Improved Identification of Differentially Expressed Genes Using Pareto Set based Pruning

Identification of differentially expressed genes (DEGs) from microarray datasets is one of the most important analyses for microarray data mining. Popular algorithms such as statistical t-test, fold change, and rank product can be improved by considering other features of differentially expressed genes. We proposed a parameter-free nondominated Pareto set based gene pruning algorithm for pruning non-differentially expressed genes before applying standard DEG identification algorithms. All genes are mapped to a feature space composed of average differences of gene expression and average expression levels and it is observed that differentially expressed genes tend to be located in boundary regions. Experiments on 17 Gene Omnibus Database (GEO) datasets showed that Pareto gene pruning can significantly improve popular algorithms such as t-test, rank product, and fold change in terms of prediction accuracy and AUC values with improvements ranging from 11% to 50% in terms of the number of identified true DEGs.