A hybrid bat and grey wolf optimizer for gene selection in cancer classification

DNA microarray is a technique in which a chip containing numerous DNA codes is used for the expression estimation of an extensive number of genes simultaneously. These genes are arranged in a table or data format. The gene expression data can be employed in pattern recognition algorithms to differentiate between samples obtained from healthy individuals and those with cancer. However, recognizing biomarkers' patterns from gene selection data is considered challenging because of its huge dimensionality and the presence of noisy, irrelevant, and unwanted genes, leading to mislearning process and, thus, declining in the classification performance. Therefore, in this paper, an intelligent gene selection approach is proposed on the basis of robust minimum redundancy maximum relevancy as the filter and hybrid improved bat algorithm (BA) with grey wolf optimizer (GWO) (BA-GWO). The BA-GWO is introduced to determinate a limited number of biomarker genes that significantly enhance the classification performance. In this approach, the k-nearest neighbor algorithm was employed for the classification task. The proposed BA-GWO is mainly introduced to improve the BA search agents' performance in searching for the best candidate gene subset that carries the biomarkers for cancer classification. Furthermore, the BA-GWO is designed to enhance both exploitation and exploration capabilities while ensuring a balanced approach and preventing stagnation in local optima. The primary function of this proposed approach is to enhance the solutions acquired through the BA by utilizing them as the initial population for the GWO. The proposed approach is evaluated using ten widely recognized microarray datasets in the experimental stage, including CNS, Colon, Leukemia 3c, Leukemia 4c, Leukemia, Lung Cancer, Lymphoma, MLL, Ovarian, and SRBCT. The performance of the hybridization of BA and GWO, as well as recent and base optimization algorithms, is evaluated. Afterward, the hybrid versions are compared with their individual optimization algorithms. Moreover, the hybridization algorithms are compared with each other. For further validation, the proposed approach performance is compared with twelve state-of-the-art comparative methods in terms of accuracy and the selected genes. The findings indicate that the proposed approach yields superior outcomes in two out of eight datasets, while also delivering highly competitive results in the remaining datasets.