Abstract
Transcriptional profiling has classified glioblastoma (GBM) into four subtypes, termed proneural, neural, classical, and mesenchymal, with different genomic characteristics. However, we observed that the across platform concordance of 130 GBM with expression profiles generated using RNA sequencing as well as Affymetrix U133A microarrays rate, was only 76%, suggesting lack of robustness. Redefined gene signatures to only include upregulated genes improved the concordance of classification from 76% to 86%. Integrating subtypes with genomic alterations verified previously observed associations, such as significant MYC, CDK4, and PDGFRA alterations in proneural and EGFR abnormalities in the classical subtype. Interestingly, we found that GBMs classified as mesenchymal showed significant reduction in tumor purity and a related increase in the proportion of stromal cells. Since GBM have frequently been described as undergoing a proneural-to-mesenchymal transition, we asked the question how these tumor-associated normal cells relate to GBM recurrence, and whether these are increased relative to their primary original tumor. To address this we analyzed RNA profiles from 45 pairs of primary-recurrent GBM, and we found that 20 of the 45 tumors retained their original subtype after treatment, 7 switched from one subtype to another, and the remaining 18 showed features of multiple subtypes including their original class. Three recurrent GBMs exclusively expressed the mesenchymal signature and were found to have increased levels of microglia relative to their primary counterparts. This confirmed that tumor microenvironment, including stromal cells and microglia, may play an integral role in shaping the mesenchymal signature. In conclusion, we have deconvoluted the "mesenchymal" subtype and its transition from other subtypes. Our new gene signatures led to more robust and accurate subtypes which highlighted the contribution of the tumor microenvironment to gliomagenesis.