- Construction of the ovarian cancer co-expression modules
The WGCNA package in R language was used to perform clustering analyses of ovarian cancer samples from TCGA dataset shown in Figure S1a. Of the 537 samples, 525 revealed no significant differences in the clustering analysis. The corresponding clinical parameters are shown in Figure S1b. The power value, representing the most critical parameter, was screened out to form a scale-free network, which influenced the scale independence and mean connectivity of the co-expression module. In our study, we set β = 3 (scale-free = 0.94) as the power value with a scale independence of up to 0.9, and higher mean connectivity (Figure S2). Subsequently, β = 3 was used to construct the co-expression modules, of which nine different ovarian cancer modules were identified (including a grey module) (Table S1). Genes with similar expression patterns were placed in one module using average linkage clustering, and the first 25% most variant genes were used from the 525 samples (Figure 1a). Figure 1b shows the correlation between the co-expression modules and clinical traits. As is commonly understood, clinical-stage, clinical-grade, and the presence of lymphatic invasion correlate with ovarian cancer prognoses. For this study, we chose modules with gene numbers >200 and p-values <0.001 to study further. Therefore, we selected turquoise modules, which showed the most relevance with ovarian cancer lymphatic invasion and blue modules, which showed the most relevance with ovarian cancer stages (Figure 1c). In addition, gene significance plots vs. module memberships are shown in Figure 1d. From the blue and turquoise modules, 1648 genes were further analyzed using Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) analyses, as shown in Table S2-5. Notably, the biologic pathway analysis showed that these genes significantly correlated with immune responses, indicating that immune responses participate in ovarian cancer progression and prognoses (Figure S3).
- Identification of a novel ovarian cancer prognostic signature
A univariate Cox regression analysis was used to investigate the prognostic role of the1648 candidate genes from turquoise and blue modules. The top 15 genes with a p-value of <0.001 were used for further analyses (Table 1). In addition to the univariate cox regression analysis, the Kaplan-Meier method was used to predict overall survival (OS) of the 15 candidate genes. The results showed that all of the genes were significantly correlated with ovarian cancer prognoses and were consistent with the results of the univariate cox regression analyses (Figure S4). Next, multivariate cox regression analyses were used to construct prognostic signatures, and 4 genes were chosen using the following equation: Risk score = (0.11483 × CH25H expression) + (0.22472 × HSPB7 expression) – (0.28916 × LOC158830 expression) + (0.21726 × PPM2C expression).
According to the signature risk scores, patients were divided into two groups, namely high-risk and low-risk groups. By investigating the prognostic value of the signatures, the high-risk group was found to have shorter survival times than the low-risk group (Figure 2a, p <0.001). Stratified survival analyses showed that the prognostic signature significantly correlated with OS in ovarian cancer patients according to the clinical parameters (Figure S5). These analyses indicated that the signatures could precisely predict prognoses and did not need the clinical parameter information. Moreover, to investigate the accuracy of the identified signature, a 3-year receiver operating characteristic (ROC) curve analysis was performed. The ROC of the signature was 0.683, which was significantly higher than that of the prognostic-related clinical parameters (Figure 2b). Clinically, the CA-125 gene was found to be a highly sensitive biomarker for ovarian cancer diagnoses. However, the ROC of the CA-125 gene was 0.572, which was significantly lower than that of the signature (Figure 2c). A nomogram was constructed to predict the clinical survival of ovarian cancer patients by combining the signature with other clinical parameters (Figure 2d). The signature risk score was found to be closely correlated with the successful outcomes of primary therapies (Figure 2e). Through the examination of venous invasion, tumor residual disease, and clinical stages in ovarian cancer patients, the signature could precisely predict ovarian cancer prognoses.
In Table 2, the univariate and multivariate Cox regression analysis was used to test whether this signature could act as an independent prognostic factor for ovarian cancer. The results showed that it could act as an independent factor when adjusted for age, stage, grade, tumor residual disease, and lymphatic and venous invasion. To clarify the relationship between the signature and clinical parameters, the samples were divided into two groups. The signature was found to be significantly correlated with the clinical-stage, residual tumor size, venous invasion, therapeutic outcome, and patient cancer status (Table 3).
To investigate the distinct biologic features between the high-risk and low-risk groups, DEGs with a fold change of >1.5 and a p-value of <0.05 were chosen. Using DAVID Gene Functional Classification software, we analyzed 170 candidate genes. The results indicated that the biologic process (BP) enrichments were significantly correlated with immune responses (Figure S6). The results showed that the identified prognostic signature might be closely related to ovarian cancer patient immune responses, thus, providing potential immunotherapy for these patients.
- Validation of the prognostic signature with independent cohorts of ovarian cancer
To validate the signature of the ovarian cancer datasets derived from the The Cancer Genome Atlas (TCGA) dataset, we downloaded GSE26193, GSE63885, and GSE18520 from the GEO database to represent three independent cohorts. According to the signature risk scores, the patients were divided into two groups. In these three datasets, patients with higher risk scores showed worse prognoses as expressed by shorter OS times than patients with lower risk scores (Figure 3). According to the Cox regression analysis of the GSE26193 dataset, this signature was an independent prognostic factor for ovarian cancer (Table S6).
In Table S7, we analyzed the relationship between the signature and the clinical parameters. However, in the GSE26193 and GSE63885 datasets, no clinical parameters were found that correlated with the signature, which could have been due to the limited number of samples in these two datasets.
- The prognostic signature correlates with immune cell expression in the ovarian cancer microenvironment.
Tumor microenvironments play an important role in regulating ovarian cancer progression. According to the above analyses, we found that the prognostic signature might have a close relationship with the immune responses of ovarian cancer patients. In Figure 4a, we showed that stromal cell expression in the tumor microenvironment correlated with ovarian cancer prognoses. The Pearson correlation analysis showed that the signature was positively correlated with stromal scores, estimation of stromal and immune cells in malignant tumor tissues using expression data (ESTIMATE) scores, and neutrophil and resting mast cell expression with an R >0.2 (Figure 4b). According to the signature risk score, the samples were divided into two groups. The relationship between immune cells and the signature were tested, and the results are shown in the bar charts of Figure 4c. These two methods showed similar results, which proved that the signature correlated significantly with stromal score expression, ESTIMATE scores, and neutrophils and resting mast cell expression. In addition, we selected a few immune checkpoint-related genes to further investigate the signature relationships (Figure 4d). The results showed some immune-related genes, such as LGALS3, PDCD1, IL6, IL6ST, CD163, FCGR2B, MSR1, HAVCR2, ICOS, IL10, and CCL2 were significantly correlated with the signature.
- The correlation between the signature and the immune status of ovarian cancer patients
The immune status of patients is well-known to play an important role in cancer progression. We investigated the correlation between ovarian cancer patient clinical parameters and immune cell expression using the CIBERSORT algorithm using a one-way ANOVA analysis. We found that of the immune cells, M2 macrophages were most correlated with ovarian cancer progression, including cancer stage, the success of primary therapy outcomes, and venous invasion, as shown in Figure 5a. In addition, the OS analysis of 20 immune cell types showed that M0 macrophages and monocytes were significantly correlated with the prognoses of ovarian cancer patients, as shown in Figure S7. Therefore, macrophages might play an important role in ovarian cancer progression.
Recently, TMB has been found to be an important factor in cancer progression and immunity with increasing attention. According to our analyses, ovarian cancer has been shown to have a high mutation rate. Over 90% of the samples showed gene mutations (Figure S8a). Among these samples, the TP53, TTN, and MUC16 (namely CA-125) genes had the highest mutation rates (Figure S8b). Co-occurrence was seen in some of the mutated genes (Figure S8c). Missense mutations had the largest classification variance, while single nucleotide polymorphisms (SNPs) had the largest ovarian cancer type variance (Figure S8d). Further analysis showed that patients with higher TMBs had better prognoses (Figure 5b). According to TCGA gene mutations, we analyzed correlations among the top 20 mutational genes and the signature and found that no significant correlations existed (Table S8). Therefore, we could classify ovarian cancer patient statuses more specifically. Next, we performed a combined analysis of the signature with TMB expression in ovarian cancer patients. The OS of ovarian cancer patients with higher risk scores and lower TMB expression had the worst prognoses (Figure 5c). Moreover, the relationship of the signature or TMB expression with immune cells showed that the signature was significantly correlated with the expression of resting memory CD4 T cells, activated memory CD4 T cells, M0 macrophages, M2 macrophages, resting mast cells, activated mast cells, and neutrophils (Figure 5d). However, TMB expression correlated with memory B cells, resting memory CD4 T cells, and M1 macrophages (Figure 5e). Overall, these results demonstrated that the effects of the signature and TMB expression on immune cells were very different. Therefore, we combined the signature and TMB expression and found that memory B cells, resting memory CD4 T cells, M2 macrophages, resting mast cells, and neutrophils were closely correlated (Figure 5f). The expression of resting memory CD4 T cells, M2 macrophages, and neutrophils were positively correlated with the status of signatures combined with TMB expression in ovarian cancer patients.