A novel framework integrating a coupled mixing rule with deep learning for toxicity prediction and environmental risk assessment of antibiotic mixtures.

Co-exposure of multiple antibiotics in the environment is widespread, and accurate prediction of antibiotic mixture toxicity is essential for ecological risk assessment. However, existing models often lack adaptive selection of mixing rules and fail to adequately capture key structural information about toxigenic compounds. In this study, a novel coupled mixing rule was developed to optimize mixture descriptor construction. A Transformer-deep neural network (Transformer-DNN) was further established for multitask prediction of antibiotic mixture toxicity. Results indicated that the model's prediction accuracy improved by 12.8% when employing coupled mixing rules, compared to traditional rules. Transformer-DNN achieved high predictive performance for antibiotic mixture toxicity across seven toxicity endpoints simultaneously (average R2 = 0.936). Compared with the classical models that use a uniform conventional mixing rule throughout the modeling process, the Transformer-DNN achieved a 26.0% improvement in predictive accuracy and a 53.9% reduction in prediction error. Preliminary risk assessment indicated that antibiotics considered individually non-risk were assessed as moderate risk in 72.2% and high risk in 16.7% of the mixtures upon combination. Overall, this study improves the prediction of antibiotic mixture toxicity through innovations in both mixing rules and model architecture and highlights the importance of incorporating mixture effects into future environmental risk assessment frameworks.