Ferroptosis-Related Gene Model to Predict Overall Survival of Ovarian Carcinoma

Background Ovarian cancer (OC) is the eighth most common cause of cancer death and the second cause of gynecologic cancer death in women around the world. Ferroptosis, an iron-dependent regulated cell death, plays a vital role in the development of many cancers. Applying expression of ferroptosis-related gene to forecast the cancer progression is helpful for cancer treatment. However, the relationship between ferroptosis-related genes and OC patient prognosis is still vastly unknown, making it still a challenge for developing ferroptosis therapy for OC. Methods The Cancer Genome Atlas (TCGA) data of OC were obtained and the datasets were randomly divided into training and test datasets. A novel ferroptosis-related gene signature associated with overall survival (OS) was constructed according to the training cohort. The test dataset and ICGC dataset were used to validate this signature. Results We constructed a model containing nine ferroptosis-related genes, namely, LPCAT3, ACSL3, CRYAB, PTGS2, ALOX12, HSBP1, SLC1A5, SLC7A11, and ZEB1, and predicted the OS of OC in TCGA. At a suitable cutoff, patients were divided into low risk and high risk groups. The OS curves of the two groups of patients had significant differences, and the time-dependent receiver operating characteristics (ROCs) were as high as 0.664, respectively. Then, the test dataset and the ICGC dataset were used to evaluate our model, and the ROCs of test dataset were 0.667 and 0.777, respectively. In addition, functional analysis and correlation analysis showed that immune-related pathways were significantly enriched. Meanwhile, we also integrated with other clinical factors and we found the synthesized clinical factors and ferroptosis-related gene signature improved prognostic accuracy relative to the ferroptosis-related gene signature alone. Conclusion The ferroptosis-related gene signature could predict the OS of OC patients and improve therapeutic decision-making.


Introduction
Ovarian cancer (OC) is the eighth most common cause of cancer death and the second cause of gynecology cancer death in women around the world [1]. Among all types of OCs, epithelial OC (EOC) accounts for over 95% of all ovarian malignancies [2,3]. OC is heterogeneous and the etiology remains complicated and uncertain [4,5]. Risk factors include inherited risk, obesity, age, perineal talc use, etc. [3,6]. e prognosis of OC relies on the stage and early prevention. Over the past years, improved screening, surgery, and treatment methods have contributed largely to the increase of survival. However, survival rates for OC have changed modestly for decades, even in developed countries such as America and Canada [3]. Approximately 70% of OCs are diagnosed at an advanced stage and have a relatively low 5-year survival rate of 30% [7]. Uncertain etiologic factors and low survival rate of OC make the finding of novel therapeutic strategies and models urgent.
Ferroptosis, first coined in 2012, is an iron-dependent and reactive oxygen species (ROS) reliant form of regulated cell death (RCD) [8,9]. Emerging evidence shows that ferroptosis acts like a nexus between metabolism, redox biology, and human health [10]. In recent years, ferroptosis has been exhibiting huge potential of triggering cancer cell death by regulating the mechanism of iron metabolism, amino acid and glutathione metabolism, and ROS metabolism, particularly for eradicating aggressive malignancies that are resistant to conventional therapies [10]. Lately, ferroptosis has been reported to play a vital role in the progression of OC and genes like stearoyl-CoA desaturase 1 could protect OC cells from ferroptosis cell death [11,12]. TAZ-ANGPTL4-NOX2 axis regulates ferroptosis cell death and chemoresistance in EOC [13]. On the other hand, ferroptosis-regulator gene glutathione peroxidase 4 (GPX4) is highly associated with tumorigenesis and progression [14,15]. erefore, ferroptosis can be a potential and powerful target for cancer therapy. However, the relationship between ferroptosis-related genes and OC patient prognosis is still vastly unknown, making it still a challenge for developing ferroptosis therapy for OC.
In this paper, we downloaded OC patient samples from publicity datasets TCGA and ICGC. After preprocessing the data, we constructed a prognostic model composed of nine ferroptosis-related genes in TCGA training set and validated it in TCGA test dataset and ICGC dataset. Further, we conducted functional annotation to discover the possible mechanisms. Finally, restricted median survival (RMS) analysis was applied to combine and evaluate the clinical information and the constructed model. e results showed that the combination had stronger power than the risk model only.

Data Collection and Preprocessing.
All datasets used in this study were publicly available and the workflow of this work is shown in Figure 1.
e count data of OC were obtained from e Cancer Genome Atlas (TCGA) (https:// tcga-data.nci.nih.gov/tcga/). A total of 377 OC patient samples with corresponding clinical information were available in TCGA. e detail information of clinical data about 377 samples is shown in Table 1. For raw count data, we first transformed the Ensembl IDs to gene symbols and protein-coding gene was selected for this research. en, we computed the transcripts per kilobase million (TPM) values, which were more comparable between samples. Meanwhile, the expression data and clinical information of OC were downloaded from International Cancer Genome Consortium (ICGC) resource (https://dcc.icgc.org/). Finally, according to the previous literatures [16][17][18][19], 60 ferroptosisrelated genes were collected and listed in Supplementary Table S1.

Construction of Risk Model and Ferroptosis-Related
Feature Signature. After data preprocessing, 50% of samples were randomly divided into training set (containing 189 OC samples) and another 50% were allocated as validation set (containing 188 OC samples). First, ferroptosis-related genes with prognostic values were identified by univariate cox analysis of overall survival (OS) in the training set selected in TCGA data. e coxph function in the survival R package was used, and p < 0.15 was selected as the threshold. Finally, 15 ferroptosis-related genes were screened (Supplementary Table S2). Further, feature selection was conducted by the randomForestSRC R package. e random forest algorithm was used for ranking the importance of prognostic genes. Only genes with variable relative importance >0.4 were identified as the final signature. en, we performed multivariate cox analysis on the final signature obtained from the random forest algorithm. Finally, using a linear combination in the training datasets, a formula for the risk score was established. e hazards model was constructed as follows: where N is the number of gene, exp is the expression value of gene, and coef is the coefficient of gene in the multivariate cox analysis.

e Robustness Verification of the Gene Signature in
Internal and External Datasets. Risk score and overall survival (OS) analysis were performed using the coxph function in the survival R package. e sensitivity and specificity of the model were assessed by the receiver operating characteristic (ROC) curve, drawn by using the timeROC R package, and were used for analyzing the prognosis prediction of 1 year, 3 years, and 5 years [20]. en, to verify the stability of the model obtained, the performance of the model was evaluated in TCGA test dataset and ICGC cohort.

Estimation of the Abundance of Immune Cell Populations.
In this study, 24 tumor-infiltrating immune cells (TIICs) from the literature [21] that included two categories of adaptive immunity and innate immunity were used to calculate the infiltration level of specific immune cell using Single-Sample Gene Set Enrichment Analysis (ssGSEA) algorithm. In brief, ssGSEA applied gene signatures expressed by immune cell populations to individual cancer samples and we used ssGSEA algorithm to estimate the infiltration levels of 24 kinds of TIICs in OC samples. In our research, the ssGSEA algorithm was implemented in the gsva R package.

Functional Annotation Analysis.
In the training dataset, patients with OC were divided into two groups, including high risk and low risk groups, according to the optimal cutoff value. To identify the potentially altered pathways between high risk and low risk groups, the Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analysis and Gene Ontology (GO) analysis were applied for gene set annotation, and GSEA algorithm was applied to identify the key pathways and biological process by using the R package "clusterProfiler." 2.6. Statistical Analysis. Statistical analysis was performed using the R software (3.6.2 version, https://cran.r-project. org/). Student's t-test was used to evaluate the difference between different groups. Chi-squared test was used to compare the differences in different proportions. e ssGSEA scores between two groups were compared by Mann-Whitney test with p values (adjusted by the BH method). e Kaplan-Meier method was applied to perform OS analysis. e differences of OS between two groups were assessed by two-sided log rank tests. p value <0.05 was regarded statistically significant.

Identification of Nine Ferroptosis-Related Genes.
In this paper, 60 ferroptosis-related genes were processed by randomForestSRC R package for gene feature selection. Ferroptosis-related genes with relative importance >0.4 were considered as the final signature. e relationship between the error rate and the number of classification trees is shown in Figure 2(a). After ranking these genes according to the importance of out of bag, 9 top ferroptosis-related genes are shown in Figure 2   Journal of Oncology weighted by their coefficients from multivariate cox analysis, a hazard model was constructed as a formula: Riskscore e risk score of each sample was calculated using the above method. e patients in TCGA training cohort were divided into high risk group (n � 58) and low risk group (n � 131) according to the optimal cutoff value determined by survminer package in R. As the Kaplan-Meier curves show in Figure 3(a), people in high risk group have a higher probability of death than those in the low risk group (p < 0.0001). e ROC analysis is shown in Figure 3(b) and the ROC curves reach 0.654 at 1 year, 0.664 at 3 years, and 0.69 at 5 years. And the detailed risk score, survival information, and ferroptosis-related genes' expression are displayed (Figures 3(c)-3(e)).

Validation of the Nine Ferroptosis Genes' Signature Using the Test Dataset.
e robustness of the model was examined in the test dataset from TCGA cohort (n � 188), including 88 samples in high risk group and 100 samples in low risk group according to the same risk formula. Patients in higher risk group had poorer survival time than those in low risk group, consistent with the former results ( Figure 4(a)). e AUC of time-dependent ROC in 1 year, 3 years, and 5 years is 0.7, 0.667, and 0.612, respectively ( Figure 4(b)). e detailed risk score, survival information, and ferroptosis-related genes' expression also are displayed (Figures 4(c)-4(e)).

Validation of the Nine Ferroptosis-Related
Genes' Signature in ICGC Cohort. To further test the robustness of the constructed model, patients (n � 93) from the ICGC cohort were categorized into high risk group (40 samples) and low risk group (53 samples) according to the same risk formula above. e survival curves show the patients in high risk group had lower survival probability than patients in low risk group (p < 0.0001) ( Figure 5(a)). e AUC of the model was 0.693 at 1 year, 0.777 at 3 years, and 0.718 at 5 years ( Figure 5(b)). e detailed risk scores, survival information, and nine ferroptosis-related genes' expression in ICGC cohort are shown (Figures 5(c)-5(e)).

Independent Prognostic Factor of the Gene Signature.
We carried out univariate and multivariate cox analysis to determine whether the gene signature was an independent prognostic predictor. Applying univariate cox regression analysis, we found the risk score was significantly associated with OS in the training dataset, test dataset, and the ICGC cohort (HR � 2.657, 95% CI � 1.823-3.872, p < 0.001; HR � 1.887, 95% CI � 1.287-2.768, p < 0.001; HR � 3.115, 95% CI � 1.914-5.069, p < 0.001, respectively) ( Table 2). After correction for other confounding factors by the multivariate cox regression analysis, the risk score still proved to be an independent predictor for OS (HR � 1.767, 95% CI � 1.155-2.704, p � 0.009; HR � 1.944, 95% CI � 1.271-2.973, p � 0.002; HR � 3.06, 95% CI � 1.865-5.021, p < 0.001, respectively). In addition, the ferroptosis-related gene model also was assessed on the clinical factors, including age, stage, and grade tumor status of the tumor, and the Kaplan-Meier analyses revealed that patients in the high risk of death group had significantly shorter OS compared with patients in the low risk of death group in the training dataset, test dataset, and ICGC dataset (p < 0.05) ( Figure S1).

e Relationship between Risk Scores and Immune Status.
Considering ferroptosis was strongly associated with immune status, we further explored the correlations between risk scores and immune status using the ssGSEA method. e different subpopulations of immune cells were divided into adaptive immunity cells and innate immunity cells. First, the correlation analysis between the nine ferroptosisrelated genes and risk scores and the abundance of immune cells are shown in Figure 6(a). e results showed that the risk scores and the nine ferroptosis-related genes meet strong correlations with most of the immune cells, such as eosinophils, iDC, macrophages, neutrophils, NK cells, Tem, Tgd, and 1 cells, suggesting strong connections between the nine ferroptosis-related genes and immune status. en, Journal of Oncology heatmap and the boxplot of ssGSEA scores of adaptive immunity cells and innate immunity cells between high risk patients and low risk patients in TCGA training cohort are shown in Figures 6(b) and 6(c). In addition, the nine ferroptosis-related genes in high risk group and low risk group were compared and the results are shown in Figure 6(d). e expression level of nine ferroptosis-related genes was significantly different in high risk group and low risk group. Among them, the expression levels of ACSL3, ALOX12, CRYAB, LPCAT3, PTGS2, SLC1A5, and ZEB1 were higher in the high risk group, while the levels of HSBP1 and SLC7A11 were lower in the high risk group. We further verified the above results in TCGA test set and ICGC cohort. e results showed that the risk scores of patients also had close positive correlations with eosinophils, iDC, macrophages, neutrophils, NK CD56dim cells, NK cells, Tem, and Tgd, while risk scores had negative correlations with NK CD56 bright cells, pDC, and TFH ( Figures S2(a)-S2(d)). ICGC results also showed that the risk scores and the nine ferroptosis-related genes had strong correlations with most of the immune cells, such as eosinophils, 2 cells, Tgd cells, cytotoxic cells, pDC, and 1 cells (Figure S3(a)-S3(d)). Above all, we can summarize that the risk score and the nine ferroptosis-related genes were associated with multiple immune cells.

Functional Analysis.
Gene Set Enrichment Analysis (GSEA) was conducted to find the key pathways and biological functions that differentiate the different groups. First, the volcano map and heatmap between two groups are drawn in Figures 7(a) and 7(b). en, KEGG analysis and GO analysis were conducted and the results showed that the DEGs were mainly enriched in cell adhesion molecules, complement and coagulation cascades, ECM-receptor interaction, JAK-STAT signaling pathway, MAPK signaling pathway, PI3K-Akt signaling pathway, and so on, which were not only iron-related but also immune-related. Interestingly, DEGs between high risk group and low risk group also were enriched in several immune-related GO terms such as adaptive immune response, immune responseactivating cell surface receptor signaling pathway, immune response-activating signal transduction, lymphocyte mediated immunity, regulation of cell growth, regulation of immune effector process, and so on, suggesting that the signature may be involved in these pathways and thus influence the survival of OC.

Combining Riskscore with Clinical Characteristics.
In addition to Riskscore, we also affirmed that clinical characteristics (i.e., tumor status) served as independent prognostic factors, which could have complementary values (Table 2). To further improve the prognostic accuracy, we combined Riskscore with the major clinical variables using the coefficients generated from multivariate cox regression analysis in the TCGA training cohort and generated a new integrative model IRiskscore as follows: IRiskscore � 5.578 × Riskscore + 2.240 × tumor status. However, due to the lack of tumor status information in ICGC cohort, the integrated model of IRiskscore was further applied to the TCGA training cohort and test cohort where full clinical information was available. Significant improvement in estimation of restricted mean survival (RMS) was achieved with the continuous form of IRiskscore relative to Riskscore (C-index: 0.67 vs 0.62 in the TCGA training cohort, p < 0.05; C-index: 0.69 vs 0.6 in TCGA test cohort, p < 0.001; Figures 8(a) and 8(b)).

Discussion
OC is still a challenging disease to human beings, especially women, with high incidence and morbidity. In recent years, large efforts have been made to unveil the etiology and mechanism in order to expand the landscape of OC therapeutic [22,23]. Selective induction of cancer cell death is the most effective therapy method of malignant tumor [24]. Increasing evidence showed that ferroptosis plays a vital role in tumorigenesis and cancer therapeutics [10,17]. However, the number of ferroptosis-related researches in OC is still very small and the systematic analysis of OC has yet to be elucidated. In the present study, we first constructed a prognostic model integrating nine ferroptosis-related genes gene signature improved prognostic accuracy relative to the ferroptosis-related gene signature alone.
In this study, the constructed prognostic model was composed of nine ferroptosis-related genes and they were reported to be involved in the development of several diseases.
LPCAT3, an enzyme that converts lysophosphatidylcholine to phosphatidylcholine in the liver, could maintain the systemic homeostasis and participate in the phospholipid remodeling and intestinal stem cell growth and tumorigenesis [25,26]. ACSL3, an androgen-responsive gene involved in the generation of fatty acyl-CoA esters, could promote intratumoral steroidogenesis in prostate  Journal of Oncology Journal of Oncology cancer cells [27]. CRYAB, a member of the small heat shock protein family, could regulate several signaling pathways including PI3K/AKT and ERK pathways in cancers [28,29]. PTGS2, also named cyclooxygenase-2, targeting the PGE2/ NF-kappaB pathway, could promote the proliferation and serve as an anti-inflammatory drug target in OC [30,31]. ALOX12, a member of a nonheme lipoxygenase family of dioxygenases, plays a crucial role in ALOX12-12HETE-GPR31 signaling axis and was dysregulated in recurrence of hepatocellular carcinoma [32]. ALOX12 is also required for p53-mediated tumor suppression through a distinct ferroptosis pathway [33]. Lin28A could enrich HSBP1 and upregulate its expression and then regulate the stem-like properties of OC [34]. SLC1A5 protects patients with nonserous OC from recurrent disease, presumably by means of biological mechanisms that are unrelated to cytotoxic drug sensitivity [35]. e SLC7A11-encoded cystine transporter supplies cells with cysteine which is a key source of GSH [36]. Antisense lncRNA As-SLC7A11 suppresses epithelial ovarian cancer progression mainly by targeting SLC7A11 [37]. ZEB1, best known for driving an epithelialto-mesenchymal transition (EMT) in cancer cells to promote tumor progression, is required by tumor-associated macrophages (TAMs) for their tumor-promoting and * * * * * * * * * *  chemotherapy resistance functions in a mouse model of ovarian cancer [38].
Based on the prognostic model, we divided patients into high risk groups and low risk groups from TCGA and ICGC cohort and then risk scores were calculated using the formula. After validation of the prognostic model, ssGSEA method was used for identifying the relationship between ferroptosis and tumor immunity. Interestingly, the immune cells and these nine genes were significant between high risk groups and low risk groups. Although the mechanisms of OC still remain largely unknown, the research we performed took an insight into several pathways in OC based on the concept of ferroptosis and immune status. Based on functional analysis, KEGG pathway enrichment analysis showed the DEGs were mainly enriched in cell adhesion molecules [39], JAK-STAT signaling pathway [40], MAPK signaling pathway [41], PI3K-Akt signaling pathway [42], ECM-receptor interaction [43], complement and coagulation cascades [44,45], focal adhesion [46][47][48], and so on, which were not only iron-related but also immune-related. Interestingly, DEGs between high risk group and low risk group were found enriched in several immunerelated GO terms such as adaptive immune response [49], immune response-activating cell surface receptor signaling pathway [50], immune response-activating signal transduction, lymphocyte mediated immunity, regulation of cell growth, regulation of immune effector process, and so on.
To our knowledge, this ferroptosis-related gene signature has not been previously reported and it will provide assistance to clinical practice. First, in this model, we only need targeted sequencing based on specific genes which greatly  reduces the financial burden of patients. Second, it also does not require the identification of somatic mutations and copy number variation in patients. ird, we can detect the expression of these genes by single cell sequencing in circulating tumor cells to patients who are poor candidates for surgery. In addition, we also integrated with other clinical factors and we found the synthesized clinical factors and ferroptosis-related gene signature improved prognostic accuracy relative to the ferroptosis-related gene signature alone, which may become routinely used in the future. However, there are still several limits of our present study. First, all data processed in this study were publicity data. e real world data need to be warranted to verify our results. Second, although we have tested the robustness of our model several times, the intrinsic weakness is still inevitable. Finally, experimental studies need to be carried out to investigate the functional roles and confirm the presence of gene products by immunohistochemistry of the nine genes in OC in future work.
In summary, this study constructed a model containing 9 ferroptosis-related genes. e model was validated to be associated with OS in the TCGA training set, TCGA test set, and ICGC cohort. e ssGSEA method demonstrated that ferroptosis had a tight link with tumor immunity but needs further experimental validation.

Conclusions
is is the first study to report a novel ferroptosis-related prognostic model to predict OS of OC.

Abbreviations
HR: Hazard ratio CI: Confidence interval.

Conflicts of Interest
e authors report no conflicts of interest in this work.

Authors' Contributions
Liuqing Yang, Saisai Tian, and Yun Chen performed data analysis and data interpretation; Chenyun Miao, Ying Zhao, and Ruye Wang prepared the manuscript and performed statistical analysis; Qin Zhang designed the experiments and prepared the manuscript. All authors reviewed the manuscript. Figure S1: Kaplan-Meier estimates of the overall survival of patients with different clinical factors (age, tumor status, stage, and grade) in training set, test set, and ICGC set. Figure S2: Comparison of the ssGSEA scores between different risk groups in the TCGA test set. (a) e risk score between the nine ferroptosis-related genes and different immune cells. (b) Heatmap of the different groups and components. (c) Detailed risk scores and comparison in high risk group and low risk group. (d) e expression level and comparison of nine ferroptosis-related genes in high risk group and low risk group. e meaning of the statistical difference is as follows: * p < 0.05, * * p < 0.01, and * * * p < 0.001. Figure S3. Comparison of the ssGSEA scores between different risk groups in the ICGC cohort. (a) e risk score between the nine ferroptosis-related genes and different immune cells. (b) Heatmap of the different groups and components. (c) Detailed risk scores and comparison in high risk group and low risk group. (d) e expression level and comparison of nine ferroptosis-related genes in high risk group and low risk group. e meaning of the statistical difference is as follows: * p < 0.05, * * p < 0.01, and * * * p < 0.001. Table S1: ferroptosis-related genes.