Most glioblastoma studies incorporate the layer of tumor molecular subtype based on the four-subtype classification system proposed in 2010. Nevertheless, there is no universally recognized and convenient tool for glioblastoma molecular subtyping, and each study applies a different set of markers and/or approaches that cause inconsistencies in data comparability and reproducibility between studies. Thus, this study aimed to create an applicable user-friendly tool for glioblastoma classification, with high accuracy, while using a significantly smaller number of variables. The study incorporated a TCGA microarray, sequencing datasets, and an independent cohort of 56 glioblastomas (LUHS cohort). The models were constructed by applying the Agilent G4502 dataset, and they were tested using the Affymetrix HG-U133a and Illumina Hiseq cohorts, as well as the LUHS cases. Two classification models were constructed by applying a logistic regression classification algorithm, based on the mRNA levels of twenty selected genes. The classifiers were translated to a RT-qPCR assay and validated in an independent cohort of 56 glioblastomas. The classification accuracy of the 20-gene and 5-gene classifiers varied between 90.7-91% and 85.9-87.7%, respectively. With this work, we propose a cost-efficient three-class (classical, mesenchymal, and proneural) tool for glioblastoma molecular classification based on the mRNA analysis of only 5-20 genes, and we provide the basic information for classification performance starting from the wet-lab stage. We hope that the proposed classification tool will enable data comparability between different research groups.
Keywords: biomarkers; classical; classification tool; glioblastoma; mesenchymal; molecular subtyping; proneural.