Prediction of the toxicity of complex antibiotic mixtures with hormetic effects by an interpretable machine learning model.

Abstract

Hormesis, a phenomenon frequently found in some low-dose chemical mixtures and their individual components in the environment, remains a formidable challenge in environmental risk assessment. To address this, a novel interpretable machine learning model, DynaHorm-ERT (Dynamic Hormesis Extremely Randomized Trees), was developed by using micro-dynamic features (extracted from molecular dynamics simulations) and hormetic effect features and applied to predict the combined toxicity of complex antibiotic mixtures (five β-lactam antibiotics, four aminoglycoside antibiotics and chloramphenicol) on a freshwater organism Vibrio qinghaiensis sp.-Q67. The results showed that most of the antibiotics and their mixtures exhibited hormesis with J-shaped concentration-response curves characterized by "low-concentration stimulation and high-concentration inhibition". Interestingly, some complex mixtures containing non-hormetic components also exhibited significant hormetic characteristics. The established model successfully and accurately predicted their joint toxicity. SHAP analysis indicated that the intensity of the hormetic effect and micro-dynamic constraints were the core mechanisms of driving toxicity, and the model demonstrated excellent predictive performance by capturing the potency of ligands within the receptor pocket and the degree of their dynamic constraints. Applicability domain analysis and external validation further confirmed the reliability of the model in assessing mixtures exhibiting hormesis. The obtained results highlight the importance of incorporating mechanistic features into the construction of reliable predictive models, which provides new insights and a new method for the ecotoxicological risk assessment of mixtures with hormetic effects.