A Molecular Predictor Reassesses Classification of Human Grade II/III Gliomas

Abstract

Diffuse gliomas are incurable brain tumors divided in 3 WHO grades (II; III; IV) based on histological criteria. Grade II/III gliomas are clinically very heterogeneous and their prognosis somewhat unpredictable, preventing definition of appropriate treatment. On a cohort of 65 grade II/III glioma patients, a QPCR-based approach allowed selection of a biologically relevant gene list from which a gene signature significantly correlated to overall survival was extracted. This signature clustered the training cohort into two classes of low and high risk of progression and death, and similarly clustered two external independent test cohorts of 104 and 73 grade II/III patients. A 22-gene class predictor of the training clusters optimally distinguished poor from good prognosis patients (median survival of 13-20 months versus over 6 years) in the validation cohorts. This classification was stronger at predicting outcome than the WHO grade II/III classification (P≤2.8E-10 versus 0.018). When compared to other prognosis factors (histological subtype and genetic abnormalities) in a multivariate analysis, the 22-gene predictor remained significantly associated with overall survival. Early prediction of high risk patients (3% of WHO grade II), and low risk patients (29% of WHO grade III) in clinical routine will allow the development of more appropriate follow-up and treatments.

Figure 1. Gene expression heatmap and overall survival of WHO grade II/III glioma patients.

Map of the gene expression levels from the 27-gene list used to generate a classification clearly identifying a high risk cluster containing most of the deceased patients of the training cohort.

Figure 2. Incidence on overall survival and comparison of WHO grade II/III versus gene expression-based PAM predictor classification methods.

The WHO classification (dotted lines) and the gene expression-based clusters or predicted classes (solid lines) were compared in a Kaplan-Meier analysis of training (A) and validation (B, C) cohorts. For life expectancy comparison, the Kaplan-Meier curves for overall survival were superimposed for training, validation and mixed cohorts (D, E). A parametric regression model of the overall survival of mixed cohorts was superimposed assuming a Weibull-distributed fit.