A stacking ensemble machine learning model for evaluating cardiac toxicity of drugs based on in silico biomarkers

Abstract

This study addresses the critical issue of drug-induced torsades de pointes (TdP) risk assessment, a vital aspect of new drug development due to its association with arrhythmia and sudden cardiac death. Existing methodologies, particularly those reliant on a single biomarker derived from CiPA O'Hara-Rudy (CiPAORdv1.0) ventricular cell model without the hERG dynamic as input to the individual machine learning model, have limitations in capturing the complexity inherent in the comprehensive range of factors influencing drug-induced TdP risk. This study aims to overcome these limitations by proposing a stacking ensemble machine learning approach by integrating multiple in silico biomarkers derived from the CiPAORdv1.0 with hERG dynamic characteristics. The ensemble machine learning model consisted of three artificial neural network (ANN) models as baseline model and support vector machine (SVM), logistic regression (LR), random forest (RF), and extreme gradient boosting (XGBoost) models as meta-classifier. The highest AUC score of 1.00 (0.90-1.00) for high risk, 0.97 (0.84-1.00) for intermediate risk, and 1.00 (0.87-1.00) for low risk were obtained using seven biomarkers derived from the CiPAORdv1.0 with hERG dynamic characteristics. Furthering our investigation, we explored the model's robustness by incorporating interindividual variability into the generation of in silico biomarkers from a population of human ventricular cell models. This study also enabled an analysis of TdP risk classification under high clinical exposure and therapeutic scenarios for several drugs. Additionally, from a sensitivity analysis, we revealed four important ion channels, namely, CaL, NaL, Na, and Kr channels that affect significantly the important biomarkers for TdP risk prediction.

FIGURE 1

A whole schematic of a stacking ensemble machine learning model with fivefold cross‐validation for predicting cardiac toxicity of drugs into high, intermediate, and low risk using in silico biomarkers.

FIGURE 2

The SHAP values of 16 testing drugs from three base line models; (a) The SHAP value from ANN with one hidden layer; (b) The SHAP value from ANN with two hidden layers; (c) The SHAP value from ANN with three hidden layers.

FIGURE 3

Variations in the top seven most influential in silico biomarkers for TdP risk classification in response to changes in ion channel conductance. The conductance of ion channels is adjusted to −1%, normal (unchanged), and +1% of their maximum values.

FIGURE 4

Top seven influential biomarkers simulated under various drug concentrations of 1 to 30× C _max for 16 testing drugs. The gray‐shaded regions represent the lower‐upper bounds of 95% confidence interval range of biomarkers from 10 samples from each drug. The bold lines in the middle of shaded areas represent the average value of the biomarkers. Finally, the color coding represents the TdP risk of drugs: Red for high‐risk drugs, blue for intermediate‐risk drugs, and green for low‐risk drugs.