bioRxiv preprint doi: https://doi.org/10.1101/145268; this version posted June 2, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.

Driver pattern identification over the gene co-expression of drug response in ovarian cancer by integrating high throughput genomics data

Xinguo Lu,1 Jibo Lu,1 Bo Liao,1 and Keqin Li1,2 1College of Computer Science and Electronic Engineering, Hunan University, Lushan Nan Rd., Changsha 410012, China; 2Department of Computer Science, State University of New York, New Paltz, NY 12561, USA.

The multiple types of high throughput genomics data create a potential opportunity to identify driver pattern in ovarian cancer, which will acquire some novel and clinical biomarkers for appropriate diagnosis and treatment to cancer patients. However, it is a great challenging work to integrate omics data, including somatic mutations, Copy Number Variations (CNVs) and gene expression profiles, to distinguish interactions and regulations which are hidden in drug response dataset of ovarian cancer. To distinguish the candidate driver genes and the corresponding driving pattern for resistant and sensitive tumor from the heterogeneous data, we combined gene co-expression modules and mutation modulators and proposed the identification driver patterns method. Firstly, co-expression network analysis is applied to explore gene modules for gene expression profiles via weighted correlation network analysis (WGCNA). Secondly, mutation matrix is generated by integrating the CNVs and somatic mutations, and a mutation network is constructed from this mutation matrix. The candidate modulators are selected from the significant genes by clustering the vertex of the mutation network. At last, regression tree model is utilized for module networks learning in which the achieved gene modules and candidate modulators are trained for the driving pattern identification and modulator regulatory exploring. Many of the candidate modulators identified are known to be involved in biological meaningful processes associated with ovarian cancer, which can be regard as potential driver genes, such as CCL11, CCL16, CCL18, CCL23, CCL8, CCL5, APOB, BRCA1, SLC18A1, FGF22, GADD45B, GNA15, GNA11 and so on, which can help to facilitate the discovery of biomarkers, molecular diagnostics, and drug discovery.

Ovarian cancer is known as a complex disease of the International Cancer Genome Consortium (ICGC), etc., have genome. During tumorigenesis, many factors including produced a large volume of data and provided us opportunity genomic affection and epigenomic affection contribute to to integrate different level of gene expression data to identify pathological gene expression changes. Hence, acquiring the candidate driver genes and driver pathways and better etiopathogenesis and chemoresponse of cancer faces a great understand the cancer at the molecular level. challenge. And the achievement can be utilized to deploy the During exploring the large volume of genomics high-performance diagnosis and treatment of cancer patients. abnormalities generated from large-scale cancer projects, In spite of many studies presented for ovarian cancer many computational and statistical methods have been prognosis, currently, there is no completely validated clinical proposed to search for this mutation driver patterns. Adib et model for predicting ovarian cancer prognosis and drug al. predicted the potential specificity of several putative response. Therefore, it remains an important research issue to biomarkers by using gene expression microarray profile to identify prognostic and predictive driver patterns for examine their expression in a panel of epithelial tissues and improving ovarian cancer treatment. tumors (Adib et al. 2004). Youn et al. proposed a new Cancer genomes possess a large number of aberrations method to identify cancer driver genes by accounting for the including somatic mutations and copy number variations functional impact of mutations on proteins (Youn et al. 2010). (CNVs). Many mutations contribute to cancer progression Tomasetti et al. combined conventional epidemiologic from the normal to the malignant state. The researches have studies and genome-wide sequencing data to infer a number shown that some aberrations are vital for tumorigenesis and of driver mutations which is essential for cancer development most cancers are caused by a small number of driver (Tomasetti et al. 2014). Jang et al. identified differentially mutations developed over the course of about two decades expressed proteins involved in stomach cancer (Cheng et al. 2015; Vogelstein et al. 2013). The detection of carcinogenesis through analyzing comparative proteomes these mutations with exceptionally high association between between characteristic alterations of human stomach the copy number variations, somatic mutations and gene adenocarcinoma tissue and paired surrounding normal tissue expression can ascertain disease candidate genes and (Jang et al. 2004). Xiong et al. proposed a general framework potential cancer mechanisms. Several large scale cancer in which the feature (gene) selection was incorporated into genomics projects, such as the Genomic Data Commons pattern recognition to identify biomarkers (Xiong et al. 2001). (GDC), The Cancer Genome Atlas (TCGA), and Jung et al. identified novel biomarkers of cancer by bioRxiv preprint doi: https://doi.org/10.1101/145268; this version posted June 2, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.

combining bioinformatics analysis on gene expression data and clinically annotated gene expression, CNVs, and somatic and validation experiments using patient samples and mutation datasets from TCGA. The experimental results explored the potential connection between these markers and show that in the driver patterns many of the identified the established oncogenes (Jung et al. 2011). Logsdon et al. candidate modulators are known to be involved in biological identified a new category of candidate tumor drivers in meaningful processes associated with ovarian cancer, which cancer genome evolution which can be regarded as the can be regard as potential driver genes. selected expression regulators (SERs)-genes driving deregulated transcriptional programs in cancer evolution, Results and uncovered a previously unknown connection between We conducted the driver patterns identification and the cancer expression variation and driver events by using a driving regulatory analysis as following procedure. Firstly, novel sparse regression technique (Logsdon et al. 2015). A co-expression network were constructed and the gene cancer genome embodies thousands of genomics modules for gene expression profiles were explored. Then, abnormalities such as single nucleotide variants, large CNVs and somatic mutations were integrated to build the segment variations, structural aberrations and somatic mutation network. The candidate modulators were selected mutations (Chan et al. 2012; Ding et al. 2015). The from the clusters of the vertex of network. A regulation corresponding malignant state reflects aberrant copy procedure was explored for candidate modulators and gene numbers (CN) and/or mutations and the expression patterns modules using network learning combined with regression in which the mutation and copy number patterns are tree model. At last, local polynomial regression fitting model embedded (Wei et al. 2016). Integration of copy was applied to conduct dosage-sensitive and dosage-resistant number/mutation data and gene expression data was analysis. proposed to discover driver genes by quantifying the impact of these aberrations on the transcriptional changes (Zhang et Data preprocessing al. 2016; Kumar et al. 2016; Alderton et al. 2011). The To select a subset of genes whose aberration/expression available implementations for the integrative analysis of profiles were significantly different between sensitive and genomics data includes regression methods, correlation resistant tumors, we applied differential analyses to the gene methods and module network methods (Huang et al. 2011). expression, and CNV data of the two groups. For differential Driver patterns, including driver genes, driver mutations, gene expression analysis and to rank CNV genes, we used driver pathways and core modules, which are considered as the EMD approach: EMDomics, which is designed cancer biomarkers, are supposed to promote the cancer especially for heterogeneous data. We selected genes with a progression. Thus, identification of driver patterns is vital for q-value<0.1. For CNV data, first, we mapped genes to the providing insights into carcinogenic mechanism. The CNV regions to obtain CNV genes; then, we used the EMD integration of gene expression, copy number and somatic approach for the differential analysis and ranking of the CNV mutations data to identify genomics alterations that induce genes. We also calculated the frequency of amplification and changes in the expression levels of the associated genes deletion for each gene in the two groups and selected genes becomes a common task in cancer mechanism and drug for which the difference between their frequencies is more response studies. However, there is still a tough work to than 20%. We used a threshold of log2 copy number ratio of integrate information across the different heterogeneous 0.3/-0.3 to call amplified/deleted genes. For somatic mutation omics data and distinguish driver patterns which can promote data, we also calculated the frequency of mutations for all the cancer cell to propagate infinitely. Hence, we proposed a genes across the samples in each groups and selected the driver pattern identification method over the gene genes that were mutated in more than 2% of the tumors. co-expression of drug response in ovarian by integrating high throughput genomics data. In the method, we integrated Gene modules with co-expression patterns different level genomic data including gene expression, copy After the data preprocessing, 2690 genes were selected for number and somatic mutation to identify drivers of resistance co-expression gene modules exploring. We constructed a and sensitivity to anti-cancer drugs. First, weighted weighted correlation network and identified the correlation network analysis is applied to acquire the co-expression modules, which was conducted using co-expression gene modules. These modules are utilized as WGCNA (Osterhoff et al. 2014; Liu et al. 2016), associated initial inputs for the following module networks learning. with cis-platinum sensitive and resistant. The gene Then, mutation network is constructed from the mutation co-expression modules and the corresponding hierarchical matrix generated by integrating the CNVs and somatic clustering dendrograms of these genes are shown in Fig. 1. mutations. The candidate modulators are selected from the The color bars correspond to the clusters of genes which can clusters of the vertex of the mutation network. Finally, an be seen as gene module. We identified thirty-nine modules optimization model is used to identify the driving pattern and which are described in thirty-nine colors as turquoise, blue, explore the modulator regulatory. We used publicly available bioRxiv preprint doi: https://doi.org/10.1101/145268; this version posted June 2, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.

Figure 1. Gene co-expression modules for gene expression profiles. A network heatmap plot (interconnectivity plot) of a gene network together with the corresponding hierarchical clustering dendrograms, with dissimilarity based on topological overlap, together with assigned module colors. Table 1. The top 10 co-expression modules module color turquoise blue brown yellow green magenta red black purple greenyellow gene counts 357 288 145 130 115 78 62 61 29 23 s brown, yellow, green and so on. The top 10 co-expression i across the samples in the mutation matrix, i.e. misi a . modules are shown in Table 1.The first row represents the i1 color of the modules, and the second row represents the The network vertex weight is defined as hm m/ , where number of genes in the corresponding module. The rest genes ii are in grey. These grey genes are not clustered into any m is the number of samples. Also the edge weight vij is modules. The modules list is shown in Additional file 5. defined as the number of samples that exactly one of the pair Candidate modulators selected by integrating is mutated divided by the number of samples that at least one CNVs and somatic mutations of the pair is mutated. We selected a subgroup of significant genes by clustering For CNVs data, the mapping of CNVs can promote the the vertex of the mutation network. These genes can be search for causal links between genetic variation and disease regarded as candidate modulators. We identify a list of 624 susceptibility. Thus we mapped gene expression to CNVs initial candidate modulators for the mutation genes regions to obtain CNVs genes, for a specific gene in a given (Additional file 6). These modulators includes 40 somatic sample, the value of amplified is assigned the number 1, the mutations. The gene expression heatmap of the top 50 deleted is assigned the number -1, the normal is assigned the candidate modulators is shown in Fig. 2. number 0; for somatic mutation data, we first obtained mutated genes and mapped the correlations between somatic Dosage-sensitive genes and dosage-resistant genes mutations and their residing genes. Then, a mutation matrix and driver patterns A is generated by integrating the CNVs and somatic We conducted the module networks learning with the gene mutations. The mutation matrix is binary: if any variation modules and candidate modulators. It generally assumed that region in a given gene of the particular sample is in a the modulators most likely regulate the expression of the statistically significant variation or any mutation arises in the genes in the corresponding gene modules. Owing to the gene specific gene in the given sample, the value of mutation is modulators would have distortions in amount of samples and assigned the number 1; otherwise the value is assigned the they are likely to be the roots of the regression (drive) trees, number 0. The matrix rows and columns correspond to the modulator gene are generally treat as a driver gene. samples and genes, respectively. A mutation network (MN) Through the module network learning, we obtained 456 was constructed according to the mutation matrix. For a modules and 93 modulators (Additional file 7). 5 out of these mutated gene, mi describes the number of mutations in gene modulators are from somatic mutations. Remarkably, several bioRxiv preprint doi: https://doi.org/10.1101/145268; this version posted June 2, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.

Figure 2. The gene expression heatmap of the top 50 candidate modulators. modules shared the same modulators, in which only one genes (DRGs). Therefore, we obtained 44 DSGs and 49 modulator is derived from somatic mutations. DRGs (Additional file 8). The LRG1, HYDIN, POLRMT We applied a local polynomial regression fitting model and TNFSF10 gene data scatter diagram for CNVs or with the R package loess to the scatter diagram for 93 somatic mutations vs. gene expression with local polynomial modulators with 324 samples. DS threshold is set to (-0.25, regression fitting model are shown in Fig. 3, in which 0.25) as the filtering criteria for dosage-sensitive genes HYDIN is derived from somatic mutations. (DSGs) and the others were selected as dosage-resistant

Figure 3. Gene data scatter diagram with local polynomial regression fitting model. Each red point represents an independent sample. The black line is the linear regression result, and the blue curve is the LOESS regression result. (A) LRG1. (B) HYDIN. (C) POLRMT. (D) TNFSF10. bioRxiv preprint doi: https://doi.org/10.1101/145268; this version posted June 2, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.

Each red point represents an independent sample. The the right side are dosage-sensitive genes. In the first driver black line is acquired from the linear regression model, and pattern with color pale blue, there are 15 genes in gene the blue curve is resulted from the local polynomial co-expression module 3 and 13 gene modulators. In the regression fitting model. In Fig. 3A and Fig. 3B the slope of second driver pattern with color green, there are 7 genes in the black lines are close to 0, so the dosage sensitivity score module 30 and 18 gene modulators. In third pattern with are nearly equal to 0. In Fig. 3C and Fig. 3D the slope's color red, 6 genes are in module 31 and 14 gene modulators absolute value of the black line are above 1, the blue curves are for regulatory. Remarkably, it shows that CCL11 regulate are mostly overlapped with the black line, so the dosage all the 3 modules; SLC18A1 and C9orf82 regulate the sensitivity score are above 1. Therefore, LRG1 and HYDIN module 3 and module 30; PLXDC1, CCL16, CDC34 and are dosage-resistant genes, whereas POLRMT and TNFSF10 HYDIN regulate the module 3 and module 31; ZFHX4 are dosage-sensitive genes. regulates the module 30 and module 31. Hence, these We identified three major driver patterns in which the top important regulatory genes regulate many gene modules, and 3 gene modules with the largest number of regulatory genes these gene modulators are highly correlated with ovarian (modulators). Fig. 4 shows these driver patterns including cancer (Nolen et al. 2010; Willis et al. 2016; Priebe et al. gene modulators and the corresponding gene modules. Gene 2008; Manabe et al. 2011; Kanska et al. 2016; Morrison et al. modulators were shown in the left side with blue background 2016; Chugh et al. 2016). ovals are the dosage-resistant genes. Gene modulators are in

Figure 4. The 3 major driver patterns for dosage-sensitive and dosage-resistant. The graph depicts inferred regulators (left; ovals with blue background represents the DRGs, and ovals with white background represents the DSGs) and their corresponding regulated gene modules (right). the DSGs are mainly enriched in receptor binding, signaling Enrichment analysis for driver patterns pathway, template transcription, cell proliferation, cell We applied the Gene Ontology (GO) category to expand the differentiation, tissue development, cell division, etc. results of DRGs and DSGs with the R package Compared with DRGs, the GO annotated to DSGs are mostly clusterProfiler, in which the significance threshold are set as management functions and basic functions of the cell and p-value < 0.05 and q-value < 0.05 separately. The top 12 tissue. Fig. 5 shows the top 12 biological process annotation enriched GO terms for DRGs and DSGs are presented in analysis results on GO terms for DRGs and DSGs, Table 2 and Table 3, respectively. The DRGs are mainly respectively. The length of the bar represents the number of enriched in receptor binding, signaling pathway, cell selected genes annotated onto this GO term, and the colors of migration, cellular response, cell chemotaxis, activator the bar represents the significance levels of enrichment. The activity, etc. These processes are highly relevant to the trait of GO enrichment analysis results for DRGs and DSGs are in tumor and contribute to the chief progression of tumor. And bioRxiv preprint doi: https://doi.org/10.1101/145268; this version posted June 2, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.

Additional file 9 and Additional file 10, respectively. From progression by arousing and adjusting the inflammatory GO biological process enrichment analysis, we achieved diseases, which sequently reconstruct the cancer CCL11, CCL16, CCL18, CCL23, CCL8, CCL5, APOB, microenvironment. Moreover, CCL5 also boost metastasis in BRCA1 play extremely significant role in DRGs, and EEF2, ovarian and breast cancer cells (Zsiros et al. 2015; Soria et al. GNA11, GNA15, MED1, MYH8 and TCAP act very 2008). APOB mutation is associated with some human important role in DSGs. CCL11, CCL16, CCL18, CCL23, cancer types such as steatosis, liver, hypocholesterolemia CCL8 and CCL5 have effective chemical attraction on (Cefalu et al. 2013; Williams et al. 2016). BRCA1 mutation eosinophil catalyzing their accumulation at allergic confers high risks of ovarian and breast cancer, encodes a inflammation sites. CCL11 has been shown to negatively tumor suppressor (Rosenberg et al. 2016; Weren et al. 2016). regulate neurogenesis with certain human tumors (Nolen et al. EEF2 is a key components for protein synthesis, and 2010; Ku et al. 2017). CCL16 is a prominently ubiquitously expressed in normal cells (Kaul et al. 2011). expressed in the liver, but also in ovarian and breast cancer GNA11 is associated with Congenital Hemangioma (Ayturk (Manabe et al. 2011; Gantsev et al. 2013). CCL18 is et al. 2016). GNA15 is associated with inhibition of associated with some human cancer types including ovarian proliferation, activation of apoptosis and differential effects cancer (Urquidi et al. 2012; Wang et al. 2015). CCL23 plays (Zanini et al. 2015). MED1 is associated with some a role in subclinical systemic inflammation and associated biological process such as prostate cancer cell growth (Liu et with atherogenesis (Ignacio et al. 2016). CCL8 is al. 2010). MYH8 is a key components for muscle strengthened in stromal fibroblasts at the tumor border and in development and regeneration (Tonami et al. 2013). TCAP is tissues at which breast cancer cells incline to metastasize involved with energy regulation and metabolism, and is such as the lungs and the brain (Farmaki et al. 2016; Zsiros et implicated in the regulation of stress-related behaviors al. 2015). CCL5 is a chemokine that boosts cancer (Aquila et al. 2016). Table 2. The top 12 enriched GO terms for DRGs. GO-BP Term pvalue qvalue gene chemotaxis 3.43E-09 2.12E-06 CCL11/CCL16/CCL18/CCL23/CCL8/CCL5 chemotaxis 4.76E-09 2.12E-06 CCL11/CCL16/CCL18/CCL23/CCL8/CCL5 mononuclear cell migration 1.95E-08 5.80E-06 CCL11/CCL16/CCL18/CCL23/CCL8/CCL5 chemokine-mediated signaling 3.4E-08 5.87E-06 CCL11/CCL16/CCL18/CCL23/CCL8/CCL5 lymphocyte migration 3.67E-08 5.87E-06 CCL11/CCL16/CCL18/CCL23/CCL8/CCL5 neutrophil chemotaxis 3.95E-08 5.87E-06 CCL11/CCL16/CCL18/CCL23/CCL8/CCL5 cellular response to -1 4.91E-08 6.25E-06 CCL11/CCL16/CCL18/CCL23/CCL8/CCL5 neutrophil migration 6.05E-08 6.25E-06 CCL11/CCL16/CCL18/CCL23/CCL8/CCL5 myeloid leukocyte migration 6.31E-08 6.25E-06 CCL11/CCL16/CCL18/CCL23/CCL8/CCL5/AZU1 granulocyte chemotaxis 1.09E-07 9.41E-06 CCL11/CCL16/CCL18/CCL23/CCL8/CCL5 cellular response to tumor 1.16E-07 9.41E-06 APOB/CCL11/CCL16/CCL18/CCL23/CCL8/CCL5/ BRCA1 granulocyte migration 1.84E-07 1.36E-05 CCL11/CCL16/CCL18/CCL23/CCL8/CCL5

Table 3. The top 12 enriched GO terms for DSGs. GO-BP Term pvalue qvalue gene dopamine receptor signaling 2.38E-06 0.001887 PALM/GNA15/GNA11/KLF16 skeletal muscle contraction 0.000127 0.032678 EEF2/TCAP/MYH8 multicellular organismal movement 0.000239 0.032678 EEF2/TCAP/MYH8 musculoskeletal movement 0.000239 0.032678 EEF2/TCAP/MYH8 trabecula morphogenesis 0.000255 0.032678 MED1/SBNO2/UBE4B DNA-templated transcription 0.000289 0.032678 MED1/MED16/POLR2E/MED24/POLRMT phospholipase C-activating 0.000289 0.032678 GNA15/GNA11 natural killer cell chemotaxis 0.000352 0.03489 CCL7/CCL4 negative regulation of keratinocyte 0.000498 0.038042 MED1/KDF1 androgen receptor signaling 0.000544 0.038042 MED1/MED16/MED24 cardiac muscle tissue morphogenesis 0.000544 0.038042 MED1/TCAP/UBE4B response to radiation 0.000691 0.038042 PPP1R1B/CCL7/GNA11/MYC/ECT2/UBE4B bioRxiv preprint doi: https://doi.org/10.1101/145268; this version posted June 2, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.

Figure 5. The analysis of GO biological process annotation analysis for DSGs and DRGs. (A)The GO biological process enrichment of DRGs. (B) The GO biological process enrichment of DSGs hormones are key regulators of metabolism, growth and other Regulatory pathway analysis for driver patterns body systems, MED1, MED16, MED24 and MYC are play a To further elucidate the biological pathway of gene role in this pathway (see Fig. 9). MYC is involved in cell modulators in the driver pattern, we performed the Kyoto cycle progression, apoptosis, and cellular transformation and Encyclopedia of Genes and Genomes (KEGG) pathway is related to many tumors (Kim et al. 2015). Apoptosis is an enrichment analysis of the results of DSGs and DRGs list evolutionarily conserved signaling pathway, which plays a with the R package clusterProfiler separately. The top 12 fundamental role in regulating cell number and eliminating pathways for DRGs and DSGs are presented respectively in damaged or redundant cells (Zeestraten et al. 2013), Table 4 and Table 5. For DRGs, the most significant GADD45B, TNFSF10, LMNB2 are enriched in this pathway canonical pathways are mainly bound up with Chemokine (see Fig. 10). TNFSF10 is associated with some human signaling, -cytokine receptor interaction, cancer types including ovarian cancer (Charbonneau et al. Parkinson's disease, Breast cancer, onset diabetes, Asthma, 2013). LMNB2 acts as regulators of cell proliferation and Prion diseases, etc. And the pathways for DSGs are major differentiation, regards as a cancer risk biomarker in several related to Thyroid hormone signaling, Apoptosis, Cytosolic cancer subtypes (Sakthivel et al. 2016). Cell cycle DNA-sensing, Amoebiasis, Chagas disease, Cell cycle, progression is completed through a reproducible sequence of WNT signaling, etc. Remarkably for DRGs, events. It is a significant pathway associated with many Cytokine-cytokine receptor interaction is embedded in the diseases including ovarian cancer (Taylor_Harding et al. Chemokine signaling pathway. For these two pathways, gene 2013). GADD45B and MYC are enriched in this pathway CCL11, CCL16, CCL18, CCL23, CCL8 and CCL5 have a (see Fig. 11). GADD45B is implicated in some responses to key function (see Fig. 6). play a role in the cell injury including cell cycle checkpoints, apoptosis and migration of many cells during development and are critical DNA repair (Salvador et al. 2013). WNT signaling pathway to nervous system development (Li et al. 2004). Cytokine is required for basic developmental processes and play an interactions play an important role in health and are vital to important role in human stem cells and cancers (Krausova et many cancer during immunological and inflammatory al. 2014). MYC and TBL1XR1 are enriched in this pathway responses in disease. It can lead to antagonist, additive, or (see Fig. 12). The KEGG pathway enrichment analysis result synergistic activities in keeping physiological functions such for DRGs and DSGs are shown in Additional file 11 and 12, as body temperature, feeding and somnus, as well as in respectively. anorectic, ardent fever, and sleepiness neurological representations of acute and chronic disease (Lippitz et al. Conclusions 2013). Parkinson's disease is a pathway specifically The main purpose of this research was to distinguish the associated to diseases of the central nervous system, candidate driver genes and the corresponding driving NDUFB5 and SLC18A1 are enriched in this pathway (see mechanism for resistant and sensitive tumor from the Fig. 7). NDUFB5 and SLC18A1 are associated with multiple heterogeneous data. We presented a machine-learning human diseases (Willis et al. 2016; Chen et al. 2014). Breast approach to integrate somatic mutations, CNVs and gene cancer is the major cause of cancer death among the female expression profiles to distinguish interactions and worldwide, gene BRCA1 and FGF22 are enriched in this regulations for dosage-sensitive and dosage-resistant genes pathway (see Fig. 8). FGF22 is required for brain of ovarian cancer. We developed a general framework for development and associated with hereditary and neoplastic integrating high dimensional heterogeneous omics data and disease (Heroult et al. 2014). For DRGs, the thyroid can potentially lead to new insight into many related studies bioRxiv preprint doi: https://doi.org/10.1101/145268; this version posted June 2, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.

Table 4. The top 12 enriched GO terms for DSGs. pathway pvalue gene Chemokine signaling pathway 4.80E-06 CCL11/CCL16/CCL18/CCL23/CCL8/CCL5 Cytokine-cytokine receptor 4.58E-05 CCL11/CCL16/CCL18/CCL23/CCL8/CCL5 Parkinson’s disease 0.052537 NDUFB5/SLC18A1 Breast cancer 0.053866 BRCA1/FGF22 Other types of O-glycan biosynthesis 0.056251 MFNG Ribosome 0.060691 RPS15/RPL19 Vitamin digestion and absorption 0.061213 APOB Maturity onset diabetes of the young 0.066151 HNF1B NOD-like receptor signaling 0.070724 MFN1/CCL5 Phototransduction 0.071064 RCVRN Asthma 0.078387 CCL11 Prion diseases 0.088067 CCL5

Table 5. The top 12 enriched GO terms for DSGs. pathway pvalue gene Thyroid hormone signaling 0.000438 MED1/MED16/MED24/MYC Apoptosis 0.009077 GADD45B/TNFSF10/LMNB2 Cytosolic DNA-sensing pathway 0.017271 POLR2E/CCL4 NF-kappa B signaling pathway 0.036048 GADD45B/CCL4 Amoebiasis 0.036744 GNA15/GNA11 Chagas disease 0.041034 GNA15/GNA11 Cholinergic synapse 0.048583 PIK3R5/GNA11 Cytokine-cytokine receptor 0.05237 CCL7/CCL4/TNFSF10 Cell cycle 0.058258 GADD45B/MYC FoxO signaling pathway 0.065054 GADD45B/TNFSF10 mediated proteolysis 0.069434 CDC34/UBE4B Wnt signaling pathway 0.074816 MYC/TBL1XR1

Figure 6. The six genes are enriched in the top 2 significant pathways (red, up-regulated). (A) Chemokine signaling pathway. (B) Cytokine-cytokine receptor interaction. bioRxiv preprint doi: https://doi.org/10.1101/145268; this version posted June 2, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.

Figure 7. Parkinson disease pathway(red, up-regulated). NDUFB5 and SLC18A1 are enriched in this pathway. NDUFB5 is a gene alias of the complex 1 deficiency(Cx1) and SLC18A1 is a gene alias of the vesicular amine transporter(VMAT).

Figure 8. Breast cancer pathway (red, up-regulated). BRCA1 and FGF22 are enriched in this pathway. bioRxiv preprint doi: https://doi.org/10.1101/145268; this version posted June 2, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.

Figure 9. thyroid hormones pathway (red, up-regulated). MED1, MED16, MED24 and MYC are enriched in this pathway.

Figure 10. Apoptosis pathway (red, up-regulated). GADD45B, TNFSF10, LMNB2 are enriched in this pathway. TRAIL is a gene alias of TNFSF10, LMNB2 is a gene alias of the lamin. bioRxiv preprint doi: https://doi.org/10.1101/145268; this version posted June 2, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.

Figure 11. Cell cycle pathway (red, up-regulated). GADD45B and MYC are enriched in this pathway.

Figure 12. Wnt signaling pathway (red, up-regulated). MYC and TBL1XR1 are enriched in this pathway. at the systems level. In the framework, weighted correlation vertex of network; the regression tree model is utilized for network analysis (WGCNA)is applied to the co-expression module networks learning in which the obtained gene network analysis; the mutation network is constructed by modules and candidate modulators are trained for the integrating the CNVs and somatic mutations and the initial modulators regulatory mechanism; finally, a local candidate modulators is selected from the clustering the polynomial regression fitting model is applied to identify bioRxiv preprint doi: https://doi.org/10.1101/145268; this version posted June 2, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.

dosage-sensitive and dosage-resistant driver patterns. From analysis is shown in Fig. 1. As can be seen in Fig. 1, the the Gene Ontology and pathway enrichment analysis, we procedure takes these datasets as inputs and then produces a obtained some biologically meaningful gene modulators, short list of genes as candidate divers. The method contains such as CCL11, CCL16, CCL18, CCL23, CCL8, CCL5, the following steps: (1) extract a pool of candidate APOB, BRCA1, SLC18A1, FGF22, GADD45B, GNA15, modulators as initial driver genes, which have significant GNA11 and so on, which can be conducive to appropriate CNVs/Somatic mutations in ovarian cancer samples and diagnosis and treatment to cancer patients. achieve the initial gene modules from the gene expression by weighted correlation network analysis (WGCNA); (2) Methods construct the heterogeneous network by module networks learning and evaluate the regulatory mechanism between We integrated gene expression, CNV, and somatic these different omics data by liner regression model; (3) mutation data to identify the driver patterns for resistant and identify driver patterns and modulators sensitive tumors. The overview of the overall integrative

Figure 13. the scheme of driver patterns identification. acquire the initial gene modules for gene expression profiles via WGCNA network analysis (on the left)and extract a pool of candidate modulators as initial driver genes integrate CNVs and Somatic mutations in ovarian cancer samples (on the right); (2) construct the heterogeneous network by module networks learning and evaluate the regulatory relationships by liner regression model;(3) identify modulators interactions in the constructed regulatory networks and analyze the driver patterns through and gene ontology analysis and pathway enrichment analysis. interactions in the constructed regulatory networks using Affymatrix Genome-Wide Human SNP Array 6.0 platform, gene ontology and pathway enrichment analysis. and somatic mutation data from whole exome sequencing data obtained using Illumina Genome Analyzer DNA Resources and datasets Sequencing. The clinical data of these patients are analyzed We downloaded gene expression data from the Agilent to determine cis-platinum chemotherapy response samples 244K Custom Gene Expression platform, CNV data from the and these tumor samples are classified into resistant and bioRxiv preprint doi: https://doi.org/10.1101/145268; this version posted June 2, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.

sensitive groups, sensitive tumors have a platinum-free space In which the network structure is not only based on the two of six months or more after the last primary treatment, no sign genes with direct adjacency correlation, but also on the of residue or relapse, and the follow-up will at least six indirect adjacency relationship. months later; and resistant tumors are recurred within six The gene expression network is constructed based on months after the last treatment (Prado_2014). According to matrix Ω. Gene modules are detected by unsupervised the above definition, we identified 93 platinum-resistant and hierarchical clustering method. Give the qth gene 231 platinum-sensitive primary tumors from the TCGA co-expression module, the eigengene expression is E()q . The website, for which the gene expression profiles, CNV, and correlation coefficient for the gene x and the module somatic mutation data were available as well. The 324 i samples list are shown in Additional file 1. The gene can be defined as: expression data are listed in Additional file 2. The CNVs data qq   kii c o r x E  ,  (5) are listed in Additional file 3. The somatically mutated genes q are listed in Additional file 4. Where ki  1,1. Gene co-expression modules analysis using Modulators identification using module networks WGCNA learning Weighted correlation network analysis is a framework for From the initial gene co-expression modules, we focused on co-expression analysis to explore gene modules of high the candidate modulators' regulatory by regression tree correlation to external sample traits. The gene expression model. Each candidate modulator is connected to a group of correlation networks are based on quantitative measurements genes by maximizing the normal-gamma scoring function correlations. Given an nm expression matrix X where the which describes the qualitative behavior of a candidate row indices correspond to genes ( in1,, ) and the modulator that regulate the gene expression modules. A column indices ( jm1, , ) correspond to samples: regression tree is formed with two basic blocks: decision nodes and leaf nodes. Each node corresponds to one of the XXX11121 m  selected modulators above. Our learning is an iterative XXX process. In each iteration, a regulation procedure is found for X  21222 m (1)  each gene module and then each gene reassigned to the  corresponding module due to prediction with the best XXXnnnm12 regulatory function. We searched for the model with the The co-expression similarity S is applied to calculate the highest score by using the Expectation Maximization (EM) ij algorithm. The scoring function defined as: adjacency matrix: log(,)log(|)log()p C Np CNp N (6) Sij cor(,) x i x j (2) Where C is the candidate data and N the model structure of Where S is the absolute value of the correlation coefficient the network. The first term is the candidate data for a given ij model which meets normal gamma distribution. The second between gene i ( ) and gene j ( jn1,, ) in the part is a penalty score on network complexity. The EM algorithm ensures the reliability of the model until expression profiles matrix X. The value of Sij is in [0, 1]. convergence to a local maximum score. The essence of the The adjacency function ij is defined by soft threshold of algorithm consists of two steps: the first procedure is learning the best regulation program (regression tree) for each module. co-expression similarity Sij by soft threshold  as The tree is constructed from the root to its leaves. The second following: procedure is to find the association rules for corresponding  ijikikpower(,) SS (3) module. Where  >1. To measure the topological structure with Dosage-sensitive and dosage-resistant analysis adjacency function, topological overlap matrix (TOM) is From the module networks learning, the candidate genes applied to build the matrix , ij were generated. Then we applied a local polynomial au regression fitting model to conduct dosage-sensitive and w  ij ij dosage-resistant analysis. In the model, a predict function ij min(m , m ) 1 a i j ij (4) was adopted to obtain n isometric points on the LOESS curve (Yan_2017). To achieve the monotonicity of the copy m , u a a i ik ij ik kj number expression interrelation a function is defined as M: k, k i k bioRxiv preprint doi: https://doi.org/10.1101/145268; this version posted June 2, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.

2 n j for ovarian cancer using gene-expression microarrays. MS (7)  ij British Journal of Cancer 90: 686–692. nn(1) ji11 Youn A, Simon R. 2010. Identifying cancer driver genes in In which tumor genome sequencing studies. Bioinformatics 27:

 1, ssji 175–181.  Tomasetti C, Marchionni L, Nowak MA, Parmigiani G, S sijji s 0, (8) Vogelstein B. 2014. Only three driver gene mutations 1, ss  ji are required for the development of lung and colorectal Where n is the number of samples. To quantize the cancers. Proceedings of the National Academy of association between gene expression and CNVs (somatic Sciences 112: 118–123. mutations), another linear fitting model was adopted and a straight line with slope of K was obtained. The slope of linear Jang JSJ, Cho HY, Lee YJ, Ha WS, Kim HW. 2004. The model can reflect the relationship of variables in LOSS differential proteome profile of stomach cancer: model to some extent. Hence the dosage sensitivity (DS) identification of the biomarker candidates. Oncology score can be defined as: Research Featuring Preclinical and Clinical Cancer D S = M K (9) Therapeutics 14: 491–499. Herein, for a gene, a larger DS indicates stronger dosage Xiong M, Fang X, Zhao J. 2001. Biomarker identification by sensitivity. feature wrappers. Genome Research 11: 1878–1887. Jung Y, Lee S, Choi HS, Kim SN, Lee E, Shin Y, Seo J, Kim Competing interests B, Jung Y, Kim WK, et al. 2011. Clinical validation of The authors declare that they have no competing interests. colorectal cancer biomarkers identified from Data access bioinformatics analysis of public expression data. Clinical Cancer Research 17: 700–709. The CNV, mutation and gene expression data are used in this paper can be downloaded from Broad GDAC Firehose website Logsdon BA, Gentles AJ, Miller CP, Blau CA, Becker PS, (http://gdac.broadinstitute.org/runs/analyses__2014_10_17/dat Lee SI. 2015. Sparse expression bases in cancer reveal a/OV/20141017). tumor drivers. Nucleic Acids Research 43: 1332–1344. Acknowledgements Chan KCA, Jiang P, Zheng YWL, Liao GJW, Sun H, Wong J, The authors are grateful to Peter Langfelder's team who Siu SSN, Chan WC, Chan, SL,et al. 2012. Cancer provided the weighted correlation network analysis method genome scanning in plasma: Detection of with WGCNA package, Prof Ripley B. D. who provided the tumor-associated copy number aberrations, Local Polynomial Regression Fitting analysis method with single-nucleotide variants, and tumoral heterogeneity by LOESS package, Dana Pe'er Lab who provided the Module massively parallel sequencing. Clinical C. Networks learning analysis method with CONEXIC Toolkit Ding J, McConechy MK, Horlings HM, Ha G, Chan FC, and Guangchuang Yu's team who provided the Functional Funnell T, Mullaly SC, Reimand, J Bashashati A, Bader enrichment analysis method with clusterProfiler package. This work was supported by the National Natural Science GD, et al. 2015. Systematic analysis of somatic Foundation grant of China and China Scholarship Council mutations impacting gene expression in 12 tumour types. grant. Nature Communications 6, 8554. References Wei PJ, Zhang D, Xia J, Zheng CH. 2016. LNDriver: identifying driver genes by integrating mutation and Cheng F, Zhao J, Zhao Z. 2015. Advances in computational expression data based on gene-gene interaction network. approaches for prioritizing driver mutations and BMC Bioinformatics 17: 221. significantly mutated genes in cancer genomes. Briefings Zhang J, Zhang S. 2016. The discovery of mutated driver in Bioinformatics 17: 642–656. pathways in cancer: Models and algorithms. IEEE/ACM Vogelstein B, Papadopoulos N, Velculescu VE, Zhou S, Transactions on Computational Biology and Diaz LA, Kinzler KW. 2013. Cancer genome landscapes. Bioinformatics. 1: 1. Science 339: 1546–1558. Kumar RD, Swamidass SJ, Bose R. 2016. Unsupervised Adib TR, Henderson S, Perrett C, Hewitt D, Bourmpoulia D, detection of cancer driver mutations with Ledermann J, Boshoff C. 2004. Predicting biomarkers parsimony-guided learning. Nature Genetics 48: 1288–1294. bioRxiv preprint doi: https://doi.org/10.1101/145268; this version posted June 2, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.

Huang N, Shah PK, Li C. 2011. Lessons from a decade of Pharmacotherapy 67: 363–366. integrating cancer copy number alterations with gene Urquidi V, Kim J, Chang M, Dai Y, Rosser CJ, Goodison S. expression profiles. Briefings in Bioinformatics 13: 2012. CCL18 in a multiplex urine-based assay for the 305–316. detection of bladder cancer. PLoS ONE 7: 37797. Osterhoff M, Frahnow T, Seltmann A, Mosig A, Neunübel K, Wang Q, Tang Y, Yu H, Yin Q, Li M, Shi L, Zhang W, Li D, Sales S, Sampaio J, Hornemann S, Kruse M, Pfeiffer A. Li L. 2012. CCL18 from tumor-cells promotes epithelial 2014. Identification of gene-networks associated with ovarian cancer metastasis via mTOR signaling pathway. specific lipid metabolites by weighted gene Molecular Carcinogenesis 55(11), 1688–1699 (2015). co-expression network analysis (WGCNA). Ignacio RMC, Gibbs CR, Lee ES, Son DS. 2016. Differential Experimental and Clinical Endocrinology & Diabetes chemokine signature between human preadipocytes and 122: P098. adipocytes. Immune Network 16: 189. Liu J, Jing L, Tu X. 2016. Weighted gene co-expression Farmaki E, Chatzistamou I, Kaza V, Kiaris, H. 2016. A CCL8 network analysis identifies specific modules and hub gradient drives breast cancer cell dissemination. genes related to coronary artery disease. BMC Oncogene 35: 6309–6318 (). Cardiovascular Disorders 16: 54. Zsiros E, Dangaj D, June CH, Kandalaft LE, Coukos G. 2015. Nolen BM, Lokshin AE. 2010. Targeting CCL11 in the Ovarian cancer chemokines may not be a significant treatment of ovarian cancer. Expert Opinion on barrier during whole tumor antigen dendritic-cell Therapeutic Targets 14: 157–167. vaccine and adoptive t-cell immunotherapy. Willis S, Villalobos VM, Gevaert O, Abramovitz M, OncoImmunology 5: 1062210. Williams C, Sikic BI, Leyland-Jones B. 2016. Single Soria G, Ben-Baruch A. 2008. The inflammatory chemokines gene prognostic biomarkers in ovarian cancer: A CCL2 and CCL5 in breast cancer. Cancer Letters 267: meta-analysis. PLOS ONE 11: 0149183. 271–285. Priebe A, Buckanovich RJ. 2008. Ovarian tumor vasculature Cefalu AB, Pirruccello JP, Noto D, Gabriel S, Valenti V, as a source of biomarkers for diagnosis and therapy. Gupta N, Spina R, Tarugi P, Kathiresan S, Averna MR. Expert Review of Obstetrics & Gynecology 3: 65–72. 2013. A novel APOB mutation identified by exome Manabe S, Iwase A, Goto M, Kobayashi H, Takikawa S, sequencing cosegregates with steatosis, liver cancer, and Nagatomo Y, Nakahara T, Bayasula, Nakamura T, hypocholesterolemia. Arteriosclerosis, Thrombosis, and Hirokawa W, Kikkawa F. 2011. Expression and Vascular Biology 33: 2021–2025. localization of CXCL16 and CXCR6 in ovarian Williams JK, Katapodi MC, Starkweather A, Badzek L, endometriotic tissues. Archives of Gynecology and Cashion AK, Coleman B, Fu MR, Lyon D, Weaver MT, Obstetrics 284: 1567–1572. Hickey KT. 2016. Advanced nursing practice and Kanska J, Zakhour M, Taylor-Harding B, Karlan BY, research contributions to precision medicine. Nursing Wiedemeyer WR. 2016. Cyclin e as a potential Outlook 64: 117–123. therapeutic target in high grade serous ovarian cancer. Rosenberg SM, Ruddy KJ, Tamimi RM, Gelber S, Schapira Gynecologic Oncology 143: 152–158. L, Come S, Borges VF, Larsen B., Garber JE, Partridge Morrison E, Wai P, Leonidou A, Bland P, Khalique S, Farnie AH. 2016. BRCA1 and BRCA2 mutation testing in G, Daley F, Peck B, Natrajan R. 2016. Utilizing young women with breast cancer. JAMA Oncology 2: functional genomics screening to identify potentially 730. novel drug targets in cancer cell spheroid cultures. Weren RDA, Mensenkamp AR, Simons M, Eijkelenboom A, Journal of Visualized Experiments 118: e54738-e54738. Sie AS, Ouchene H, van Asseldonk M, Gomez-Garcia Chugh S, Meza J, Sheinin YM, Ponnusamy MP, Batra SK. EB, Blok MJ, de Hullu JA, et al. 2016. Novel BRCA1 2016. Loss of n-acetylgalactosaminyltransferase 3 in and BRCA2 tumor test as basis for treatment decisions poorly differentiated pancreatic cancer: augmented and referral for genetic counselling of patients with aggressiveness and aberrant ErbB family glycosylation. ovarian carcinomas. Human Mutation 38: 226–235. British Journal of Cancer 114: 1376–1386. Kaul G, Pattan G, Rafeequi T. 2011. Eukaryotic elongation Kuć P, Charkiewicz R, Klasa-Mazurkiewicz D, Milczek T, factor-2 (eEF2): its regulation and peptide chain Mroczko B, Nikliński J, Laudański P. 2017. Profiling of elongation. Cell Biochemistry and Function 29: selected angiogenesis -related genes in serous ovarian 227–234. cancer patients. Advances in Medical Sciences 62: Ayturk UM, Couto JA, Hann S, Mulliken JB, Williams KL, 116–120. Huang AY, Fishman SJ, Boyd TK, Kozakewich HPW, Gantsev SK, Umezawa K, Islamgulov DV, Khusnutdinova Bischoff J, Greene AK, Warman ML. 2016. Somatic EK, Ishmuratova RS, Frolova VY, Kzyrgalin SR. 2013. activating mutations in GNAQ and GNA11 are The role of inflammatory chemokines in lymphoid associated with congenital hemangioma. The American neoorganogenesis in breast cancer. Biomedicine & Journal of Human Genetics 98: 1271. bioRxiv preprint doi: https://doi.org/10.1101/145268; this version posted June 2, 2017. The copyright holder for this preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.

Zanini S, Giovinazzo F, Alaimo D, Lawrence B, Pfragner R, requirements shape ovarian cancer progression. Cancer Bassi C, Modlin I, Kidd M. 2015. GNA15 expression in Research 73: 1749. small intestinal neuroendocrine neoplasia: Functional Salvador JM, Brown-Clay JD, Fornace AJ. 2013. Gadd45 in and signalling pathway analyses. Cellular Signalling 27: stress signaling, cell cycle control, and apoptosis. In: 899–907. Advances in Experimental Medicine and Biology: 1–19 Liu G, Sprenger C, Wu PJ, Sun S, Uo T, Haugk K, Epilepsia Krausova M, Korinek V. 2014. Wnt signaling in adult KS, Plymate S. 2010. MED1 mediates androgen intestinal stem cells and cancer. Cellular Signalling 26: receptor splice variant induced gene expression in the 570–579. absence of ligand. Oncotarget 1: 288–304. Prado CMM, Baracos VE, Xiao J, Birdsell L, Stuyckens K, Tonami K, Hata S, Ojima K, Ono Y, Kurihara Y, Amano T, Park YC, Parekh T, Sawyer MB. 2014. The association Sato T, Kawamura Y, Kurihara H, Sorimachi H. 2013. between body composition and toxicities from the Calpain-6 deficiency promotes skeletal muscle combination of doxil and trabectedin in patients with development and regeneration. PLoS Genetics 9(8), advanced relapsed ovarian cancer1. Applied Physiology, 1003668 (). Nutrition, and Metabolism 39: 693–698. D’Aquila AL, Hsieh AHR, Hsieh AHM, Almeida RD, Yan Z, Liu Y, Wei Y, Zhao N, Zhang Q, Wu C, Chang Z, Xu Lovejoy SR, Lovejoy DA. 2016. Expression and actions Y. 2017. The functional consequences and prognostic of corticotropin-releasing factor/diuretic hormone-like value of dosage sensitivity in ovarian cancer. Mol. peptide (CDLP) and teneurin c-terminal associated BioSyst 13: 380–391. peptide (TCAP) in the vase tunicate, ciona intestinalis: Antagonism of the feeding response. General and Comparative Endocrinology. Li Q, Shirabe K, Kuwada JY. 2004. Chemokine signaling regulates sensory cell migration in zebrafish. Developmental Biology 269: 123–136. Lippitz BE. 2013. Cytokine patterns in patients with cancer: a systematic review. The Lancet Oncology 14: 218–228. Chen M, Wu L, Wu F, Wittert GA, Norman RJ, Robker RL, Heilbronn LK. 2014. Impaired glucose metabolism in response to high fat diet in female mice conceived by in vitro fertilization (IVF) or ovarian stimulation alone. PLoS ONE 9: 113155. Heroult M, Ellinghaus P, Ince S, Ocker M. 2014. Fibroblast receptor signaling in cancer biology and treatment. Current Signal Transduction Therapy 9: 15–25. Kim T, Jeon YJ, Cui R, Lee JH, Peng Y, Kim SH, Tili E, Alder H, Croce CM. 2015. Role of MYC-regulated long noncoding RNAs in cell cycle regulation and tumorigenesis. JNCI Journal of the National Cancer Institute 107: 505–505. Zeestraten BA, Reimers M, Schouten P, Schouten P, van VC, Kuppen P, Liefers G. 2013. The prognostic value of the apoptosis pathway in colorectal cancer: A review of the literature on biomarkers identified by immunohistochemistry. Biomarkers in Cancer, 13. Charbonneau B, Block MS, Bamlet WR. 2013. Risk of ovarian cancer and the NF- b pathway: Genetic association with IL1a and TNFSF10. Cancer Research 74: 852–861. Sakthivel KM, Sehgal P. 2016. A novel role of lamins from genetic disease to cancer biomarkers. Oncology Reviews 10(2). Taylor-Harding B, Agadjanian H, Orsulic S, Walsh C, Karlan BY, Wiedemeyer WR. 2013. Abstract 1749: Cell cycle