A 35-gene signature discriminates between rapidly- and slowly-progressing glioblastoma multiforme and predicts survival in known subtypes of the cancer

Abstract

Background: Gene expression can be employed for the discovery of prognostic gene or multigene signatures cancer. In this study, we assessed the prognostic value of a 35-gene expression signature selected by pathway and machine learning based methods in adjuvant therapy-linked glioblastoma multiforme (GBM) patients from the Cancer Genome Atlas.

Methods: Genes with high expression variance was subjected to pathway enrichment analysis and those having roles in chemoradioresistance pathways were used in expression-based feature selection. A modified Support Vector Machine Recursive Feature Elimination algorithm was employed to select a subset of these genes that discriminated between rapidly-progressing and slowly-progressing patients.

Results: Survival analysis on TCGA samples not used in feature selection and samples from four GBM subclasses, as well as from an entirely independent study, showed that the 35-gene signature discriminated between the survival groups in all cases (p<0.05) and could accurately predict survival irrespective of the subtype. In a multivariate analysis, the signature predicted progression-free and overall survival independently of other factors considered.

Conclusion: We propose that the performance of the signature makes it an attractive candidate for further studies to assess its utility as a clinical prognostic and predictive biomarker in GBM patients. Additionally, the signature genes may also be useful therapeutic targets to improve both progression-free and overall survival in GBM patients.

Keywords: Chemoradiation resistance pathways; Glioblastoma multiforme; Prognostic genes; Risk groups.

Fig. 1

Sample selection for the identification of prognostic genes in glioblastoma multiforme. PFS: progression-free survival (days); OS: overall survival (days); adjuvant treatment: chemotherapy and radiation

Fig. 2

Cross-validated error rates of R-SVM in each recursive steps. *The number of features used for SVM classification in each step. Parameters for SVM: kernel = linear, cost = 10, and 5% cross-validation. The red star represents the level at which the minimal cross-validation error was achieved

Fig. 3

Kaplan-Meier plots for low-risk and high-risk groups of GBM patients that received adjuvant chemotherapy and radiotherapy. The patients were classified based on PI score. a PFS plots and b OS plots of risks groups from 118 TCGA patients not used in the feature selection. c OS plots of risk groups from 380 TCGA patients with OS times. d OS plots of risks groups from the Murat et al. data set used for validation. The two numbers in the topright corner of each plot represents the total number of patients in each risk group and the number of patients who experienced progression or death within the follow-up periods, respectively

Fig. 4

Kaplan-Meier progression-free survival plots for risk groups of patients in each subtype of GBM. The patients were classified based on PI score. The two numbers in the topright corner of each plot represents the total number of patients in each risk group and the number of patients who experienced progression or death within the follow-up periods, respectively

Fig. 5

Kaplan-Meier overall survival plots for risk groups of patients in each subtype of GBM. The patients were classified based on PI score. The two numbers in the topright corner of each plot represents the total number of patients in each risk group and the number of patients who experienced progression or death within the follow-up periods, respectively

Fig. 6

Analysis of the subnetwork formed from the interaction between signature genes. a The subnetwork from the STRING database. b Enriched pathways in the subnetwork