Developing a prognostic gene expression biomarker for re-irradiation in recurrent glioblastoma

Introduction: Glioblastoma is the most common primary, malignant brain tumor. Despite aggressive treatment, the prognosis remains poor, and recurrence is nearly universal. Radiotherapy improves overall survival (OS) in patients with newly diagnosed glioblastoma (ndGBM). Re-irradiation (reRT) improves progression-free survival (PFS) when added to bevacizumab in recurrent GBM (rGBM), but an OS advantage in unselected populations has not been demonstrated. Here we test the hypothesis that gene expression profiling can identify glioblastomas with better prognosis after re-irradiation delivered at the time of disease recurrence. Methods: A retrospective cohort of 98 IDH-wildtype, CNS WHO grade 4 GBM tissue samples (n=43 ndGBM, n=55 rGBM) from 83 patients who ultimately underwent reRT at the time of any tumor recurrence was analyzed using a barcode-based RNA hybridization platform with a custom 291-gene panel representing pathways underlying GBM growth and therapeutic response. Univariate analysis identified 60 genes associated with outcomes after reRT. A ridge regression model was trained using gene expression data to predict OS after reRT normalized to OS from diagnosis. Model outputs were mapped to reRT scores, which were dichotomized (high versus low) using the maximally selected rank statistic. Results: The median time between RT courses was 19.9 mo. Median OS after reRT was 9.74 mo. Median OS after reRT for tumors with high versus low reRT scores was 12.5 versus 8 mo (p<0.001). The hazard ratio for death after reRT for tumors with low reRT scores compared to tumors with high reRT scores was 2.47 ([95% CI 1.47-4.16], p<0.001). Using only gene expression data from ndGBM, the median OS after reRT for tumors with high versus low reRT scores was 10.5 versus 5.7 mo (p=0.022; HR for low reRT scores: 2.20, [95% CI 1.09-4.45], p=0.027). Using only gene expression data from rGBM, the median OS after reRT for tumors with high versus low reRT scores was 16.82 versus 9.18 mo (p=0.0014; HR for low reRT scores: 3.55, [95% CI 1.55-8.17], p=0.0028). There were no differences in MGMT promoter methylation status or demographic characteristics between tumors with low versus high reRT scores. ReRT scores were calculated for a second cohort of 10 patients with spatially sampled ndGBM (n=6-19 samples/patient). By a mean-rating, one-way random effects, absolute agreement model, the intraclass correlation coefficient estimate for the reRT score was 0.89 ([95% CI 0.76-0.97], p<0.001), which indicates high score concordance for regionally distinct samples from within individual tumors. Conclusions: Gene expression data from ndGBM or rGBM should be considered as a stratification variable for reRT clinical trials. The model reported here requires validation in additional cohorts to determine its prognostic value.