Identification of Relevant Genes from Microarray Experiments based on Partial Least Squares Weights: Application to Cancer Genomics

Abstract

In microarray genomics expression data obtained from hybridization of different cancer tissue samples or samples from cancer and normal cellular conditions, it is of interest to identify differentially expressed genes for follow-up validation studies. One approach to the analysis of genomics expression data is to first reduce the dimensionality using partial least squares (PLS), which has been useful in various cancer microarray data applications (Nguyen and Rocke, Bioinformatics 18:39–50, 2002a). PLS involves reducing the dimensionality of the gene expression data matrix by taking linear combinations of the genes/predictors (referred to as PLS components). However, the weights assigned to each gene and at each dimension are nonlinear functions of both the genes and outcome/response variable, making analytical studies difficult. In this paper, we propose a new measure for identifying relevant genes based on PLS weights called random augmented variance influence on projection (RA-VIP). We compare the relative performance of RA-VIP in terms of its sensitivity and specificity for identifying truly informative (differentially expressed) genes to two previously suggested heuristic measures, both based on PLS weights, namely the variable influence on the projection (VIP) and the PLS regression B-coefficient (denoted B-PLS). The methods are compared using simulation studies. We further illustrate the proposed RA-VIP measure on two microarray cancer genomics data sets involving acute leukemia samples and colon cancer and normal tissues.