A Computational Study Of Drug Molecules Based On Chemoinformatics Methods

Increasingly large, diverse and publicly available data sets (especially in the field of medicinal chemistry) are being generated using various high-throughput screening, combinatorial chemistry, and parallel syntheses technologies. For large accumulations of these bioactive data, an important question considered is how to extract useful information from them. In the present dissertation, the chemoinformatics methods (the application of informatics methods to the solution of chemical problems) are proposed to address this issue.Drug design is the most important field for using chemoinformatics methods. There are several reasons:Firstly, the costs and time involved with drug discovery and development are great; Secondly, experimental methods recently introduced in the drug design process produce enormous amount of data that have to be analyzed.In this dissertation, we mainly focus on the applications of the classification, regression and combinational methods including three-dimensional quantitative structure-activity relationship, molecular docking and molecular dynamics simulations to address two issues in drug design:(1) Perform activity prediction in order to screening chemicals for further development into chemical probes with desirable characteristics;(2) Understand the machinery of biological processes between small molecules and their targets at the molecular level. The main content of this dissertation can be described as follows in detail:(1) Variable selection algorithm coupled with random forests is used to develop the classification models based on the RSV inhibitors and human β3-AR agonists. Both models present more than90%predictive accuracies for the external prediction sets, respectively, which can be used to screening the desirable compounds.(2) The Random Forest algorithm is firstly employed to the prediction of a series of190structural diverse FBPase inhibitors based on the Mold2molecular descriptors and the genetic algorithm is used to extract the proper subset of descriptors. Finally, the built model not only gives the satisfactory prediction, but also identifies the most important descriptors which play a crucial role in FBPase inhibitory activity. In addition, a genetic algorithm-support vector machine coupled approach is proposed for optimizing the2D molecular descriptor subset generated for a series of P2Y12antagonists, with the statistical performance and efficiency of the model being simultaneously enhanced by SVM kernel-based nonlinear projection. This model produces the r2cv of0.83and r2pred of0.81for the cross-validation and test sets, respectively, which should be helpful for screening, prediction and optimization of potential P2Y12antagonists prior to chemical synthesis in drug development.(3) Several reliable computational models in combination with three-dimensional quantitative structure-activity relationship, molecular docking and molecular dynamics, are developed based on P2Y12antagonists,5-HT6receptor ligands, and inhibitors of FIXa. MK-2as well as PKCθ, which not only exhibit satisfied statistics but also provide several possible mechanism interpretations from a molecular-level perspective. The models can offer a theoretical guide to the modification of drug molecules.In summary, both the proposed classification and regression methods yield the high predictive performances, which can be used to screening the potential drug molecules in early drug development. In addition, the agreement between three-dimensional quantitative structure-activity relationship, molecular docking and molecular dynamics simulations proves the rationality of the developed models, which provide an insight into some instructions for further synthesis of highly potent drug molecules.