Given the importance of signaling in the biology of gliomas, dividing these tumors into subsets by the pattern of coordinate signaling pathway activation may have practical implications for choice of therapies and for interpretation of patient responses in existing clinical trials. In order to clarify the net activation of signaling pathways we used a targeted proteomic analysis to determine not only the levels but the posttranslational modifications associated with signaling activity. The intrinsic cellular heterogeneity of gliomas is masked by the methods used in this study since both the proteomic and TCGA genomic analyses are performed on homogenized tissue, blending the characteristics of the cells together. Additionally, the small sample size and selection of proteins in this study limits the statistical power to define protein correlations and sample assignments. Nonetheless, we find common features defining three basic groups by both genomic and protein analysis, illustrating the high complementarity between protein signaling activity, transcriptomal signature and genomic alteration in GBM. The observed enrichment of EGFR
genomic alterations with transcriptome pattern could mean that signaling activity directly influences the transcriptome and/or that both signaling and the transcriptome patterns are part of a common underlying phenotype. Comparing unsupervised and supervised clustering results, it is likely that only a portion of the transcriptomal features distinguishing unsupervised clusters are associated with signaling, either directly or indirectly. In fact, unsupervised clustering identifies four clusters and broader phenotypic and genetic differences distinguishing these four transcriptomal groups have been reported.
Analysis of the downstream signaling components of the PDGF proteomic group revealed generally lower PI3K/Akt activity than in the EGFR glioma group although S6 phosphorylation was paradoxically high. Histologic analysis of the tumors provided an explanation by demonstrating that strong pS6 immunopositivity was localized in reactive astrocytes rather than tumor cells per se (data not shown), concordant with recent observations of mTOR activation in reactive astrocytes under experimental conditions of injury.
These cells were more common in PDGF-class tumors. We found a trend for treated tumors to be in the PDGF proteomic class and PDGFB levels were significantly higher in treated compared to untreated tumors. It is possible that some of the features of the PDGF signaling pattern are influenced by prior treatment though it is unlikely that this accounts for the genotypic differences in this proteomic tumor class, such as the paucity of EGFR amplification, chr7 gain and Ink4a/ARF locus deletion. Comparison with treated samples in TCGA is complicated by the fact that the current dataset contains few treated cases and of these, many are secondary GBM which would arguably be assigned to the PDGFRA co-cluster by their common Proneural signature.
The histology of the TCGA samples was uniformly GBM, but this is a histologically heterogeneous tumor type. A priori, it is possible that the transcriptomal classes identified in this analysis could be related to tumor sampling or microenvironment. However, this does not appear to be the case since well-defined genetic lesions are enriched in specific tumor classes and there is no evidence for regional localization of mutations in glioma as a general phenomenon. Although clinical and pathologic data are limited and a more detailed review of this information is underway, there appeared little clinical or histological differences between the three groups identified in this analysis. PDGFRA-cocluster tumors in TCGA occurred in younger patients, and there were small but significant differences in the amount of associated necrosis and inflammatory cells (data not shown).
It remains to be established whether ligand-driven PDGF signaling is common among tumors in the transcriptomal PDGFRA-cocluster and whether this is functionally important. We have shown that PDGFB ligand levels are highly variable in GBM, are associated with receptor activation, and are not correlated with mRNA expression. The PDGFRA-cocluster transcriptomal class shares features with the Proneural group of gliomas identified by Phillips et al using transcription analysis, and is characterized by genes expressed during normal cortical oligodendrocyte development such as olig2, Sox2 and doublecortin and signaling pathways involved in that process as well, such as PDGF and SHH. While the PDGFRA-cocluster group is enriched for Proneural signature genes, it is important to note that the original signature was derived from a dataset of mixed histologies and the analysis designed specifically to resolve a prognostic signature. Therefore the exact relationship between our PDGF proteomic class, the PDGFRA co-cluster, and gliomas harboring Proneural signature is unclear and will need to be further investigated.
Although PDGFRA amplification predominated in the PDGFRA-cocluster transcriptomal group, a full 15% showed amplified EGFR and another 15% showed amplified MET. From the prevalence of PDGF signaling we found at the protein level one might hypothesize the existence of concurrent PDGF signaling in EGFR- and MET- amplified tumors in this class. In fact, we found two such tumors in our proteomic analysis: one EGFR- and one MET-amplified, both with high levels of PDGFB, phosphorylation of PDGFRβ and an overall signaling pattern matching the PDGF proteomic class. It is unclear in these cases whether the level of PDGF pathway activation is functionally important, perhaps in a subpopulation of cells. It is notable that 6 tumors in TCGA show focal amplification of both PDGFRA and another RTK: four cases sharing focal co-amplification of EGFR, and two cases sharing focal co-amplifications of PDGFRA and MET.
In conclusion, our findings support a division of GBMs into three classes according to patterns of signal transduction pathway activation. These patterns reflect, in part, mutually exclusive signaling involving EGFR, PDGF RTK activation or NF1 silencing. Both the transcriptomal and proteomic classes were imperfectly related to genotype, suggesting that molecular assays used in patient stratification and clinical trial analysis should include measures of PDGF ligand and receptor phosphorylation as well as NF1 expression. Notch signaling was prominently associated with the EGFR class at the protein level, an observation which was not predicted by mRNA expression levels of Notch pathway members in EGFR-altered tumors from TCGA. Whether one or more non-EGF/PDGF RTKs are contributing to NF1 tumors is uncertain, but the finding that NF1-silenced tumors show elevated MET, HGF and IRS1 at the transcriptomal level and validation of IRS1 at the protein level suggest IGF and or MET signaling may be contributory. Further refinement of GBM subclasses will likely come from direct investigations of these and other signaling proteins, as well as investigation of newly described recurrent mutations in GBM such as ERBB2 and IDH1. The current study provides an initial architecture for such subclasses and suggests the potential for class-directed therapies.