Evaluation of Mutual Information, Genetic Algorithm and SVR for Feature Selection in QSAR Regression

Abstract

Feature selection has become increasingly important for quantitative structure-activity relationship (QSAR) studies. In the present article, we evaluate three state-of-the-art feature selection algorithms, namely mutual information (MI), genetic algorithm (GA), and support vector machine regression (SVR)-based recursive feature elimination (SVR-RFE), in the reduction of high dimensional feature space for QSAR regression. We used SVR to evaluate the performance of these feature selection algorithms. In addition, we present a simple but very efficient iterative strategy for optimizing parameters for SVM-RFE algorithm. All three algorithms can effectively reduce the number of features and often achieve improved performance.