Integrating Experimental Toxicology and Machine Learning to Model Levonorgestrel-Induced Oxidative Damage in Zebrafish.

Levonorgestrel (LNG), a synthetic progestin widely used in pharmaceuticals, is increasingly recognized as an emerging aquatic contaminant capable of exerting adverse biological effects beyond endocrine disruption. Acting in a xenobiotic-like manner, LNG may perturb redox homeostasis and induce oxidative stress in non-target species. To elucidate these mechanisms, this study integrates experimental toxicology with supervised machine learning to characterize tissue-specific and dose-time related oxidative responses in adult Zebrafish (Danio rerio). Fish were exposed to two environmentally relevant concentrations of LNG (0.312 µg/L; LNG-L and 6.24 µg/L; LNG-H) and a solvent control (LNG-C) for 24, 48, and 96 h in triplicate static bioassays. Redox biomarkers-superoxide dismutase (SOD), catalase (CAT), glutathione peroxidase (GPx), and malondialdehyde (MDA)-were quantified in liver and muscle tissues. LNG-H exposure elicited a time-dependent increase in SOD activity, variable CAT responses, and a marked elevation in hepatic GPx, with sustained MDA levels indicating persistent lipid peroxidation. Five classification algorithms (Logistic Regression, Multilayer Perceptron, Gradient-Boosted Trees, Decision Tree and Random Forest) were trained to discriminate exposure outcomes based on biomarker profiles; GBT yielded the highest performance (96.17% accuracy), identifying hepatic GPx as the most informative feature (AUC = 0.922). Regression modeling via Extreme Gradient Boosting (XGBoost) further corroborated the dose- and time-dependent predictability of GPx responses (R² = 0.922, MAE = 0.019). These findings underscore hepatic GPx as a sentinel biomarker of LNG-induced oxidative stress and demonstrate the predictive utility of machinelearning-enhanced toxicological frameworks in detecting and modeling sublethal contaminant effects with high temporal resolution in aquatic systems.