A non-parametric and information theoretic algorithm for identifying differentially expressed genes in multiclass RNA-seq samples

Abstract

Gene expression in the whole cell can be routinely measured by microarray technologies or recently by using sequencing technologies. Using these technologies, identifying Differentially Expressed Genes (DEGs) among multiple phenotypes is one of the most important tasks in biology. Thus many methods for detecting DEGs between two groups has been developed. For example, T-test and relative entropy are used for detecting the difference between two probability distributions. When more than two phenotypes are considered, these methods are not applicable and other methods such as ANOVA F-test and Kruskal-Wallis are used for finding DEGs in the multiclass data. However, ANOVA F-test assumes a normal distribution and it is not designed to identify DEGs where gene are expressed distinctively in each of phenotypes. Kruskal-Wallis method, a non-parametric method, is more robust but sensitive to outliers. This thesis proposes a non-parametric and information theoretical approach for identifying DEGs. Our method can identify DEGs in the multiple class data and is less sensitive to outliers. In extensive experiments with simulated and real data, our method outperformed existing tools. In addition, a web service is implemented for the analysis of multi-class data: http://biohealth.snu.ac.kr/software/degselection