Cardiotoxicity is damage to the heart by adverse effect of certain chemical compounds, which may cause malignant ventricular arrhythmias and even incur cardiac sudden death. Numbers of drugs were withdrawn from the market because of their cardiac toxic effects. Various in-silico Quantitative Structure-Activity/Property Relationship (QSAR/QSPR) models have been built to predict the cardiotoxicity during early drug designed stages. By combining the rapid accumulation of experimental results, it now being possible to build up more sophisticated models for highly accurate cardiotoxicity prediction. A novel molecular graph convolution neural network (MGCNN) model based on the latest constructed compound-assay database targeting at the hERG channel has been built in this study. The MGCNN model was generated by altering the number of convolution layer (CL) from 1 to 5. Random forest (RF) and Support Vector Machine (SVM) models assemble Extendedconnectivity fingerprint (ECFP) were built to have a direct comparison with the MCGNN model. The MGCNN model with 2 CLs achieved the best performance whereas the RF and SVM models with ECFP have a more stable performance over the preset radius 1 to 5.
These results suggested a novel potential approach for highly accurate cardiotoxicity prediction which can be applied in the early stage of drug development to improve the efficiency of drug development process and reduce the risk of drug failure and withdrawal.