crisis, and evasion of p53-mediated apoptosis during DNA damage and repair supports cell cycle progression. Together with altered c-Myc regulation, these events outline the progression toward malignancy. These mechanisms are determined as a function of a set of SeOvCa genes that may be applied for screening serous ovarian tumors in the future. Such expression tools additionally include hub-derived genes that delineate the specificity of involved pathways across tumors. Such approaches open up newer opportunities in the design and application of multiple gene profiling as a guide in the decision making and transitioning to a more personalized ovarian cancer treatment.

No potential conflicts of interest were disclosed.

We thank Dr. G.C. Mishra, Director, National Centre for Cell Science (Pune, India) for the encouragement and support, the publicly available TCGA and IST databases applied in this study, and Avinash Mali and Prasad Chaskar for the technical assistance.

Grant Support: Department of Biotechnology, Government of India, New Delhi, grant no. BT/PR11465/Med/30/145/2008 (S.A. Bapat). A.P. Kusumbe and R.S. Kalra receive a research fellowship from the Council of Scientific and Industrial Research, New Delhi.

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

1
Swanton
C
,
Caldas
C
. 
Molecular classification of solid tumours: towards pathway-driven therapeutics
.
Br J Cancer
2009
;
100
:
1517
22
.
2
Sotiriou
C
,
Pusztai
L
. 
Gene expression signatures in breast cancer
.
N Engl J Med
2009
;
360
:
790
800
.
3
Kobel
M
,
Kalloger
SE
,
Boyd
N
, et al
. 
Ovarian carcinoma subtypes are different diseases: implications for biomarker studies
.
PLoS Med
2008
;
5
:
e232
.
4
Lawrenson
K
,
Gayther
SA
. 
Ovarian cancer: a clinical challenge that needs some basic answers
.
PLoS Med
2009
;
6
:
e25
.
5
Bapat
SA
,
Mali
AM
,
Koppikar
CB
, et al
. 
Stem and progenitor-like cells contribute to the aggressive behavior of human epithelial ovarian cancer
.
Cancer Res
2005
;
65
:
3025
9
.
6
McLendon
R
,
Friedman
A
,
Bigner
D
. 
Comprehensive genomic characterization defines human glioblastoma genes and core pathways
.
Nature
2008
;
455
:
1061
8
. .
7
Kilpinen
S
,
Autio
R
,
Ojala
K
, et al
. 
Systematic bioinformatic analysis of expression levels of 17,330 human genes across 9,783 samples from 175 types of healthy and pathological tissues
.
Genome Biol
2008
;
9
:
R139
. .
8
Bild
AH
,
Yao
G
,
Chang
JT
, et al
. 
Oncogenic pathway signatures in human cancers as a guide to targeted therapies
.
Nature
2006
;
439
:
353
7
.
9
Margolin
AA
,
Nemenman
I
,
Basso
K
, et al
. 
ARACNE: an algorithm for the reconstruction of gene regulatory networks in a mammalian cellular context
.
BMC Bioinformatics
2006
;
7
Suppl 1
:
S7
. .
10
Wu
J
,
Vallenius
T
,
Ovaska
K
, et al
. 
Integrated network analysis platform for protein-protein interactions
.
Nat Methods
2009
;
6
:
75
7
.
11
Kurrey
NK
,
Jalgaonkar
SP
,
Joglekar
AV
, et al
. 
Snail and Slug mediate radio- and chemo-resistance by antagonizing p53-mediated apoptosis and acquiring a stem-like phenotype in ovarian cancer cells
.
Stem Cells
2009
;
27
:
2059
68
.
12
Yasui
DH
,
Peddada
S
,
Bieda
MC
, et al
. 
Integrated epigenomic analyses of neuronal MeCP2 reveal a role for long-range interaction with active genes
.
PNAS
2007
;
104
:
19416
21
.
13
Thomas
PD
,
Campbell
MJ
,
Kejariwal
A
, et al
. 
PANTHER: a library of protein families and subfamilies indexed by function
.
Genome Res
2003
;
13
:
2129
41
. .
14
Wani
AA
,
Sharma
N
,
Shouche
YS
, et al
. 
Nuclear-mitochondrial genomic profiling reveals a pattern of evolution in epithelial ovarian tumor stem cells
.
Oncogene
2006
;
25
:
6336
44
.
15
Kusumbe
AP
,
Mali
AM
,
Bapat
SA
. 
CD133-expressing stem cells associated with ovarian metastases establish an endothelial hierarchy and contribute to tumor vasculature
.
Stem Cells
2009
;
27
:
498
508
.
16
Kusumbe
AP
,
Bapat
SA
. 
Cancer stem cells and aneuploid populations within developing tumors are the major determinants of tumor dormancy
.
Cancer Res
2009
;
69
:
9245
53
.
17
Staebler
A
,
Karberg
B
,
Behm
J
, et al
. 
Chromosomal losses of regions on 5q and lack of high-level amplifications at 8q24 are associated with favorable prognosis for ovarian serous carcinoma
.
Genes Chromosomes Cancer
2006
;
45
:
905
17
.
18
Fejzo
MS
,
Dering
J
,
Ginther
C
, et al
. 
Comprehensive analysis of 20q13 genes in ovarian cancer identifies ADRM1 as amplification target
.
Genes Chromosomes Cancer
2008
;
47
:
873
83
.
19
Takano
M
,
Kudo
K
,
Goto
T
, et al
. 
Analyses by comparative genomic hybridization of genes relating with cisplatin-resistance in ovarian cancer
.
Hum Cell
2001
;
14
:
267
71
.
20
Scotto
L
,
Narayan
G
,
Nandula
SV
, et al
. 
Identification of copy number gain and overexpressed genes on chromosome arm 20q by an integrative genomic approach in cervical cancer: potential role in progression
.
Genes Chromosomes Cancer
2008
;
47
:
755
65
.
21
Ghoussaini
M
,
Song
H
,
Koessler
T
, et al
. 
Multiple loci with different cancer specificities within the 8q24 gene desert
.
J Natl Cancer Inst
2008
;
100
:
962
6
.
22
Pomerantz
MM
,
Ahmadiyeh
N
,
Jia
L
, et al
. 
The 8q24 cancer risk variant rs6983267 shows long-range interaction with MYC in colorectal cancer
.
Nat Genet
2009
;
41
:
882
4
.
23
Li
Y
,
Meng
G
,
Guo
QN
. 
Changes in genomic imprinting and gene expression associated with transformation in a model of human osteosarcoma
.
Exp Mol Pathol
2008
;
84
:
234
9
.
24
Yamamoto
A
,
Sakurai
H
. 
The DNA-binding domain of yeast Hsf1 regulates both DNA-binding and transcriptional activities
.
Biochem Biophys Res Commun
2006
;
346
:
1324
9
.
25
Lin
M
,
Morrison
CD
,
Jones
S
, et al
. 
Copy number gain and oncogenic activity of YWHAZ/14-3-3ζ in head and neck squamous cell carcinoma
.
Int J Cancer
2009
;
125
:
603
11
.
26
Ishikawa
N
,
Takano
A
,
Yasui
W
, et al
. 
Cancer-testis antigen lymphocyte antigen 6 complex locus K is a serologic biomarker and a therapeutic target for lung and esophageal carcinomas
.
Cancer Res
2007
;
67
:
11601
11
.
27
Vucic
EA
,
Brown
CJ
,
Lam
WL
. 
Epigenetics of cancer progression
.
Pharmacogenomics
2008
;
9
:
215
34
.
28
Chan
TA
,
Glockner
S
,
Yi
JM
, et al
. 
Convergence of mutation and epigenetic alterations identifies common genes in cancer that predict for poor prognosis
.
Plos Med
2008
;
5
:
e114
.
29
Ohm
JE
,
McGarvey
KM
,
Yu
X
, et al
. 
A stem cell-like chromatin pattern may predispose tumor suppressor genes to DNA hypermethylation and heritable silencing
.
Nat Genet
2007
;
39
:
237
42
.
30
Lee
PS
,
Teaberry
VS
,
Bland
AE
, et al
. 
Elevated MAL expression is accompanied by promoter hypomethylation and platinum resistance in epithelial ovarian cancer
.
Int J Cancer
2009
;
26
:
1378
89
.
31
Berchuck
A
,
Iversen
ES
,
Luo
J
, et al
. 
Microarray analysis of early stage serous ovarian cancers shows profiles predictive of favorable outcome
.
Clin Cancer Res
2009
;
15
:
2448
55
.
32
Nakanishi
H
,
Suda
T
,
Katoh
M
, et al
. 
Loss of imprinting of PEG1/MEST in lung cancer cell lines
.
Oncol Rep
2004
;
12
:
1273
8
.
33
Dekel
B
,
Metsuianim
S
,
Schmidt-Ott
KM
, et al
. 
Multiple imprinted and stemness genes provide a link between normal and tumor progenitor cells of the developing human kidney
.
Cancer Res
2006
;
66
:
6040
9
.
34
Sadr-Nabavi
A
,
Ramser
J
,
Volkmann
J
, et al
. 
Decreased expression of angiogenesis antagonist EFEMP1 in sporadic breast cancer is caused by aberrant promoter methylation and points to an impact of EFEMP1 as molecular biomarker
.
Int J Cancer
2009
;
124
:
1727
35
.
35
Wang
Y
,
Hayakawa
J
,
Long
F
, et al
. 
“Promoter array” studies identify cohorts of genes directly regulated by methylation, copy number change, or transcription factor binding in human cancer cells
.
Ann N Y Acad Sci
2005
;
1058
:
162
85
.
36
Frigola
J
,
Munoz
M
,
Clark
SJ
, et al
. 
Hypermethylation of the prostacyclin synthase (PTGIS) promoter is a frequent event in colorectal cancer and associated with aneuploidy
.
Oncogene
2005
;
24
:
7320
6
.
37
Pontén
F
,
Jirström
K
,
Uhlen
M
. 
The Human Protein Atlas–a tool for pathology
.
J Pathol
2008
;
216
:
387
93
. .
38
Yan
T
,
Schupp
JE
,
Hwang
HS
, et al
. 
Loss of DNA mismatch repair imparts defective cdc2 signaling and G([2]) arrest responses without altering survival after ionizing radiation
.
Cancer Res
2001
;
61
:
8290
7
.
39
Kulkarni
AA
,
Kingsbury
SR
,
Tudzarova
S
, et al
. 
Cdc7 kinase is a predictor of survival and a novel therapeutic target in epithelial ovarian carcinoma
.
Clin Cancer Res
2009
;
15
:
2417
25
.
40
Biswas
S
,
Trobridge
P
,
Romero-Gallo
J
, et al
. 
Mutational inactivation of TGFBR2 in microsatellite unstable colon cancer arises from the cooperation of genomic instability and the clonal outgrowth of transforming growth factor β resistant cells
.
Genes Chromosomes Cancer
2008
;
47
:
95
106
.
41
Hanahan
D
,
Weinberg
RA
. 
The hallmarks of cancer
.
Cell
2000
;
100
:
57
70
.
42
Peiró
G
,
Diebold
J
,
Löhrs
U
. 
CAS (cellular apoptosis susceptibility) gene expression in ovarian carcinoma: Correlation with 20q13.2 copy number and cyclin D1, p53, and Rb protein expression
.
Am J Clin Pathol
2002
;
118
:
922
9
.
43
Zhou
Z
,
Flesken-Nikitin
A
,
Corney
DC
, et al
. 
Synergy of p53 and Rb deficiency in a conditional mouse model for metastatic prostate cancer
.
Cancer Res
2006
;
66
:
7889
98
.

Supplementary data