Representative feature selection of molecular descriptors in QSAR modeling

Abstract

Abstract Quantitative structure-activity relationship (QSAR) has been widely applied to many fields such as molecular toxicity detection and biological activity predictions. The screening of molecular descriptors (MDs) is a preliminary stage for building the QSAR model. An appropriate MDs set significantly impacts on the performance of QSAR model. Traditional screening of MDs is usually executed by artificial selection strongly depending on expertise and experience, which is difficult to cover all valuable cross-information among these dispersed and weak chemical information representative MDs. In this paper, we proposed a Representative Feature Selection (RFS) method that is capable of forming a representative set of MDs for QSAR model by calculating the Euclidean distances and Pearson correlation coefficients. The experimental result revealed that RFS effectively selects representative features from the feature space with information redundancy, and RFS enhances the performances of QSAR model.