Toward Reducing HERG Affinities for Dat Inhibitors with a Combined Machine Learning and Molecular Modeling Approach

Abstract

The human Ether-à-go-go-Related (hERG) potassium channel plays a critical role in repolarization of cardiac action potentials. hERG blockade by drugs may result in abnormal heart rhythms such as the potentially fatal Torsades de Pointes. Experimental methods to identify compounds with high hERG affinities are time-consuming and costly. Compulsory hERG screening by the FDA has resulted in an increase of publicly available hERG binding data. Utilizing machine learning algorithms, we constructed structure activity relationship regression models to predict hERG binding affinity. The models were trained on molecules with known hERG affinities from the ChEMBL database. We filtered out erroneous points among the low affinity data that were found to affect prediction accuracy. We compared the performance of the eXtreme gradient boosting (XGBoost) and the random forest (RF) algorithms against several test sets and found that XGBoost slightly outperformed RF. The XGBoost models were then used to predict the hERG affinities of a series of dopamine transporter (DAT) inhibitors. The predictions showed good correlation with experimentally measured affinities. In addition, we implemented a machine-learning based classifier to identify compounds which appear to have extremely low affinity for hERG and are therefore unlikely to be well-predicted by the consensus regression model. In coordination with ligand-based approaches, we have used a protein structure-based method to characterize binding modes of these ligands against both our hDAT models and the recent cryo-EM structure of hERG, and to identify clues to improve hDAT affinities while reducing hERG affinities. Together, we established a combined ligand-based and protein structure-based computational framework, which is integrated with experimental approaches, to predict hERG affinities. We specifically applied it to facilitate the rational optimization of lead compounds targeting DAT for the treatment of psychostimulant use disorders.