Mixture of experts for multitask learning in cardiotoxicity assessment

Abstract: In recent years, the integration of Artificial Intelligence and Machine Learning methods with biochemical and biomedical research has revolutionized the field of toxicology, significantly advancing our understanding of the toxicological effects of chemicals on biological systems. Cardiovascular diseases remain the leading global cause of death. The constant exposure to multiple chemicals with potential cardiotoxic effects, including environmental contaminants, pesticides, food additives, and drugs, can significantly contribute to these adverse health outcomes. Traditional methods for assessing chemical hazards and their impact on biological function heavily rely on experimental assays and animal studies, which are often time-consuming, resource-intensive, and limited in scalability. To overcome these limitations in silico methods have emerged as indispensable tools in toxicological research, reducing the need for traditional in vivo testing and conserving valuable resources in terms of time and cost. In this study, Artificial Intelligence methods are used as first-tier components within an Integrated Approach to Testing and Assessment. We explored the potential benefits of using Multitask Neural Networks, where multiple levels of cardiotoxicity information are combined to enhance model performance. Multitask learning, based on specific architectures such as Mixture of Experts (MoE), showed promising results and surpasses the performance of single-task baseline models. When predicting a holdout set, multitask model achieved high performance on twelve different endpoints related to cardiotoxicity defined by Adverse Outcome Pathways Network. The best developed model achieved a balanced accuracy of 78%, a sensitivity of 80%, and a specificity of 76% across all endpoints in the holdout set. Scientific contribution: An advanced multitask model was developed to predict cardiotoxicity mechanisms induced by small molecules. The model demonstrates broad mechanistic coverage and achieves performance comparable to, or exceeding, state-of-the-art methods. These results suggest that the model could serve as a valuable first-tier component in advanced New Approach Methodologies for prioritizing chemicals for further testing.