Development of models for predicting Torsade de Pointes cardiac arrhythmias using perceptron neural networks

TL;DR

This study develops a 3D spectral data-activity relationship model using perceptron neural networks to predict the risk of Torsade de Pointes arrhythmias caused by drug candidates, aiding early toxicity screening.

Contribution

Introduces a novel 3D-SDAR method combined with neural networks for predicting TdP risk and identifying molecular features responsible for cardiotoxicity.

Findings

Achieved 99.2% training accuracy with the model.

Correctly predicted 70.3% of positive TdP drugs in external validation.

Generated toxicophores associated with TdP risk.

Abstract

Blockage of some ion channels and in particular, the hERG cardiac potassium channel delays cardiac repolarization and can induce arrhythmia. In some cases it leads to a potentially life-threatening arrhythmia known as Torsade de Pointes (TdP). Therefore recognizing drugs with TdP risk is essential. Candidate drugs that are determined not to cause cardiac ion channel blockage are more likely to pass successfully through clinical phases II and III trials (and preclinical work) and not be withdrawn even later from the marketplace due to cardiotoxic effects. The objective of the present study is to develop an SAR model that can be used as an early screen for torsadogenic (causing TdP arrhythmias) potential in drug candidates. The method is performed using descriptors comprised of atomic NMR chemical shifts and corresponding interatomic distances which are combined into a 3D abstract space matrix. The method is called 3D-SDAR (3 dimensional spectral data-activity relationship) and can be interrogated to identify molecular features responsible for the activity, which can in turn yield simplified hERG toxicophores. A dataset of 55 hERG potassium channel inhibitors collected from Kramer et al. consisting of 32 drugs with TdP risk and 23 with no TdP risk was used for training the 3D-SDAR model.An ANN model with multilayer perceptron was used to define collinearities among the independent 3D-SDAR features. A composite model from 200 random iterations with 25% of the molecules in each case yielded the following figures of merit: training, 99.2 %; internal test sets, 66.7%; external (blind validation) test set, 68.4%. In the external test set, 70.3% of positive TdP drugs were correctly predicted. Moreover, toxicophores were generated from TdP drugs. A 3D-SDAR was successfully used to build a predictive model for drug-induced torsadogenic and non-torsadogenic drugs.