Malignant Peritoneal Mesothelioma Interactome with 417 Novel Protein-Protein Interactions

Malignant peritoneal mesothelioma (MPeM) is an aggressive cancer affecting the peritoneal lining of the abdominal cavity and intra-abdominal organs, with a median survival of ~2.5 years . We constructed an ‘MP eM interactome’ with over 4 00 computationally predicted protein-protein interactions (PPIs) and over 4,700 known PPIs of 59 literature-curated genes whose activity affects MPeM. Known PPIs of the 59 MPeM-associated genes were derived from BioGRID and HPRD databases. Novel PPIs were predicted by applying the HiPPIP algorithm, which computes features of protein pairs such as cellular localization, molecular function, biological process membership, genomic location of the gene, and gene expression in microarray experiments, and classifies the pairwise features as interacting or non-interacting based on a random forest model. 75.6% of the interactome and 65% of the novel interactors in it were supported by transcriptomic evidence in rodent and human peritoneal mesothelioma/mesothelial cell lines and tumor specimens. 152 drugs targeted 427 proteins in the MPeM interactome. Comparative transcriptome analysis of peritoneal mesothelioma-associated versus drug-induced gene expression profiles revealed 39 repurposable drugs, out of which 29 were effective against peritoneal/pleural mesothelioma and/or peritoneal metastasis/primary peritoneal cancer in clinical trials, animal models or cell lines. Functional modules of chromosomal segregation, transcriptional deregulation, positive regulation of IL-6 production and hematopoiesis were identified from the interactome. Genes with tissue-specific expression in 2 sites of extramedullary hematopoiesis (spleen and thymus) and those correlated with unfavorable prognosis in liver, renal, pancreatic and lung cancers were noted. MPeM interactome showed extensive overlap with the malignant pleural mesothelioma (MPM) interactome and MPM cell line expression profiles. Our findings demonstrate the utility of the MPeM interactome in discovering systems-level functional links among MPeM genes and generating clinically translatable results such as repurposed drugs. somatic cancer-related mutations in 4 MPeM patients. 7 Chirac et al. performed comparative genomic hybridization (CGH) on samples collected from 33 French MPeM patients. 8 Foster et al. performed DHPLC (denaturing high-performance liquid chromatography) and sequencing analysis on 29 MPeM samples. 9 Hung et al. examined molecular features of 26 diffuse malignant peritoneal mesothelioma using targeted next-generation sequencing. 10 Pillai et al. investigated MUC1 expression and other prognostic factors in relation to survival in 42 MPeM patients using immunohistochemistry. 11 Varghese at al. performed gene expression analyses on tumor samples collected from 41 MPeM patients who underwent surgical cytoreduction and received regional intraoperative chemotherapy perfusion. 12 Zaffaroni et al. investigated the expression of survivin and the effects of survivin knockdown in 32 MPeM surgical specimens and an established MPM cell line respectively.


Introduction
Internal organs, such as the heart and lungs, and body cavities, such as the thoracic and abdominal cavities, are covered by a thin, slippery layer of cells called the "mesothelium". Mesothelioma is a rare but highly aggressive cancer that originates from this lining, is of types pericardial (heart), pleural (lung) and peritoneal (abdomen) mesothelioma; it is usually found in advanced stages and has a median survival of one year. 1 Mesothelioma is intricately linked with exposure to asbestos but with a long latency period of few decades between exposure and the occurrence of the disease and does not have a non-invasive pre-malignant phase unlike other cancers. The focus of this work is on the genetics and biological mechanisms of MPeM.
The peritoneum is a serosal membrane made up of two layers of mesothelial cells lining the abdominal cavity and intra-abdominal organs. MPeM affects this peritoneal lining and grows within the peritoneal space. 1 Patients may exhibit symptoms such as weight loss, shortness of breath, chest and abdominal pain, increased abdominal girth and peritoneal effusion between the ages of 40-65 years. 1 In contrast to pleural mesothelioma, where 80% of the cases are linked to asbestos exposure, only 8% of peritoneal mesothelioma cases have a history of asbestos exposure. 2 Factors predisposing patients to MPeM are less clear. 2 MPeM appeared to associate with germline mutations more frequently than MPM. 1 MPeM was more apparent among patients with a history of abdominal surgeries rather than asbestos exposure. 1 In this case, chronic inflammation resulting from the procedure is presumed to have induced MPeM. Protracted inflammation of the peritoneum itself, or chronic peritonitis, seem to also pose risk. 2 MPeM has a higher median survival rate than pleural mesothelioma (31 months versus 14 months), 3 and is currently treated with a combination of pemetrexed and cisplatin, which elicits complete/partial responses in 26% (i.e. overall response rate) and disease stabilization (i.e. stable disease rate) in 45% of the patients. 1 Several studies have speculated that the genetic factors driving peritoneal mesothelioma may be different from those driving pleural mesothelioma. One observation supporting this hypothesis is that 5p15 and 7p12 chromosomal gains were more common in peritoneal than in pleural mesothelioma. 4 Multiple studies have examined the genetic underpinnings of MPeM. Hung et al. showed rearrangements in the gene ALK among a subset of MPeM patients having no asbestos fibres, history of therapeutic radiation and other genetic aberrations commonly found in MPeM tumors such as those in BAP1, SETD2 and NF2. 5 Joseph et al. sequenced 510 cancerrelated genes in 13 MPeM patients. 6 Ugurluer et al. analyzed 236 somatic cancer-related mutations in 4 MPeM patients. 7 Chirac et al. performed comparative genomic hybridization (CGH) on samples collected from 33 French MPeM patients. 8 Foster et al. performed DHPLC (denaturing high-performance liquid chromatography) and sequencing analysis on 29 MPeM samples. 9 Hung et al. examined molecular features of 26 diffuse malignant peritoneal mesothelioma using targeted next-generation sequencing. 10 Pillai et al. investigated MUC1 expression and other prognostic factors in relation to survival in 42 MPeM patients using immunohistochemistry. 11 Varghese at al. performed gene expression analyses on tumor samples collected from 41 MPeM patients who underwent surgical cytoreduction and received regional intraoperative chemotherapy perfusion. 12 Zaffaroni et al. investigated the expression of survivin and the effects of survivin knockdown in 32 MPeM surgical specimens and an established MPM cell line respectively. 13 The next step to discovering biological mechanisms is to understand how these genes play a role in the disease.
Protein-protein interactions (PPIs) drive the biological processes in cells including signal transduction, formation of cellular structures and enzymatic complexes. The molecular mechanisms of disease are often revealed by the PPIs of disease-associated genes. For example, the involvement of transcriptional deregulation in pleural mesothelioma pathogenesis was identified through mutations detected in BAP1 and its interactions with proteins such as HCF1, ASXL1, ASXL2, ANKRD1, FOXK1 and FOXK2. 14 PPI of BAP1 with BRCA1 was central to understanding the role of BAP1 in growth-control pathways and cancer; BAP1 was suggested to play a role in BRCA1 stabilization. 15,16 Studies on BAP1 and BRCA1 later led to clinical trials of the drug vinorelbine as a second line therapy for MPM patients, and the drug was shown to have rare or moderate effects in MPM patients. 17,18 Despite being critical to unravelling novel disease mechanisms and drugs, ~75% of estimated PPIs are currently unknown and several disease-associated genes have no known PPIs. More than 600,000 PPIs are said to exist in the human interactome 19 and only ~150,000 PPIs are known from PPI repositories such as HPRD 20 and BioGRID 21 . Detecting the PPIs using experimental techniques such as co-immunoprecipitation (Co-IP) 22,23 is prohibitively laborious and time-consuming at large scale. Tens of thousands of PPIs are being added into the interactome through systematic high throughput studies with yeast two hybrid (Y2H) system 24 and AP-MS 25 . Despite this, a large part of the interactome remains unknown. We have previously developed a computational model called HiPPIP (high-precision protein-protein interaction prediction) that was deemed highly accurate by computational evaluations, and experimental validations of 18 predicted PPIs, where all the tested pairs were shown to be true PPIs. 26,27 We gained several high-impact biological insights from systems-level analysis of the interactome of disease-specific proteins which also included predicted PPIs. We published 2,156 novel PPIs related to MPM, schizophrenia, rheumatoid arthritis and congenital heart disease. 26,[28][29][30] . We constructed the MPM interactome from which we experimentally validated 5 novel PPIs of MPM-associated genes and identified 5 repurposable drugs using comparative transcriptome analysis; biological validity of the interactome was shown by the fact that more than 85% of this interactome was supported by genetic variant, transcriptomic and proteomic evidence related to MPM. 28 Our analysis of the schizophrenia interactome resolved the apparent disconnect between schizophrenia GWAS genes and risk genes discovered in the pre-GWAS era, by showing that they belong to common pathways. 26 We showed interactome overlap between schizophrenia and rheumatoid arthritis that may explain their shared pathology and inverse epidemiological relationship, 29 a link between cilia, neuronal function and neurological disorders, 31 the central role of cilia in congenital heart disease, 30 and the role of mitochondria in hypoplastic left heart syndrome. 32 We were able to shortlist 12 repurposable drugs based on an integrated computational analysis of the schizophrenia drug-protein interactome; 33 two of these drugs are in clinical trials as adjunctive therapy for schizophrenia (ClinicalTrials.gov ; ). Most recently we showed how the genes associated with malignant pleural mesothelioma identified by various high throughput studies were functionally linked within the mechanistic framework of the protein interactome. 34 In this work, we constructed the 'MPeM interactome' by assembling the known and computationally predicted PPIs of the genes associated with malignant peritoneal mesothelioma, and studying pathways and biological processes associated with it. We analysed this interactome in the context of peritoneal mesothelioma transcriptomic datasets, tissue specificity of the constituent genes and their prognostic significance in other related cancers. We extended the MPeM interactome to include the drugs that target any of its proteins and analyzed it to identify a shortlist of 29 drugs that can potentially be repurposed for MPeM.

Results
We compiled a list of 59 genes associated with malignant peritoneal mesothelioma (MPeM) from nine studies; 5-13 these genes harboured mutations, copy number aberrations, rearrangements or showed expression correlated with poor prognosis in MPeM patients or reduced cell survival or less favorable response to drugs in MPeM surgical specimens (Supplementary File 1). PPIs of these genes were collected from HPRD 20 (Human Protein Reference Database) and BioGRID 21 (Biological General Repository for Interaction Datasets). The HiPPIP algorithm described in our earlier work was applied to MPeM genes to discover hitherto unkown PPIs. 35 HiPPIP computes features of protein pairs such as cellular localization, molecular function, biological process membership, genomic location of the gene, and gene expression in microarray experiments, and classifies the pairwise features as interacting or non-interacting based on a random forest model. 26 The 'MPeM interactome' assembled in this manner contained 4,747 known PPIs and 417 novel PPIs connecting 58 MPeM-associated genes to 2,747 known interactors and 306 novel interactors ( Fig. 1 and Supplementary File 2). The 59 th MPeMassociated gene ADAM3A had neither known nor novel PPIs.
The number of known and computationally predicted novel PPIs for each of the MPeM genes are shown in Fig.  2 and Table 1; the novel interactors are listed in Table 1. 13 genes had 10 or less interactions each and 73 novel PPIs were predicted for all of the genes combined. There are 21 hub-genes that had more than 75 known PPIs each and 160 novel PPIs were predicted for all of the genes combined.   Table X), is shown. Legend: dark blue square-shaped nodes: MPeM candidate genes; red nodes/edges: novel interactors/interactions; light blue nodes and blue edges: known interactors/interactions.

Functional modules in the MPeM interactome
We used the HumanBase toolkit 36 (https://hb.flatironinstitute.org/) to identify functional modules in the MPeM interactome. HumanBase employs shared k-nearest-neighbors and the Louvain community-finding algorithm to cluster the genes sharing the same network neighborhoods and similar Gene Ontology (GO) biological processes, into functional modules. 14 modules were detected of which 11 had more than 4 proteins each; 13 modules were enriched for the following biological processes:

Tissue-specificity of the genes in the MPeM interactome
We checked whether any of the genes in the interactome showed tissue-specific expression in the peritoneum using BaseSpace Correlation Engine. 42,43 In this case, a gene was considered to be specific to a particular tissue, if the decrease in its expression in each of the other tissues relative to the tissue of interest is greater than 0.8, among a list of tissues ranked by expression intensity (i.e. specificity index > 0.8). By this definition, only 6 genes were considered to exhibit peritoneum-specific expression based on data from array-based gene expression studies. Nevertheless, OVGP1 which was predicted as a novel interactor of the MPeM-associated gene DPYD exhibited moderate specificity for peritoneum (specificity index = 0.57).
In order to examine whether the interactome showed preferential enrichment for any specific subtype of peritoneal mesothelioma, we computed its overlap with genes found to be differentially expressed in biphasic versus epithelial peritoneal mesothelioma tumor specimens and vice versa. 49 Significant enrichment was found with biphasic mesothelioma (118 genes, P-value = 2.17E-19, odds ratio = 2.25), but not with epithelial mesothelioma (15 genes). This overlap included 4 genes predicted as novel interactors of 4 MPeM-associated genes (MPeM genes are shown in bold): ARID1A-TAF12, PIK3CA-LYPLA1, EPHB1-MRPL3 and KEAP1-LONP1. Out of the 15 genes in the interactome that showed specific expression in the epithelial subtype, 2 were predicted as novel interactors of MPeM-associated genes: VEGFB-SF1 and SUZ12-PLA2G2A.
Diffuse MPeM is known to share similar clinical presentation, morphology and immunostaining profiles with ovarian/primary peritoneal serous carcinoma (OC/PPC), and may hence be indistinguishable from the latter. 50 Gene expression signatures characterizing these two tumors have been identified in an attempt to elucidate the molecular differences distinguishing them from one another. 50 We computed the overlap of the MPeM interactome with these expression profiles. Out of the 12 genes in the interactome found to be differentially expressed in OC/PPC versus diffuse MPeM (including the MPeM-associated gene ESR1), 3 were predicted as novel interactors of MPeM-associated genes: HRAS-IGF2, JUN-TACSTD2 and CHEK2-SUSD2. 8 genes including the MPeM-associated gene KDR were found to be differentially expressed in diffuse MPeM versus OC/PPC.
A recent study had examined the gene expression profiles from the lungs of mice exposed to asbestos fibers (crocidolite and tremolite), an asbestiform fiber (erionite) and a mineral fiber (wollastonite). 51 Crocidolite, tremolite and erionite are capable of inducing lung cancer and mesothelioma in human and animal models. 51 On the other hand, wollastonite is a low pathogenicity fiber that shows no association with the incidence of lung cancer and mesothelioma in humans, or carcinogenesis in animal models. 52 The MPeM interactome showed high statistical significance for genes differentially expressed upon exposure to crocidolite (322 genes, P-value = 3.5E-13, odds ratio = 1.44) and marginal significance for genes differentially expressed upon wollastonite exposure (23 genes, P-value = 0.044, odds ratio = 1.43). No enrichment was observed with erionite (13 genes) and tremolite (64 genes).
In summary, these overlap studies (a) ascertained the biological validity of the MPeM interactome by showing that it contained genes that were biologically relevant to MPeM in the context of rodent xenograft/cell line mesothelioma models and human mesothelial cell lines, (b) highlighted genes that were specific to MPeM subtypes and those that help in distinguishing MPeM from other morphologically and histogenetically similar tumors, and (c) suggested a mechanistic framework to examine genes that appeared to be linked to MPeM etiology, based on transcriptomic evidence .

Disease associations
We systematically analyzed the overlap of the MPeM interactome with prognostic genes of 20 cancer types. Data for correlation of gene expression and fraction of the patient population surviving after treatment for various cancer types were taken from Pathology Atlas. 53 Genes with log-rank P-value < 0.001 were considered to be prognostic. A positive correlation of high gene expression with reduced patient survival is considered unfavourable prognosis, whereas its correlation with increased patient survival is considered favourable. The interactome showed significant enrichment of genes associated with both prognosis for 6 cancer types, and either one type of prognosis for 10 other cancer types. The most significant enrichments were noted for unfavorable prognosis in liver cancer (P-value = 1.35E-25, odds ratio = 1.43), renal cancer (P-value = 3.57E-18, odds ratio = 1.31), pancreatic cancer (P-value = 4.18E-09, odds ratio = 1.51) and lung cancer (P-value = 4.01E-07, odds ratio = 1.62). The most significant enrichments for favourable prognosis were noted in testis cancer (P-value = 7.7E-03, odds ratio = 2.72), breast cancer (P-value = 8.23E-03, odds ratio = 1.29), thyroid cancer (P-value = 0.013, odds ratio = 1.41), and melanoma (P-value = 0.02, odds ratio = 1.81).
We studied the association of the interactome genes with diseases in the DisGeNET database. 54 The top-5 diseases that were associated with MPeM were prostatic neoplasms (odds ratio = 1.98), mammary neoplasms (odds ratio = 2.48), stomach neoplasms (odds ratio = 2.41), liver carcinoma (odds ratio = 2.34) and lung neoplasms (odds ratio = 2.32), all at P-value < 1E-15. Several novel interactors were found to be associated with these diseases.

Potentially repurposable drugs for malignant peritoneal mesothelioma
We followed the established approach of comparing drug-induced versus disease-associated differential expression 57 to identify drugs for MPeM. For this, we used the software suite BaseSpace Correlation Engine (https://www.nextbio.com), 42,43 which allows users to study the effect of diseases and/or drugs on thousands of pre-processed publicly available gene expression datasets and has helped to identify repurposable drug candidates for schizophrenia 33 46 ). Then, we compiled a list of chemical compounds whose differential gene expression profiles (drug versus no drug) were negatively correlated with at least one of the 5 peritoneal mesothelioma differential gene expression datasets (disease versus control). Following this methodology, we identified 22 drugs negatively correlated with BCA induced peritoneal mesothelioma, 19 correlated with O-NT induced peritoneal mesothelioma, 24 correlated with spontaneous malignant mesothelioma, 5 correlated with G-MDSCs from the spleens of AB12 xenografts and 7 correlated with AB12 graft infiltrating neutrophils. Although in each case there were some genes that were differentially expressed in the same direction for both the drug and disorder, the overall effect on the entire transcriptome had an anti-correlation (Supplementary Files 5-9). We identified 39 drugs as potentially repurposable candidates for MPeM. Literature review supported the biological validity of 29 (74%) out of these 39 drugs (Fig. 5), including 2 drugs (paclitaxel: NCT04000906 and imatinib: NCT00402766) that are already in clinical trials for MPeM. The evidence supporting these drugs is:  Two drugs are part of the standard therapy for mesothelioma 58 (pemetrexed and vinorelbine).  One drug has shown activity against peritoneal mesothelioma, pleural mesothelioma and peritoneal metastasis (irinotecan). Although ineffective as a single agent, 59 irinotecan elicited modest response rates and showed an acceptable toxicity profile in malignant pleural mesothelioma clinical trials, and produced an inhibitory effect on mesothelioma cell lines in combination with a p53 activator. [60][61][62] Irinotecan was shown to be effective and tolerable in gastric cancer patients with peritoneal seeding. 63 Importantly, irinotecan in combination with cisplatin showed efficacy and tolerability against peritoneal mesothelioma in a clinical setting. 64  Two drugs have shown activity against peritoneal mesothelioma in clinical settings and peritoneal metastasis in clinical trials/cell lines (paclitaxel and sirolimus). Paclitaxel appeared to be ineffective against MPM both as a single agent as well as in combination with other agents such as filgrastim and cisplatin. [65][66][67] However, it has led to the complete remission of an MPeM patient for 20 months, 68 and clinical efficacy and an acceptable toxicity profile in ovarian, pancreatic and gastric cancers with peritoneal metastasis or serosal exposure. 69,70 Sirolimus inhibited proliferation and increased cell death in MPM cell lines in combination with cisplatin, inhibited epithelial-to-mesenchymal transition in peritoneal mesothelial cell lines, and showed clinical efficacy in a patient with benign multicystic peritoneal mesothelioma. 71-73  Twelve drugs have shown activity against malignant pleural mesothelioma in clinical trials, animal models or cell lines (epirubicin, panobinostat, doxorubicin, imatinib, vinblastine, idarubicin, azacitidine, vorinostat, dactinomycin, acetylcysteine, staurosporine and quercetin). Epirubicin has shown a modest 10-20% response rate in malignant mesothelioma patients. 74 Panobinostat has been shown to have an inhibitory effect in mesothelioma cell lines and tumors in murine xenograft models. 75 Doxorubicin has shown tolerable toxicity and improvement in the quality of life of MPM patients. 76 Imatinib has shown both cytotoxicity and apoptosis in PDGFRB-positive mesothelioma cell lines, and has shown modest clinical efficacy in malignant mesothelioma patients. 77,78 Vinblastine has shown improved progression-free survival rate and acceptable toxicity in combination with methotrexate and platinum. [79][80][81] Both idarubicin and dactinomycin showed more cytotoxicity in MPM cell lines compared with pemetrexed and cisplatin. 82 Azacitidine has shown inhibitory activity against malignant mesothelioma in clinical trials. 83 Vorinostat induced apoptosis in mesothelioma cell lines. 84 However, it did not provide any therapeutic benefit in patients with pleural mesothelioma. 85 Acetylcysteine and quercetin have shown dose-dependent inhibition and time-and dosedependent inhibition in malignant mesothelioma cell lines respectively. [86][87][88] Staurosporine was effective against mesothelioma tumors in murine xenograft models. 89  6 drugs have shown activity against primary peritoneal cancer or peritoneal metastasis in other cancers (ruxolitinib, daunorubicin, dasatinib, topotecan, dexamethasone and nintedanib). The gene expression profile induced by ruxolitinib was shown to be negatively correlated with all the 5 peritoneal mesothelioma datasets. Ruxolitinib in combination with paclitaxel has been shown to inhibit tumor growth in a mouse model of advanced ovarian cancer with peritoneal metastasis. 90 Daunorubicin induced side effects and showed no clinical activity against MPM. 91 However, treatment with daunorubicin led to complete remission of a gastric Kaposi's sarcoma patient with peritoneal metastasis. 92 Dasatinib showed inhibitory activity against peritoneal 13 metastasis in a murine xenograft model of gastric cancer. 93 Topotecan demonstrated clinical efficacy and tolerability in primary peritoneal carcinoma. 94 Dexamethasone protected peritoneal mesothelial cells from epithelial-to-mesenchymal transition by acting on MAPK, GSK-3β and SNAI1. 95 Nintedanib has been shown to inhibit peritoneal fibrosis in a mouse model by blocking mesothelial-to-mesenchymal transition. 96  4 drugs have shown activity both against malignant pleural mesothelioma and peritoneal metastasis or sclerosis (methotrexate, resveratrol, everolimus and genistein). Methotrexate was shown to be clinically active against MPM in combination with gemcitabine, and against peritoneal metastasis in advanced gastric cancer. 79,[97][98][99] Everolimus induced AMPK/p38-mediated apoptosis in MPM cell lines and showed clinical activity in encapsulating peritoneal sclerosis. [100][101][102] Resveratrol has produced an inhibitory effect on mesothelioma cell lines as a single agent as well as in combination with clofarabine, and has inhibited the adhesion of ovarian cancer cells to peritoneal mesothelial cells in vitro. 26,[103][104][105] Genisten has shown inhibitory activity against matrix metalloproteinases in malignant mesothelioma cell lines, and against peritoneal metastasis in intestinal adenocarcinomas in Wistar rats. 106,107  Two drugs have been shown to be effective in pleural/peritoneal effusions (mitoxantrone and vincristine) Mitoxantrone has shown only modest clinical activity in malignant mesothelioma. 108 However, this drug has shown activity against pleural effusions in cancer patients. 109 Vincristine was active against malignant peritoneal effusions in a murine xenograft model of Ehrlich ascites carcinoma. 110

Discussion
MPeM constitutes a substantial percentage (15%-20%) of all mesothelioma diagnoses and is additionally notable for its distinction from pleural mesothelioma, such as its limited association with asbestos exposure and proportionally higher prevalence in comparison with MPM among patients with germline mutations and without a history of asbestos exposure (25% versus 7% 111 ). 1 In addition to this, three other factors that underscore the importance of MPeM-centric studies include the highly variable pattern of disease progression that it exhibits, risk for postoperative morbidity and mortality, and the increasing prevalence of peritoneal cases among mesothelioma patients without occupational exposure (given the current scenario in which the population of asbestos-exposed individuals is diminishing). 1 42% of MPeM patients survived for at least 5 years after cytoreductive surgery (CRS); however, instances of accelerated disease progression and death and survival for a long time with active disease have also been reported, indicating the variable nature of MPeM outcomes. 1 A postoperative morbidity (complications and adverse events arising from the surgical procedure) of 35% and a mortality rate varying from 2% to 6% have been reported in patients who received CRS or hyperthermic intraoperative peritoneal chemotherapy (HIPEC), indicating the variable nature of surgical outcomes. 1 Given the unique features of MPeM and its fatal nature, it is imperative that the molecular mechanisms underlying this disease are expeditiously discovered.
Although multiple studies have examined the genetic basis of MPeM, 5-13 a mechanistic framework to interrogate its underlying cellular mechanisms and the functional consequence of the genes identified in these studies in an integrative manner is still lacking. In this study, we employed the protein interactome to discover the biological coherence of MPeM-associated genes identified in disparate studies, noting that MPeM may develop when PPIs are perturbed by genetic mutations or aberrant expression of MPeM-associated genes/proteins, ultimately leading to disrupted cellular functions. 112 Extensive inter-connectivity and intra-connectivity of the network components in the human PPI network may augment the spreading of the effects of such perturbations to other proteins, encoded by genes that may/may not harbor any disease-associated alterations, through the network of their interactions, posing deeper implications for disease development, e.g. comorbidity, shared genetics and symptomatology among diseases and phenotypic responses to drugs. 112 We constructed the MPeM interactome with more than 4,700 known interactions and more than 400 novel interactions of MPeM-associated genes collected from 9 studies, and conducted functional enrichment and transcriptome-based analyses to establish the biological validity of the interactome, gain valuable insights into MPeM etiology and identify drugs that may be repurposed to treat MPeM.
A key finding from our study was that more than 75% of the interactome as a whole, and more than 60% of the novel interactors that were predicted to interact with MPeM-associated genes by the HiPPIP algorithm, had MPeM-related transcriptomic aberrations in rodent and human cell line models. This demonstrates the validity of our interactome-based method due to two reasons: we (a) identified other genes that were previously identified in MPeM-related gene expression studies using an unbiased approach from the MPeM interactome, and (b) showed that they were closely connected to a curated set of MPeM-associated genes and that they could be functionally relevant to MPeM etiology in an interdependent manner. Another key finding was that more than 70% of the repurposable drugs that we identified using comparative transcriptome analysis has been shown to be effective against peritoneal mesothelioma, pleural mesothelioma, peritoneal metastasis and/or primary peritoneal cancer in clinical trials, animal models or cell lines. The drug-associated expression profiles analyzed in this study were induced in a wide variety cell lines. The effect of the proposed drugs should be specifically examined in human peritoneal mesothelioma cell lines or animal models.
Human orthologs of spleen-specific and thymus-specific mouse genes showed significant enrichment in the interactome, instead of major abdominal organs. This finding seemed to be in line with the identification of hematopoiesis as a functional module enriched in the MPeM interactome. Both spleen and thymus are hematopoietic sites regulating the proliferation, maturation and activation of lymphoid cells outside of the bone marrow (extramedullary hematopoiesis). Expansion of myeloid cells in the spleen through the process of extramedullary hematopoiesis resulting in splenomegaly has been observed in BAP1 knockout mice. 113 Further investigations may be necessary to elucidate the functional consequences of this phenomenon in MPeM. 56% (33) of the MPeM-associated genes that we had used for interactome construction were involved in chromosomal events such as copy number gain/loss, gene loss, deletion and gene rearrangement. In line with this, chromosome segregation was identified as the most enriched functional module in the interactome. We identified 10 novel interactors involved in chromosome segregation, out of which 7 were predicted to interact with MPeM-associated genes involved in chromosomal events ( As noted earlier, a higher prevalence of MPeM compared to MPM has been observed among carriers of germline mutations. 1 These mutations (e.g. in BAP1 and TP53) may predispose MPeM patients to multiple other cancers. 1 We identified a significant enrichment of genes, whose elevated expression positively correlates with (a) unfavourable prognosis in liver, renal, pancreatic and lung cancers and (b) favourable prognosis in testis, breast, thyroid and skin cancers, in the MPeM interactome. Prolonged survival has been reported in mesothelioma patients with germline mutations, indicating that these mutations may sometimes mediate inhibition of aggressive tumor growth. 1 Additionally, prolonged survival has been associated with the occurrence of other cancers in carriers of germline mutations. 1 Genetic screening to identify aberrations in prognostic genes may guide clinicians in patient stratification (based on their survival time and risk of developing other cancers) and in devising better treatment strategies.
We showed extensive overlap of the MPeM interactome with the MPM interactome, interconnections between MPeM-associated genes and MPM-associated genes in the interactome and significant enrichment of the MPeM interactome in MPM cell lines. Further investigations focused on these findings may provide clues on the etiological diversification and common genetic underpinnings of MPeM and MPM.
In summary, our study provides a network-level view of MPeM-associated genes and their functional consequences. The MPeM interactome will serve as a functional landscape to integrate and analyse MPeM related multi-omics data and will inform and catalyze future investigations on the genetic basis of MPeM and biomedical studies seeking to improve clinical interventions in MPeM.

Compilation of MPeM-associated genes and prediction of novel interactions
A list of 59 MPeM-associated genes that harboured mutations, copy number aberrations, rearrangements or showed expression correlated with poor prognosis in MPeM patients or reduced cell survival or less favorable response to drugs in MPeM surgical specimens was compiled from eight studies. [5][6][7][8][9][10][11][12][13] Novel PPIs of the proteins encoded by these genes were predicted using the HiPPIP model that we developed. 35 Each MPeM protein (say N1) was paired with each of the other human proteins say, (M1, M2,…Mn), and each pair was evaluated with the HiPPIP model. 35 The predicted interactions of each of the MPeM proteins were extracted (namely, the pairs whose score is >0.5, a threshold which through computational evaluations and experimental validations was revealed to indicate interacting partners with high confidence). The interactome figures were created using Cytoscape. 114

Identification of functional modules
Functional gene modules were extracted using the HumanBase toolkit 36 (https://hb.flatironinstitute.org/). HumanBase uses shared k-nearest-neighbors and the Louvain community-finding algorithm to cluster the genes sharing the same network neighborhoods and similar GO biological processes into functional modules. The pvalues of the terms enriched in the modules are calculated using Fisher's exact test and Benjamini-Hochberg method.

Functional enrichment analysis
Biological process (Gene Ontology 115 ), pathway (Reactome 116 ) and disease (DisGeNET 54 ) enrichments were computed using WebGestalt. 37 WebGestalt computes the distribution of genes belonging to a particular functional category in the input list and compares it with the background distribution of genes belonging to this functional category among all the genes that belongs to any functional category in the database selected by the user. Statistical significance of functional category enrichment is computed using Fisher's exact test, and corrected using the Benjamini-Hochberg method for multiple test adjustment. Annotations with FDR-corrected p-value < 0.05 were considered significant.

Tissue-specific expression analysis
Tissue-specificity of the genes in the MPeM interactome were checked using TissueEnrich. 117 The analysis was based on tissue-specific genes compiled from GTEx and Mouse ENCODE. 39,40 This included 'tissue-enriched genes' with at least 5-folds higher mRNA levels in a particular tissue compared to all the other tissues, 'groupenriched genes' with at least 5-folds higher mRNA levels in a group of 2-7 tissues and 'tissue-enhanced genes' with at least 5-folds higher mRNA levels in a particular tissue compared to average levels in all tissues.

Network overlap analysis
Statistical significance of the overlaps between genes in the MPeM and MPeM interactomes was computed based on hypergeometric distribution.

Identification of prognostic cancer genes
Data for correlation of gene expression and fraction of patient population surviving after treatment of 20 cancer types was taken from Pathology Atlas. 53 Genes with log-rank P-value < 0.001 were considered to be prognostic. Unfavorable prognosis indicates positive correlation of high gene expression with reduced patient survival.

Identification of repurposable drugs
The list of chemical compounds whose gene expression profiles correlated negatively with 5 gene expression datasets associated with peritoneal mesothelioma were compiled using the BaseSpace correlation software (https://www.nextbio.com) (List 1). The datasets considered were granulocytic myeloid-derived suppressor cells (G-MDSCs) from spleens of mice bearing AB12 mesothelioma grafts versus naive neutrophils, neutrophils infiltrating AB12 mesothelioma tumor grafts versus naive bone marrow derived neutrophils (GSE43254 44 ), BCA induced peritoneal mesothelioma versus non-transformed mesothelial cell line, O-NT induced peritoneal mesothelioma versus non-transformed mesothelial cell line (GSE4682 45 ) and spontaneous malignant mesotheliomas from 2-year-old rats versus normal mesothelial Fred-PE cells (GSE47581 46 ). Next, we identified drugs that targeted at least one gene in in the MPeM interactome using Drug Bank. 118 We then compared list 1 and list 2 to identify the drugs that not only target proteins in the interactome but are also negatively correlated with MPeM-associated gene expression profiles.