A multi-objective optimization algorithm for gene selection and classification in cancer study

In recent years, non-clinical diagnostics and predictive modeling in statistical genomics have gained increased attention, particularly in the analysis of microarray gene expression data. One of the key advantages of using microarray data is its ability to analyze the expression levels of thousands of genes simultaneously. A key challenge remains the efficient identification of a small subset of informative genes that are statistically correlated with specific groups of Messenger Ribonucleic Acid (mRNA) tissue samples, followed by a meaningful biological interpretation of the findings. This paper introduces a two-stage hybrid multi-objective optimization (MOO) algorithm for feature selection and classification of mRNA samples into their respective groups. The proposed method enhances the classification performance of Support Vector Machines (SVMs) by framing gene selection as a MOO problem. Initially, a filter method-using either the t-test or F-test-was employed to eliminate noisy genes, which are prevalent in microarray gene expression data. Subsequently, the genes selected by the filter methods were further refined through a MOO approach, applying Pareto optimality criteria to identify a minimal yet optimal gene subset. The effectiveness of the proposed method was evaluated using both simulated and published high-dimensional microarray datasets, considering out-of-bag (OOB) accuracy, misclassification error rates, and several other performance metrics. The results demonstrated that the proposed method is robust, achieving high prediction accuracy with minimal gene subsets. Furthermore, it outperformed existing methods in its ability to select a small number of gene biomarkers that are strongly correlated with the biological response class.