Drug safety assessment by machine learning models.

Area of Science:

• Cardiovascular Pharmacology
• Computational Toxicology
• Drug Safety Science

Background:

• Drug-induced Torsades de pointes (TdP) poses a significant risk in drug development.
• Accurate prediction of TdP is essential for patient safety and regulatory approval.
• Preclinical assays provide valuable data for assessing cardiac liability.

Purpose of the Study:

• To evaluate machine learning models for predicting drug-induced TdP risks.
• To utilize preclinical data, specifically from the rabbit ventricular wedge assay, for TdP risk prediction.
• To validate the utility of computational approaches in drug safety assessment.

Main Methods:

• Development and training of a random forest model using preclinical data.
• Dataset generated from the rabbit ventricular wedge assay.
• Performance evaluation using leave-one-drug-out cross-validation on 28 drugs from the Comprehensive In Vitro Proarrhythmia Assay initiative.
• Uncertainty assessment via stratified bootstrap.

Main Results:

• The random forest model demonstrated utility in predicting drug-induced TdP risks.
• Leave-one-drug-out cross-validation provided an unbiased estimation of model performance.
• Stratified bootstrap quantified the uncertainty in model predictions.

Conclusions:

• Machine learning approaches are effective for predicting drug-induced TdP risks from preclinical data.
• The developed methods can be adapted to other preclinical protocols.
• This approach serves as a valuable supplementary tool in comprehensive drug safety assessments.