A 3D Multitask Graph Neural Network Framework for Ames Mutagenicity Prediction

The long, complex, and costly process of drug development makes early, reliable toxicity assessment a central challenge in pharmaceutical research. Accurately predicting drug‐induced toxicity and mutagenicity is crucial for guiding decision‐making and reducing late‐stage attrition. Deep learning approaches can predict the results of the Ames mutagenicity test, a standard screening method; however, most existing models rely on molecular descriptors that limit structural representation, lack interpretability, and do not account for Ames results of individual bacterial strains. Graph neural networks (GNNs) are chemically intuitive and overcome these limitations. This work presents a deep learning, multitask GNN model for predicting Ames mutagenicity. The proposed approach outperforms several existing models, including those from the Ames/QSAR International Challenge, on the dataset presented in this work. Furthermore, the model is highly interpretable. Explainability analysis of atom‐level contributions reveal that the model is capable of learning known structural alerts associated with mutagenicity, along with additional molecular patterns and 3D structural properties. Importantly, these insights are captured in a strain‐specific manner, highlighting the value of a multitask approach. Finally, the model is able to identify previously unreported molecular substructures potentially associated with mutagenicity, supporting its utility in drug design and development.