In Silico Prediction Of Drug Induced Cardiotoxicity By Machine Learning Approaches And Its Application

ObjectiveDrug induced adverse cardiac reaction is one of the most common drug toxicity and accounts for the withdrawal on use of multitudinous post-market drugs.And side effects especially cardiotoxicity aroused by TCM also have been reported widely.Thus,it is meaningful to assess drug cardiotoxicity and predict TCM for cardiotoxicity alerting.MethodsIn this study,we developed cardiotoxicity prediction models to discriminate cardiotoxic compounds from non-cardiotoxic ones by machine learning approaches.Firstly,cardiotoxic compounds were collected from the adverse drug reaction database,and subsequently the corresponding non-cardiotoxic compounds were generated to construct the training set and the test set.Secondly,we used MOE(version,2009)and PaDEL-Descriptor(version,2.21)softwares to calculate and optimize descriptors,and three machine learning algorithms including Logistic regressiont,Random Forest and k-nearest neighbors were applied to develop six models.Besides,in order to improve the predictive capability of the models,three types of fingerprints(EState,MACCS and SubFP)descriptors were further introduced on the basis of the two best classification models which we chose,and six new classification models were rebuilt.Finally,the two best models were utilized to predict the ingredients of 20 Traditional Chinese Medicines with clear reported cardiotoxicity.ResultsAfter systematic evaluation,PaDEL_LFS_LR+MACCS(PLL+MACCS)and PaDEL_LFS_LR+SubFP(PLL+SubFP)fingerprints were considered as the two most reliable models.Six compounds(Linoleic acid,palmitic acid,stearic acid,beta-sitosterol,sitosterol,oleic acid)of the chemical constituentscontained in 20 Chinese herbs were worthy of further study.ConclusionIn this study,in silico models exhibiting higher sensitivity and specicity for drug induced cardiotoxicity to discriminate cardiotoxic compounds from non-cardiotoxic ones were developed,and PLL+MACCS and PLL+SubFP performed best.Subsequently,the two best models were utilized to predict the 708 ingredients of 20 TCMs with clear reported cardiotoxicity.Prediction results pointed out that six compounds(Linoleic acid,palmitic acid,stearic acid,beta-sitosterol,sitosterol,oleic acid)of the chemical constituents contained in 20 Chinese herbs were worthy of further study.In general,this study can yet be regarded as a reliable and accurate tool for discovering novel cardiotoxicity of compounds in silico prediction,which can provide an effective reference for preclinical cardiotoxicity evaluation of drugs.