Generic relative relations in hierarchical gene expression data classification

Abstract

Relative Expression Analysis (RXA) plays an important role in biomarker discovery and disease prediction from gene expression profiles. It deliberately ignores raw data values and investigates only the relative ordering relationships between a small group of genes. The classifiers constituted on that concept are therefore robust to small data perturbations and normalization procedures, but above all, they are easy to interpret and analyze. In this paper, we propose a novel globally induced decision tree in which node splits are based on the RXA methodology. We have extended a simple ordering with a more generic concept that also explores fractional relative relations between the genes. To face up to the newly arisen computational complexity, we have replaced the typical brute force approach with an evolutionary algorithm. As this was not enough, we boosted our solution with the OpenMP parallelization, local search components calculated on the GPU and embedded ranking of genes to improve the evolutionary convergence. This way we managed to explore in a reasonable time a much larger solution space and search for more complex but still comprehensible gene-gene interactions. An empirical investigation carried out on 8 cancer-related datasets shows the potential of the proposed algorithm not only in the context of accuracy improvement but also in finding biologically meaningful patterns.