Metabolic/Proteomic Signature Defines Two Glioblastoma Subtypes With Different Clinical Outcome

Abstract

Glioblastoma (GBM) is one of the deadliest human cancers. Because of the extremely unfavorable prognosis of GBM, it is important to develop more effective diagnostic and therapeutic strategies based on biologically and clinically relevant subclassification systems. Analyzing a collection of seventeen patient-derived glioblastoma stem-like cells (GSCs) by gene expression profiling, NMR spectroscopy and signal transduction pathway activation, we identified two GSC clusters, one characterized by a pro-neural-like phenotype and the other showing a mesenchymal-like phenotype. Evaluating the levels of proteins differentially expressed by the two GSC clusters in the TCGA GBM sample collection, we found that SRC activation is associated with a GBM subgroup showing better prognosis whereas activation of RPS6, an effector of mTOR pathway, identifies a subgroup with a worse prognosis. The two clusters are also differentiated by NMR spectroscopy profiles suggesting a potential prognostic stratification based on metabolic evaluation. Our data show that the metabolic/proteomic profile of GSCs is informative of the genomic/proteomic GBM landscape, which differs among tumor subtypes and is associated with clinical outcome.

Figure 1. Gene expression analysis of 17 patient-derived GSCs identifies two distinct clusters.

(A) Hierarchical clustering of GSCs using the top 1,000 most variable genes/probes in our gene expression datasets. (B) Hierarchical clustering of GSCs using a subset of the genes/probes in A, selected among the highest-score Gene Sets resulting from GSEA on genes/probes differentially regulated between Cluster 1 and Cluster 2. The colored left-side bar annotates the correspondent Gene Sets clusters. (C) Hierarchical clustering of the top 1,000 most variable common genes found in our GSC (yellow and brown rectangles) samples and in Schulte GSr/GSf (grey and black rectangles) samples. Heatmaps throughout the figure display differences in expression levels in log2 scale (red + 3, black = 0, green = −3).

Figure 2. 1D¹H NMR distinguishes two metabolic profiles of GSCs.

(A) Unsupervised clustering of GSCs based on the levels of the analyzed metabolites. Samples are annotated with their matching gene expression clusters (colored boxes on the left) with the exception of sample #112. (B) GSr-/GSf-like ratios of NMR signal intensity means were converted to log2 and expressed as fold change (FC). Metabolites displaying a statistically significant (p < 0.05, Student’s t-test) difference between GSr- and GSf-like are indicated by asterisks. (C) Boxplots of raw NMR signal intensities of metabolites measured in GSf-like (cluster 1, yellow) and GSr-like (cluster 2, brown) samples.

Figure 3. Phosphoproteomic analysis supports genomic and metabolic clustering of GSCs.

Boxplots of standardized levels of selected RPPA analytes displaying a statistically significant (p < 0.05, Wilcoxon rank sum test) difference between GSCs of GSr- or GSf-like gene expression cluster. Based on the absolute difference between RPPA levels measured in GSr- and GSf-like GSCs, endpoints were divided in two groups, i.e. GSr-like (brown, high in GSr-like) and GSf-like (yellow, high in GSf-like), respectively. RPPA endpoints are sorted in descending order starting from the one showing the highest levels in GSr-like GSCs. Each antibody is annotated (right side) with the corresponding p value. Outliers are shown as circles and names of the major pathways defining GSr- or GSf-like cells are reported for both clusters.

Figure 4. RPPA data from TCGA GBM samples.

Boxplots of standardized levels of RPPA analytes measured in P (proneural, n = 41, pink) and M (mesemchymal, n = 29, red) samples selected from the full cohort (n = 214) of GBM tumors that was subjected to RPPA analysis within the TCGA consortium. The RPPA endpoints reported here were selected based on the presence, in the TCGA data, of analytes matching the GSf- (A, yellow box), or GSr-like (B, brown box) lists defined by our RPPA analysis (see methods sections for details). P values from unpaired t test are reported on top of each endpoint plot.

Figure 5. GSf- and GSr-like RPPA endpoints ability to predict TCGA patient survival.

Table plots of hazard ratios (HR) for expression levels of individual RPPA analytes, selected as distinctive of GSf- (A) and GSr-like (B) groups. For each RPPA endpoint, a Cox regression model was applied using the intensity levels (z score) and the subtyping as covariates. In order to generate a multi-gene prognostic index (mPI), the full selection of either GSf- (C) or GSr-like (D) RPPA analytes were included as covariates in a Cox model and table plots of corresponding hazard ratios are shown. Confidence intervals (95% CI) are reported for all HR together with logrank test p values and, for A and B, p values for the RPPA levels. The cohort of patients used for performing Cox survival analyses includes all TCGA patients with known Verhaak subtype (n = 111).

Figure 6. Combined levels of phospho-SRC and phospho-RPS6 predict survival of TCGA GBM patients.

GSf- (yellow) or GSr-like (brown) phenotypes were assigned to patients based on the combined expression of either SRC and RPS6 pS235-36 (A) or SRC pY527 and RPS6 pS235-36 (B) see methods section for details. Patients fulfilling the SRC/RPS6 pS235-36 median expression cut-off criterion were 58 per group while for the SRC pY527/RPS6 pS235-36 were 53 per group out of a total 211 TCGA patients analyzed. Log-rank p values, hazard ratio (with 95% confidence interval) and relative p values for the group variable are reported inside each plot.