Feature selection based on activation of signaling pathways applied for classification of samples in microarray studies

Abstract

This paper presents a new method of deriving features for sample classification based on massive throughput data such as microarray gene expression studies. The number of features in these studies is much bigger than the number of samples thus strong reduction of dimensionality is essential. Standard approaches attempt to select subsets of features (genes) realizing highest association with the target, and they tend to produce unstable and non-reproducible feature sets. The purpose of this work is to improve feature selection by using prior biological knowledge of potential relationships between features, available e.g., in signaling pathways databases. We first identify most activated pathways and then derive pathway-based features based on expression of the up- and down-regulated genes in the pathway. We demonstrate performance of this approach using real microarray data. © 2012 Springer-Verlag Berlin Heidelberg.