Study On The Prediction Of Drug Toxicity Based On The Molecular Structural Characteristic

Evaluation of drug toxicity is both the major content of preclinical evaluation of drug safety and an important subject closely relating with human health and drug fate. It has become the attention focus of drug administration organizations and pharmaceutical companies in the world. Study on how to evaluate the drug toxicity immediately, accurately and quickly in the early stage of drug development will contribute to shortening experiment cycle, saving money, increasing success frequency and protecting human health. For the laborious testing, huge investment, low sensitivity and the requirement of many animals, traditional toxicological experiments can hardly satisfy the need of high throughput screening in nowaday drug development. It is in very urgent need to develop new methods for toxicity evaluation. The fast development of pharmaceutical analysis promotes the progress of toxicology and provides rich experimental datum of toxicity. Computer sciences and informatics were employed to analyze and dig the existed experimental datum, the prediction of drug toxicity based on the molecular structural characteristic came into being, where the quantitative structure-toxicity relationship (QSTR) models were built. Due to the great advantages over the traditional toxicological experiment in that the toxicity prediction based on the molecular structural characteristic is fast, economic and independent of experimental conditions, it has been playing a more and more important role in the drug toxicity evaluation during the early stage of drug development.Aiming at the fact that most of the existing QSTR models have poor predictive potential, this dissertation developed a new systematic scheme to construct mechanistic model, mode of action model and statistical model for three different kinds of toxicity data sets, respectively. The aim of this scheme is to improve the predictive accuracy of QSTR models and contribute to the toxicity evaluation and high throughput screening of drug candidates in the early stage of drug development. The major results obtained from these studies have been outlined as follows:1. Quasi-mechanistic models were built for toxicity prediction. According to the influences of electronic and steric effects on toxicity, two novel molecular structural descriptors, lone-pair electrons index (LEI) and molecular volume index (MVI) were developed to characterize the electronic and steric effects, respectively. Based on these two structural descriptors, quasi-mechanistic models were built for three toxicity data sets. The results show that quasi-mechanistic models can predict the toxicity of chemicals accurately and have clear physic meanings. The performances of these QSTR models are superior to those reported by literatures.2. Mode of action (MOA)-based QSTR models were built for the prediction of acute toxicity in the fathead minnow. At first, the chemicals in the training set were clustered into different subsets according to the mode of toxic action. Each subset has only one kind of MOA. Then local QSTR models were constructed for each subset and global QSTR model was built for the whole training set with several MOAs. The predictive abilities of local models and global model were compared, and the influence of the reliability of MOA determination on the performance of local model was also investigated. It was indicated that the performances of local models based on the single MOA were improved largely over those of the global model.3. A new scheme called as "clustering first, and modeling then" was developed to build and validate the QSTR statistical models. The data sets on baseline toxicity with unknown toxic mechanism and mode of action were employed to validate this scheme. Local QSTR models were firstly constructed for the subsets resulted from clustering of the training set according to structural similarity. The compounds in the test set were classified into the corresponding subsets just as those of the training set, and then the prediction was carried out by the relevant local model for each subset. The predictive performances of local models were superior to those of global model, which confirmed the rationality and feasibility of the scheme of "clustering first, and modeling then".