Explainable Machine Learning Model for Predicting Ozone Reaction Rate Constants of Aromatic Compounds in Water
Quantitative structure-activity relationship (QSAR) models were developed to predict the reaction rate constants (kO3) of aromatic compounds with ozone in water. Molecular descriptors were screened using a combination of genetic algorithm and stepwise regression. Multiple linear regression (MLR), support vector machine (SVM), and projection pursuit regression (PPR) were employed to construct local models. The PPR model exhibited superior performance with a goodness-of-fit R2 of 0.923, leave-one-out cross-validation Q2LOO of 0.836, and external validation Q2ext of 0.873. The model was interpreted using SHapley Additive exPlanations (SHAP), revealing that ozone attack is hindered by the presence of dssC (=C<) fragments and chlorine atoms. The applicability domain was characterized using Williams plots. Tree manifold approximation and projection (TMAP) was used to visualize structural similarity and diversity, and Arithmetic Residuals in K-groups Analysis (ARKA) identified potential activity cliffs. The model adheres to OECD principles for QSAR validation, providing a robust tool for predicting kO3 of untested or novel aromatic compounds and extendable to other environmental applications.