A machine learning platform for genotype-specific cardiotoxicity risk prediction using patient-derived iPSC-CMs

Drug-induced Torsades de Pointes (TdP) has led to withdrawal of several drugs from the market. Individuals with inherited cardiac channelopathies are at increased risk due to their underlying electrophysiological vulnerability.

We aimed to develop a machine learning (ML) platform for disease-specific cardiotoxicity using patient-specific induced pluripotent stem cell-derived cardiomyocytes (iPSC-CMs) combined with high-throughput microelectrode array (MEA) recordings.

We generated genetically confirmed and phenotypically characterized iPSC-CMs from patients with long QT syndrome (LQTS) and Brugada syndrome (BrS). These cells were exposed to 28 compounds with varying TdP risk levels. Electrophysiological responses including field potential duration, corrected field potential duration, beat period and amplitude were measured using MEA. These data were used to train and compare machine learning models, including artificial neural networks (ANN), random forest, and XGBoost. Model performance was optimized by grid search and evaluated by fivefold cross-validation.

The ANN model trained on LQTS iPSC-CMs achieved the highest accuracy (area under the curve [AUC] = 0.94). BrS cell lines showed hypersensitivity to calcium channel blockers, while LQTS lines exhibited heightened responses to potassium channel inhibitors. Previously ambiguous compounds were reclassified based on disease-specific electrophysiological profiles, demonstrating the platform's utility in genotype-specific cardiotoxicity risk assessment.

This study presents a scalable and individualized approach for cardiotoxicity screening using well-characterized patient-derived iPSC-CMs. The platform enhances drug safety prediction, supports regulatory evaluation, and advances precision medicine in arrhythmia risk assessment.