Evaluation of cardiac pro-arrhythmic risks using the artificial neural network with ToR-ORd in silico model output

Abstract

Torsades de pointes (TdP) is a type of ventricular arrhythmia that can lead to sudden cardiac death. Drug-induced TdP has been an important concern for researchers and international regulatory boards. The Comprehensive in vitro Proarrhythmia Assay (CiPA) initiative was proposed that integrates in vitro testing and computational models of cardiac ion channels and human cardiomyocyte cells to evaluate the proarrhythmic risk of drugs. The TdP risk classification performance using only a single TdP metric may require some improvements because of information limitations and the instability of generalizing results. This study evaluates the performance of TdP metrics from the in silico simulations of the Tomek-O'Hara Rudy (ToR-ORd) ventricular cell model for classifying the TdP risk of drugs. We utilized these metrics as an input to an artificial neural network (ANN)-based classifier. The ANN model was optimized through hyperparameter tuning using the grid search (GS) method to find the optimal model. The study outcomes show an area under the curve (AUC) value of 0.979 for the high-risk category, 0.791 for the intermediate-risk category, and 0.937 for the low-risk category. Therefore, this study successfully demonstrates the capability of the ToR-ORd ventricular cell model in classifying the TdP risk into three risk categories, providing new insights into TdP risk prediction methods.

Keywords: Tomek–O’Hara Rudy ventricular in silico cell model; artificial neural networks; explainable artificial intelligence; grid search; torsade de pointes.

FIGURE 1

Proposed torsades de pointes (TdP) risk evaluation method. This study utilized 2,000 Hill samples from in vitro patch-clamp data provided by Li et al. (2019). These samples were subsequently integrated into the ToR–ORd in silico ventricular cell model, given 28 different drugs at 4 maximum concentration levels (C_max 1-4), resulting in 2,000 samples for 9 key TdP features. The average sample values were calculated for each sample based on the results obtained at the four distinct drug concentration levels. Then, the dataset was segmented into 2 parts: a training dataset consisting of 24,000 samples (12 drugs × 2,000 samples) used in the artificial neural network (ANN) classifier model and a testing dataset consisting of 32,000 samples (16 drugs × 2,000 samples) for evaluation through 10,000 testing iterations.

FIGURE 2

Schematic diagram of the 10,000 test algorithms based on the the study by Jeong et al. (2022).

FIGURE 3

ANN architecture optimized by grid search (GS) hyperparameter optimization.

FIGURE 4

Feature importance ranking based on the SHapley Additive exPlanation (SHAP) values from explainable artificial intelligence (XAI) with the ANN, which was optimized by GS hyperparameter optimization.

FIGURE 5

Distribution of area under the curves (AUCs) based on the TdP risk using 9 features of 16 test drugs; yellow, red, and green are the AUC distribution for high, intermediate, and low risks, respectively.