蛋白质结构功能预测中若干关键算法的分析比较

发布时间：2019-02-21 15:44

【摘要】：随着测序技术的快速发展,蛋白质序列数量与已测定结构与功能的蛋白质数量的差距越来越大,迫切需要通过理论计算方法实现蛋白质结构功能的预测。目前,许多有效的方法被提出来研究蛋白质序列、结构和功能之间的关系,但不同方法在解决蛋白质结构功能研究中具有偏好性。因此,本文主要围绕蛋白质结构功能研究中方法展开,系统地比较分析了不同的特征提取方法、特征挑选方法和预测算法在蛋白质结构类、蛋白质紊乱、蛋白质分子伴侣、蛋白质溶解度和RNA结合蛋白质的预测中效率。主要研究内容如下:1、简要介绍了蛋白质研究的研究背景及意义、蛋白质的组成、结构和物理化学性质,并简述了常用的数据库及本文采用的数据集,为本文的研究提供了理论和数据基础。2、分析比较了蛋白质结构功能预测中氨基酸约化和特征提取方法。根据522种氨基酸性质将20种氨基酸约化成k类,提取蛋白质6类不同信息,结合支持向量机比较分析了氨基酸约化与信息提取方法在蛋白质结构功能预测中的效率。结果表明,在蛋白质结构类和蛋白质分子伴侣预测中,最好采用氨基酸的转向倾向类性质约化20种氨基酸,再提取蛋白质的顺序特征,而蛋白质溶解度的预测则偏向于蛋白质的RCTD特征提取方法。3、分析比较了蛋白质结构功能预测中特征挑选方法。本章选取了基于互信息的特征挑选方法、基于支持向量机的特征挑选方法等16种,结合K近邻预测算法比较分析了特征挑选方法在蛋白质结构功能预测中的效率。结果表明,基于非线性支持向量机的特征挑选方法在蛋白质结构类预测、蛋白质溶解度预测、蛋白质分子伴侣预测和蛋白质溶解度预测中表现最好,经过挑选后特征的准确率提升了13.16%-71%,尤其是蛋白质的k-mer特征和PSSM特征。4、分析比较了蛋白质结构功能预测中预测算法。本章选取了线性判别分析算法、主成分分析判别算法等7种预测算法,并比较分析了不同预测算法在蛋白质结构功能中效率。结果表明,在蛋白质结构类预测中,SVM预测算法表现最好,尤其与蛋白质PRseAAC特征结合,预测准确率达到99.15%;选择PCADA、CART、PLSDA、KNN或者SVM算法可以较准确地预测蛋白质的分子伴侣;在蛋白质紊乱预测中,KNN预测算法与蛋白质RCTD特征结合表现最好,准确率达到了94.75%;蛋白质溶解度预测应选取PSSM特征,结合PLSDA和PCADA预测算法;而在预测RNA结合的蛋白质时,采用GO特征和CART算法的组合或者GO特征和PLSDA算法的组合,都能获得较好的预测准确率。
[Abstract]:With the rapid development of sequencing technology, the gap between the number of protein sequences and the number of proteins with measured structure and function is increasing. Therefore, it is urgent to realize the prediction of protein structure and function by theoretical calculation. At present, many effective methods have been proposed to study the relationship among protein sequence, structure and function, but different methods have a preference in solving the problem of protein structure and function. Therefore, this paper mainly focuses on the research methods of protein structure and function, and systematically compares and analyzes different feature extraction methods, feature selection methods and prediction algorithms in protein structure class, protein disorder, protein molecular chaperone. Protein solubility and RNA binding protein prediction efficiency. The main research contents are as follows: 1. The background and significance of protein research, the composition, structure and physicochemical properties of protein are briefly introduced, and the commonly used databases and the data sets used in this paper are briefly introduced. 2. The methods of amino acid reduction and feature extraction in the prediction of protein structure and function were analyzed and compared. According to the properties of 522 amino acids, 20 kinds of amino acids were reduced to k class, and 6 kinds of different information of protein were extracted. The efficiency of amino acid reduction and information extraction methods in predicting protein structure and function was compared and analyzed with support vector machine (SVM). The results showed that in the prediction of protein structure and protein molecular chaperone, it was better to use the conversion tendency of amino acids to reduce 20 kinds of amino acids, and then extract the sequence characteristics of proteins. However, the prediction of protein solubility tends to the RCTD feature extraction method of protein. 3. The feature selection method in protein structure and function prediction is analyzed and compared. In this chapter, 16 feature selection methods based on mutual information and support vector machine are selected, and the efficiency of feature selection in protein structure and function prediction is compared with K-nearest neighbor prediction algorithm. The results show that the feature selection method based on nonlinear support vector machine performs best in protein structure prediction, protein solubility prediction, protein molecular chaperone prediction and protein solubility prediction. The accuracy of the selected features is improved by 13.16- 71, especially the k-mer features and PSSM features of proteins. 4. The prediction algorithms in the prediction of protein structure and function are analyzed and compared. In this chapter, seven prediction algorithms, such as linear discriminant analysis (LDA) and principal component analysis (PCA), are selected, and the efficiency of different prediction algorithms in protein structure and function is compared and analyzed. The results show that the SVM prediction algorithm is the best in protein structure class prediction, especially combined with protein PRseAAC features, and the prediction accuracy is 99.15. The molecular chaperones of proteins can be predicted accurately by choosing PCADA,CART,PLSDA,KNN or SVM algorithms, and in the prediction of protein disturbance, the combination of KNN prediction algorithm with protein RCTD features is the best, and the accuracy is 94.75%. In predicting protein solubility, PSSM features should be selected, combined with PLSDA and PCADA prediction algorithms, while the combination of GO feature and CART algorithm or GO feature and PLSDA algorithm can obtain better prediction accuracy when predicting RNA bound proteins.
【学位授予单位】：浙江理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：Q51

【相似文献】