S1.03 * GLIOBLASTOMA SUBTYPES DEFINED BY QUANTITATIVE IMAGING MAP TO DIFFERENT CANONICAL SIGNALING PATHWAYS

Abstract

We sought to discover subtypes of glioblastoma (GBM) defined by quantitative imaging features and to identify their canonical signaling pathways. Preoperative MR imaging from 121 Stanford patients with de novo, focal, unilateral GBM were analyzed. Two board-certified neuroradiologists and a neurosurgeon reached consensus in delineating Regions-of-Interest (ROIs) around areas of enhancement in each T1 post-contrast MR. We extracted 138 quantitative image features representing signal intensity and morphology of each lesion. We then used Consensus Clustering to define subtypes in the Stanford group. Next, Prediction Analysis for Microarrays (PAM) and the In-Group Proportion (IGP) statistic were used to validate the reproducibility and robustness of our subtype classification in a second cohort of subjects from The Cancer Genome Atlas (TCGA, n = 145). Finally, to map imaging subtypes to particular molecular signaling pathways, we performed Significance Analysis of Microarrays (SAM) on pathway activity estimates derived from analysis of TCGA tumor copy number and RNA sequencing data with the PARADIGM algorithm. Consensus Clustering analysis of the training set identified three clusters whose validity was confirmed in the TCGA cohort (p < 0.0001, p< 0.0001, p = 0.015). Cluster sizes for the TCGA validation cohort (i.e. 45%, 21% and 35%) were similar to those in the Stanford training set (57%, 17% and 26%). Cluster 1 was associated with 49 overexpressed pathways (FDR < 5%) including Wnt, hypoxia, apoptosis, cell proliferation, and angiogenesis signaling pathways while Cluster 2 was characterized by downregulation of these pathways (FDR < 5%). No pathway was significantly associated with Cluster 3. We identified three distinct, robust clusters of focal, unilateral GBM defined by quantitative image signatures. Subtypes identified in a homogeneous Stanford training set were validated in the heterogeneous multi-institutional TCGA data set. Two of our three clusters had significant associations with canonical signaling pathways.