Development of convolutional neural network-based QSAR model to predict cardiotoxicity and principal component analysis, fingerprint analysis

Small chemicals that block a potassium ion channel result in a prolonged QT interval, which can have serious cardiotoxic effects and is a major factor in drug development failures. To develop the drug successfully, quantitative prediction of human-ether-a-go-go-related (hERG) blockers is essential for designing drug candidates without the risk of cardiotoxicity. We built a convolutional neural network (CNN)-based quantitative structure–activity relationships (QSAR) model to predict cardiotoxicity. The statistical parameters of mean squared error (MSE) were 0.001, the mean absolute error (MAE) was 0.016, and the correlation coefficient (Q2) was 0.99 for the training dataset. The MSE was 0.62, the MAE was 0.65, and the predicted correlation coefficient (R2) was 0.70 for the test dataset. Further, we explored principal component (PC) analysis, t-SNE, scaffold analysis, active cliff, fingerprint analysis and chemical analyses to identify molecular similarity. We discovered that adding an acidic oxygen/aliphatic oxygen (hydroxyl group) reduces hERG inhibition and increases lipophilicity. The fragments are furan, sulfonamide, methanesulfonamide, p-chlorophenyl, p-fluorophenyl, and ethyl(heptyl)amino groups increased the hERG risk. Finally, we conclude that the QSAR model in combination with the convolutional neural network (CNN) offers a potentially novel approach for quantitatively predicting the cardiotoxicity of drug candidates.