Interpretable machine learning models for predicting hormesis of binary antibiotic mixtures on Vibrio qinghaiensis sp.-Q67.

The hormesis phenomenon of antibiotic mixtures hinders environmental risk assessment owing to the lack of predictive models for chemical mixtures. This study developed a novel machine learning classification approach to predict the hormesis effect of 275 binary antibiotic mixtures on Vibrio qinghaiensis sp.-Q67. Time-dependent toxicities (12, 18, and 24 h) indicate that 81.1 % of mixtures demonstrate hormesis at 24 h (compared with 54.1 % at 12 h), revealing a time-dependent amplification. A machine learning framework incorporating nine algorithms was developed to distinguish between hormesis and non-hormesis responses. For each exposure duration, 27 base models and 243 ensemble models were developed via stacked generalisation. The Random Forest model achieved 92.8 % accuracy (area under the receiver operating characteristic curve > 0.985), outperforming eight algorithms. Shapley additive explanations method employs a game-theoretic approach for feature interpretation. This analysis revealed three critical molecular determinants, namely, SsssN, SpMax6_Bhs and nHCsatu, which collectively explain the bioactivity profiles of antibiotic mixtures. The results revealed that steric hindrance, electron delocalisation and the hydrogen-bond network were key factors in predicting the biological activity of antibiotic mixtures exhibiting the hormesis effect. This study introduces a machine learning framework for predicting the hormesis effect of antibiotic mixtures, providing vital tools for ecological risk assessment and more intelligent management of aquatic contaminants.