Evaluating the Prognostic Accuracy of Biomarkers for Glioblastoma Multiforme Using The Cancer Genome Atlas Data

Abstract

Background: Glioblastoma multiforme (GBM) is the most common and aggressive primary brain tumor. Previous studies on GBM biomarkers focused on the effect of the biomarkers on overall survival (OS). Until now, no study has been published that evaluates the performance of biomarkers for prognosing OS. We examined the performance of microRNAs, gene expressions, gene signatures, and methylation that were previously identified to be prognostic. In addition, we investigated whether using clinical risk factors in combination with biomarkers can improve the prognostic performance.

Methods: The Cancer Genome Atlas, which provides both biomarkers and OS information, was used in this study. The time-dependent receiver operating characteristic (ROC) curve was used to evaluate the prognostic accuracy.

Results: For prognosis of OS by 2 years from diagnosis, the area under the ROC curve (AUC) of microRNAs, Mir21 and Mir222, was 0.550 and 0.625, respectively. When age was included in the risk prediction score of these biomarkers, the AUC increased to 0.719 and 0.701, respectively. The SAMSN1 gene expression attains an AUC of 0.563, and the "8-gene" signature identified by Bao achieves an AUC of 0.613.

Conclusions: Although some biomarkers are significantly associated with OS, the ability of these biomarkers for prognosing OS events is limited. Incorporating clinical risk factors, such as age, can greatly improve the prognostic performance.

Keywords: MGMT methylation; Prognostic accuracy; The Cancer Genome Atlas; gene signature; glioblastoma; microRNA; survival analysis.

Figure 1.

Kaplan-Meier product-limit survival curve of overall survival among all patients with GBM in The Cancer Genome Atlas. The median survival time among these patients is about 14 months. GBM indicates glioblastoma multiforme.

Figure 2.

Time-dependent ROC curves of Mir21 based on (A) ROC^C/D(Y; t) and (B) ROC^I/D(Y; t) by or at 15, 24, 33, and 42 months. AUC indicates area under the ROC curve; KPS, Karnofsky Performance Score; ROC, receiver operating characteristic; TPR, true-positive rate.

Figure 3.

Time-dependent AUC of Mir21 based on ROC^C/D(Y; t). AUCs were calculated up until 5 years (60 months) after diagnosis. The Mir21 + age classifier has the best overall accuracy (largest AUC) to prognose cumulative overall survival events over the entire 60-month period, followed by the Mir21 + KPS classifier. AUC indicates area under the ROC curve; GBM, glioblastoma multiforme; KPS, Karnofsky Performance Score; ROC, receiver operating characteristic.

Figure 4.

Time-dependent ROC curves of Mir222 based on (A) ROC^C/D(Y; t) and (B) ROC^I/D(Y; t) by or at 15, 24, 33, and 42 months. AUC indicates area under the ROC curve; KPS, Karnofsky Performance Score; ROC, receiver operating characteristic; TPR, true-positive rate.

Figure 5.

Time-dependent AUC of Mir222 based on ROC^C/D(Y; t). AUCs were calculated up until 5 years (60 months) after diagnosis. The Mir222 + age classifier has the best overall accuracy (largest AUC) to prognose cumulative overall survival events over the entire 60-month period. There is a crossover between the time-dependent AUC curve for Mir222 and Mir222 + KPS. This indicates that during the earlier follow-up period (~<30 months), incorporating KPS in the prognosis will result in better prognostic accuracies than only using the Mir222 expression; however, for prognosis after 30 months, including KPS will not improve the overall prognostic performance. AUC indicates area under the ROC curve; GBM, glioblastoma multiforme; KPS, Karnofsky Performance Score; ROC, receiver operating characteristic.

Figure 6.

Time-dependent AUC of SAMSN1 based on ROC^C/D(Y; t). AUCs were calculated up until 5 years (60 months) after diagnosis. The SAMSN1 + age classifier has the best overall accuracy (largest AUC) to prognose cumulative overall survival events over the entire 60-month period, followed by the SAMSN1 + KPS classifier. AUC indicates area under the ROC curve; GBM, glioblastoma multiforme; KPS, Karnofsky Performance Score; ROC, receiver operating characteristic.

Figure 7.

Time-dependent AUC of 4 gene signatures reported by Bao et al based on ROC^C/D(Y; t). The 8-gene signature has the best overall accuracy (largest AUC) to prognose cumulative overall survival events over the entire 60-month period. The performance of the 17-gene and 61-gene signatures is similar as the time-dependent ROC curves for them are close to each other over the 60-month period. AUC indicates area under the ROC curve; GBM, glioblastoma multiforme; KPS, Karnofsky Performance Score; ROC, receiver operating characteristic.