Here, we show that genomic profiling defined four subtypes of tumors with a common morphologic diagnosis of GBM. The reproducibility of this classification was demonstrated in an independent validation set suggesting that it is highly unlikely that these GBM tumor subtypes are a spurious finding due to technical artifact, chance or bias in TCGA sample qualification criteria. The importance of detecting these subtypes lies in the different therapeutic approaches that different subtypes may require. Furthermore, it is possible that GBMs in specific subtypes develop as the result of different etiologies or different cells of origin. Studying GBMs in the light of subtypes therefore may accelerate our understanding of GBM pathology. A larger sample set might describe additional subtypes for which we lack the power to detect. Additionally, we provide the community with the means to identify the tumor subtypes prospectively [http://tcga-data. nci.nih.gov/docs/publications/gbm_exp/
In addition to validating the subtype in other human GBM datasets, we identified gene expression patterns of xenografts highly comparable to Proneural, Classical, and Mesenchymal tumors. However, identification of comparable cell line models was not as easily achievable (data not shown). For example, there is a relative lack of EGFR amplification and EGFRvIII mutants in cell lines models, potentially lost or selected against during the culturing process. The identification of valid subtype counterparts in xenografts represents an important contribution toward our ability of studying GBM subtypes, in particular for modeling and predicting therapeutic response.
One of the most important aspects of this work is the unprecedented ability to examine molecularly-defined tumor subtypes for correlations with both genome-wide DNA copy number events and sequence-based mutation detection for 601 genes. While a mechanistic explanation of subtype is beyond the scope of this manuscript, our cross-platform analyses highlight a number of important characteristics of each subtype and hint at cell of origin. For example, the Proneural subtype was associated with younger age, PDGFRA
mutations, all of which have previously been associated with secondary GBM (Arjona et al., 2006
; Furnari et al., 2007
; Kleihues and Ohgaki, 1999
; Watanabe et al., 1996
; Yan et al., 2009
). Most known secondary GBMs classified as Proneural (). In a previous study, most grade III gliomas as well as 75% of lower grade gliomas from the validation sets classified as Proneural or Neural (Phillips et al., 2006
). While it is outside the scope of the current manuscript to establish the etiology of the classes, the Proneural TCGA class was enriched both for secondary GBM established by prior lower-grade histology and for IDH1
mutations which are known to be prevalent in secondary GBM. Other tumors in this class which appear to be clinically de novo (primary) may share common pathogenesis with secondary GBM and might arise from lower grade lesions which are clinically silent. Alternatively, Proneural GBM tumors may arise from a progenitor or neural stem cell that can also give rise to oligodendrogliomas thereby sharing similar characteristics. High similarity with a purified oligodendrocytic signature and previous work identifying high expression of PDGFRA
in cells of the SVZ give credence to this hypothesis (Jackson et al., 2006
The identity of the Classical subtype is defined by the constellation of the most common genomic aberrations seen in GBM, with 93% of samples harboring chromosome 7 amplifications and 10 deletions, 95% showing EGFR amplification and 95% showing homozygous deletion spanning the Ink4a/ARF locus. This class also shows a distinct lack of additional abnormalities in TP53, NF1, PDGFRA or IDH1.
In the current study we also confirm the presence of a Mesenchymal subtype characterized by high expression of CHI3L1
(Phillips et al., 2006
). A striking characteristic of this class was the strong association with the recently reported high frequency of NF1
mutation/deletion and low levels of NF1
mRNA expression overall. Inherited NF1
mutations are associated with a variety of tumors, including neurofibromas, which reportedly have a Schwann cell-like origin (Zhu et al., 2002
). Although Schwann cells are not present in the central nervous system, the Mesenchymal class expresses Schwann cell markers such as the family S100A as well as microglial markers. The higher percentage necrosis and associated inflammation present in these samples is potentially linked to the mesenchymal phenotype through an expression signature including genes from wound healing and NF-κB signaling.
Samples in the Neural subtype are unequivocally GBMs by morphology by light microscopy and contain mutation and DNA copy number alterations. Their expression patterns are recognizable as the most similar to samples derived from normal brain tissue, and their signature is suggestive of a cell with a differentiated phenotype. This is confirmed by the association with neural, astrocytic and oligodendrocytic gene signatures.
Cellular organization and differentiation in the brain has been intensively investigated yet there is much to be discovered. It is therefore striking to find the clear relationships between subtypes of GBM and cellular lineages as demonstrated here (). It is possible that a common cell of origin, such as the previously proposed neural stem cell (Galli et al., 2004
), exists for all GBMs, and that the classes presented here result from distinct differentiation paths. However, the presence of precursor cells with self replicating ability in the brain, such as cells expressing stem cell markers and PDGFRA
(Jackson et al., 2006
) suggests that multiple stem cell-like populations exist. While there is a clear need for conclusive evidence supporting this hypothesis, it is at least striking to find the same genes as markers of two of the four classes lending support for a difference in cell of origin. This is further supported by the specific characteristics of the Mesenchymal and Neural class. Establishing the cell of origin of GBM is critical for establishing effective treatment regimens (Sanai et al., 2005
Given the set of characteristic subtype abnormalities, we deem it unlikely that patients transition between subtypes during different stages of their disease. This is substantiated by several samples in the Murat et al dataset, that did not switch between subtype after recurrence.
An association was observed between the Proneural subtype and age and a trend towards longer survival. Furthermore, our data suggest that Proneural samples do not have a survival advantage from aggressive treatment protocols. Importantly, a clear treatment effect was observed in the Classical and Mesenchymal subtypes. Profiling-based classification may therefore have highest clinical relevance in suggesting different therapeutic strategies. It appears that the simple classification into these four subtypes carries a rich set of associations for which there is no existing diagnostic test. We envision that the next generation of biomarker assays for GBM could include a molecular test for subtype and linked molecular genetics for key genetic events including NF1
amplification (i.e. genetic events that are best assayed on the DNA level) and MGMT
methylation status. Additionally, early evidence suggests that subclasses differ measurably by signal transduction pathways such that protein biomarkers might be easily measured (Brennan et al., 2009
). Future studies should further elucidate the intricate relationship between tumor subtypes, treatment sensitivity and MGMT
GBM is one of the most feared of all of human diseases both for its near uniformly fatal prognosis and associated loss of cognitive function as part of the disease process. For those facing the diagnosis there are few biomarkers of favorable prognosis and accordingly few therapies strongly influencing disease outcome. This comprehensive genomic- and genetic-based classification of GBM should lay the groundwork for an improved molecular understanding of GBM pathway signaling that could ultimately result in personalized therapies for groups of GBM patients.