應用支援向量機解蛋白質雙硫鍵預測及藥物結構活性量化回歸模型建構

Abstract

Support Vector Machine (SVM) is widely adopted in the field of machine learning and pattern recognition, and recently the application of SVM techniques to bioinformatics is also very promising. In this dissertation, we applied SVM to two important issues in bioinformatics: protein disulfide connectivity prediction and quantitative-structure activity relationship (QSAR) model construction. For disulfide connectivity prediction, we implemented an algorithm which infers pair-wise bonding probability by SVM, and introduced a descriptor which derived from the sequential distance between oxidized cysteines (DOC). From the analysis of prediction, it revealed that the prediction accuracy is improved with the addition of this descriptor DOC. Furthermore, we developed a two-level prediction model to integrate protein local and global information. The experimental results showed that the prediction accuracy is greatly enhanced. These results are compared with those of previous studies, and a prediction web-service is also provided on the internet. For QSAR model construction, we developed an approach to build QSAR models by selecting the hypothetical descriptor pharmacophore (HDP) with generic evolutionary method (GEM) and correlating the descriptors to activities with SVM. Experimental results of 5 public datasets indicated that our approach is comparable to those of previous studies. Additionally, we incorporated k-means and hierarchical clustering methods to cluster compounds into subsets and construct specific QSAR model for each cluster. The experimental results show that compounds with particular structural features are successfully clustered into the same subset, and the prediction accuracy was enhanced using specific models build by these clusters.