【作者】 王强；

【导师】 王明会；

【作者基本信息】 中国科学技术大学 ， 生物医学工程， 2015， 硕士

【摘要】 可逆蛋白质磷酸化是一种重要的蛋白质翻译后修饰,它调节着许多生物细胞过程。因此,鉴定蛋白质底物与蛋白质激酶之间相互作用非常重要,它帮助我们理解调节过程和磷酸化机制。在本研究中,蛋白质底物与蛋白质激酶相互作用为磷酸化网络中的链路,因此相互作用的预测也就是链路的预测。鉴定蛋白质底物与蛋白质激酶之间作用关系常用方法是通过生物学实验,虽然其准确性高但效率低下,不能对大量的磷酸化数据进行快速有效鉴定。为了解决这一问题,机器学习方法被用来预测蛋白质底物与蛋白质激酶之间的相互作用关系。基于机器学习的方法用大量数据预测潜在的蛋白质底物与蛋白质激酶之间相互作用,然后通过实验去验证那些可能性较高的相互作用,从而可以减少鉴定的时间和提高工作效率。在本文中,已知的蛋白质底物与蛋白质激酶之间相互作用关系被用来构建磷酸化网络,在此基础上,蛋白质底物序列信息和蛋白质激酶序列信息也被加入进来,用于预测磷酸化网络中潜在的蛋白质底物与蛋白质激酶相互作用链路。首先利用蛋白质底物序列信息和磷酸化网络信息计算蛋白底物的相似性矩阵,然后把该相似性矩阵作为评价蛋白质底物与蛋白质激酶之间关系的特征,并把特征数据作为支持向量机的输入数据。这一过程被称为SVM-Net,该算法可用于预测磷酸化网络中蛋白底物与蛋白质激酶之间相互作用链路。除了SVM-Net算法,我们还设计了一种网络算法LapRLS-AWN,它是在拉普拉斯最小二乘法算法上发展而来。LapRLS-AWN算法通过蛋白质底物序列信息、蛋白质激酶信息和磷酸化网络信息预测磷酸化网络中蛋白质底物与蛋白质激酶相互作用链路,并通过AWN方法解决磷酸化网络中孤立蛋白质底物节点问题。在本文蛋白质底物与蛋白质激酶之间的链路预测性能分析中,我们首先把不同的信息作为算法的输入数据,并比较它们对预测性能的影响。结果显示蛋白质底物与蛋白质激酶间相互作用的磷酸化网络信息和蛋白质激酶序列信息对预测性能的提升有很大作用。其次,比较了SVM-Net和LapRLS-AWN与现有的BDT、 SVM的预测性能,结果显示其性能明显优于现有算法。最后,比较SVM-Net和LapRLS-AWN的预测结果显示,运用孤立节点处理方案和蛋白质激酶序列信息可以有效提高算法的预测性能。更多还原

【Abstract】 Reversible protein phosphorylation is an important post-translational modification, which regulates various biological cellar processes. Therefore, it is important to identify the protein substrate-protein kinase interactions, which are helpful for us to understand regulation processes and the phosphorylation mechanism. In this study, protein substrate-protein kinase interactions are links in phosphorylation network, therefore the interactions prediction are links prediction. Experiment method is the common method, which is used to identify interaction between protein substrate and protein kinase. Experiment method has high accuracy, however, it has low efficiency. In order to overcome the disadvantage of it, Machine learning method is used to predict the interactions between protein substrate and protein kinase. Machine learning method uses mass data to predict the potential interactions between protein substrate and protein kinase. Then, experiment-based method is used to verify those protein substrate-protein kinase interactions which have high probability. This procedure can reduce identification time and improve working efficiency.In this work, known interactions between protein substrate and protein kinase are used to construct phosphorylation network. Based on phosphorylation network, protein substrate sequence information and protein kinase sequence information are included to predict the potential interaction links between protein substrate and protein kinase in phosphorylation network. First, protein substrate sequence information and phosphorylation network information are used to calculate protein substrate similarity matrix, which is used as feature data for substrate-kinase interaction prediction. And this feature data is used as input data for Support Vector Machine and then an algorithm SVM-Net was developed. SVM-Net is used to predict interaction links between protein substrate and protein kinase in phosphorylation network.Except for SVM-Net algorithm, we also developed a new network-based algorithm LapRLS-AWN, which is developed from Laplacian Regularized Least Squares algorithm. LapRLS-AWN algorithm uses protein substrate sequence information, protein kinase information and phosphorylation network information to predict interaction links between protein substrate and protein kinase in phosphorylation network. Furthermore, AWN method is also used to solve the isolated protein substrate node in phosphorylation network.In this study, we first use different information as input data for algorithms, and compare the performance of these algorithms. The result shows that phosphorylation network between protein substrate and protein kinase interaction and protein kinase sequence information are helpful for improving the prediction performance. Second, we compare the prediction performance of SVM-Net and LapRLS-AWN with BDT and SVM, the result shows that the prediction performance of SVM-Net and LapRLS-AWN is better than SVM and BDT. Finally, we compared the prediction performace of SVM-Net with LapRLS-AWN, the result shows that treatment of isolated nodes and protein substrate sequence information can improve the prediction result.更多还原