Environmental Assessment of Chemicals: Artificial Intelligence for Predicting Persistence, Bioaccumulation, and Toxicity Properties

Early assessment of the persistence, bioaccumulation, and toxicity (PBT) of chemicals is a major challenge for environmental protection and international regulatory frameworks. The objective of this study is to compare the effectiveness of three graph-based deep learning architectures—a graph neural network (GNN), a message passing network (MPNN), and a graph attention network (GAT)—for the binary classification of molecules as PBT or non-PBT.We compiled a regulatory dataset comprising 5,130 molecules annotated from public sources, such as ECHA and international POP lists. Molecular graphs were generated from SMILES using RDKit. The three models were implemented in PyTorch Geometric with homogeneous hyperparameters. The experiments were conducted with a scaffold split ratio of 80/10/10 and 10-fold cross-validation. Performance was evaluated using accuracy, AUC-ROC, and F1-score. Interpretability was examined using GAT model attention maps and atomic contribution analysis. The MPNN model achieves the best overall performance (Accuracy = 0.92; ROC-AUC = 0.94; F1 = 0.91), followed by GAT (Accuracy = 0.89; ROC-AUC = 0.93). The basic GNN performs less well (Accuracy = 0.82; ROC-AUC = 0.89). The GAT model provides more detailed atomic explanations thanks to attention weights, while the MPNN stands out for its predictive accuracy. The dataset includes annotations from heterogeneous experimental sources, which may introduce noise into the labels. The models rely solely on 2D graphs, without 3D conformational information. MPNN models can accelerate PBT pre-screening and help prioritize substances for experimental testing. GATs provide useful interpretations for understanding the substructures associated with PBT properties. This study provides the first reproducible and systematic comparison of GNN, MPNN, and GAT models applied to a large regulatory dataset dedicated to PBT, analyzing both performance and interpretability. These results highlight the potential of graph-based QSAR models for regulatory PBT screening and environmental risk assessment.