3.1. Patient characteristics and collection of GRGs
From TCGA database, we downloaded the complete data of 583 OV patients which contained clinical information and the expression profiles of RNA sequencing. We manually searched for glycolysis-related gene sets from MSigDB version 6.2 and referenced relevant literature. Five related gene sets (REACTOME_GLYCOLYSIS, HALLMARK_GLYCOLYSIS, GO_GLYCOLYTIC_PROCESS, KEGG_GLYCOLYSIS_GLUCONEOGENESIS, and BIOCARTA_GLYCOLYSIS_ PATHWAY) were downloaded, and 443 genes were obtained. After excluding duplicate genes, the corresponding 386 genes remained for the subsequent analysis. The integrated clinical data and list of the GRGs are shown in Tables 1 and S1, respectively.
3.2. Construction of the glycolysis-related risk signature
To examine the prognostic value of the 386 GRGs, 11 genes were obtained by preliminary screening, which was performed by univariate Cox regression analysis with the criterion of adjusted p < 0.05. Finally, nine genes (ISG20, CITED2, PYGB, IRS2, LHX9, PC, ANGPTL4, TGFBI, and DDIT4) significantly correlated with OS (adjusted p < 0.05) after filtration using LASSO and multivariable Cox regression analyses (Figure 1A, B and C, Table 2). And then a predictive signature based on GRGs was constructed and the formula to assess the risk of every patients survival was showed as follows: risk score = (−0.25414) × ISG20 expression level + 0.07897 × CITED2 expression level + 0.11769 × PYGB expression level + 0.09112 × IRS2 expression level + 0.06399 × ANGPTL4 expression level + 0.04811 × TGFBI expression level + 0.03555 × LHX9 expression level + 0.05593 × PC expression level + 0.05907 × DDIT4 expression level. Additionally, Figure 1D showed the mutation of the nine selected GRGs of 583 OV patients in the cBioPortal database. And we used the gene model calculated every patient’s risk score in the training set on the basis of the nine-GRG expression levels). The results suggested that patients with high risk scores were associated with a worse prognosis, when compare to patients with low risk scores (P < 0.0001, log-rank test; Figure 2A and 2B). The areas under the curve (AUC) for 3-, 5-, and 10-year OS were 0.709, 0.762, and 0.808, respectively (Figure 2C). The heatmap revealed the expression patterns of the 9 GRGs between the two different prognostic patient groups (Figure 2D).
3.3. Evaluation of the predictive capability of the nine-GRG risk signature
After constructing the GRG predictive model, we selected three datasets to verify its prediction performance. The sets for validation including GSE63885, GSE26193, and GSE30161 datasets with 101, 107, and 58 patients with OV, respectively (Figure 3). The AUCs of 3-, 5-, and 10-year OS were respectively 0.716, 0.767, and 0.819 in the GSE26193 dataset; 0.808, 0.8, and 0.853 in the GSE30161 dataset; and 0.636, 0.722, and 0.815 in the GSE63885 dataset. Survival analysis revealed that our risk signature performed well in the validation sets. And the survival differences between the satisfied high-risk and low-risk groups were all statistically significant (P value of <0.05). The distribution of risk scores and survival statuses of the patients with OV in the three sets are also displayed in Figure 3.
3.4. Risk score generated from the nine-gene signature as an independent prognostic indicator
The exploration of independent predictive factors is through univariate analysis of clinical factors and risk models combined with multivariate regression analysis. Table 3 showed that, in addition to age and tumor stage, our GRG risk model could independently predict the OS according to the results from univariate analysis (HR [95% CI], 2.334 [1.817−2.997]; P < 0.001) and multivariate analysis (HR [95% CI], 2.361 [1.830−3.047]; P < 0.001), referring to the statistical standard of adjusted P value of <0.05. Furthermore, the AUC of the nine-gene signature was higher than that of any single clinicopathological variable (Figure 4A). The findings in the present study suggest that the gene model had independently and effectively predictive ability in the survival prediction for OV patients. To develop a quantitative method that can predict the OS of patients with OV, a nomogram was constructed. The predictors included risk score and age (Figure 4B).
3.5. Validation of the nine-GRG signature in predicting survival using Kaplan-Meier curves
The clinical features of age, histological grade, and tumor stage represent predictive prognostic factors of OV after the performance of univariate Cox regression analysis of OS. Kaplan-Meier curves revealed that clinical features that showed consistent results, namely patient age of >65 years and disease stages III and IV, were associated with poor prognosis (Figures 4C–4D).
For the purpose of testing whether our nine-GRG signature can play a role in different TNM stages, histological grades, and ages, a subgroup analysis was performed for each clinical feature. After the analyses of Kaplan-Meier survival, the risk signature was demonstrated to be stably prognostic power and applicable to the OV patients when they were stratified in to different age and TNM stage groups (Figures 5A and 5B). However, when the patients were stratified into high-grade (grades 3 and 4) and low-grade (grades 1 and 2) subgroups, the risk score of the eight-gene signature remained an independent prognostic indicator in the high-grade subgroup (P < 0.001), but not in the low-grade subgroup (P = 0.067; Figure 5C). The risk model showed more effective prediction in the patients groups of high-grade OV.