Machine learning approach to evaluate TdP risk of drugs using cardiac electrophysiological model including inter-individual variability

Introduction: Predicting ventricular arrhythmia Torsade de Pointes (TdP) caused by drug-induced cardiotoxicity is essential in drug development. Several studies used single biomarkers such as qNet and Repolarization Abnormality (RA) in a single cardiac cell model to evaluate TdP risk. However, a single biomarker may not encompass the full range of factors contributing to TdP risk, leading to divergent TdP risk prediction outcomes, mainly when evaluated using unseen data. We addressed this issue by utilizing multi-in silico features from a population of human ventricular cell models that could capture a representation of the underlying mechanisms contributing to TdP risk to provide a more reliable assessment of drug-induced cardiotoxicity. Method: We generated a virtual population of human ventricular cell models using a modified O’Hara-Rudy model, allowing inter-individual variation. [Formula: see text] and Hill coefficients from 67 drugs were used as input to simulate drug effects on cardiac cells. Fourteen features ( [Formula: see text] , [Formula: see text] , [Formula: see text] , [Formula: see text] , [Formula: see text] , [Formula: see text] , [Formula: see text] , [Formula: see text] , [Formula: see text] , [Formula: see text] , [Formula: see text] , [Formula: see text] , qNet, qInward) could be generated from the simulation and used as input to several machine learning models, including k-nearest neighbor (KNN), Random Forest (RF), XGBoost, and Artificial Neural Networks (ANN). Optimization of the machine learning model was performed using a grid search to select the best parameter of the proposed model. We applied five-fold cross-validation while training the model with 42 drugs and evaluated the model’s performance with test data from 25 drugs. Result: The proposed ANN model showed the highest performance in predicting the TdP risk of drugs by providing an accuracy of 0.923 (0.908–0.937), sensitivity of 0.926 (0.909–0.942), specificity of 0.921 (0.906–0.935), and AUC score of 0.964 (0.954–0.975). Discussion and conclusion: According to the performance results, combining the electrophysiological model including inter-individual variation and optimization of machine learning showed good generalization ability when evaluated using the unseen dataset and produced a reliable drug-induced TdP risk prediction system.