A hybrid filter-wrapper gene selection method for cancer classification

The advent of DNA microarray technology has paved the way to providing increased opportunities to the molecular biologists to analyze the expression level of thousands of genes (features) in one experiment. The gene expression level provides the possibility of diagnosing various diseases such as cancer. In this regard, several computational techniques such as pattern classification approaches can be applied. However, the existence of a huge quantity of genes and very few patients' samples available hinders the classifier or machine learning techniques from producing accurate classification results. Most of these genes are irrelevant and redundant, which may deteriorate the classification performance. Therefore, gene selection is needed to select the most relevant genes. This paper proposes hybrid filter-wrapper gene selection method using Minimum Redundancy Maximum Relevancy (MRMR) as the filter approach and flower pollination algorithm (FPA) as the wrapper approach. MRMR was used to find the most important genes from all genes in the gene expression data, and FPA is employed in order to locate the most informative gene subset from the reduce set that obtained by MRMR. To test the accuracy and performance of the study's proposed method, extensive experiments are conducted and three microarray datasets are used. They include Colon, Breast, and Ovarian. A similar procedure has been performed on the Genetic algorithm (GA) in comparison with the proposed method (MRMR-FPA) in this study. The results concluded that the MRMR-FPA can be used as an alternative method to address the gene selection problem.