Towards better understanding of feature-selection or reduction techniques for Quantitative Structure–Activity Relationship models

A Quantitative Structure–Activity Relationship (QSAR) is a linear or non-linear model, which relates variations in molecular descriptors to variations in the biological activity of a series of active and/or inactive molecules.For this article, different feature-selection or reduction methods were all coupled with Partial Least Squares (PLS) modeling during the selection of features. A PLS model was also built with the entire set of molecular descriptors and was used as a reference to check the reliability and the performance of the different feature-selection methods. To evaluate the ability of the different feature-selection methods, they were performed on two data sets.

► Molecular structure determined and conformational analysis performed. ► Molecular descriptors are calculated based on above structure determination. ► Several feature-selection methods are applied and linked to PLS modeling technique. ► QSAR models were built with PLS and MLR (when only a few descriptors were selected). ► The Replacement Method was found to select few variables and led to good models.