A Multitask Graph Neural Network Framework for AMES Mutagenicity Prediction

The process of drug screening and development is long, complex, and costly, making early and reliable assessment of toxicity a central challenge in pharmaceutical research. In this context, the ability to accurately predict drug-induced toxicity in general, and mutagenicity in particular, is crucial for guiding decision-making and reducing late-stage attrition. In particular, deep learning approaches can be utilized to predict the results of the Ames mutagenicity test, which is commonly employed for toxicity screening in drug development. Recent work has shown that deep learning multitask approaches that account for the contributions of individual bacterial strains improve mutagenicity prediction. However, most existing models rely on molecular descriptors, which are limited in their representation of molecular structure and properties and often lack interpretability. Graph neural networks are chemically intuitive and overcome many of these limitations. Here, we present a deep learning, multitask graph neural network model for predicting Ames mutagenicity. The proposed approach outperforms existing models, including those from the Ames/QSAR International Challenge, and demonstrates high sensitivity for identifying mutagenic compounds. In addition, our model is highly interpretable. GNNExplainer analysis of atom-level contributions revealed that the model learned known structural alerts associated with mutagenicity, as well as additional molecular patterns and contextual features, including three-dimensional structural properties. Importantly, these insights were captured in a strain-specific manner, highlighting the value of a multitask approach for modeling mutagenic mechanisms. Finally, the model identified previously unreported molecular substructures potentially associated with mutagenicity, supporting its utility in drug design and development.