AbstractsComputer Science

There are four versions for the disulfide prediction problem, including chain classification, bonding state of each cysteine, connectivity of each cysteines pair and disulfide bonding pattern. Among these problems, the prediction of disulfide bonding pattern is the most difficult, especially for the proteins with four or more disulfide bonds. In that case, a large amount of training proteins should be collected in order to obtain a reliable prediction model. In this thesis, we propose a novel algorithm to solve the disulfide bonding pattern problem for the proteins that the bonding states of cysteines are given. Our method first predicts the labels of cysteines with global information, and then uses the label prediction results to predict the whole disulfide bonding pattern. As the experimental results show, our method achieves a higher accuracy than other previous methods for given the same training dataset. Furthermore, to improve the accuracy of our method, we use the CSP method and the BKS table to build our hybrid model, which was proposed by Chen et al.