Prediction of second-order rate constants for reactions between contaminants and reactive species by a ChemBERTa-based QSAR model.
Accurate prediction of second-order rate constants (k) for reactions between contaminants and reactive species (RS) is essential for understanding transformation pathways and optimizing advanced oxidation/reduction processes (AOPs/ARPs). However, existing QSAR models mostly rely on manually engineered molecular descriptors and have limited capability in capturing complicated structure-reactivity relationships, while the application of pre-trained Transformer-based molecular language models in k prediction remains largely unexplored. In this study, a deep learning-based QSAR framework (ChemBERTa-FC) was developed to predict k values for reactions of water contaminants with HO•, SO4•-, and eaq-. The model integrates a pre-trained molecular language model (ChemBERTa) for representation learning with Fully Connected (FC) layers for regression, enabling end-to-end prediction directly from SMILES without manual feature engineering. The proposed model achieves high predictive performance across all three RS: for HO•, the model yields RMSE values of 0.052 (training) and 0.088 (test); for SO4•-, RMSEtest and R2test reach 0.108 and 0.586, respectively; for eaq-, the model exhibits consistently low error and balanced performance across the full reactivity range. SHAP analysis reveals RS-specific attribution patterns aligned with reaction mechanisms, while Pearson correlation shows that most learned embeddings are not linearly explainable by traditional descriptors, indicating the capture of higher-order structure-activity relationships. Applicability domain analysis further confirms the reliability of model predictions within the defined chemical space. Overall, this work establishes a transferable and interpretable deep learning framework for k prediction and provides new insights into the molecular determinants of contaminant reactivity, supporting the rational design of water treatment processes.