Studies On Several Issues In Informatics Of Pharmacotoxicogenomics

Pharmacotoxicogenomics, offers new technologies to drug safety assessment etc., extends the coverage of pharmaceutical sciences, and brings about new challenges to researchers in pharmaceutical analysis. Not a few issues, including the minimum training sample size estimation, applicability domain, and how to improve model performance etc., are still left unaddressed in class comparison and class prediction analyses based on microarrays, the main technology in pharmacotoxicogenomics. Thus we studied on several issues mentioned above in this thesis, with the aim to provide methodological guidance for the generalization of pharmacotoxicogenomics. The main contents and academic contributions of this thesis are summarized as follows:1. Using aflatoxin B1 as an example, we studied on the capability of microarrays in detecting liver toxicity based on class comparison analysis. It was found that microarrays can provide markers much more sensitive than that of traditional biochemical analysis and histopathology. The results might be of great help for microarray-based early detection of liver toxicity.2. An SSNR-based protocol for estimating the minimum number of training samples required in class prediction analysis was proposed. Compared to what has been published, the method proposed in this study was much easier and more pragmatic, and can estimate required sample size correctly.3. It was found that the issue of applicability domain may do not exist for microarray-based prediction models. In other words, model performance can not be improved by defining applicability domain. The results can provide invaluable guidance for class prediction analysis based on microarrays.4. An index-prediction confidence that can effectively estimate the performance of models on an individual sample was proposed in this study. Based on prediction confidence, we further developed a microarray-based individualized prediction protocol. It was found that model performance can be apparently improved by defining prediction confidence.5. Using toxicogenomic dataset obtained from 8 drugs as an example, the above findings about training sample size estimation, applicability domain and prediction confidence were applied to class prediction analysis in pharmacotoxicogenomics. The results confirmed the feasibility of using pharmacotoxicogenomics in drug safety assessment. Moreover, the satisfactory comparability of different microarray platforms in respect of liver toxicity prediction was further ascertained, which laid the foundation for meta-analysis of datasets profiled from different microarray platforms.