{"paper_id":"15a8661c-131f-4443-b93f-f06d83d67da7","body_text":"Page 1/25\nIdentication and validation of shared genes and\nkey pathways in endometriosis and endometriosis-\nassociated ovarian cancer by weighted gene co-\nexpression network analysis and machine learning\nalgorithms\nYongming Du \nNingbo First Hospital\nYichao Hu \nNingbo First Hospital\nFubin Zhang \nNingbo First Hospital\nTianhong Zhu \nNingbo First Hospital\nYutao Guan  (  guanyutao@nbdyyy.com )\nNingbo First Hospital\nResearch Article\nKeywords: endometriosis, endometriosis-associated ovarian cancer, WGCNA, machine learning\nalgorithms\nPosted Date: February 7th, 2023\nDOI: https://doi.org/10.21203/rs.3.rs-2542861/v1\nLicense:   This work is licensed under a Creative Commons Attribution 4.0 International License.  \nRead Full License\n\nPage 2/25\nAbstract\nBackground: Endometriosis is a widespread disease in reproductive age. Epidemiological studies reported\nthat patients with endometriosis had an increased risk of developing endometriosis-associated ovarian\ncancer (EAOC). The present study aimed to identify shared genes and key pathways that commonly\ninteracted between EAOC and endometriosis.\nMethods: The expression matrix of ovarian cancer and endometriosis were collected from the Gene\nExpression Omnibus database. The weighted gene co-expression network analysis (WGCNA) was used to\nconstruct co-expression gene network. Functional enrichment analyses were conducted to clarify the\npotential regulatory mechanisms. Protein-protein interaction (PPI) network and machine learning\nalgorithms were applied to identify characteristic genes. CIBERSORT deconvolution algorithm was used\nto explore the difference in tumor immune microenvironment. Receiver operating characteristic curves\nwere utilized to assess the clinical diagnostic ability of hub genes. Furthermore, diagnostic nomogram\nwas constructed and evaluated for supporting clinical practicality.\nResults: We identied 262 shared genes between EAOCand endometriosis via WGCNA analysis. They\nwere mainly enriched in cytokine-cytokine receptor interaction, which may be considered a common\nmechanism between EAOC and endometriosis. After PPI network and machine learning algorithms, we\nrecognized two characteristic genes (EDNRA, OCLN) and established a nomogram that presented an\noutstanding predictive performance. The hub genes demonstrated remarkable associations with\nimmunological functions. OCLN were highly upregulatedin ovarian cancer compared to non-tumor\ntissues, while expression levels of EDNRA were signicantly downregulated in ovarian cancer samples.\nSurvival analysis indicated that dysregulated expressions of EDNRA and OCLNwere closely correlated\nwith prognosis of ovarian cancer patients. GSEA analyses revealed that the two characteristic genes were\nmainly enriched in the cancer- and immune-related pathways. Gene drug interaction analysis found 15\ndrugs compound that interacted with the hub genes.\nConclusion: We identied two hub genes (EDNRA, OCLN) and constructed a nomogram to predict the risk\nof EAOC based on WGCNA analyses and machine learning algorithms. They can be used as effective\npredictive biomarkers for detecting EAOC. Our ndings pave the way for further investigation of potential\ncandidate genes and will aid in improving the diagnosis and treatment of EAOC in endometriosis\npatients.\nIntroduction\nEndometriosis is a disease caused by the implantation of active and growing ectopic endometrium in the\novary, commonly in women of childbearing age, which is benign but with invasive and recurrent\nmalignancy-like behavior [1, 2]. Epidemiological studies reported that patients with endometriosis,\nespecially ovarian endometriosis cysts, had an increased risk of developing epithelial ovarian cancer,\nsignicantly higher than the general population, and 0.7%-2.5% of endometriosis was associated with\n\nPage 3/25\nmalignant transformation to endometriosis-associated ovarian cancer (EAOC) [3, 4]. The main clinical\nfeatures of patients at risk of endometriosis malignancy are: (i) a long history of endometriosis (> 10\nyears); (ii) having endometriosis at an early age; (iii) a history of endometriosis-related infertility and/or\ninfertility treatment; and (iv) ovarian endometriosis cysts. Sampson originally proposed the criteria of\nEAOC caused by endometriosis included that the tumor needed to be situated next to unequivocal foci of\nendometriosis, no other primary tumor present, and the presence of tissues similar to endometrial stroma\nsurrounding epithelial glands [5]. Growing number of evidences demonstrated that atypical endometriosis\nwas a precancer lesion for ovarian cancer [6–8]. EAOC patients have a relatively early age of onset and\nearly clinical stage, with the proportion of stage I patients accounting for approximately 54.2%. Age\nfactors, economic status, depression, pelvic inammatory disease, and endometriosis-related infertility\nare high-risk factors for EAOC.\nThe development of sequencing technologies and bioinformatics has made it possible to investigate the\ngenetic pathogenesis of disease-disease interactions. This study aimed to identify the co-expression\ngenes modules and shared genes of EAOC and endometriosis via weighted gene co-expression network\nanalysis (WGCNA) [9] and machine learning algorithms. WGCNA enables the reconstruction of gene\nregulatory networks and the identication of modules of genes with similar expression patterns. By\nintegrating gene expression data with other omics data, WGCNA provides a comprehensive view of gene\ninteractions and can help to uncover the underlying biological pathways involved in complex diseases\n[10, 11]. Machine learning algorithm has been utilized to identify characteristic genes for several diseases\n[12, 13]. Support vector machine - recursive feature elimination (SVM-RFE) [14] algorithm eliminates\nsuperuous variables and retains only those variables that are relevant to the results. Least absolute\nshrinkage and selection operator (LASSO) [15] logistic regression is commonly utilized to narrow down\nvariables and identify the most appropriate λ  for categorization. Random forest (RF) [16] is used to work\nwith high-dimensional data, construct predictive model and predict the importance of each variable. It is\ngenerally accepted that bioinformatics and machine learning algorithms can be effectively employed to\nidentify shared genes and biomarkers for forecasting EAOC.\nOur study revealed that EAOC and endometriosis shared 262 genes. They were mainly enriched in\ncytokine-cytokine receptor interaction, which may be considered a common mechanism between EAOC\nand endometriosis. After machine learning algorithm in the conrmation cohort, we identied two hub\ngenes (EDNRA, OCLN) and established a nomogram. They can be used as effective predictive biomarkers\nfor detecting occult EAOC in endometriosis patients. Additionally, the dysregulated degree of immune cell\ninltration and association of hub genes with immune cells were initially discussed.\nMaterial And Methods\nData sources and preprocessing\nThe expression proles of ovarian carcinoma were collected from the Gene Expression Omnibus (GEO)\ndatabase. In total, four available datasets were collected via R package “GEOquery”. The high throughput\n\nPage 4/25\nsequencing dataset GSE157153 including 29 ovarian carcinoma tissues and 37 endometriosis tissues\nwas used to explore the shared genes between endometriosis and endometriosis-associated ovarian\ncancer (EAOC). GSE5108 contained 11 ectopic endometrial samples and 11 eutopic endometrial tissues,\nGSE18520 consisted of 53 ovarian adenocarcinoma tissues and 10 normal ovary samples. Both of them\nwere utilized to conrm the shared genes. Furthermore, the expression matrix of malignant ovarian\nsamples from The Cancer Genome Atlas (TCGA) dataset and expression prole of normal ovarian tissues\nfrom the Genotype-Tissue Expression (GTEx) also served as a validation dataset. The study owchart is\nshown in Fig. 1.\nWGCNA\nWe constructed co-expression gene networks of endometriosis and EAOC using WGCNA in order to\ninvestigate the underlying biological processes and common genes involved in disease progression. The\ndata were preprocessed by calling the \"WGCNA\" function package, and the \"goodsamplegenes()\" function\nwas used to check whether there were abnormal genes in the data and select appropriate thresholds to\nreject outliers. The clinical feature data were loaded, and the gene-feature clustering tree was drawn to\ncomplete the construction of the co-expression network and the selection of modules. The\n\"pickSoftThreshold\" algorithm chose an appropriate \"soft\" threshold ( β ) to calculate the adjacency matrix\nand to obtain a biologically signicant scale-free network. The clustering dendrogram was drawn for\nmodule selection, and the minimum module capacity and module shear height were set to merge\nmodules with high similarity. The EAOC gene set in the modules that were highly correlated with ovarian\ncarcinoma. Scatter plots of gene signicance (GS) and module membership (MM) in the signicant co-\nexpression modules were plotted to clarify the signicance of genes in the selected modules.\nFunctional enrichment analysis\nGene set enrichment analysis (GSEA) was performed to clarify the potential regulatory mechanisms\nresponsible for the disease progression of endometriosis to EAOC, and the gene set\n(“c2.cp.kegg.v7.4.symbols.gmt”) were obtained from the Broad Institute database. Gene Ontology (GO)\nand Kyoto Encyclopedia of Genes and Genomes (KEGG) functional enrichment were also conducted.\nGene Set Variation Analysis (GSVA) was utilized to reveal the differences in signaling pathways between\nthe endometriosis and EAOC by \"GSVA\" R package. The selected gene set “h.all.v7.4.symbols.gmt” was\ngot from the Molecular Signature Database (MSigDB). All these analyses were performed via R packages\n“limma”, “clusterProler”, “gseaplot2”, and “ggplot2”.\nConrmation of shared genes in endometriosis and EAOC\nTo conrm the shared genes in endometriosis and EAOC, we applied the differentially expressed gene\n(DEGs) analysis in conrmation dataset GSE157153 including both endometriosis and EAOC tissues with\nthe cutoff value of adjusted p < 0.05 and |Log2 (fold change)| >1 via “limma” package. The overlap of\nabove shared genes from WGCNA analyses of GSE5108 and GSE18520 and DEGs in the conrmation\n\nPage 5/25\ncohort GSE157153 were presented by a Venn diagram using EVenn (http://www.ehbio.com/test/venn/#/)\nonline tool.\nIdentication of hub genes from protein-protein interaction\nnetwork\nThe STRING database was used for protein interaction analysis for a set of shared genes and the results\nwere imported into the Cytoscape software to investigate the target modules and hub genes from the\nprotein-protein interaction (PPI) network. An interaction score with a condence level of 0.4 was used as\nthe cutoff criterion for construction and unconnected nodes were hidden. Then we utilized four different\nalgorithms (Closeness, Betweenness, MCC, and Degree) via the CytoHubba plugin to identify the top 30\nhub shared genes.\nMachine learning algorithm\nThree machine learning algorithms including SVM-RFE, RF, and LASSO regression analysis were used to\ninvestigate the effectiveness of 23 shared DEGs via packages of “msvmRFE.R”, “caret”, and “glmnet”.\nThese algorithms were performed to carry out the binary classication of the data through supervised\nlearning method. Receiver operating characteristic (ROC) curves were calculated based on 23 common\ngenes from the GSE157153 transcriptome dataset, and area under the ROC curves (AUCs) were used to\nevaluate the diagnostic ability of commonly shared genes from the three classication algorithms. The\ndataset was randomly divided into training cohort and test cohort in a ratio of 4:1.\nNomogram construction\nThe common genes obtained as a result of the three classication algorithms were combined to\nconstruct the nomogram via “rms” package. The consistence between the nomogram − predicted\nprobability of nonadherence and the actual results were evaluated by calibration curves. We also used\ndecision curve analysis (DCA) to summarize the performance of our shared gene constructed model and\nto evaluate whether this model has utility in supporting clinical decisions. All these analyses were\nperformed via R packages “VRPM”, “rmda”, “ggDCA”, “rms”, and “ggplot2”.\nExpression prole and survival analysis of shared genes\nbetween endometriosis and EAOC\nWe calculated the different expression of shared genes between endometriosis and EAOC in the\nGSE157153 dataset. Then, these expression proles were evaluated in the GSE18520 and TCGA-OV\ndatabase. Boxplot was utilized to present the differential expression of these genes. Furthermore, the\noverall survival of these shared genes was compared by Kaplan-Meier method according to the package\n“survminer” and “survival”. An adjusted p-value < 0.05 was considered statistically signicant.\nTumor microenvironment cell inltration analyses\n\nPage 6/25\nTo quantify the relative proportion of immune cells in endometriosis cohort and EAOC cohort, CIBERSORT\ndeconvolution algorithm was used to evaluate the immune inltration proportion of 22 immune cell\nsubtypes in accordance to the gene expression values. Spearman correlation analysis was used to\ndisplay the association between immune inltration proportion of immune cells and the expression of\nhub genes.\nPredicting regulatory targets of transcription factor\nThe JASPAR TF binding site prole database was used to identify transcription factors (TFs) that bound\nto shared genes. Additionally, shared genes regulation-related miRNA that negatively correlated with\nprotein expression were predicted by TarBase (v8.0). The topology networks of TFs-gene and miRNA-gene\nwere constructed based on the NetworkAnalyst platform.\nEvaluation of clinical diagnostic ability of shared genes\nROC curves and AUCs were utilized to assess whether shared genes could well distinguish EAOC patients\nfrom control group via “pROC” package in GSE157153. Additionally, GSE18520 and TCGA-OV datasets\nwere used to assess and explore the shared genes with the most clinical diagnostic ability.\nThe protein expressions and shared genes between\nendometriosis and EAOC\nThe Human Protein Atlas (HPA) database (https://www.proteinatlas.org/) mapped all the human proteins\nin cells, tissues and organs by integrating multiple omics technologies. We searched the representative\nimages of immunohistochemistry showing shared genes expression in endometriosis and EAOC tissues.\nGene–Drug interaction analysis\nAlthough there is promising development of novel therapeutic drugs for ovarian cancer, the prognosis of\novarian cancer is often poor. In the present study, we searched the drug-gene interaction database\n(DGIdb) for potential drugs that interacted with common genes between endometriosis and EAOC.\nResults\nIdentication of signicant modules via WGCNA\nWe executed WGCNA to identify co-expressed gene modules in endometriosis- and EAOC-related\ndatasets. Based on scale independence of > 0.8, the best soft threshold ( β ) was 4 for GSE5108 to ensure\nbiologically signicant scalefree topology and we found the same soft threshold for GSE18520 (Fig. 2A,\n2B). GSE5108 and GSE18520 co-expression networks were separated into 14 and 6 modules,\nrespectively. Figure 2C and Fig. 2D presented the co-expression modules clustering in endometriosis and\novarian cancer. Each branch represented each gene, and genes clustered into the same module were\nassigned the same color. As depicted in Fig. 2F, correlation heatmaps of module eigengene values with\nclinical traits revealed that the Blue and Brown modules were highly correlated with EAOC, while Black\n\nPage 7/25\nand Grey modules were predominantly associated with endometriosis (Fig. 2E). Therefore, we identied\nand intersected the hub modules to obtain the shared genes between endometriosis and EAOC patients.\nFinally, 362 common genes were selected for further investigation.\nGO and KEGG pathway enrichment analysis\nGO and KEGG pathway enrichment were used based on these 362 common genes between\nendometriosis and ovarian cancer. GO enrichment analysis consisted of biological process (BP), cell\ncomposition (CC) and molecular function (MF) analysis. BP mainly correlated with leukocyte cell-cell\nadhesion and regulation of leukocyte proliferation. CC was mainly associated with apical part of cell and\ncollagen-containing extracellular matrix. The MF was mainly related to peptidase regulator activity and\nendopeptidase regulator activity (Fig. 3A). The KEGG pathway showed that these shared genes were\nmainly enriched in the cytokine-cytokine receptor interaction, intestinal immune network for lgA\nproduction, and IL-17 signaling, indicating that these common genes may be involved in tumorigenesis\nand tumor progression in ovarian cancer (Fig. 3B).\nConrmation of shared genes between endometriosis and\nEAOC\nTo conrm shared genes between endometriosis and EAOC, we performed the differentially expressed\ngene (DEGs) analysis in conrmation dataset GSE157153, which included both endometriosis and EAOC\ntissues, using the cutoff value of adjusted p < 0.05 and |Log2 (fold change)| >1. We identied 1316\nupregulated and 1662 downregulated DEGs from these data (Fig. 3C). Heatmap showed the top 30 up-\nregulated and down-regulated DEGs in endometriosis and EAOC (Fig. 3D). The shared genes in\nendometriosis and EAOC were then intersected with these DEGs. Finally, we obtained a total of 116\ncommon genes, which may play crucial role in both endometriosis and EAOC (Fig. 3E).\nProtein-protein interaction network construction\nA PPI network of potential targets was constructed via the STRING online database, and a total of 116\ncommon genes were visualized by Cytoscape software. Using the CytoHubba plugin and four distinct\nalgorithms (Closeness, Betweenness, MCC, and Degree), we identied the top 30 shared genes (Fig. 3F).\nFinally, 23 node genes were selected from this network for subsequent analysis.\nMachine learning algorithms for hub genes selection\nThree classication algorithms were applied in this study to further investigate shared genes between\nendometriosis and EAOC for potential biomarkers from 23 potential genes. Twelve genes were found to\nhave the lowest binominal deviation using the LASSO regression algorithm (Fig. 4A, 4B). SVM-RFE\nanalysis revealed that the model incorporating four genes owing the lowest error (Fig. 4D). Based on the\nimportance score provided by the RF algorithm, the top ten genes were chosen and ordered. (Fig. 4C, 4E).\nFinally, two hub genes (EDNRA and OCLN) were selected for the nomogram construction and diagnostic\nability evaluation. A Venn diagram was used to display the intersected genes of three machine learning\nalgorithms (Fig. 4F).\n\nPage 8/25\nNomogram construction and assessment of diagnostic\nability of hub genes\nThe diagnostic nomogram was constructed and presented that the expression of two characteristic\ngenes contributed to the clinical diagnosis of EAOC. By integrating the scores corresponding to the\nrelative expression of each gene, the probability of a patient being diagnosed with EAOC was calculated\n(Fig. 5A). Calibration plots demonstrated the diagnostic nomogram's strong prognostic value (Fig. 5B). To\nevaluate the clinical practicality of our characteristic genes, DCA was performed, which integrated the\npreferences of the patients or clinicians into consideration (Fig. 5C). In this analysis, clinical judgment of\nthe benets and harms related to diagnostic model will be made. We compared EDNRA, OCLN and\nnomogram and found that nomogram demonstrated the best overall net benet across a wide and\npractical range of threshold probabilities. This study may help clinicians make more accurate diagnoses\nof patients with EAOC.\nROC curves and AUCs were utilized in GSE157153 to evaluate the diagnostic performance of EDNRA and\nOCLN. The results showed each gene to have exceptional predictive performance as shown below.: OCLN\n(AUC: 0.991, 95%CI: 0.976–1.000, Fig. 5D); EDNRA (AUC: 0.922, 95%CI: 0.850–0.993, Fig. 5E); Nomogram\n(AUC: 0.993, 95%CI: 0.982–1.000, Fig. 5F). The result showed that the nomogram consisting of EDNRA\nand OCLN presenting the best predictive performance and acting a potential diagnostic biomarker for\nEAOC. GSE18520 was utilized to validate the diagnostic value and the results showed that the AUCs of\nOCLN, EDNRA, and Nomogram were 0.789, 0.860, and 0.900 respectively via ROC curve analyses (Fig. 5J-\nL). Furthermore, TCGA-OV was also used to evaluate the predictive performance of EDNRA and OCLN.\nThe ROC analyses yielded the AUCs of 0.995, 0.867, and 0.996 for OCLN, EDNRA, and Nomogram\nrespectively, demonstrating good clinical diagnostic ability (Fig. 5G-I).\nImmune inltration analysis\nTo compare the difference in immune inltration landscape between endometriosis and EAOC patients,\nwe evaluated the inltration levels of the immune cells, immune checkpoint molecular. CIBERSORT\ndeconvolution algorithm demonstrated that the immune inltration levels of plasma cells, T cells\nfollicular helper, Macrophages M1, and Mast cells activated were signicantly elevated while CD8+ T\ncells, NK cells activated, macrophages M2, and mast cells resting were remarkably decreased in EAOC\ngroup (Fig. 6A). Correlation analysis showed that T cells CD4 memory resting had the greatest positive\ncorrelation with CD8+ T cells, while B cells memory displayed the highest negative correlation with T cells\nregulatory (Tregs) (Fig. 6D). Furthermore, spearman correlation analyses were conducted to investigate\nthe expression values of hub genes with immune cell inltration degree and the results showed that\nOCLN was positively correlated with Tfh, macrophages M1, and mast cells activated, while negatively\ncorrelated with CD8+ T cells, mast cells resting, and macrophages M2 (Fig. 6B). EDNRA was positively\ncorrelated with mast cells resting, NK cells activated, and CD8+ T cells, and negatively correlated with\nplasma cells, macrophages M1, T cells follicular helper (Tfh), and mast cells activated (Fig. 6C).\n\nPage 9/25\nGSVA of hallmark gene sets between EAOC and endometriosis and GSEA functional analysis of hub\ngenes\nGSVA of hallmark gene sets between EAOC and endometriosis patients was performed in the GSE157153\ndataset (Fig. 7C). The results showed that the MYC_TARGETS_V1, PI3K_AKT_MTOR_SIGNALING, and\nP53_PATHWAY were activated, while NOTCH_SIGNALING, and TGF_BETA_SIGNALING were suppressed in\nthe progression from endometriosis to EAOC. Furthermore, GSEA analyses were utilized to evaluate\nsignaling pathways involved in the hub genes. We found that EDNRA and OCLN were mainly enriched in\nthe tumor-related pathways, including P53 signaling pathway, calcium signaling pathway, cell cycle, focal\nadhesion, and MAPK signaling pathway (Fig. 7A, 7B). While OCLN were also associated with enrichment\nin immune-related pathways, like systemic lupus erythematosus, cytokine–cytokine receptor interaction,\nchemokine signaling pathway, and autoimmune thyroid disease (Fig. 7B).\nIdentication of TFs and miRNAs interacting with shared\ngenes\nTo investigate the regulatory mechanisms of shared genes OCLN and EDNRA in endometriosis and EAOC,\nwe used NetworkAnalyst tool to explore the TFs and miRNAs. In the TF-hub genes network, we found that\nOCLN may be regulated by HNF4A, JUND, and SRF. EDNRA may be regulated by STAT3, FOXL1, and\nCREB1(Additional le 1A). Additionally, miRNA-hub genes network demonstrated that has-mir-129-2-3p\nmay play core regulatory role. Through network analysis, the TFs and miRNAs connected to two\ncharacteristic genes were shown in Additional le 1B.\nThe expression proles and protein expression of shared\ngenes\nWe compared the expression proles of the hub genes EDNRA and OCLN in patients with endometriosis\nand EAOC, the results revealed that the expression values of EDNRA were profoundly downregulated\nwhile the expression proles of OCLN were signicantly upregulated in ovarian cancer (Fig. 8A, 8B). The\nabove ndings were also evaluated and validated in the GSE18520 (Fig. 8C, 8D) and TCGA-OV cohorts\n(Fig. 8E, 8F). To further investigate the protein expression of EDNRA and OCLN, we searched the\nrepresentative images of immunohistochemistry stainings from HPA and found that OCLN proteins\nexpression levels were signicantly elevated in ovarian carcinoma tissues when compared to normal\novarian tissues, while EDNRA were remarkably overexpressed in normal ovarian samples when compared\nto ovarian carcinoma samples (Fig. 8I), which were consistent with the expression patterns of mRNA. To\nevaluate the correlation of expression proles of hub genes with prognosis of EAOC, Kaplan-Meier curves\nwere plotted to compare the overall survival difference between high- and low-expression groups. The\nresults demonstrated that EAOC patients in TCGA-OV dataset with upregulated expression of OCLN and\nEDNRA showed signicant worse prognosis than those with downregulated expression of EDNRA and\nOCLN (Fig. 8G, 8H).\nPrediction of potential therapeutic drugs\n\nPage 10/25\nDrug-gene interaction analysis was performed to investigate the potential drugs that interact with two\nshared genes between endometriosis and EAOC via the DGIdb database. We constructed drug-\ncorresponding hub genes networks and found 15 potential drugs, including sitaxentan, ambrisentan, and\nbosentan. These potential drugs derived from hub genes may play a crucial role in elucidating the\nmechanism of genes and cell function in relation to receptor sensitivity (Additional le 2).\nDiscussion\nEndometriosis is a widespread disease in reproductive age. It causes chronic inammatory reaction that\nmay result in the formation of scar tissue within the pelvis. The expression of cytokines, which were\nresponsible for the inammatory response and its regulation, was shown to be affected by the\ndysfunction of the immune system. The activation of immune cells triggered signaling pathways, caused\nthe release of inammatory cytokines and accumulated multiple cell types at the site of inammation\n[17–19]. In the endometriotic environment, higher levels of IL-17, which modulates immune cells and\nperpetuates chronic inammation, have been detected [20–22]. Recent research demonstrated that the\nimmune system actively contributed to the pathogenesis of ovarian cancer, and that local and systemic\nimmune responses determined tumor growth and clinical outcomes. Immunotherapy for gynecological\ncancers may assist in reversing immunosuppression and lymphocyte depletion brought on by previous\ntreatments. Interestingly, the KEGG pathway showed that these sheard genes were mainly correlated with\nthe cytokine-cytokine receptor interaction, intestinal immune network for lgA production, and IL-17\nsignaling pathway. The GO enrichment analysis demonstrated the biological functions of these common\ngenes were enriched in regulation of immune cells proliferation. These results indicated that the immune\nsystem played a crucial role in the progression of endometriosis and EAOC.\nWGCNA is a systematic biological method that uses a topological overlap measure to construct a\nnetwork of genes based on their expression proles. WGCNA has been widely used to identify gene-gene\ninteractions, study gene expression patterns, and identify gene regulatory networks. It has also been\napplied to a variety of biological problems, from cancer to cardiovascular disease. Herein, there were 14\nendometriosis-related and 6 EAOC-related gene modules via WCGNA. The blue and brown modules\nexhibited the highest correlation and were deemed EAOC-related modules, whereas the black and grey\nmodules were applied as endometriosis-related key modules. By intersecting the shared genes of\nendometriosis-related and EAOC-related gene clusters, a total of 362 genes were obtained. Limma\nanalysis was utilized to identify DEGs in the conrmation cohort consisting both endometriosis and\nEAOC patients. Then, we selected 23 DEGs after PPI network construction and analysis. After intersection\nof three machine learning methods (SVM-RFE, LASSO, and RF), two hub DEGs (EDNRA and OCLN) were\nchosen for nomogram construction and diagnostic ability evaluation for endometriosis patients with\nEAOC. First, we identied two diagnostic biomarkers (EDNRA and OCLN) by integrating the scores\ncorresponding to the expression of hub genes. Then, we performed external validation using two\nadditional datasets, which revealed that the two hub genes had exceptional predictive performance for\nEAOC in endometriosis patients, indicating their potential diagnostic roles during the tumorigenesis and\nprogression of EAOC.\n\nPage 11/25\nEndothelin (EDN) was initially isolated from cultured porcine aortic endothelial cells [23, 24]. EDN\nreceptors, namely endothelin A receptor (EDNRA) and endothelin B receptor (EDNRB), which belonged to\nthe G protein-coupled receptor family [25], mediated the biological actions of EDN. EDNRA was expressed\nin both vascular and non-vascular tissues and regulated a wide range of physiological processes [26, 27].\nA study on chemotherapeutic resistance of ovarian cancer found that the activation of EDNRA by EDN1\npromoted a direct interaction between β -arr1 and β -catenin to regulate epigenetic modications driving\ncancer chemoresistance onset. It also reported that the overexpression of EDNRA correlated with poor\nprognosis, indicating EDNRA as a potential predictive marker of chemoresistance [28]. The survival rate\nof EAOC patients with a high expression of EDNRA was also found to be lower in the present study.\nTight junctions are protein structures that regulate the transcellular transport of water, ions, and\nmacromolecules. OCLN, the rst tight junctions protein identied, is considered a necessary integral\nprotein for the structure and function of tight junctions [29]. Paku ł a et al. demonstrated that the\nexpression of junctional proteins, including OCLN, E-cadherin, and desmoglein, in peritoneal mesothelial\ncells determined the integrity of the peritoneal mesothelium, which played critical role for the invasion of\novarian cancer [30]. Tobioka et al. reported that in human endometrial carcinoma, OCLN expression\ndecreased proportionally with tumor grade and correlated with myometrial invasion and lymph node\nmetastases [31]. Forced expression of OCLN in human cervical carcinoma cells enhanced the sensitivity\nof the cells to differently triggered apoptotic stimuli and thus inhibited the tumorigenic potencies of the\ntransformed cells and suppressed tumor development [32]. Martin et al. found that knockdown of OCLN\nin human breast cancer cells lead to the profoundly increased invasive potential cells. In addition, the\nabsence of OCLN expression was linked to metastatic disease. And they concluded that OCLN was\nnecessary for the proper function of tight junctions in human breast epithelial cells [33].\nLocal and systemic immune responses determined tumor growth and clinical outcomes and had a\nsignicant impact on the pathogenesis of ovarian cancer. As a result, we compared the proportion of\nimmune cell inltrating in EAOC patients and found that the immune inltration levels of plasma cells,\nTfh, macrophages M1, and mast cells activated were signicantly elevated, while CD8+ T cells, NK cells,\nmacrophages M2, and mast cells resting were remarkably decreased. Previous research demonstrated\nthat CD8+ T cells were essential for tumor suppression and linked to better prognoses in a variety of\ncancers [34]. NK cells are granular lymphocytes that play critical roles in the immune surveillance of\ntumors. In patients with cancer, the degree of NK cell inltration in tumor tissues was prognostic, and a\nreduced NK cell function was associated with worse outcome [35, 36]. The recent advancements in\novarian cancer research showed that macrophage could enhance ovarian tumor progression, but due to\nits high plasticity, it can also inhibit tumor growth by remodeling the M2-like phenotype to M1-like\nphenotype, and the plasticity of macrophage can be exploited in ovarian cancer treatment [37, 38]. Tfh\nare a subset of CD4+ T cells identied for the rst time in the human tonsil. They play an essential role in\nprotective immunity by assisting B cells in the production of antibodies against foreign pathogens.\nRecent studies nd that mast cells have both tumor-promoting and tumor-inhibiting effects, and\ncompartmentalized by the microenvironment responsible for this effect [39]. Chan et al. demonstrated\n\nPage 12/25\nthat high degree of mast cell inltration was correlated with better prognosis than low degree of\ninltration in women with advanced ovarian cancer [40]. Moreover, our study revealed that there was\npositive correlation between EDNRA expression level and activated CD8+ T cells and NK cells, indicating\nthat EDNRA was associated with enhancing the antitumor immune response. Combined with the\ndownregulated expression levels of EDNRA in EAOC patients, our study indicated that decreased EDNRA\nlevels were correlated with a suppressive immunophenotype in patients with EAOC. OCLN expression\nlevel showed positive correlations with immune inltration degree of Tfh, macrophages M1, and mast\ncells activated, Meanwhile, it was negatively correlated macrophages M2. Similar to EDNRA, the above\nresults also suggest that OCLN is associated with an activated immunophenotype. It is partially\nconrmed that the increased OCLN expression level may be related to the enhanced activation of the\nimmune system when considering the expression pattern of OCLN in EAOC patients. However, the\nfunctions of these hub genes in EAOC need further investigate by vitro and vivo experiments.\nThere were some limitations in this study. Firstly, the diagnostic biomarkers were constructed only on the\nbasis of public databases without real-world data validation. Therefore, further verication with\nimprovement in sample size, sequencing data, and clinical information is needed in the future. Secondly,\ndue to the distinct experimental designs for each dataset, the combination of various gene expression\ndatasets may be biased.\nConclusion\nWe identied two shared genes (EDNRA, OCLN) and developed a diagnostic nomogram to predict the risk\nof EAOC patients with endometriosis via WGCNA analysis and machine learning algorithms. Cytokine-\ncytokine receptor interaction was found to be a common signaling pathway between EAOC and\nendometriosis. The correlation of hub genes with immune cells as well as the dysregulated level of\nimmune cell inltration were initially discussed. Collectively, the results of our research lay the\ngroundwork for further exploration of possible candidate genes and aid in the improvement of EAOC\ndiagnosis and treatment. However, more research is required to determine the exact mechanisms by\nwhich these two hub genes affecting the development and progression of EAOC from endometriosis.\nAbbreviations\nAUCs: Area under the ROC curves\nEAOC: Endometriosis-associated ovarian cancer \nGSEA: Gene set enrichment analysis \nGO: Gene Ontology \nGSVA: Gene set variation analysis\n\nPage 13/25\nGEO, Gene Expression Omnibus\nGTEx: Genotype-Tissue Expression\nKEGG: Kyoto Encyclopedia of Genes and Genomes\nLimma: Linear models for microarray data \nLASSO: Least absolute shrinkage and selection operator \nRF: Random forest\nSVM-RFE: Support vector machine-recursive feature elimination \nTCGA: The Cancer Genome Atlas \nWGCNA: Weighted gene co-expression network analysis\nTFs: Transcription factors\nROC: Receiver operating characteristic curve\nDeclarations\nEthics approval\nNot applicable\nConsent for publication\nNot applicable\nAvailability of data and materials\nThe data during the present study is available from the corresponding authors on reasonable request.\nFunding\nThis work was supported by grants from Medical Health Science and Technology Project of Zhejiang\nProvincial Health Commission (2022KY303).\nAuthor contributions\nYM Du, YC Hu and FB Zhang conducted the data analyses; YM Du and TH Zhu drafted the manuscript;\nYM Du, YC Hu, and YT Guan contributed to manuscript review. All the authors approved the nal\nmanuscript.\n\nPage 14/25\nAcknowledgments \nNone.\nConict of Interest \nThe authors have no conict of interest to disclose.\nReferences\n1. Taylor HS, Osteen KG, Bruner-Tran KL, Lockwood CJ, Krikun G, Sokalska A, Duleba AJ. Novel\ntherapies targeting endometriosis. Reprod Sci. 2011;18:814-823.\n2. Burney RO, Giudice LC. Pathogenesis and pathophysiology of endometriosis. Fertil Steril.\n2012;98:511-519.\n3. Pearce CL, Templeman C, Rossing MA, Lee A, Near AM, Webb PMet al.Berchuck A. Association\nbetween endometriosis and risk of histological subtypes of ovarian cancer: a pooled analysis of\ncase-control studies. Lancet Oncol. 2012;13:385-394.\n4. Hermens M, van Altena AM, Nieboer TE, Schoot BC, van Vliet H, Siebers AG, Bekkers RLM. Incidence\nof endometrioid and clear-cell ovarian cancer in histological proven endometriosis: the ENOCA\npopulation-based cohort study. Am J Obstet Gynecol. 2020;223:107 e101-107 e111.\n5. SAMPSON JA. ENDOMETRIAL CARCINOMA OF THE OVARY, ARISING IN ENDOMETRIAL TISSUE IN\nTHAT ORGAN. Archives of Surgery. 1925;10:1-72.\n. Wiegand KC, Shah SP, Al-Agha OM, Zhao Y, Tse K, Zeng Tet al.Huntsman DG. ARID1A mutations in\nendometriosis-associated ovarian carcinomas. N Engl J Med. 2010;363:1532-1543.\n7. Fukunaga M, Nomura K, Ishikawa E, Ushigome S. Ovarian atypical endometriosis: its close\nassociation with malignant epithelial tumours. Histopathology. 1997;30:249-255.\n. Ogawa S, Kaku T, Amada S, Kobayashi H, Hirakawa T, Ariyoshi Ket al.Nakano H. Ovarian\nendometriosis associated with ovarian carcinoma: a clinicopathological and immunohistochemical\nstudy. Gynecol Oncol. 2000;77:298-304.\n9. Langfelder P, Horvath S. WGCNA: an R package for weighted correlation network analysis. BMC\nBioinformatics. 2008;9:559.\n10. Guo X, Xiao H, Guo S, Dong L, Chen J. Identication of breast cancer mechanism based on weighted\ngene coexpression network analysis. Cancer Gene Ther. 2017;24:333-341.\n11. Feng S, Xu Y, Dai Z, Yin H, Zhang K, Shen Y. Integrative Analysis From Multicenter Studies Identies a\nWGCNA-Derived Cancer-Associated Fibroblast Signature for Ovarian Cancer. Front Immunol.\n2022;13:951582.\n12. Yu R, Zhang J, Zhuo Y, Hong X, Ye J, Tang S, Zhang Y. Identication of Diagnostic Signatures and\nImmune Cell Inltration Characteristics in Rheumatoid Arthritis by Integrating Bioinformatic Analysis\nand Machine-Learning Strategies. Front Immunol. 2021;12:724934.\n\nPage 15/25\n13. Zheng PF, Chen LZ, Liu P, Pan HW, Fan WJ, Liu ZY. Identication of immune-related key genes in the\nperipheral blood of ischaemic stroke patients using a weighted gene coexpression network analysis\nand machine learning. J Transl Med. 2022;20:361.\n14. Sanz H, Valim C, Vegas E, Oller JM, Reverter F. SVM-RFE: selection and visualization of the most\nrelevant features through non-linear kernels. BMC Bioinformatics. 2018;19:432.\n15. Tibshirani R. Regression shrinkage and selection via the lasso: a retrospective. Journal of the Royal\nStatistical Society: Series B (Statistical Methodology). 2011;73:273-282.\n1. Hu J, Szymczak S. A review on longitudinal data analysis with random forest. Brief Bioinform. 2023;\n17. Vallvé-Juanico J, Houshdaran S, Giudice LC. The endometrial immune environment of women with\nendometriosis. Hum Reprod Update. 2019;25:564-591.\n1. Zhang T, De Carolis C, Man GCW, Wang CC. The link between immunity, autoimmunity and\nendometriosis: a literature update. Autoimmun Rev. 2018;17:945-955.\n19. Harada T, Iwabe T, Terakawa N. Role of cytokines in endometriosis. Fertil Steril. 2001;76:1-10.\n20. Hirata T, Osuga Y, Takamura M, Saito A, Hasegawa A, Koga Ket al.Taketani Y. Interleukin-17F\nincreases the secretion of interleukin-8 and the expression of cyclooxygenase 2 in endometriosis.\nFertil Steril. 2011;96:113-117.\n21. Ahn SH, Edwards AK, Singh SS, Young SL, Lessey BA, Tayade C. IL-17A Contributes to the\nPathogenesis of Endometriosis by Triggering Proinammatory Cytokines and Angiogenic Growth\nFactors. J Immunol. 2015;195:2591-2600.\n22. Shi FD, Ljunggren HG, La Cava A, Van Kaer L. Organ-specic features of natural killer cells. Nat Rev\nImmunol. 2011;11:658-671.\n23. Inoue A, Yanagisawa M, Kimura S, Kasuya Y, Miyauchi T, Goto K, Masaki T. The human endothelin\nfamily: three structurally and pharmacologically distinct isopeptides predicted by three separate\ngenes. Proc Natl Acad Sci U S A. 1989;86:2863-2867.\n24. Yanagisawa M, Kurihara H, Kimura S, Tomobe Y, Kobayashi M, Mitsui Yet al.Masaki T. A novel potent\nvasoconstrictor peptide produced by vascular endothelial cells. Nature. 1988;332:411-415.\n25. Davenport AP, Hyndman KA, Dhaun N, Southan C, Kohan DE, Pollock JSet al.Maguire JJ. Endothelin.\nPharmacol Rev. 2016;68:357-418.\n2. Bondurand N, Dufour S, Pingault V. News from the endothelin-3/EDNRB signaling pathway: Role\nduring enteric nervous system development and involvement in neural crest-associated disorders.\nDev Biol. 2018;444 Suppl 1:S156-S169.\n27. Kohan DE, Rossi NF, Inscho EW, Pollock DM. Regulation of blood pressure and salt homeostasis by\nendothelin. Physiol Rev. 2011;91:1-77.\n2. Rosanò L, Cianfrocca R, Tocci P, Spinella F, Di Castro V, Caprara Vet al.Bagnato A. Endothelin A\nreceptor/ β -arrestin signaling to the Wnt pathway renders ovarian cancer cells resistant to\nchemotherapy. Cancer Res. 2014;74:7453-7464.\n\nPage 16/25\n29. Furuse M, Hirase T, Itoh M, Nagafuchi A, Yonemura S, Tsukita S, Tsukita S. Occludin: a novel integral\nmembrane protein localizing at tight junctions. J Cell Biol. 1993;123:1777-1788.\n30. Paku ł a M, Witucka A, Uruski P, Radziemski A, Moszy ń ski R, Szpurek Det al.Ksi ąż ek K. Senescence-\nrelated deterioration of intercellular junctions in the peritoneal mesothelium promotes the\ntransmesothelial invasion of ovarian cancer cells. Sci Rep. 2019;9:7587.\n31. Tobioka H, Isomura H, Kokai Y, Tokunaga Y, Yamaguchi J, Sawada N. Occludin expression decreases\nwith the progression of human endometrial carcinoma. Hum Pathol. 2004;35:159-164.\n32. Osanai M, Murata M, Nishikiori N, Chiba H, Kojima T, Sawada N. Epigenetic silencing of occludin\npromotes tumorigenic and metastatic properties of cancer cells via modulations of unique sets of\napoptosis-associated genes. Cancer Res. 2006;66:9125-9133.\n33. Martin TA, Mansel RE, Jiang WG. Loss of occludin leads to the progression of human breast cancer.\nInt J Mol Med. 2010;26:723-734.\n34. St Paul M, Ohashi PS. The Roles of CD8(+) T Cell Subsets in Antitumor Immunity. Trends Cell Biol.\n2020;30:695-704.\n35. Albertsson PA, Basse PH, Hokland M, Goldfarb RH, Nagelkerke JF, Nannmark U, Kuppen PJ. NK cells\nand the tumour microenvironment: implications for NK-cell function and anti-tumour activity. Trends\nImmunol. 2003;24:603-609.\n3. Morvan MG, Lanier LL. NK cells and cancer: you can teach innate cells new tricks. Nat Rev Cancer.\n2016;16:7-19.\n37. Zhang Z, Huang L, Brayboy L. Macrophages: an indispensable piece of ovarian health. Biol Reprod.\n2021;104:527-538.\n3. Puthenveetil A, Dubey S. Metabolic reprograming of tumor-associated macrophages. Ann Transl\nMed. 2020;8:1030.\n39. Ch'ng S, Wallis RA, Yuan L, Davis PF, Tan ST. Mast cells and cutaneous malignancies. Mod Pathol.\n2006;19:149-159.\n40. Chan JK, Magistris A, Loizzi V, Lin F, Rutgers J, Osann Ket al.Samoszuk M. Mast cell density,\nangiogenesis, blood clotting, and prognosis in women with advanced ovarian cancer. Gynecol Oncol.\n2005;99:20-25.\nFigures\n\nPage 17/25\nFigure 1\nStudy owchart.\n\nPage 18/25\nFigure 2\nIdentication of modules linked to clinical features of endometriosis in the GSE5108 dataset and ovarian\ncancer in the GSE18520 dataset by WGCNA. (A, B) Determining the appropriate “soft” threshold ( β ) that\nsatised a scale-free network. (C, D) Co-expression module clustering in endometriosis (C) and ovarian\ncancer (D). Each branch represented each gene, and genes clustered into the same module were assigned\nthe same color. (E, F) Heap of module–trait relationships in endometriosis (E) and ovarian cancer (F). (G,\n\nPage 19/25\nH) Scatter plots for the relationships of module membership in the black module and blue module with\ngene signicance for (G) endometriosis and (H) ovarian cancer.\nFigure 3\nFunctional enrichment analysis and identication of DEGs between endometriosis and EAOC. (A) GO\nenrichment analysis. (B) KEGG pathways enrichment analysis. (C) Volcano diagram shows the DEGs in\nendometriosis and EAOC. (D) Heatmap shows the top 30 up-regulated and down-regulated DEGs in\nendometriosis and EAOC. (E) A total of 116 common genes were identied after intersection of genes\n\nPage 20/25\nfrom three datasets. (F) 23 DEGs were selected for further analysis based on the intersection of genes\nfrom four algorithms.\nFigure 4\nMachine learning algorithms for hub genes selection. (A) In the LASSO regression model, minimum\nstandard was adopted to obtain the value of the super parameter λ  by 10-fold cross-validation. Twelve\ngenes were selected as the potential biomarkers with the lowest binominal deviation. (B) The overview of\n\nPage 21/25\nLASSO coecients. (C) Random forest construction. The number of trees used to build the random forest\nwas specied as 200 before making the prediction. (D) Four genes were identied based on SVM-RFE\nwith the lowest error. (E) The top 10 genes were selected based on the importance score of RF algorithm.\n(F) Two hub genes (OCLN, EDNRA) were nally identied via intersection of three machine learning\nalgorithms.\nFigure 5\n\nPage 22/25\nNomogram construction and evaluation of diagnostic ability of hub genes. (A) Nomogram construction\nfor diagnosing EAOC based on two characteristic genes. (B) Calibration curve. (C) Decision curve analysis\n(DCA). (D-F) ROC curves and AUCs were utilized in GSE157153 to evaluate the diagnostic performances\nof OCLN (D), EDNRA (E), and Nomogram (F). (G-I) ROC curves and AUCs were applied in TCGA-OV to\nevaluate the diagnostic performances of OCLN (G), EDNRA (H), and Nomogram (I). (J-L) ROC curves and\nAUCs were used in GSE18520 to assess the diagnostic performances of OCLN (J), EDNRA (K), and\nNomogram (L).\n\n\nPage 23/25\nFigure 6\nImmune cell inltration between endometriosis and EAOC groups. (A) The inltrating levels of 22 immune\ncell types between endometriosis and EAOC groups via CIBERSORT deconvolution algorithm. (B-C)\nSpearman analyses were conducted to investigate the correlation of OCLN (B) and ENDRA (C) with\nimmune cell inltration degree. (D) Correlation among 22 immune cells types, where red and blue denote\npositive and negative correlations, respectively, and white denotes no correlation between the designated\nimmune cell populations.\n\n\nPage 24/25\nFigure 7\nGSVA of hallmark gene sets between EAOC and endometriosis and GSEA functional analysis of hub\ngenes. (A, B) GSEA identied signaling pathways involved in the characteristic genes of ENDRA (A) and\nOCLN (B). (C) GSVA of hallmark gene sets was conducted to reveal the differences in signaling pathways\nbetween EAOC and endometriosis patients.\n\n\nPage 25/25\nFigure 8\nThe expression proles and validation of the two hub genes by immunohistochemistry analysis.\n(A, B) The expression proles of EDNRA (A) and OCLN (B) in EAOC and endometriosis patients in\nGSE157153 dataset. (C, D) The expression proles of EDNRA (C) and OCLN (D) in EAOC and control\ngroups in GSE18520 dataset. (E, F) The expression proles of EDNRA (E) and OCLN (F) in EAOC and\ncontrol groups in TCGA-OV cohort. (G, H) Kaplan-Meier curves were plotted to compare the overall\nsurvival difference between high- and low-expression groups with regard to EDNRA (G) and OCLN (H). (I)\nImmunohistochemistry of two hub genes in ovarian cancer tissues and control tissues from the Human\nProtein Atlas database.\nSupplementary Files\nThis is a list of supplementary les associated with this preprint. Click to download.\nAdditionalle1.tif\nAdditionalle2.tif","source_license":"CC0","license_restricted":false}