Abstract 4937: Factors in NSCLC stage and outcome: Integrated genome-wide study

Abstract

Abstract Genome wide data has transformed disease diagnostic and prognostic model development through the identification of molecular profiles. Enhanced molecular insight can be achieved by multiple -omic profile integration. Accordingly, we used a novel multistep integrative unsupervised approach, the network phenotyping strategy (NPS), for discovery of disease diagnostic/prognostic panels in resected NSCLC frozen tumor samples (n=81; 29 Stage IA/52 Stage IB). Whole genome gene expression and DNA methylation data was generated using Illumina BeadChips. In NPS step 1, the most common (>18% of the cohort) hypo- and hyper-methylated genes (lowest/highest β-score; n=31) were identified. In step 2, a subset of these genes was found to reside in networked multiple loci, co-protected against genetic variation by extreme (high n=10 or low n=6) incorporation energy costs, based on an exome entromic analysis. In step 3, maximal spanning tree reduction of a network of all 120 possible expression level relationships for these 16 genes, weighted by the correlation coefficients quantifying co-regulation, was used to identify the most informative sub-network, an 8-partite graph K8. For each partition, cases were dichotomized into groups where gene 1 was expressed at least 1.5x higher or lower relative to gene 2. In step 4, K8 cycle-decomposition revealed only 8 distinct gene expression patterns (C1-C8) representing nominally different molecular NSCLC subtypes. Characterization of individual tumor expression patterns was computed as the patient's difference vector (ΔC) from C1-C8. Stage IA tumor gene expression patterns (overexpression of all 16 genes) most closely matched C1 (true positive rates 0.8 - 0.93 and false positive rates 0.2 - 0.07, ROC 0.9 in 10-fold cross-validation). For stage IB, these 16 gene expression levels were lower and heterogeneous. In step 5, we generated OS prediction models. The ΔCs were used in an alternating decision tree algorithm. Two rule sets were found, one for the high gene expression Stage IA tumors and the second for Stage IB. In a 10-fold cross-validation, the log-rank test for equality of survivor functions resulted in optimal OS separation at 1650 days (p<0.0005). These results demonstrate that: 1) graph theory provides tools for handling complex data relationships without loss of analytic power; and 2) networks of genome loci co-protected against variant incorporation form a filter identifying the function in experimental -omic data. Here, the 16 gene panel we identified includes members of 3 common lung tumorigenesis pathways and/or were co-regulated by miRNAs independently found to be associated with prognosis in our cohort. The integrative NPS facilitated the epigenetic identification of a biomarker gene set, stage classification by overall expression levels and disease outcome prediction by the expression patterns captured by C1-C8. Supported in part by NIH 5P50 CA090440, P30 CA047904, UPMC Institutional Funds. 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 4937. doi:1538-7445.AM2012-4937