Study On Prediction Of Drug Toxicity

The issue of drug safety is contributed to about30%failure rate of drug discovery and development (DDD). Study on development of new methods with high predictive capability of toxicity is desirable for reducing the high cost of DDD, and it has become a hot topic for toxicology, pharmaceutical analysis, computational chemistry and systems biology, etc.Recently, traditional toxicological experiments are increasingly replaced by the structure based methods (e.g. Quantitative Structure-Activity Relationship, QSAR) and systems biology based methods (e.g. Toxicogenomics) due to their high cost, long cycle and the great demand of animals. For example, QSAR has gained wide application in the prediction of toxicity in the early phase of DDD because of its independence of experiment, high throughput and low cost. However, the prediction accuracy of QSAR models remains disappointingly low especially when the chemical structure is diverse or the toxicological mechanism of compound is different from the training set. With respect to the Toxicogenomics, it has got wide application, and models built on this technique can help us understand better the related toxic mechanism, but it has disadvantages in that it is expense in cost and the statibility of predictive models built under this technique is controversial.As a result, we have developed and improved several prediction methods, based on the QSAR and Toxicogenomics technique mentioned above. Considering the low prediction accuracy yielded by the available in silico models, which is a general problem in the area of computational toxicology, this study developed four new efficient methods for prediction of toxicity as following:1. An improved Decision Forest (IDF) was proposed for the prediction of drug toxicity. Based on two genomic data sets, IDF got higher accuracy than those of many popular methods, and it seems more stable, suggesting that it is promising in dealing with high dimensional data sets.2. Several ensemble methods were proposed based on SVM,kNN and NC (Nearest Centroids) algorithms. Results showed that, by using these ensemble methods, the accuracy can go up more than3%on classification problem, and the ensemble of SVM seems the best.3. A serial of prediction models were built to predict the hepatotoxicity by means of cross-tissue prediction based on rat’s blood genomic data sets. The models were also evaluated using three independent data sets. Results showed that the models are predictive of hepatotoxicity, with accuracy as high as92%. Moreover, we also found6genomic indicators from the blood genomic data that can serve as biomarkers of hepatotoxicity.4. A method based on SVM and Particle Swarm (PS) algorithm was proposed to identify whether a compound is P-gp substrate or not. It is of important to predict if a compound is P-gp substrate or not for the study of ADME/T properties of drug candidates. Results showed that a higher accuracy (-90%) was achieved comparing with the literature by using the proposed method, and the discriptors contained in our models are more interpretable.