Machine-Learning Prediction of Drug-Induced Cardiac Arrhythmia: Analysis of Gene Expression and Clustering.

Abstract

A marked delay in the electrical repolarization of heart ventricles is characterized by prolongation of the Q-T wave (QT) interval on a surface electrocardiogram. Such a delay can lead to potentially life-threatening cardiac arrhythmia (torsades de pointes). Such prolongation is also a widely accepted cardiac safety biomarker in drug development. Current preclinical drug-safety assays include patch clamp analysis to evaluate drug-related blockade of cardiac repolarizing ion currents. Recently reported patch clamp assay results have shown predictive sensitivities and specificities in the ranges of 64%-82% and 75%-88%, respectively. In this project, we use a support vector machine classifier to find mean sensitivities and specificities of 85% and 90%, respectively, across 77 drug subclassifications. Clustering by gene expression profile similarities shows that drugs known to prolong the QT interval do not always form distinct groups, but the number of groups is limited. The most common biological network links associated with these groups involve genes linked with fatty acid metabolism, G proteins, intracellular glutathione, immune responses, apoptosis, mitochondrial function, electron transport, and mitogen-activated protein kinases. These results suggest that machine-learning analysis of gene expression and clustering may augment cardiac safety predictions for improving drug-safety assessments.