QCD-MLP: A Machine-Learning Framework for Predicting Transformation Products of Organic Micropollutants in Radical-Mediated Systems

Understanding the transformation products (TPs) of organic micropollutants in radical-mediated reactions is essential for improving water treatment and environmental risk assessment. This study introduces the quantum-chemical-descriptor-based machine-learning prediction (QCD-MLP) model to predict TPs in hydroxyl (^•OH) and chlorine (Cl^•) radical-mediated systems. QCD-MLP, trained on the publicly available elementary radical-reaction database (RMechDB), leverages quantum-chemical descriptors derived from dispersion-corrected density functional theory (DFT-D) to capture atomic-level reactivity, achieving AUC values of 93% for^•OH and 89% for Cl^•systems across 345 and 150 reactions, respectively. Global-level SHapley Additive exPlanations (SHAP) analysis identified the nuclear magnetic shielding constant (NMR) as the dominant factor in^•OH-mediated reactions, while steric effects governed Cl^•reactivity. Local-level SHAP interpretation within molecular groups highlighted Fukui values as key predictors in alcohols and aromatics, showing the model’s ability to differentiate radical interactions based on molecular properties. Benchmarking against experimental and computational data confirmed the model’s precision in predicting major TPs, including meta-hydroxylated dimethyl phthalate and hydroxylated phenol derivatives. While QCD-MLP excels at radical addition and H-atom abstraction, it currently excludes downstream reactions such as ring-opening or oxidative fragmentation. Unlike conventional models, QCD-MLP offers a scalable and interpretable framework for TPs identification in natural and engineered processes.