Machine Learning-Based Prediction of Bioaccumulation and Ecotoxicity of Emerging Contaminants in Aquatic Ecosystems

Pharmaceuticals, pesticides, personal care products, and industrial additives are emerging contaminants (ECs) threatening aquatic ecosystems and are persistent, bioaccumulative, and have toxic effects that remain unseen by traditional tests. The objective of this study was to create and test machine learning (ML)-based predictive models to predict bioaccumulation and ecotoxicity of ECs with more constraints than the traditional QSAR approaches. ECOTOX, PubChem, OECD QSAR Toolbox, and literature sources were then reviewed to come up with a list of 350 compounds with the following descriptors of logKow, molecular weight, solubility, and TPSA. Four ML algorithms (Random Forest, Support Vector Machine, XGBoost, and Deep Neural Networks) were trained and optimised, and performance was compared to QSAR models. The interpretability of the models was improved using SHAP analysis. The findings indicated that XGBoost performed better than the rest of the models (R 2 = 0.87; ROC-AUC = 0.90), whereas the performance of the Random Forest and SVM was similarly strong, whereas DNNs were likely to overfit. The ML models were significantly better at bioaccumulation, acute toxicity, and chronic toxicity +20.8, +15.4, and +17.3, respectively (p < 0.05). SHAP identified logKow, molecular weight, and TPSA to be key predictors with external validation using contaminants such as triclosan and bisphenol A being used to validate the generalizability. The research suggests the use of ML in a regulatory framework to improve ecological risk assessment, decrease animal tests, and save the aquatic environment.