Abstract Glioblastoma multiforme (GBM) is the most common primary malignant brain tumor in adults and also the most aggressive, with a median survival time of approximately 1 year. Differences in age and pathology have long been recognized as key factors in determining prognosis, and, more recently, comprehensive mRNA profiling has been undertaken to establish the molecular basis underlying prognosis. Though early mRNA expression-based gene sets represented unbiased queries into molecular differences in gliomas, they only uncovered differences in survival ex post facto1, 2. Newer gene sets capitalize upon advances in technology, yet they suffer for their unwieldy size and ambiguous biological meaning3. Though the myriad gene sets proposed share little overlap with each other, recent work has demonstrated that the use of interaction networks can consolidate these differences. To determine the molecular basis of prognostic differences in GBM directly, we employed an existing computational framework, CRANE, to exhaustively search protein interaction networks and identify molecular signatures distinguishing short-term (<225 days) and long-term (>635 days) survivors of GBM in The Cancer Genome Atlas gene expression microarray dataset. Using CRANE, we identified a 50 gene subnetwork signature of survival in GBM, which we validated in an independent dataset of 91 patients (80% classification accuracy). We further hypothesized that the use of interaction networks consolidates variation in mRNA expression into more stable features at higher phenotypic levels. To test this hypothesis, we examined protein expression levels of the 50 network targets in an independent cohort of 16 patients (10 patients: survival<9 months; 6 patients: survival>18 months) using label-free proteomics. Of the 17 network targets identified in common with our 50 subnetwork signature genes, 8 were significantly differentially expressed between short-term and long-term survivors, and expression levels of 3 of the proteins (STAT3, Calnexin, and HSPA9) predicted phenotype (LTS vs. STS) with 75% or more accuracy (12 or 13 correct out of 16). Our conclusion is that data integration using interaction networks can drive discovery of a stable subnetwork signature exhibiting clear differences in protein expression. This manageable subset of proteomic targets will prove invaluable in the design of cost-effective clinical assays (e.g. IHC) suitable for retrospective analysis from FFPE tumor tissue in existing biobanks, as well as for prospective studies predicting patient survival - a much desired endpoint for clinicians and patients alike. Citation Format: {Authors}. {Abstract title} [abstract]. In: Proceedings of the 103rd Annual Meeting of the American Association for Cancer Research; 2012 Mar 31-Apr 4; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2012;72(8 Suppl):Abstract nr 4936. doi:1538-7445.AM2012-4936