Application Of Support Vector Machine (SVM) For Prediction Of Drug Metabolism And Drug Inhibitory Activity

The support vector machine (SVM) is a new method of data mining, which has been widely applied in many fields. Using this method we constructed several computational model of CYP3A4 inhibition based on the experimental IC50 values of 42 inhibitors and few molecular descriptors represented the physiochemical and structural property of compound. These models were further evaluated by the cross validation and the independent test set. The model with the highest predictive ability reached the cross validated correlation coefficient r2 of 0.92 and the predictive coefficient r2 of test set of 0.73. The predictive ability of model obtained using SVM is better compared with the models derived from the ANN and PLS methods. The results indicate that SVM can also be severed as a promising statistic tool for quantitative structure- property/ activity relationship (QSPR/QSAR) studies. We also built some model to predict the Km values of P450 substrates and a 3D-QSAR model for a series of pyrimidine derivatives. We selected hundreds of Km values of different P450 enzyme substrates from many references. We used different regression methods to build QSAR models and compare SVM models with the others. Malaria kills over 1 million people annually. It was reported that the Plasmodium falciparum methionine aminopeptidase 1b inhibitors could be potential antimalaria agents. This three dimensional quantitative structure activity relationship(3D-QSAR) model for a series of pyrimidine derivatives was developed to gain insights into design for potential new antimalaria agents. It is found the Comparative Molecular Field Analysis (CoMFA) method gives good results while the Comparative Molecular Similarity Indices Analysis(CoMSIA) is less satisfactory. The CoMFA method yields a cross–validated correlation coefficient q2 =0.617, non-cross-validated R2=0.975, SE(Standard error)=0.086 and the F ratio=94.518. It indicates that the model possesses a high predictability. The model we build may be of value in facilitating design of new potent antimalaria agents.