A simplified integrated molecular and immunohistochemistry-based algorithm allows high accuracy prediction of glioblastoma transcriptional subtypes.

Glioblastomas (GBM) can be classified into three major transcriptional subgroups (proneural, mesenchymal, classical), underlying different molecular alterations, prognosis, and response to therapy. However, transcriptional analysis is not routinely feasible and assessment of a simplified method for glioblastoma subclassification is required. We propose an integrated molecular and immunohistochemical approach aimed at identifying GBM subtypes in routine paraffin-embedded material. RNA-sequencing analysis was performed on representative samples (n = 51) by means of a “glioblastoma transcriptional subtypes (GliTS) redux” custom gene signature including a restricted number (n = 90) of upregulated genes validated on the TCGA dataset. With this dataset, immunohistochemical profiles, based on expression of a restricted panel of gene classifiers, were integrated by a machine-learning approach to generate a GliTS based on protein quantification that allowed an efficient GliTS assignment when applied to an extended cohort (n = 197). GliTS redux maintained high levels of correspondence with the original GliTS classification using the TCGA dataset. The machine-learning approach designed an immunohistochemical (IHC)-based classification, whose concordance was 79.5% with the transcriptional- based classification, and reached 90% for the mesenchymal subgroup. Distribution and survival of GliTS were in line with reported data, with the mesenchymal subgroup given the worst prognosis. Notably, the algorithm allowed the identification of cases with comparable probability to be assigned to different GliTS, thus falling within overlapping regions and reflecting an extreme heterogeneous phenotype that mirrors the underlying genetic and biological tumor heterogeneity. Indeed, while mesenchymal and classical subgroups were well segregated, the proneural types frequently showed a mixed proneural/classical phenotype, predicted as proneural by the algorithm, but with comparable probability of being assigned to the classical subtype. These cases, characterized by concomitant high expression of EGFR and proneural biomarkers, showed lower survival. Collectively, these data indicate that a restricted panel of highly sensitive immunohistochemical markers can efficiently predict GliTS with high accuracy and significant association with different clinical outcomes. The authors developed a novel simplified assay for glioblastoma transcriptional classification on formalin-fixed-paraffin-embedded tissue samples. On such dataset, immunohistochemical profiles, based on expression of a restricted panel of gene classifiers, were integrated by machine learning approach to generate a glioblastoma transcriptional signature based on protein quantification that allowed to efficiently assign transcriptional subgroups to an extended cohort. Correlations with both histopathological features and clinical outcome have been also performed.