Machine Learning Approach to Predict hERG Inhibition Potential of New Chemical Entities

Abstract

Human Ether a-go-go related gene (hERG) or KCNH2 encodes for the potassium channel Kv11.1, which expressed at the cardiac myocites are responsible for the conductance of K+1 ions for the normal cardiac action potential. Various proof of the concept studies have highlighted that different drugs have tendency to trap inside the basal cavity of the hERG thus, leading to the prolongation of the cardiac action potential which is reflected as a prolonged interval between the Q and T wave at the surface of the electrocardiogram. This drug-induced QT interval prolongation may lead to cardiac arrhythmia known as Torsades de Pointes (TdP). Previously, various drugs have been withdrawn from the market due to their potential hERG liability. Thus, as per FDA guidelines, a drug candidate has to pass the hERG liability test before its approval. Therefore, it is crucial to identify the 2D and 3D structural features to probe the hERG liability of the new chemical entities. In order to fill this void, this work proposes various machine learning models to classify the publicly available databases of hERG as hERG blockers or non-blockers. A data set of 21488 compounds of hERG inhibitors has been collected from publicly available databases. A total of 4021 2D attributes were computed using Alva Desc and Morgan fingerprints were calculated using Rdkit. Various Machine learning classification models such as K Nearest Neighbor (KNN), Support Vector Machine (SVM), Decision Tree, were used to classify chemical entities into hERG blockers and non-blockers. K Nearest Neighbor (KNN) classify molecular compounds into blockers and non-blockers with the accuracy of 76% , Support Vector Machine (SVM) with 74% accuracy while SVM non-linear with 79% and Decision Tree with accuracy of 75%. Finally selected SVM (non-linear) model with 79% could be further integrated with 3D structural attributes of the hERG blockers to further probe the notorious drug features. The selected SVM model's high accuracy indicates its efficiency in differentiating between the two groups. Building on this achievement, the integration of hERG blocker 3D structural qualities into the SVM model offers an opportunity to further investigate key pharmacological features.