Predict The ADME/T Properties Of Drugs Using Supporting Vector Machine Based Method

Pharmacokinetic characteristic of drug, namely the absorption, distribution, metabolism and excretion (ADME) together with toxicity, is a crucial factor for a drug stepping into market successfully. Traditionally, such an important characteristic was not got more than enough notice, as a result, the probability of drug candidates successfully hitting market is less than one-tenth. Therefore it is necessary to evaluate ADME/T properties of drug candidates at the beginning of the drug research and development so as to reduce elimination rate, and reduce the cost of drug development. Due to the great difficulty of gaining the ADME/T properties through high throughput screening experiment, it must be a valuable way to predict those properties by the aid of computer modeling.Nowadays several kinds of methods based on computer modeling have been employed to forecast the ADME/T properties of drugs. These methods include Quantitative Structure-Activity Relationships (QSAR), Quantitative Structure-Property Relationships (QSPR) and so on. Another relatively new method, the Supporting Vector Machine (SVM) method, has also be applied broadly in many fields attributed to its excellent performance in dealing with mode recognizing related to small sample, non-linear or high dimensions.In this paper, SVM classification method combined with Genetic Algorithm (GA) was employed to predict the ADME/T properties. It was shown that GA is an effective algorithm to remove redundant descriptors and increase the calculation power of the model while screening variations. At the same time, the calculated results indicated that our SVM classification method combined with GA is superior to the methods of other researchers in forecasting the ADME/T properties. It can be concluded that SVM classification method combined with GA is a more effective one in predicting those properties.In the first chapter, the theory of SVM was introduced. The two important parameters (C and y) of the kernel function of SVM were determined by the aid of grid-search based algorithm fulfilled with the help of C language. The classification and calculation of descriptors, the principal of selecting variable using GA and concrete steps of setting parameters were also presented here.In the second chapter, the capability of SVM classification method combined with GA to identify p-glycoprotein substrates was tested. The data sets and descriptors same with that of other researchers were selected. The descriptors were further screened using GA algorithm.The 5-fold cross-validation was employed to assess the classification model and the grid-search method was used to identify the two important parameters C and y of the kernel function of SVM in the model training process. The calculated results were compared with that of other researchers who used SVM classification method combined with Recursive Feature Elimination (RFE) to build their model. It was the SVM classification method combined with GA that could be used to enhance the accuracy of prediction, reducing the number of descriptors and improving the speed of predictionIn the third chapter, the blood-brain barrier penetrating of drugs was predicted using the method suggested by us and the results were compared with that of other researchers who got them by statistical methods. Molecular descriptors are selected by GA and the number of descriptors was reduced from 37 to 17. The overall accuracy of our model is much better.In the last two chapters, our research method was further employed to investigate two systems which had never been studied by others using the similar way. These two systems are drugs binding to Human Serum Albumin (HSA) and the carcinogenic activity of drug. The overall accuracy of forecast is greater than 80.0%.