An accurate and interpretable bayesian classification model for prediction of HERG liability

Abstract

Drug-induced QT interval prolongation has been identified as a critical side-effect of non-cardiovascular therapeutic agents and has resulted in the withdrawal of many drugs from the market. As almost all cases of drug-induced QT prolongation can be traced to the blockade of a voltage-dependent potassium ion channel encoded by the hERG (the human ether-à-go-go-related gene), early identification of potential hERG channel blockers will decrease the risk of cardiotoxicity-induced attritions in the later and more expensive development stage. Presented herein is a naive Bayes classifier to categorize hERG blockers into active and inactive classes, by using a universal, generic molecular descriptor system.1 The naive Bayes classifier was built from a training set containing 1979 corporate compounds, and exhibited an ROC accuracy of 0.87. The model was validated on an external test set of 66 drugs, of which 58 were correctly classified. The cumulative probabilities reflected the confidence of prediction and were proven useful for the identification of hERG blockers. Relative performance was compared for two classifiers constructed from either an atom-type-based molecular descriptor or the long range functional class fingerprint descriptor FCFP_6. The combination of an atom-typing descriptor and the naive Bayes classification technique enables the interpretation of the resulting model, which offers extra information for the design of compounds free of undesirable hERG activity.