Exploring Cancer Biomarker Genes from Gene Expression Data via Natureinspired Multiobjective Optimization

The identification of high-dimensional cancer genesis particularly important for cancer detection and early prevention. Many studies have been proposed to address cancer sub-type diagnosis, however, these methods often suffer from low diagnostic power and poor generalization performance. In this paper, we investigate a TLBO-based multi-objective hybrid algorithm (MOTLBO) to optimize six objectives including the number of features, the accuracy, the precision, two entropy based measures: sensitivity, specificity and comprehensive index Kappa index. First, the data are dimensionalzed by a mutual information-based approach, and then we employ a fractional encoding mapping to select a subset of informative genes to optimize these target functions. Secondly, different TF values are used based on the roles in the evolutionary process to seek Pareto solutions. Finally, a perturbation-based local search method is proposed to explore neighborhood regions with sparse high-quality solutions. To demonstrate the effectiveness of the algorithm, we apply it on cancer gene expression data and compare it ten classification algorithms.