The Study On Atomic Distance-Dependent Pair-wise Statistical Potentials In Protein Structure Prediction

Protein three-dimensional structure prediction is one of the most popular and challenging topics in the field of bioinformatics. The existing methods can be roughly classified into two categories:template-based and template-free. Template-based structure prediction relies exclusively on already known structures of related proteins. We can find the most suitable structure by virtue of sequence or structure similarity, and then set it as a template to predict the structure of the target protein. However, template-free structure prediction is to predict three-dimensional protein structures directly by protein sequences through exploring and utilizing the physical and chemical laws in the protein folding. The latter has greater theoretical significance, and has always been regarded by western scholars as the'holy grail'of computational biology.For template-free structure prediction, the design and construction of energy function are the base and core of the whole work. According to extracting from already known structures or not, energy function can be divided into two types which are knowledge-based energy function and physics-based energy function. This thesis attempts to give a detailed discussion on a knowledge-based energy function named atomic distance-dependent pair wise statistical potential, and mainly focuses on its reference state. The primary work and achievements are as follows:(1) The key to designing statistical potential is what kind of reference state is used. After making comprehensive and deep investigation, we finally choose averaged, quasi-chemical approximation, finite ideal-gas, spherical noninteracting, atom-shuffle and random-walk chain reference states because of their great effectiveness. Then their characteristics and differences are analyzed and discussed in depth. Moreover their respective strengths and weaknesses are expounded emphatically.(2) A non-homologous set of 1022 crystallographic structures from protein data bank is selected to serve as statistical sample to construct six atomic distance-dependent pair wise statistical potentials with different reference states while other conditions being the same. The comparison diagrams indicate that the statistical potentials maintain a similar trend; subtle differences in the distribution may determine their performance characteristics.(3) With those statistical potentials being applied to kinds of decoy sets, the results show that the performance of statistical potentials under different decoy set is diverse. In other words, statistical potentials have preferences for specific application environment, which is a true inspiration. Maybe we can envisage the range of application at the beginning of potential design, while not being keen on its universal validity. Moreover, although the statistical potentials perform well in the native conformation detection, the results in decoy rank ordering demonstrate that the discriminatory powers of these statistical potentials are still open to further enhancing.