Explainable machine learning for predicting PFAS toxicity using molecular descriptors

Per- and polyfluoroalkyl substances (PFAS) are trace hazardous chemicals that accumulate in the human body and ecosystems, necessitating improved tools for toxicity assessment. This study developed machine learning (ML) models to predict PFAS toxicity using molecular descriptors derived from the simplified molecular input line entry system (SMILES) representations and computed with RDKit. Four ML algorithms, extreme gradient boosting machine (XGBoost), support vector machine (SVM), deep neural networks (DNN), and decision tree (DT), were evaluated, yielding AUROC values of 0.810, 0.817, 0.792, and 0.743, respectively, and average precision values of 0.610, 0.660, 0.584, and 0.443. These results demonstrate that ML models can reliably predict PFAS toxicity. To enhance interpretability, Shapley additive explanations (SHAP) were applied to the XGBoost to quantify the influence of individual molecular descriptors. Molecular weight (MolWt) was identified as the most influential variable, followed by the number of -CF<sub>2</sub> groups in the molecule. These findings were consistent with observed data distributions, indicating that the model effectively captured meaningful structure-toxicity relationships. Functional-group-level similarity maps were also generated to visualize how specific substructures affect toxicity predictions, providing complementary insight into model decision pathways. Overall, integrating ML with explainable AI enhanced the transparency, interpretation, and applicability in PFAS toxicity assessment.