Comprehensive pathway-related genes signature for prognosis and recurrence of ovarian cancer

Introduction

Ovarian cancer (OC), one of the most common female malignant tumours, is the fifth-most common cause of mortality in women with severe gynaecological issues, with an estimated 13,940 deaths and 21,750 new cases reported in the 2020 United States cancer statistics (Siegel & Miller, 2020). Because OC is usually diagnosed at a late stage and currently lacks effective treatment options, patients with OC have extremely poor prognosis and only a 30–40% 5-year survival rate (Enroth et al., 2019). Despite the moderate improvements that have been made in the diagnosis and treatment of OC over the past 30 years, OC is still a highly malignant disease that is life-threatening (Li et al., 2019). In addition, OC patients have an average relapse-free survival (RFS) of two to three years after first-line therapy and die due to chemotherapy resistance (Dinh et al., 2008). To date, several common biomarkers for diagnosing OC remain ineffective due to lack of sensitivity and specificity; these include carbohydrate antigen 125 (CA125, AUC = 0.630) (Mansha, Gill & Thomson, 2019) and human epididymis protein 4 (HE4, AUC = 0.719) (Scaletta et al., 2017). At the same time, a number of studies have identified certain genes that are significantly associated with OC patient prognosis, such as TRIM44 (Wang et al., 2018) and CENPK (Lee et al., 2015). However, there is very limited prognostic value in a single candidate biomarker, which is attributed to inconsistent sample collection, detection methods, and small sample sizes. A number of studies have shown that combined biomarkers could improve prognostic accuracy compared to single biomarkers (Barata & Rini, 2017). Thus, extensive studies have attempted to establish molecular signatures based on gene expression to predict survival of patients, including mRNA- (Zhou et al., 2019), microRNA- (Li et al., 2019), and long non-coding RNA-based signatures (Xu et al., 2019). However, only a small number of prognostic signatures have been developed, and none have been directly applied in clinical practice. Hence, there is an urgent need to identify and validate novel, highly sensitive, and specific molecular biomarkers for prognosis, monitoring, and therapy improvement.

In this study, we sought to explore signalling pathways related to ovarian tumorigenesis and attempted to identify the potential predictive abilities of genes participating in these pathways for OC prognosis and recurrence. We established a nine-gene signature panel for overall survival (OS) and a five-gene signature group for RFS using LASSO Cox regression analysis. Both signatures have powerful prognostic predictive abilities and high sensitivity and specificity. Our results revealed that both signatures were independent risk factors for OC survival and recurrence. In addition, we constructed two nomogram models encompassing signature, TNM stage, and tumour status for OS and RFS, respectively, and found that both nomogram models exhibited high potential clinical utility by decision curve analysis (DCA). In general, we constructed two reliable and superior gene signature panels as prognostic predictors of OC patient survival and recurrence.

Materials and Methods

Results

Identification of signalling pathways correlated with OC tumorigenesis

To determine related signalling pathways involved in ovarian tumorigenesis, we performed GSEA on Hallmark gene sets for five independent GEO datasets, GSE40595, GSE12470, GSE10971, GSE27651, and GSE38666. The results showed that there are multiple signalling pathways significantly enriched in OC including “MYC_TARGETS_V1”, “MTORC1_SIGNALING”, “GLYCOLYSIS”, “E2F_TARGETS”, “G2M_CHECKPOINT”, “OXIDATIVE_PHOSPHORYLATION”, “DNA_REPAIR” and “PI3K_AKT_MTOR_SIGNALING” (Figs. 1A–1E and Table S4), suggesting that these signalling pathways might play critical roles in the development of OC.

Establishment of gene signatures for overall survival and relapse-free survival

To evaluate the effect of genes in these pathways on OS and RFS, we performed univariate Cox regression analysis using TCGA data. The results showed that 48 and 44 genes were correlated with patient OS and RFS, respectively. Furthermore, the LASSO Cox selection method was implemented to establish a prognosis-predicting model (OS: Figs. 2A, 2B; RFS: Figs. 2C, 2D). Finally, a nine-gene signature panel for OS and a five-gene signature group for RFS were established by LASSO Cox regression analysis. The formula of the nine-gene signature panel for OS = (−0.0027) * expression of HMGB3 + 0.0122 * expression of PYGB + 0.0094 * expression of IDUA + 0.0092 * expression of LHX9 + (−0.0254) * expression of SLC7A11 + (− 0.0037) * expression of DPYSL4 + 0.0017 * expression of ANGPTL4 + (−0.0357) * expression of ISG20 + (−0.0353) * expression of GRM8; and the formula of the five-gene signature group for RFS = (−0.0114) * expression of SNRPA1 + (−0.0346) * expression of WARS + 0.0234 * expression of CITED2 + 0.002 * expression of ANGPTL4 + 0.0357 * expression of EFNA5 (Table S5). Next, we calculated risk scores for the OS and RFS of each patient based on the above formula. The distribution of risk scores, survival times, patient status, and expression heat maps of genes were obtained. The results showed that patients in the high-risk group had shorter OS and RFS and higher mortality than those in the low-risk group (Figs. 3A, 3B and Figs. S1A, S1B). According to these two signatures, patients were divided into high-risk or low-risk groups using the median risk score as the cut-off point (OS: median = 0.863; RFS: median = 0.457). We performed Kaplan–Meier analysis and found that patients with high risk scores had significantly shorter OS and RFS than those with low risk scores (OS: P <  0.0001, HR = 2.289, 95% CI [1.724–3.04]; RFS: P <  0.0001, HR = 1.783, 95% CI [1.465–2.264]; Figs. 4A, 4B). Furthermore, to explore the sensitivity and specificity of the risk score, we performed time-dependent ROC analysis. The results showed that for OS, the prognostic accuracy of the nine-gene signature panel was 0.640, 0.663, 0.758, and 0.891 at 1, 3, 5, and 10 years, respectively. The AUCs for RFS were 0.638, 0.722, 0.813, and 0.972 at 1, 3, 5, and 10 years, respectively (Figs. 4C, 4D). These results demonstrate that the nine-gene signature panel is a helpful indicator of OS for OC patients and that the five-gene signature group is a powerful prognostic predictor of OC patient recurrence.

Validation of the nine-gene signature panel for overall survival and five-gene signature group for relapse-free survival

To verify the prognostic value of our signatures, we investigated the independent dataset GSE17260 for validation of OS and RFS. We also performed Kaplan–Meier analysis and observed the same results as above, which showed that patients in the high-risk group had shorter OS and RFS and higher mortality than those in the low-risk group based on our nine- and five-gene-based signatures, respectively (Figs. 5A, 5B). In addition, we performed cross validation to prove the consistency of these signatures. We found the same gene sets in GSE17260 dataset (Figs. S2A, S2B), the only difference is the coefficients of each gene have changed a little. Furthermore, we also performed Kaplan–Meier analysis and observed the same results in TCGA dataset (Figs. S2C, S2D).

Prognostic value of gene signatures is independent of other clinical characteristics

To further evaluate the predictive ability of gene signatures and other clinical characteristics in terms of OS and RFS, univariate and multivariate Cox analyses were performed. As shown in Figs. 6A and 6B, the nine-gene signatures, stage, race, and tumour status were significantly correlated with OS, as seen in the univariate Cox regression analysis. Furthermore, multivariate Cox regression analysis results showed that the nine-gene signatures and tumour status were independent prognostic factors for OS. Meanwhile, univariate Cox regression analysis revealed that the five-gene signatures, staging, and tumour status were significantly associated with RFS. The five-gene signatures and tumour status were independent prognostic factors for RFS after multivariate Cox regression analysis.

Clinical application of gene signatures

We further constructed a nomogram predictive model combining gene signatures, race, stage, and tumour status to predict the probability of 1-, 3-, and 5-year OS and RFS in OC patients (Figs. 7A and 7C). The results showed that the nomogram model combined gene signatures, stage, and tumour status (OS: C-index: 0.743; RFS: C-index: 0.712). These results demonstrate that our gene signature-based nomogram approach exhibits superior performance in clinical settings. Furthermore, DCA was conducted to evaluate the clinical usefulness of the risk prediction nomogram for OS and RFS (Figs. 7B and 7D). The DCA results showed that for OS and RFS, the complex factor had a broader range of threshold probabilities and higher net benefits than single clinical characteristics or risk scores. Taken together, the DCA results demonstrate that our nomogram for OS and RFS has high potential clinical utility.

Discussion

To a certain extent, indolent and aggressive cancers cannot be accurately stratified by the TNM staging system, which mostly relies on anatomical information without any molecular biological information or features (Choudhary et al., 2019; McIntyre et al., 2017). Therefore, it is indispensable to identify novel and effective predictive biomarkers for OC prognosis. In this study, we established and validated novel prognostic gene-based signatures for OS and RFS to improve the prediction of survival and tumour recurrence in OC patients.

Increasing number of studies are gradually clarifying the key signalling pathways involved in cancer initiation, progression, and metastasis, including the PI3K-Akt signalling pathway (Mabuchi et al., 2015), MYC signalling pathway (Dang, 2012) and so on. Moreover, the genes participating in these key signalling pathways might play critical roles in tumorigenesis. Therefore, it is necessary to explore the functions and roles of these key genes. Consistently, extensive studies have confirmed several signalling pathway-related genes that are significantly associated with the prognosis of cancer patients. Cao et al. (2019) confirmed that EMT-related genes promote bladder cancer subtype transition from non-muscle invasive bladder cancer (NMIBC) to muscle invasive bladder cancer (MIBC) and affect the prognosis of bladder cancer patients. In this study, our candidate genes for OS and RFS were screened for signalling pathways affecting the development of OC, including ”MYC_TARGETS_V1”, ”MTORC1_SIGNALING”, ”GLYCOLYSIS”, ”E2F_TARGETS” and ”G2M_CHECKPOINT”. MYC is one of the most highly amplified oncogenes in many human cancers and plays a pivotal role in tumorigenesis (Dang, 2012). Mammalian target of rapamycin complex 1 (mTORC1) signalling is found to be elevated in many human cancers, including OC, and promotes cell growth, proliferation, and metabolism (Menon & Manning, 2008). Increasing evidence suggests that aerobic glycolysis is related to tumour growth and chemotherapy resistance (Chakraborty et al., 2017). The E2F family plays a significant role in cancer initiation, progression, and resistance to therapy (Kent & Leone, 2019) and E2Fs can mediate fundamental cell-cycle deregulation in high-grade serous OC (DeMeyer et al., 2009). Hence, it is a plausible inference for these genes to be associated with the survival of patients with clear cell renal cell carcinoma (ccRCC).

High-dimensional data survival models have a high risk of overfitting, which is characterised by reduced importance when applying predictors to independent datasets. To solve this problem, we applied a Cox proportional hazard model using the LASSO penalty method to optimise gene selection. This is useful for selecting genes with strong prognostic value, high expression differences, and low correlation among each other (Tibshirani, 1997). The potential of LASSO has been previously implemented in colon cancer (Dai et al., 2018), bladder cancer (He et al., 2019) and other studies. Therefore, in this study, we successfully applied this method and established a nine-gene-based signature panel for OS and a five-gene-based signature group for RFS to optimise the predictive ability of prognosis in OC patients.

Although extensive studies have advocated several molecular biomarkers for risk stratification in OC, along with traditional clinicopathologic parameters, there are still some points to be improved as follows: (1) small sample sizes need to be enlarged; (2) independent validation datasets are needed to further confirm the stability and accuracy of the established signatures; (3) it is indispensable to compare the predictive power between signatures and the existing TNM staging system; and (4) it is more valuable to explore prognostic biomarkers for OC recurrence. In our study, we established novel superior prognostic signatures for OS and RFS in the TCGA dataset and validated the prognostic value of these signatures using an independent GEO dataset, suggesting the repeatability and practicability of our two signatures. Then, we confirmed that these signatures were independent risk factors for OS and RFS by univariate and multivariate Cox analyses. In addition, we constructed a nomogram model to better predict survival of OC patients and validated its promising clinical application by DCA. In general, our findings provided a superior and useful biomarker for OS and RFS in OC patients, based on the results of the LASSO regression analysis.

Our study demonstrated that nine-gene- and five-gene-based signatures for OS and RFS, respectively, were both closely associated with the prognosis of OC patients. However, the functions of most of the genes in our signatures have not been entirely elucidated. For genes associated with OS, the chromatin-associated high-mobility group box 3 (HMGB3) protein is upregulated and has poor prognosis, and its targeted depletion can attenuate cisplatin resistance in human OC cells (Mukherjee et al., 2019). PYGB has been shown to be upregulated in OC tissue and markedly associated with poor prognosis in OC patients (Zhou, Jin & Wang, 2019). LHX9 is a biomarker in ovarian stem cells (Auersperg, 2013). SLC7A11 is an independent prognostic risk factor for OS in OC (Yin et al., 2019). Low levels of DPYSL4 are significantly correlated with poor prognosis in OC (Nagano et al., 2018). In addition, Jiang et al. (2019) suggested that DPYSL4 is significantly associated with OS in hepatocellular carcinoma. High expression of ISG20 is related to poor clinical outcomes in glioblastoma (Gao et al., 2019). GRM8 is a tumour-suppressor gene in endometrial cancer cell lines (Liang et al., 2012). For genes associated with RFS, SNRPA1 is an oncogene in colorectal cancer (Zeng et al., 2019). Yanagie et al. (2009) demonstrated that CBP/p300-interacting transactivator (CITED2) is potentially a multidrug-resistance gene, which is consistent with our results. ANGPTL4 is a novel biomarker of hypoxia in OC, and is upregulated in partial responders to chemotherapy compared to responders (McEvoy et al., 2015). Yang et al. (2019) found that EFNA5 can be used as a novel prognostic biomarker and potential therapeutic target for OC patients.

Inevitably, there are certain innate limitations to our study that need to be addressed. First, screening for these signatures was based on data from TCGA and GEO datasets without validating their prognostic value in clinical practice. Second, although these candidate genes are involved in key signalling pathways that influence prognosis in OC, the mechanisms underlying some genes in our signature panels remains unclear. Undoubtedly, further investigation of the roles of these genes in OC progression is needed. It is reported that blacks vs whites were over two-fold less likely to receive a surgery-chemotherapy, suggesting that race may affect survival status of ovarian cancer patients greatly, which was consistent with our results. In addition, higher socioeconomic status is associated with a greater probability of undergoing surgical resection and with improved survival in patients with ovarian cancer, demonstrating socioeconomic status might also affect prognosis of patients. Unfortunately, several important clinical characteristics, such as treatment modality and socioeconomic status are not available in the TCGA cohort. Finally, prospective, large-scale, multicentre studies are indispensable to verify the prognostic ability of our signatures before applying them in clinical settings.

Conclusions

In conclusion, we established a nine-gene-based signature panel for OS, and a five-gene-based signature group for RFS using a LASSO Cox regression model. These signatures could be potentially novel and independent prognostic indicators for predicting survival and disease recurrence in OC patients. Moreover, combining these signatures with current clinical indicators may serve to optimise their predictive ability and guide future clinical planning for patient monitoring.

Supplemental Information