Algorithm Research On Protein Structure Analysis

Protein structure analysis is one of the most challenging problems remaining in protein science field. Many fatal diseases are caused by protein stucture disturption; therefore the protein structure analsis can have large effect on medication development. Protein folding and protein domain prediction are two important problems in protein structure analysis. Protein folding studies focus on the structure dynamics of protein peptide, while protien domain predictions focus on the static structure. However, this two research fields both consider one central problem: how protein sequence determine the tertiary structure of protein. Recently, more and more computer science methods and technologies have been applied in the protein structure analysis field, such as signal process, distribution computation, data mining and machine learning.First, we present a wavelet approach for analysis of folding trajectory of protein Trp-cage. In this new method, in order to filter the noisy information, the folding trajectory datas are decompesed and reconstructed for nine times with the 'DB2' wavelet. The results show that our wavelet method can extract the folding intermediate states and folding events in folding trajectory effectively. By applying this new method, we have a better understanding of the folding mechanism as well as the limitation of the current force fields. Then we analyze the local contact in the unfolding trajectory of wild type and mutant lysozyme protein with the wavelet approach. Our results show that the single mutant, W62G, can indeed cause the protein lysozyme to lose some key long-range interactions and become less stable, and we also find that the ARG-TRP-ARG 'sandwich' strucutre of mutant site, TRP62, plays an important role in these long-range interactions.In this paper, we also present a BP network based method to predict the boundary of two-domain protein. In this method, the BP network has 169 nodes in input layer, 5 nodes in hidden layer and 1 node in output layer, and the network is trained with Levenberg - Marquardt algorithm. This method archives good result on two datasets, and proposes the relative importance of the input feature descriptors. By including a new descriptor PBWLI, we present another BP network based method for multi-domain protein. This network has 99 nodes in input layer, 9 nodes in hidden layer and 1 node in output layer. The prediction results show that our method can effectively predict the boundaries of multi-domain protein to some extent.