Validation of new AI-based classification method for in silico cardiac safety assessment of drugs following the CiPA framework.

Abstract

The comprehensive in vitro proarrhythmia assay (CiPA) has paved the way for integrating in silico trials into drug evaluation processes. In alignment, the International Council for Harmonization (ICH) has initiated efforts to update the ICH S7B and E14 guidelines through a structured Questions and Answers (Q&A) format. A significant challenge in this paradigm is ensuring consistent application and evaluation of diverse proarrhythmia risk prediction models across experimental systems. This study utilized the CiPAORdv1.0 model to predict cardiac toxicity, leveraging in vitro data from 28 drugs for training and validation. A modified O'Hara-Rudy model simulated a virtual population of human ventricular cell models. Seven critical features (qNet, APD50, APD90, Camax, Carest, CaTD50, CaTD90) were extracted as inputs for analysis. CiPAORdv1.0 demonstrated robust performance, achieving predictive accuracies with an area under the curve (AUC) of 1.0 for high risk and 0.95 for low-risk categories. The calibration process was enhanced using normalized Euclidean distances (R1 and R2), effectively distinguishing risk categories. Sensitivity analysis identified key drugs, ensuring a strong calibration drug set to anchor model predictions. The proposed ANN model validated the CiPAORdv1.0 framework as an effective TdP-risk prediction system, ensuring robust and lab-specific validation. This study presents a novel algorithm leveraging artificial neural networks to implement validated cardiac safety models, addressing a critical need for standardized proarrhythmia risk assessment in drug development.