GPU-accelerated Particle Swarm Optimization And Its Application In Protein Structure Prediction

With the completion of the genomic sequencing,an effective method must be developed to detect the spatial structure of a large quantity of redundant protein sequences.The determination of the spatial structure of those sequences will have a significant effect on the science such as biology and medicine.This paper has employed the ab initio prediction to predict the protein structure.Based on the presumption that the natural composition of the protein is most stable when the free energy remains at the lowest level,this method can transform PSP problem into function optimization problem.The key problem of the ab initio prediction is how to find a reliable model which can reflect the protein structure accurately and also develop an effective and accurate algorithm to work out the optimal value for the corresponding energy function.Hence,this paper has adopted the AB off-lattice model which can better reflect the structural feature of the protein.Moreover,the Particle Swarm Optimization(PSO)which is an intelligent optimization algorithm widely applied to predicting the protein structure in recent years has been used in this paper.The energy function of the model will be calculated,based on which the secondary protein structure can be predicted.Considering that PSO has the disadvantages of slow speed and low accuracy in calculating the protein energy function,this paper has offered the improvement strategy as follows.Firstly,with the aid of CLJDA,the global PSO and local PSO parallel based on GPU have been realized.The simulation experiment has been conducted to compare the PSO in the same CPU series through the basic testing function and protein sequence.It has been proven that the PSO based on GPU parallel can effectively reduce the time for calculation.By comparing the global PSO and local PSO based on GPU parallel,it has also proven that the local PSO can outperform the global PSO in terms of calculating speed and accuracy for the multi-modal function and protein energy function.Secondly,in order to solve the problem of the low accuracy of the PSO in calculating the protein energy function,this paper has proposed the hybrid learning strategy.A number of different learning strategies will be integrated into a strategy pool.During the process of iteration,the individual learning strategy will be adjusted on a real-time basis.More appropriate strategies in the strategy pool will be selected to instruct the individual to seek the global optimization.Besides retaining the parallel structure of the algorithm,this strategy can also improve the ability for global optimization during the calculation of protein energy function.As shown in the simulation experiment,the algorithm with the hybrid learning strategy can gain a lower energy value in calculating the protein energy function.