Ovarian cancer (OC) is one of the most common gynaecological malignancies. According to the Global Cancer Observatory of the World Health organization (WHO) international agency for research on cancer, a total of 207,252 new fatalities due to ovarian cancer were reported in 2020, placing it 14th out of 36 different types of tissue cancers (World Health Organization International Agency for Research on Cancer, 2020). Most ovarian malignancies originate from epithelial cells, and the most prevalent histological subtype of epithelial ovarian cancer is ovarian serous adenocarcinoma (Heintz et al., 2006). Early OC detection is the best treatment scenario, but as OC presents with nonspecific symptoms and reflects detection, most patients are given a stage III diagnosis, indicating that the disease has spread throughout the peritoneum and/or has involved the lymph nodes (Prat and FIGO Committee on Gynecologic Oncology, 2014). A multi-stage evaluation is necessary to manage OC, to determine personalized treatment, and to predict the presence of distant metastases, tumour stage and prognosis.

As a new treatment option, immunosuppressants, address the tumour microenvironment (TME) (Pitt et al., 2016). For ovarian cancer, this cutting-edge therapeutical approach is recently being studied and applied (Yang et al., 2022, 2023). Despite the fact that many variables have been demonstrated to predict the therapeutic effectiveness of immunosuppressant’s, the accuracy of this strategy still needs to be improved (Gibney et al., 2016). Tumour development, spread and invasion are all dependent on the TME (Schreiber et al., 2011; Lei et al., 2020), which contains multiple cell types, including stroma, vasculature, secretory factors, surrounding stroma and the internal environment of the tumour cells. As the TME is primarily determined by the genomic landscape of the tumour, several algorithms have been developed to predict tumour purity and estimate the abundance of tumour-infiltrating immune cells based on gene expression profiles (Tamborero et al., 2018). These include CIBERSORT, MCP, Xcell, EPIC, ESTIMATE, Timer, IPS, and QuantiSeq.

As an essential component of TME, the non-cellular network surrounding the cells, known as the extracellular matrix (ECM), is tightly associated to the pathophysiology of healthy and cancerous tissue (Frantz et al., 2010; Henke et al., 2020; Zhu et al., 2022). This renders ECM a crucial study niche for the initiation, progression, dissemination, and furthermore treatment and prognosis of epithelial ovarian cancer (Ween et al., 2011). The metabolic disruption of various ECM protein-related factors derived from epithelial cells during tumorigenesis leads to the formation of a pro-tumorigenic microenvironment that favors tumor growth and metastasis. This is followed by tumour cell-mediated ECM remodelling, which ultimately promotes the survival of tumour cells at the expense of healthy tissue (Zigrino et al., 2005). Therefore, ECM proteins, which have bidirectional effects on the generation, recurrence and metastasis of tumour cells (Valmiki et al., 2021 should be considered key players to the treatment and prognosis of tumours (Donelan et al., 2022; Zhu et al., 2022).

An artificial intelligence algorithm, Random Forest (RF) has been recently employed to predict disease progression by virtue of its high performance and interpretability (Wu et al., 2021). A convincing predictive model can be constructed by combining analysis of gene expression data with diagnostic and therapeutic data. This model could be effective at forecasting patient survival, the course of the tumour, and recurrence following various types of treatment (Lin et al., 2022; Miao et al., 2022). Big data machine learning may also also be applied. Despite the recent advances in machine learning methods for ovarian cancer survival analysis, integration of multi-omics data with immunotherapy targeting is an approach that has not been thoroughly explored (Henderson et al., 2016; Belotti et al., 2022). This approach could be advantageous for the identification of potential therapeutic targets and may lead to improved outcomes for ovarian cancer patients.

In this study we used artificial intelligence algorithms to integrate multifaceted omics data with immunotherapy targets in ovarian cancer. Specifically, we employed the Random Forest and Lasso algorithms to process gene expression and survival data from the TCGA database. The tumour risk score was calculated to construct features for predicting OC prognosis and immunotherapy efficacy.

Ovarian cancer is a malignant tumour with a high mortality rate. Because of its insidious onset, it is usually late-stage when obvious clinical symptoms appear (Prat J and FIGO Committee on Gynecologic Oncology, 2014). Patient prognosis is thus based on an accurate and reliable assessment. As the extracellular matrix is closely linked to epithelial cells, we aimed to construct a prognostic prediction model based on extracellular matrix proteins to evaluate the prognostic survival of patients with ovarian serous adenocarcinomas. We successfully filtered 15 key genes out of 1068 extracellular matrix-associated proteins, with excellent predictive ability on ovarian serous adenocarcinoma prognosis using the Random Forest and the Lasso algorithms.

We used the random forest algorithm as is a popular machine learning technique with documented outstanding performance in a wide range of predictive modeling tasks, including cancer prognosis prediction (Toth et al., 2019; Li Y et al., 2020). When analyzing tumor sequencing data, the relationship between variables and outcomes can be complex due to the volume of data. Random forest can handle a large number of input variables without overfitting, which has been a challenge for traditional models such as logistic regression or decision trees (Maroco et al., 2011; Lan et al., 2020). Additionally, traditional linear models are unable to handle the nonlinear relationship between a large amount of sequencing data and survival outcomes, but random forest can address this issue (Lee and Lim, 2019). Compared to support vector machines (SVM), random forest is less sensitive to outliers and missing data, which is important in cancer prognosis prediction, where data quality may vary and missing data is common (Pelckmans et al., 2005; Lee and Lim, 2019).

Among the 15 key genes generated by our our analysis, CSPG5, CXCL11 and ESM1 mRNAs were abundantly expressed in OC tissue compared to normal tissue. Cancer cells, fibroblasts, endothelial cells, and immune cells such as leukocytes, monocytes, and dendritic cells are primarily responsible for CXCL11 production (Gao and Zhang, 2021) CXCL11 is an effector chemokine regulating T cell recruitment that promotes effector immune cells (e.g., CD8T cells, Th1 cells, TH17 cells and antigen-presenting cells). Studies have shown that induction of CXCL10 and CXCL11 expression in breast cancer cells enhances the infiltration of CD8 T cells (Liu et al., 2011; Gao et al., 2019). In addition to its influence on tumour progression through its angiostatic effects (Strieter et al., 2006), CXCL11 is part of the CXCR7/CXCL11 axis that was shown to induce the epithelial–mesenchymal transition and metastatic behaviour of OC cells under ERα control (Benhadjeba et al., 2018). ESM1 is a soluble proteoglycan expressed by vascular endothelial cells and is associated with inflammatory cell recruitment (Hung et al., 2020). Vascular endothelial dysfunction can be brought on by high levels of ESM1 (Kalantaridou et al., 2006; Rocha et al., 2014; Sun et al., 2019), whereas it has also been shown to be closely correlated with OC development and progression (Li et al., 2023). Our analysis showed that the tumour tissue had significantly higher ESM1 expression levels than normal ovarian tissue, which may be associated with abnormally elevated cell proliferation and tumour tissue revascularization. Furthermore, CSPG5, also known as Neuroglycan C (NGC), is a protein originally associated with extracellular matrix production in the nervous system (Pintér et al., 2020) and shown to decrease first and then increase following ischemic and hypoxic injury, presumably associated with ECM damage repair (Matsui et al., 2005). Of note, two recent studies showed that expression of CSPG5 was significantly correlated with the prognosis of patients with epithelial OC (Su et al., 2021; Wang et al., 2023). In our study, higher CSPG5, CXCL11 and ESM1 expression and inflammatory cell infiltration, especially of CD8 T cells, were present in the low-risk group. The high level of immune cell environment may explain the better prognosis for overall survival in the low-risk group.

Interestingly, two proteins commonly associated with bone biology and development were identified among the ECM prognostic markers. AMBN was discovered as a tooth enamel matrix protein, playing an important role in enamel, cranial and long bone development. It was however demonstrated that AMBN was among four factors that were potential independent prognostic factors for prostate cancer (Xu et al., 2018). In our multifactorial Cox analysis, AMBN was shown to be one of the independent prognostic factors for ovarian serous adenocarcinomas. Furthermore, the analysis revealed a prognostic role for FGF23, the bone-derived hormone secreted by osteoblasts and osteocytes (Guo and Yuan, 2015). Previous studies have shown its expression alterations in breast cancer (Cekin et al., 2020) and identified that serum or plasma FGF23 concentrations are elevated in patients with advanced stage epithellal ovarian cancer (Tebben et al., 2005). According to our data, FGF23 was expressed in the tumours, and its expression levels were higher in the high-risk group.

MMP17, PI3, TLL1, ANGPTL4, and TGFBI have all been previously associated with cancer. Our analysis found that transcripts of all these five genes were expressed at higher levels in the high-risk group than in the low-risk group. MMP17 has been associated with the maintenance of normal physiological function in vascular smooth muscle (Martín-Alonso et al., 2015) and a promotive effect on tumour cells (Paye et al., 2014). Additionally, it was shown that its expression was much higher in EOC patients than in pericarcinomatous tissues (Xiao et al., 2022). PI3, also known as elafin, is an elastase-specific inhibitor that directly affects tumour suppression by inhibiting elastase (Hunt et al., 2013). High levels of PI3 are associated with severe disease severity in various cancers (Hunt et al., 2013; Longatto-Filho et al., 2021), while another TCGA-OV analysis showed its prognostic value in OC (Li J et al., 2020). Additionally, high elafin expression has been associated with unfavorable OS but better immunotherapy responses (Lu et al., 2023). Recent studies have found that TGFBI CpG islands are hypermethylated in adjacent normal colon tissue, with the corresponding sequences showing hypomethylation in colon cancer tissue, and that higher TGFBI levels are associated with poorer prognosis (Zhang H. et al., 2019). In mammals, Tolloid-like (mTLL)-1 is a BMP-1-associated protease, and BMP1/TLL1 is involved in the process of tissue remodelling in the ovary, assisting in the maturation of pre-collagen molecules and the deposition of collagen fibres (Ohnishi et al., 2005). In hepatocellular carcinoma (Matsuura et al., 2017), SNP may impact the splicing of TLL1 mRNA and result in short variants with high catalytic activity, speeding up the development of liver fibrosis and cancer. In a recent study, TGFBI, PI3, TLL1 and MMP17 were predicted to be among the 19 genes that comprise the TME-related high grade serous ovarian carcinoma prognostic genetic panel (Belotti et al., 2022). ANGPTL4 is regulated by peroxisome proliferator-activated receptor γ (PPARγ) (Aryal et al., 2019), who has been observed to be significantly increased in malignant ovarian tumours (grade 1, 2 and 3) compared to benign and borderline tumours (Zhang et al., 2005). It was also recently identified in a scRNA-seq study of ovarian cancer CAF ligands to epithelial cells (Carvalho et al., 2022). Finally, ANGPTL4 and TGFBI were identified both in a hypoxia risk model constructed to reflect the OC immune microenvironment in and predict prognosis (Wei et al., 2021), and among the genes that comprise an OC glycolysis-related gene signature (Zhang et al., 2021).

To further assess the relationship between the ECM risk score and the immune microenvironment, we assessed the abundance of multiple inflitrating cells in the immune microenvironment of these patients using multiple algorithms. Patients with more infiltrating and activated immune cells in TME may have better immunotherapeutic outcomes (Li et al., 2016). We observed higher levels of CD8 T cell infiltration in the low-risk group, suggesting enhanced immune surveillance via CD8⁺ T cells, while this implies a possible enhanced susceptibility to PD-1/PD-L1-targeted immune checkpoint therapies (Strickland et al., 2016; Iyer et al., 2021). Furthermore, the degree of macrophage infiltration was similar to both CD8 and CD4 T cell infiltration, with both showing high levels in the low-risk group. This is consistent with previous studies describing a positive correlation between T cells and macrophage infiltration levels (Desbois et al., 2020).

Previous studies have also shown that cancer cells regulate their local microenvironment to promote tumour survival, chemoresistance and evasion of immune surveillance (Kim et al., 2007) and that there is a tight association between malignant tumour cells and CAFs in promoting tumour growth and survival (Xing et al., 2010; Karagiannis et al., 2012; Chen et al., 2021). scRNA-seq analysis of high grade serous ovarian cancers, also showed that CAFs induce epithelial-mesenchymal transition (EMT) of tumor cells via TGFβ signaling, with consequent effects on chemoresistance and metastasis (Kan et al., 2020). In accordance with these, our results also showed that higher levels of CAFs were found in high-risk groups with poor prognosis.

It is known that the immune microenvironment immune cells, immunomodulating factors and immune checkpoint molecules are crucial for the immune escape of tumour cells (Charoentong et al., 2017; Zhang Y. et al., 2019). We thus developed an immunophenoscore (IPS) based on immune subpopulation infiltration and expression of immune regulatory molecules using the random forest to identify determinants of immunogenicity. Among several IPS subtypes tested, we found that the low-risk group had higher IPS and could benefit during treatment with immune checkpoint inhibitors. As the immunophenoscore is a surrogate to patients’ immunotherapeutic outcomes, our IPS results of the ECMs risk scores may only be considered of predictive value, and future studies will confirm their clinical importance.

We also explored the tumour mutation burden (TMB) changes in the TCGA-OV cohort. The ovarian cancer genome exhibits high levels of instability, as evidenced by functional cells (Stewart et al., 2011), copy number changes (Schwarz et al., 2015), and status of somatic mutations (Bashashati et al., 2013). TMB is the total number of substitutions and insertion/deletion mutations per megabase in the exon-coding region of the gene under evaluation in the tumour cell genome (Stenzinger et al., 2019). Somatic mutations may result in tumourigenesis and many somatic mutations can generate neoantigens facilitating anti-tumour immunity (Gubin et al., 2015). In a study on immunotherapy for lung cancer, researchers discovered that patients with PD-L1 1% but a subgroup of 10 mutations/Mb in the combination chemotherapy group had a better objective response rate (ORR) and median progression-free survival (median PFS, mPFS) with the immune combination regimen CheckMate 227 (Hellmann et al., 2018). This suggests that in the higher TMB population, PFS was better in the combination immunotherapy group than in the chemotherapy alone group, irrespective of PD-L1 expression. In our study, the mean TMB values were higher in the low-risk group than in the high-risk group, implying that the low-risk group may have more potential for immunotherapy. TP53, the gene encoding the tumour suppressor protein p53, is one of the most commonly mutated genes in human cancers, and driver mutations are prevalent in high-grade ovarian plasmacytoma (Ahmed et al., 2010). Chalmers and coworkers have shown that TP53 mutations were associated with high TMB (Chalmers et al., 2017). Our study similarly confirmed that in the TCGA-OV datasets TP53 mutations were the most frequent in the high- and low-risk groups, and that TP53 mutations were higher in the low-risk group than in the high-risk group. However, there are limitations to cohort-based studies. Most mutated genes are unique to each case, and in clinical treatment, patients should be treated based on their mutation sequencing results. Our analysis may provide theoretical support for the selection of immunological agents.

Our study aimed to investigate the role of the matrisome and the gene changes in the ECM in a widely studied, publicly available ovarian cancer transcriptomic and clinicopathological collection of patients. The role of matrisome has been regretfully understudied in this type or cancer that commonly goes undetected till it reaches high grades, as the ECM can influence drug resistance. We used bioinformatics and machine learning methods to investigate the TCGA-OV collection and identified several prognostic genes, some of which have also been identified by previous studies. Given the current cost-effectiveness of biotechnological approaches, rapid genetic testing tools are commonly promoted and widely applied in clinical diagnostics and treatment (Young and Argáez, 2019). These tools have improved accuracy and testing times have significantly shortened. Targeted multigene tests and genetic screening can be thus rapidly employed to assist diagnosis postoperatively and develop more effective treatment plans.

In conclusion, this study developed an ECM risk score prediction model to enable prognosis of patients with ovarian serous adenocarcinoma. We further identified the tumour microenvironment and somatic mutations using the TCGA-OV collection datasets. These results should be further validated with targeted future studies to evaluate their real predictive function and their use in personalized immunotherapy applications.