On the relevance of feature selection algorithms while developing non-linear QSARs

Abstract

Quantitative structure-activity relationships (QSARs) are mathematical models aimed at finding a quantitative relationship between a set of chemical compounds and a specific activity or endpoint, such as toxicity, chemical or physical property, biological activity, and so on. In order to find out the correlation between the chemicals and the selected endpoints, QSAR models use the so-called molecular descriptors (MDs) which encode specific chemical information or features of the molecules. The early QSAR models were based on a small set of MDs and a specific endpoint, and the correlation was usually a linear mathematical correlation. However, nowadays, QSAR models are usually non-linear and made up by thousands of chemicals and hundreds of MDs. In addition, novel QSAR models are also aimed at the prediction of different endpoints with the same model, the so-called multi-target QSAR (MT-QSAR). Due to this, nowadays many QSARs are usually developed using machine learning approaches which can model a dataset with different endpoints. Although these approaches have demonstrated to be able to solve MT-QSAR models, feature selection (FS) in these cases is a challenging task and a main point in the QSAR field. Considering these aspects, the main aim of this chapter is to analyze feature selection methods while developing non-linear QSAR models.