A support vector machine based pharmacodynamic prediction model for searching active fraction and ingredients of herbal medicine: Naodesheng prescription as an example

Abstract

The complex chemical composition of herbal medicine leads to the lack of appropriate method for identifying active compounds and optimizing the formulation of herbal medicine. One of the most commonly used method is bioassay-guided fractionation. However, if the herbal medicine was divided into many fractions, it would cost much money and time in carrying out such a full bioassay. So, can we just perform the bioassay of a few fractions, and then develop a method to predict the bioactivities of other fractions? This study is designed to try to answer the question. In this work, a support vector machine (SVM) pharmacodynamic prediction model was introduced to search active fraction and ingredients of Naodesheng prescription. The prescription was first divided into five extracts, yielding a total of 2 5 = 32 combinations. Anti-platelet aggregation experiment with SD rats was just carried out on 16 combinations. The effects of the remained 32 − 16 = 16 combinations were then predicted by the SVM model. The prediction quality was evaluated by both the rigorous jackknife test and the independent dataset validation test. Furthermore, the present method was compared with the frequently used MLR, PCR and PLSR. The present method outperforms the other 3 methods, yielding: RMSECV = 2.40, R = 0.895 by the jackknife test and RMSEP = 7.41, R = 0.910 by the independent dataset test. It indicates that the SVM prediction model has good accuracy and generalization ability. The active fraction and ingredients of Naodesheng prescription were then predicted by the model. It is believed that the present model can be extended to help search the active fraction and ingredients of other herbal medicines.