Mixture of experts for multitask learning in cardiotoxicity assessment

Area of Science:

• Toxicology
• Biochemistry
• Biomedical Research
• Artificial Intelligence
• Machine Learning

Background:

• Cardiovascular diseases are a leading cause of death globally.
• Chemicals like environmental contaminants, pesticides, food additives, and drugs pose cardiotoxic risks.
• Traditional toxicology testing is time-consuming, resource-intensive, and lacks scalability.

Purpose of the Study:

• To develop and evaluate an Artificial Intelligence (AI) model for predicting chemical cardiotoxicity.
• To explore the benefits of Multitask Neural Networks within an Integrated Approach to Testing and Assessment (IATA).
• To reduce reliance on traditional in vivo testing methods.

Main Methods:

• Utilized Artificial Intelligence (AI) and Machine Learning (ML) methods.
• Developed a Multitask Neural Network model, incorporating Mixture of Experts (MoE) architecture.
• Trained and validated the model on twelve cardiotoxicity endpoints from the Adverse Outcome Pathways Network.

Main Results:

• The multitask model outperformed single-task baseline models.
• Achieved high performance on a holdout set across multiple cardiotoxicity endpoints.
• The best model demonstrated 78% balanced accuracy, 80% sensitivity, and 76% specificity.

Conclusions:

• An advanced multitask model effectively predicts cardiotoxicity mechanisms induced by small molecules.
• The model offers broad mechanistic coverage and performance comparable to state-of-the-art methods.
• This AI model can serve as a valuable component in New Approach Methodologies (NAMs) for prioritizing chemical safety testing.