| Protein function is closely related to its spatial structure. Function prediction benefits a lot from good understand of three-dimensional protein structure similarity, which is therefore a major challenge in biological research. The research on protein structure similarity is pushed forward by either biochemistry method or geometric method, or the combination of both. Protein structure similarity which is the key for function recognition is commonly marked by various geometric features. Therefore, the dissertation presents a thorough research on protein structure partitioning model, statistical feature extraction, multi-scale feature of protein structure and geometric property of the structure distribution. The major work of the dissertation is as follow:(1) By evaluating feature extraction method of sector partitioning, a further study of surface point distribution and the uniformity in sector is presented. Thus an improved feature extraction by means of statistical method is introduced. The experimental result shows that the method proposed presents a better accuracy than conventional points counting scheme for sector partitioning. Hence the inner distribution of the sector, which indicates the detail of the protein structure, is meaningful to the classification and multi-scale property of protein structure is then validated.(2) Exploring the multi-scale property of protein structure, a multi-scale analysis by wavelet transformation for protein structure similarity comparision is proposed. By mapping the protein structure into one-dimension, wavelet transform is applied to generate features of various scales. The experiment shows the method performances better than conventional points counting scheme in the same dimension.(3) With a study of moment property of the structure distribution, 3d weight Krawtchouk moment is introduced to evaluate protein structure similarity. The experiment indicates the method serves a promising measure for protein structure similarity.
|
PDF Full Text Request