[Quantitative structure-activity relationship model for prediction of cardiotoxicity of chemical components in traditional Chinese medicines].

OBJECTIVE: Some quantitative structure-activity relationship (QSAR) models have been developed to predict cardiac toxicity of drugs, which have limited predictive power due to based on hERG channel inhibition. The objective of this study was try to develop a QSAR model based on all kinds of cardiac adverse effects, and to predict the potential cardiotoxicity of chemical components in traditional Chinese medicines (TCM).

METHODS: In this study, the compounds data of all kinds of cardiac adverse reactions were selected as the training set. The QSAR models were constructed based on 1 109 compounds with cardiotoxicity and 789 compounds without cardiotoxicity, which were available from the Toxicity Reference Database (ToxRefDB) and Side Effect Resource (SIDER) database. The ADMET Predictor software was applied to calculate and to screen the molecular descriptors, and to construct the QSAR models using support vector machine (SVM) and artificial neural networks (ANN) algorithm, respectively. The models were optimized using compound-based 10-fold cross validation. Then, the predictive performance for the potential cardiotoxicity of chemical components in TCM were assessed using external validation by 19 components in TCM with cardiotoxicity and 10 components in TCM without cardiotoxicity.

RESULTS: A total of 220 molecular descriptors were selected for modeling, and the best model using SVM algorithm contained 87 molecular descriptors. The internal validation results showed that the predictive sensitivity, specificity, the Youden's index (YI) and the Matthews correlation coefficient (MCC) were 71%, 70%, 0.41, and 0.41, respectively. The best model constructed using ANN algorithm contained 13 neurons and 87 molecular descriptors. The internal validation results showed that the predictive sensitivity, specificity, the YI and the MCC were 78%, 77%, 0.54, and 0.54, respectively. Both models were validated using external validation by the same set of 29 chemical components in TCM with or without cardiotoxicity, which were not included in the training set. The predictive performances of SVM or ANN model were as follows, respectively: sensitivity 95%, 95%; specificity 40%, 60%; and accuracy 76%, 83%.

CONCLUSION: The predictive performance of the QSAR model using ANN algorithm was better than that of the model using SVM algorithm. The external validation study of 29 chemical components in TCM illustrated that the QSAR model was applicable for screening and predicting the potential cardiotoxicity of chemical components in TCM.