Combining Multi-Dimensional Molecular Fingerprints to Predict hERG Cardiotoxicity of Compounds

Abstract

At present, drug toxicity has become a critical problem with heavy medical and economic burdens. acLQTS (acquired Long QT Syndrome) is acquired cardiac ion channel disease caused by drugs blocking the hERG channel. Therefore, it is necessary to avoid cardiotoxicity in the drug design and computer models have been widely used to fix this plight. In this study, we present a molecular fingerprint based on the molecular dynamic simulation and uses it combined with other molecular fingerprints (multi-dimensional molecular fingerprints) to predict hERG cardiotoxicity of compounds. 203 compounds with hERG inhibitory activity (pIC50) were retrieved from a previous study and predicting models were established using four machine learning algorithms based on the single and multi-dimensional molecular fingerprints. Results showed that MDFP has the potential to be an alternative to traditional molecular fingerprints and the combination of MDFP and traditional molecular fingerprints can achieve higher prediction accuracy. Meanwhile, the accuracy of the best model, which was generated by consensus of four algorithms with multi-dimensional molecular fingerprints, was 0.694 (RMSE) in the test dataset. Besides, the number of hydrogen bonds from MDFP has been determined as a critical factor in the predicting models, followed by rgyr and sasa. Our findings provide a new sight of MDFP and multi-dimensional molecular fingerprints in building models of hERG cardiotoxicity prediction.