Definition of QSPR models' scope and their effective application according to chemical structures

Abstract

We propose two methods for more reliable and accurate property estimation in the field of QSPR. One of those is a method to determine applicability domain (AD) of QSPR model numerically based on pattern recognition of chemical structures. AD is a part of chemical space in which chemical structures are predicted moderately by corresponding QSPR model. A classification model is constructed through supervised learning and its outputs indicate whether a compound is in the AD of a QSPR model or not. On the other hand, we also developed integration method of multiple QSPR models rather than selecting one of them. For each compound, a linear combination of the all outputs of applicable models is formed to minimize error of prediction. Coefficients of linear combination are determined considering model predictivity and correlativity of each pair of models. As a case study, we examined these methods with two data sets of water solubility taken from literatures.