Classification Model Based Approach to Definition of Applicability Domain of QSPR model for Aqueous Solubility

Abstract

We have investigated two methods for definition of applicability domain (AD) of quantitative structure - property relationships (QSPR) models. One of them is a range-based method which defines an AD by the ranges of several molecular descriptors. The other one is based on one-class support vector machines (OCSVM) which typically used for data domain description and outlier detection. We have built several QSPR models to evaluate robustness and stability of AD methods. We used two sets of organic compounds with experimental values of aqueous solubility; aliphatic mono alchol and herbicides. For both datasets, three regression models have been built: two partial least squares (PLS) models and an epsilon supprot vector regression (e-SVR) model. One of the PLS models was optimized via variable selection using genetic algorithm. The parameters of e-SVR model were selected via grid searching. Thus we have tested AD methods for various types of QSPR models; linier and non-linier, overfitted and optimized, local and global. As a result, we conclude using OCSVM can be suitable for definition of AD robust for various types of QSPR models.