Joint Covariate Detection on Expression Profiles for Selecting Prognostic miRNAs in Glioblastoma

Abstract

An important application of expression profiles is to stratify patients into high-risk and low-risk groups using limited but key covariates associated with survival outcomes. Prior to that, variables considered to be associated with survival outcomes are selected. A combination of single variables, each of which is significantly related to survival outcomes, is always regarded to be candidates for posterior patient stratification. Instead of individually significant variables, a combination that contains not only significant but also insignificant variables is supposed to be concentrated on. By means of bottom-up enumeration on each pair of variables, we propose a joint covariate detection strategy to select candidates that not only correspond to close association with survival outcomes but also help to make a clear stratification of patients. Experimental results on a publicly available dataset of glioblastoma multiforme indicate that the selected pair composed of an individually significant and an insignificant miRNA keeps a better performance than the combination of significant single variables. The selected miRNA pair is ultimately regarded to be associated with the prognosis of glioblastoma multiforme by further pathway analysis.