New generation of QSAR modeling for bee safety: predicting toxicity using graph neural networks and apistox data.

Accurate prediction of chemical toxicity is essential for environmental risk assessment and the protection of non-target organisms, such as pollinators. In this study, a Graph Neural Network (GNN) was employed to model the toxicity of organic compounds to bees, using molecular graphs derived from SMILES strings. The model achieved an average AUC-ROC of 0.76 on the test set and 0.97 on the training set, with low standard deviations (± 0.03), indicating consistent performance across cross-validation folds. The combined ROC curve of all test folds yielded an overall AUC-ROC of 0.802, demonstrating robust discriminative ability despite class imbalance. In this sense, additional metrics, including Matthews Correlation Coefficient (MCC), Balanced Accuracy and the confusion matrix, were also calculated. Beyond satisfactory parameters, the GNN provided interpretable results through its attention mechanism, which identified molecular substructures associated with toxicity. Notably, regions corresponding to organophosphates, polyhalogenated groups and nucleophilic functional groups were frequently highlighted chemical patterns, which are part of already known mechanisms, such as acetylcholinesterase inhibitors, for example. Additionally, lipophilic alkyl chains and aromatic groups were associated with increased metabolic persistence, suggesting enhanced bioaccumulation potential. These insights reinforce the utility of GNNs as powerful tools for predictive ecotoxicology, offering both performance and mechanistic interpretability in the identification of hazardous substances. The approach presented here supports the development of safer chemicals and contributes to the advancement of non-animal testing strategies in environmental safety assessments.