Utilizing Machine Learning Models with Molecular Fingerprints and Chemical Structures to Predict the Sulfate Radical Rate Constants of Water Contaminants

Abstract

Sulfate radicals are increasingly recognized for their potent oxidative capabilities, making them highly effective in degrading persistent organic pollutants (POPs) in aqueous environments. These radicals excel in breaking down complex organic molecules that are resistant to traditional treatment methods, addressing the challenges posed by POPs known for their persistence, bioaccumulation, and potential health impacts. The complexity of predicting interactions between sulfate radicals and diverse organic contaminants is a notable challenge in advancing water treatment technologies. This study bridges this gap by employing a range of machine learning (ML) models, including random forest (DF), decision tree (DT), support vector machine (SVM), XGBoost (XGB), gradient boosting (GB), and Bayesian ridge regression (BR) models. Predicting performances were evaluated using R2, RMSE, and MAE, with the residual plots presented. Performances varied in their ability to manage complex relationships and large datasets. The SVM model demonstrated the best predictive performance when utilizing the Morgan fingerprint as descriptors, achieving the highest R2 and the lowest MAE value in the test set. The GB model displayed optimal performance when chemical descriptors were utilized as features. Boosting models generally exhibited superior performances when compared to single models. The most important ten features were presented via SHAP analysis. By analyzing the performance of these models, this research not only enhances our understanding of chemical reactions involving sulfate radicals, but also showcases the potential of machine learning in environmental chemistry, combining the strengths of ML with chemical kinetics in order to address the challenges of water treatment and contaminant analysis.