Conformation Space Optimization Theory And Method For Ab Initio Protein Structure Prediction

Biological cells include a number of proteins.These macromolecules play an important role in living organisms.The function of a protein is determined by its spatial structure.Therefore,predicting the spatial structure of proteins is a prerequisite for understanding their biological functions and realizing gene therapy for protein folding disease.The 3D structures of proteins can be experimentally predicted by X-ray crystallography and nuclear magnetic.However,these methods are usually time-consuming and costly.Hence,predicting the protein structure from the amino acid sequence using the computer optimization method is one of the important research topics in bioinformatics.Ab initio prediction is one of the most active research topics in the field of protein structure prediction.According to the Anfinsen's dogma,the method directly searches the conformation with the global minimum of energy in the conformation space according to the physics or the knowledge-based model using the optimization algorithm.However,due to the essential complexity of protein conformational space optimization,it becomes a challenging research topic in the field of protein structure ab initio prediction.In order to find the unique natural structure in the large conformation space,it is necessary to design an efficient conformation optimization algorithm to convert the problem into a practical computing problem.In this paper,we firstly propose two abstract convex underestimate-based differential evolution(DE)approaches in the framework of evolutionary algorithms according to the abstract convex theory.Based on these two algorithms,a protein conformation space optimization algorithm using abstract convex underestimate-based multimutaion strategy is presented to slove the problem of single-domain protein conformation space optimization.In addition,a multi-domain protein domain structure assembly method is designed for the structure prediction of multi-domain proteins.The main works of this thesis are summarized as follows:(1)A differential evolution using the local abstract convex underestimation is proposed to reduce the computational cost of differential evoluation for the computationally expensive problem.In the proposed approach,the local abstract convex underestimate model is constructed based on the neighbor individuals of the trial.Then the underestimated information is extracted from the underestimate model to guide the population updating.According to the piecewise linear geometric characteristics of the underestimate model,some invalid search region is effectively excluded from the search space.In addition,the generalized descent direction of the abstract convex supporting hyperplane is employed for local enhancement.The computational cost is reduced,and the search efficiency is enhanced by using the above strategies.The experimental results show that the proposed algorithm is superior to the main-stream algorithms in terms of the computational cost,reliability,and convergence speed.(2)An abstract convex underestimate-assisted multistage differential evolution is proposed to enhance the search efficiency.The local abstract convex underestimate model is firstly constructed to obtain the underestimation of the objective function.According to the variation of the average underestimation error,the evolutionary process is divided into three stages.The different mutation strategy pools are set up for the each stage based on their characteristics.A suitable mutation strategy is automatically selected from the corresponding pool during the search process.In addition,the centroid-based mutation strategy is designed to balance the population divisity and convergence speed in the second stage.Various numerical experiments are performed to demonstrate that the proposed algorithm is effectively and superiorly.(3)A protein conformation space optimization method using abstract convex underestimatebased multimutaion strategy is proposed according to the above two algorithms.For each target conformation,a variety of different strategies are used to simultaneously generate multiple trial conformations.Several conformations related to the trial conformation are selected to contruct the abstract convex underestimate model of the energy function.According to the underestimation of the energy,the best trial conformation is chosen as the offspring conformation.The cooperation of different mutation strategy and the distance-profile assisted conformation selection can generate superior conformations and enhance the search capacity.Experimental results of various test proteins show that the proposed algorithm has a strong exploration capability and can obtain high quality 3D structures of proteins.(4)A multi-domain protein structure assembly algorithm is proposed to solve the problem of multi domain protein structure prediction.Based on the idea of "divide-and-conquer",the amino acid sequence is firstly divided into multiple domains,and then the structure of each domain is independtly predicted.All the domains are assembled together to get the full length structure of the multi-domain protein.In the assembly,the initial structure is obtained through the template identified by the structural alignment.Then the energy function is designed to guide the Monte Carlo simulation to search the conformation with the lowest energy by sampling the rigid body degrees of freedom of the domains.Finally,the Linker between domains is rebuilt to get the final model of the query protein.The multi-domain proteins with different types and number of domains are applied to verify the effectiveness of the proposed method.