Computational Modeling Researches On Protein Three Dimensional Structures

Proteins are among the most important macromolecules in living organisms. They are involved in almost all biological activities at molecular levels. The functions of proteins molecules often intimately depend on their three-dimensional structures. Increasing attentions are paid to the determination and interpretation of protein structures(for example, the international structure genomics program has been put forward in 2000). The gap between our capability to determine protein structures by experiments and the number of protein molecules without known structures are wide. To narrow this gap, and perhaps more importantly, to understand the principles underlying protein folding, many molecular modeling methods are being developed for structure prediction. In the past decade, the special blind protein structure prediction experiments — CASPs have been held for 6 rounds, which showed the current status of protein structure prediction researches. Though great progresses have been made during the past decade, it is still far from solving the problem.The ab initio protein structure prediction methods are an important class of methods,. Although currently such methods are still far from practical applications, in a certain sense successful and reproducible ab initio predictions of protein structures should be the ultimate test of our understanding of protein folding. Ab initio structure predictions are generally based on three components: an appropriate model to represent protein structures, an accurate potential energy function at the same at the same level of coarse-graining as the model, and a sampling method to efficiently search the conformation space of proteins.In the first part of our work, we have developed an efficient conformation sampling algorithm using the genetic algorithm combined with local minimization and with a niche technique. To represent the three-dimensional structure of proteins in a simplified manner, only the main-chain heavy atoms and Cp are considered, and the main chain torsion angles are represented using 6 discrete states. Two different classes of potential energy functions are used in the simulations: one is a simple Go-type potential function, in which the native state is the only global minimum ; the other is the knowledge-based, pair-wise atoms potential function DFIRE. The former energy function was employed to benchmark the efficiency of the conformation search method, and the latter was employed to explore the capability and limitations of current knowledge-based potentials in ab initio protein folding. We have run the...