Global Gene Coexpression Analysis Identifies a Poor Prognosis Signature across Tumor Types
AC Ladd, CI Dumur, CN Powers, DS Wilkinson, CT Garrett. Virginia Commonwealth University, Richmond, VA
Background: Global gene expression data are currently being used in attempts to develop cancer survival prediction tools. A predictor with power across tumor types has utility especially for use with uncommon cancers for which gene expression data sets are small and therefore molecular characterization is incomplete. Standard biostatistical methods build predictors from comparisons between two or more phenotypic classes while a novel approach, utilized here, focuses on identifying sets of genes coexpressed across all samples.
Design: Microarray gene expression data (Affymetrix GeneChips) from 136 tumors from 15 sites of origin were used to calculate coexpression measures. Two biologically relevant modules of highly coexpressed genes (mitosis and immune response), were detected in our data set. A coexpression signature comprised of the 186 genes in the two modules was tested for survival prediction against the entire set of tumors. Using Cox proportional hazard regression analysis methods (with gene expression measures as variables), a prognostic index was computed for each case which was then used to assign it to a high or low risk group. Leave-One-Out-Cross-Validation methods were used to ensure unbiased risk group assignment. Kaplan-Meier (KM) survival curves were then generated for each group.
Results: Risk prediction based on these methods was statistically significant (p=0.01). Interestingly, there was a high level of agreement between the genes in the coexpression signature and two other signatures based on undifferentiated tumors (one developed from our data set and another previously published by Chinnaiyan et al, PNAS, 2004), although KM curves generated from these signatures showed no significant differences between the two groups (p=0.12 and p=0.18, respectively).
Conclusions: While mitosis and immune response are not novel pathways to be uncovered in cancer, the coexpression analysis performed here refined the large list of genes in these pathways to those most likely contributing to their disruption. Therefore, not only are these types of methods of potential use in cancer risk assessment, but they may be useful in identifying new targets for therapeutic intervention.
Category: Special Category for 2009 - Pan-genomic/Pan-proteomic approaches to Cancer
Tuesday, March 10, 2009 9:30 AM
Poster Session III # 244, Tuesday Morning