Prediction Of Bioactivity And Toxicity

Bioactivity and toxicity of drug candidates are the two major problems to be considered in drug development. For a drug candidate, it is more likely to be marketed if its bioactivity (i.e., the ability of activating or inhibiting the target) is high and its toxicity (i.e., cardiotoxicity, hepatotoxicity, muscular toxicity, etc.) is low. Computational methods, such as Kohonen’s self-organizing map (SOM), support vector machine (SVM), and hierarchical clustering analysis (HCA), can be used to investigate structure-activity relationships (SAR). Using these methods, this dissertation focuses on four major problems in drug development.(1) Classification of reactions catalyzed by hydrolases and oxidoreductases, respectively. The Enzyme Commission (EC) classification system is the traditional method for the classification of enzymatic reactions, which is on the basis of criteria, such as reaction types, substrates, acceptor groups, etc. In this study, basised on six descriptors of the breaking and making bonds of the reaction center,311reactions catalyzed by hydrolases and651reactions catalyzed by oxidoreductases were classified using SOM, SVM and HCA methods. Classification accuracies of more than90%were obtained, which were consistent with the EC classification system.(2) Classification of inhibitors and non-inhibitors against Src tyrosine protein kinase. The inhibition of Src kinase is a reasonable strategy for diseases, such as prostate and breast cancers.686ATP competitive inhibitors of Src kinase and1941non inhibitors were collected. On the basis of23selected molecular descriptors, inhibitors were distinguished from non-inhibitors using SOM and SVM methods, with classification accuracies of more than98%.(3) Prediction of myopathy and rhabdomyolysis. Myopathy and rhabdomyolysis are the rare but severe adverse effects of drugs such as statins.232compounds inducing myopathy,117compounds not inducing myopathy,186compounds inducing rhabdomyolysis, and117compounds not inducing rhabdomyolysis were collected. On the basis of selected molecular descriptors, toxic compounds were distinguished from non toxic compounds using SOM and SVM methods, with classification accuracies of more than80%. Furthermore, it was found that charge, electronegativity and polarizability related molecular descriptors are important for the prediction of myopathy and rhabdomyolysis. It was also found that some substructures may be related to the myopathy or rhabdomyolysis toxicities.(4) Systematic analysis and prediction of activity cliffs.Activity cliffs are structurally similar compounds having large difference of activity against the same target (at least2orders of magnitude). As so far, activity cliffs have mostly been defined on the basis of Tanimoto similarity values using molecular descriptors.On the basis of matched molecular pairs (MMP), the activity cliffs for621target sets, the corresponding transformation and the exchanged substructures were systematically analyzed. For MMP-cliff generation, the size and size difference of exchanged substructures were carefully restricted. It was found that most of the transformations inducing cliffs involved only small substructures of similar size.Then, using an SVM method, computational models of activity cliffs prediction were built for nine target sets. Three new SVM kernels (i.e., substructure difference kernel, substructure-pair kernel and MMP kernel) were designed and utilized for the prediction of activity cliffs. Good prediction results were obtained for the nine target sets.In summary, chemoinformatics approaches were utilized for the prediction of bioactivity and toxicity of compounds and good results were obtained.