DNA copy number alterations are a common occurrence in all cancers. Specific chromosomal regions and focal points favor either gains or losses in DNA among cancer types 
. These amplifications and deletions are shown to include tumor suppressors and oncogenes. In addition, DNA methylation exhibits redistribution within a cancer genome 
. Often times the copy number and DNA methylation profiles are generated as a static representation of a particular cancer's whole genome aberrations. However, the amplitude of specific gene function within ovarian cancer is often highly variable between tumor samples 
. It is therefore essential to accurately determine each gene's individual functional state within its cancer environment. Here, we not only looked at whole genome patterns of copy number aberrations and methylation but also focus on sample specific CNV and methylation properties for altered genes to provide a better understanding of ovarian cancer gene functionality. We first separately present data of DNA structural variation and DNA methylation changes in ovarian tumors and then combine the two modalities with expression data to identify how these aberrations may affect individual gene function within the tumor population.
We first analyzed the DNA copy number variation of primary ovarian tumors from 42 individuals and compared our findings to The Cancer Genome Atlas data set for ovarian cancer 
. In analysis of CNV segmentation changes in our 42 tumor samples, DNA variability is shown to be most prevalent in chromosomes 1, 2, 4, 8, 9 and 19. We have shown a large variability in amplification and deletion breakpoint loci in ovarian tumors and identified chromosomal areas of frequent copy number variations. We see a high level of amplification frequency in known oncogenic regions containing MYC (chromosome 8), CCNE1 (chromosome 19) and frequent deletions are found in chromosome arms 4q, 16q, and 17p 
. Similarly to the TCGA data analysis ROMA detected high frequency copy losses in PTEN, RB1, and NF1. We also show the known but previously unreported in primary ovarian tumors amplification of the MCL region of chromosome 1q21.1-q21.2 
and previously unreported deletions in the IDH2 and IDH3 region of chromosome 15. A total of 983 genes are included in amplified and deleted regions. A strong correlation of expression with copy number variation has been reported in ovarian cancer 
. Here, we primarily focused on quantifying this correlation at varying CNV levels for the purposes of functional annotation. Most significant changes of expression occur at extreme CNV. We show that copy number variation has a strong effect on expression in primary ovarian tumors for 156 genes. Notable genes with correlated CNV and expression include MYC, CCNE1, KRAS, NDRG1, MLL4, MTSS1, C11orf30, MLH1 and CEPBG. Genes identified in the MSKCC data set were from similar genomic loci as those found in the TCGA data set. Five genes were predicted from both data sets: CCNE1, POP4, UQCRB, PHF20L1 and C19orf2. In addition isocitrate dehydrogenase isoforms IDH2 (chr15), IDH3A (chr15), and IDH3B (chr20) show correlated expression to CNV. The IDH1 and IDH2 genes are mutated in glioblastoma and AML cancer patients. IDH1 has been implicated as a prognosis positive biomarker in glioblastoma and AML IDH1/2 mutants show hypermethylation in comparison to other AML subtypes 
. All the isocitrate dehydrogenase genes exhibit deleted CNV in ovarian cancer samples (ranging from 14%–52%, ). Here, we show for both the IDH2 and IDH3A/IDH3B genes expression is mainly correlated in tumor samples exhibiting deleted copy number to normal copy number.
Next, we examined genome wide DNA methylation in ovarian tumors. DNA methylation alterations are a significant feature of the cancer genome 
. The protocadherin gene cluster of chromosome 5 has been shown to be frequently hypermethylated and silenced in various forms of cancer 
. We see broad hypermethylation for both the PCDHα and PCDHβ loci in ovarian cancer as well. We also observe increased levels of methylation in other cancer related genes e.g. ALX3 and PHOX2B both implicated in neuroblastoma 
. Decreased levels of methylation are seen in genes such as calcyphosin which exhibits oncogenic properties in endometrial cancer 
, lactate dehydrogenase which when inhibited impairs cell proliferation via the Warburg effect of aerobic glycolysis in cancer cells 
and DUB3, the CDC25A stabilizing protein ubiquitin hyrdolase 
which has been shown to rescue CDC25A from proteasomal degradation and promote an oncogenic induction response 
. Hypomethylation with subsequent cellular rise of DUB3 can therefore be a candidate for the cellular regulation of CDC25A protein levels and CDC25A linked oncogenesis.
Finally, to identify genes with a more direct genomic and epigenetic effect on the function of the cancer cell, we directed our focus on the combined gene features of copy number, methylation and expression. Tumor suppressors and oncogenes are often implicated by their transcriptional abnormalities in the cancer cell. It is of interest to understand which tumor suppressors and oncogenes play a direct role in a particular cancer among all genes affected by genomic aberrations. A certain tumor suppressor or oncogene function may be gained or silenced at varying frequencies by different epigenetic and genomic conditions within the tumor sample population. Examining these properties and their affects on gene expression can provide better insight into identifying which genes are most responsible to the pathology of the tumor. We therefore formulate a set of predictive features based on genomic and epigenetic properties of the tumor that can be indicative of altered function for tumor suppressors and oncogenes in the cancer genome (). Low expression of various tumor suppressors in cancer cells can be a result of deleted copy number or silencing by promoter hypermethylation. While amplification and promoter hypomethylation can play a role in the over expression of oncogenes. Conversely, a particular known oncogene may be deleted in a particular cancer lessening its pathogenic role within that cancer or an individual sample. Interestingly, for highly amplified genes, a high level of methylation accompanied by low expression could indicate altered tumor suppressor function in the cancer cell. In highly amplified genes, low level methylation and high expression would indicate oncogenic features in a cancer cell. We therefore utilized samples with extreme copy number variations and examined the methylation and expression changes of genes within these aberrant loci to identify potential tumor suppressors and oncogenes. Examining both our primary MSKCC tumor sample data set and the TCGA data set, we discovered 180 genes with tumor suppressor features of low expression with copy number deletion and high methylation. These features are characteristic of known classic tumor suppressors among which the established tumor suppressor RB1 (retinoblastoma protein) was captured. Additionally we find another 48 genes with elevated copy number but low expression and high methylation. For oncogenic epigenetic gene features we discover 318 genes within amplified loci and 65 within deleted copy number loci between the two data sets. Several genes discovered in ovarian cancer tumors with these specific tumor suppressor and oncogenic features have been previously implicated in other cancers and are now shown to have additional methylation and copy number variation properties. Furthermore, 25% of the genes captured with tumor suppressor and oncogenic gene features were represented in 941 MSKCC data set genes () with significant changes in methylation and expression per CNV. Seven genes were identified from both the MSKCC and TCGA data sets that contained strong correlations for methylation dependent expression exhibited at varying copy number aberrations; CDCA8, ATAD2, CDKN2A, RAB25, AURKA, BOP1 and EIF2C3. Four of these seven genes (CDCA8, ATAD2, CDKN2A, AURKA) have direct functional relationships of binding and regulation with other experimentally established oncogenes and tumor suppressors such as TP53, RB1, MYC and E2F1 
. Thereby indicating a potential functional cancer module (Figure S8
) that can be further computationally and experimentally targeted. Using genomic features specific to aberrations found in tumor sample data captures previously identified tumor suppressors and oncogenes in addition to genes associated with these biomarkers. This genomic and epigenetic function-based feature approach identified genes in cancer pathways such as endometrial cancer, ErbB signaling pathways, epithelial cell signaling and actin cytoskeleton regulation. This type of primary gene function identification approach can provide a base feature set for further machine-learning cancer network prediction protocols.
In addition, cancer genes exhibiting contradictory tumor suppressor or oncogenic epigenetic features in ovarian cancer may provide clues into the regulatory pathways within ovarian cancer. Of note, predicted within the MSKCC data set tumor suppressor features is the established oncogenic transcription factor STAT3 
. Here we see significant STAT3 deletion (≥73% sample frequency) contributing to a potential heterozygous gene copy loss in both the TCGA and MSKCC data sets. Furthermore, within samples containing a low copy number of STAT3 gene, slightly higher methylation and lower expression values are observed. This may suggest a decreased role for STAT3 in the oncogenic function within ovarian tumors. Therefore, epigenetic and genomic specific gene features are at the strength of our predictions and can be used to i
) predict novel gene functions in ovarian cancer and ii
) elucidate or verify the direct cancer functioning role for previously implicated tumor suppressors or oncogenes. We therefore decided to examine many known cancer oncogenes and tumor suppressors for varying levels of regulation among tumor samples. For instance, the ovarian cancer oncogenes CCNE1 and RAB25 
show significant methylation and expression correlation for both amplified and deleted copy number aberrations. The expression levels of these cancer functioning genes differs between samples and the modes of epigenetic regulation exhibit different levels of frequency 
. Each gene affecting the growth of the tumor is not evenly implicated in all samples. We therefore attempted to illustrate these genomic and epigenetic sample irregularities (such as observed in PLAGL1, CCNE1 and PIAS3) for many of the known ovarian cancer genes (, Tables S2
). Sample specific feature analysis of identical gene combinations and modules at amplified, neutral or deleted copy number with corresponding epigenetic regulatory features can be used to identify ovarian cancer heterogeneity and the driving genes contributing it. Application of this gene function diversity can be further studied using clinical information for each sample, thereby combining cancer gene modules with each samples' clinical features. The development of this type of knowledge base of gene features in a cancer population will better help identify subtype specific tumor function.
The continuing increase of experimental epigenetic data from various tumor samples offers the ability to computationally search for putative genes with properties in the proliferation of cancer cells. Here we performed a coarse-grained bioinformatics whole genome evaluation of epigenetic features in ovarian cancer tumor cell from two separate platforms covering over 11,500 genes. We demonstrate ovarian cancer specific epigenetic regulation of previously identified cancer genes and cancer biomarkers. Furthermore, we were also able to implicate genes with tumor suppressor and oncogenic epigenetic properties specific to ovarian cancer tumors that have not been previously reported. Examination of multiple cancer epigenetic modalities will help segregate cancer specific genes from randomly altered cancer genes and can possibly elucidate the genetic mediators of ovarian tumorigenesis. The focus on gene combinations with specific copy number aberrations per individual tumor sample plus their methylation and expression properties within those samples allows for the better understanding and eventual identification of tumor type specific cancer pathways.