Comparative Study Of Several Key Algorithms In Protein Structure And Function Prediction

With the rapid development of Next-generation sequencing technology,the gap between the protein sequences and their structures/functions is becoming larger and larger.Therefore,it is urgent necessary for us to predict protein structure and function through computational methods.Now,several effective methods have been proposed to analyze protein sequence,structures and functions.But different methods have advantage in different problems of protein structures and function.Therefore,this dissertation mainly focuses on the key algorithms in protein structure and function research,and systematically compared their performances in protein structure prediction,protein disorder prediction,protein chaperone prediction,protein solubility prediction and RNA-binding prediction.The main work is summarized as follows:1.We briefly introduced the background protein research,the composition,structures and physicochemical properties of the proteins.Then,we summarized several protein databases and introduced standard datasets used in this dissertation,which provides powerful theoretical foundation for the following research.2.We compared amino acid reduction and feature extraction methods in protein structure and function prediction.Based on the 522 amino acid properties,we reduced 20 amino acids into k classes and extracted six kinds of protein features.With help of support vector machine(SVM),we compared the efficiencies of the amino acid reduction and the feature extraction.The results indicates that the prediction of protein structure and protein chaperone performs better when using alpha and turn propensities to reduce the amino acids,and RCTD feature performs better in the prediction of protein solubility.3.We compared feature selection methods in protein structure and function prediction.This dissertation chose 16 features selection methods,such as nine feature selection methods based on mutual information and two feature selection methods.We then compared the efficiency of the feature selection methods in protein structure and function prediction with help of the K-Nearest Neighbor(KNN)prediction algorithm.The results indicates that the nonlinear support vector machine performs better in the prediction of protein structure,chaperone and solubility,and improves the prediction accuracies by 13.16%-71%,especially for the k-mer and PSSM features.4.We compared prediction methods in protein structure and function prediction.In this chapter,we used seven prediction algorithms,such as linear discriminant analysis algorithm,principal component analysis discrimination algorithm and so on,and compared their efficiency in protein structure and function prediction.The results indicates that SVM achieved the bestperformance in protein structure prediction,and the prediction accuracy reaches 99.15% when using PRseAAC features.PCADA,CART,PLSDA,KNN and SVM algorithm perform better in the chaperone prediction.As for the prediction of protein disorder,the combination of KNN and RCTD feature achieved the best performance with prediction accuracy 94.75%.The prediction of protein solubility will be better when using PLSDA and PCADA.As for the prediction of RNA-binding proteins,the combinations of GO and CART or GO and PLSDA achieved better performance.