Development Of COX-2 Inhibitor Models Using Machine Learning Methods

In this thesis, Artificial Neural Network(ANN) and Support Vector Machine(SVM) are applied to develop the inhibitor model for COX-2 inhibitors. This thesis consists of three parts. In the first chapter, a brief introduction to the background of computer-aided drug design is given and in the second chapter, the fundamental theory of ANN and SVM are briefly described. In the third chapter, ANN and SVM are applied to the development of inhibitor model for selective COX-2 inhibitors.In the first chapter, the importance and the cun'ent trend of the computer-aided drug design are given and the applications of machine learning methods, such as ANN and SVM, to the computer-aided drug design are introduced.In the second chapter, the fundamental theory of ANN and SVM is described and the molecular descriptors used to characterize the molecular structure in this thesis are given. The methods for feature preprocessing and feature selection are also described here.The third chapter is the major part. In this chapter, two Machine learning methods, ANN and SVM, are employed to model the discrimination between the activity and inactivity for COX-2 inhibitors. The data set with 476 COX-2 inhibitors is splitted into training set, test set and validation set using Kennard-Stone method. For each molecule, 459 molecular descriptor including constitutional descriptors and topological descriptors are calculated. To eliminate the descriptors of low information content, any descriptor are removed if satisfied one of the following conditions: (1) Variance or standard variance is less than 0.05; (2) Pearson's correlation coefficient with any other descriptor is above 0.90; (3) Identical value content is above 90ï¼…. Then, the remained descriptors are ranked in order of decreasing importance using the Fischer-Score as the criterion and the first 136 molecular descriptors are selected through maximization of the generation ability of SVM model. The selected descriptors are used in ANN and SVM, respectively, to develop the inhibitor models for COX-2 inhibitors. It is shown that ANN method outperforms the SVM method before feature selection and SVM method outperforms ANN method after feature selection and the final prediction accuracy of our best model is superior to the literature reported results.