Statistical Potentials And Their Application In Protein Structure Prediction

Protein structure prediction has been one of the hotspot in biology and is the most promising way to alleviate the huge gap between the number of sequences and protein structures, thus of great importance for the structural and functional studies of protein. As a popular prediction method, Ab initio protein structure prediction is mainly composed of two parts, that is, potential function and conformational sampling algorithm. Among them, Potential function is one of the constricting problems in the protein structure prediction field. Previous experiments showed that the traditional physics-based potential functions perform poorly, while the statistics-based potentials win their honor both in faster computation as well as better accuracy. Although extensive studies have been done on this topic, the current prevalent statistical potentials are still far from satisfactory. On one hand, the most commonly used distance dependent potential does not take into consideration of the real restraints imposed by peptide environment when choosing the critical reference state, which in turn hinders the improvement of accuracy. On the other hand, there is large scale of experimental protein structure data publicly available, which make it possible to design more complicated higher dimensional potential.Therefore, wefirstly tried to use the unfolded state as reference in constructing our own distance dependent statistical potential SPOUSE. Due to the more information about the basic property of polypeptide included, as well as the devoid of interaction-specific properties, SPOUSE not only theoretically unifies the protein structure prediction and the protein folding, but also elevates the performance compared with its counterparts, which is illustrated in this thesis.After that, the author further optimized the distance dependent statistical potential, and proposed a novel multidimensional orientation-dependent statistical potential name ORDER_AVE, by incorporating both distance and angular information. Through taking into consideration of the mutlibody effect, ORDER_AVE shows considerable improvement compared with SPOUSE. Also, ORDER_AVE outperforms all its analogs with higher recognition accuracy.Meanwhile, several other statistical potentials, including Soft-core Van der Vaals, hydrogen bond, burial energy, β strand packing energy and contact energy, are proposed and benchmarked their influence on prediction accuracy for three Proteins, in different folds. These potential functions, are integrated into the final prediction software, and total energy of the system are the weighted sum of those energies.Finally, the author participated in designing several conformational algorithms based on previous work, and implemented a novel protein structure prediction package using C++.Preliminary results show that our program works well with all α protein and α/β protein.Therefore, the author not only propsed two high performance statistical potentials, but also will play important role in the related fields, based on the program we wrote.