基于流形学习的蛋白质功能预测与优化

发布时间：2018-09-14 10:02

【摘要】：后基因组时代中,随着高通量实验技术的快速发展,大量的蛋白质数据被收集起来。然而,蛋白质数据与功能标注数据之间的差距却在不断变大。即使如酵母菌这样得到广泛研究的物种,其仍有近四分之一的蛋白质功能无法确定。因此,从计算角度设计出高效的蛋白质功能的自动标注方法成为生物信息学领域的重要挑战之一。此外,由高通量实验方法或计算预测方法获得的蛋白质功能标注数据包含了较高比例的假阳性和假阴性噪声,严重影响了与蛋白质功能标注相关的生物、医疗的应用效果。在本论文中,依据蛋白质相互作用网络的拓扑结构、流形学习方法和图理论知识,我们提出了三种有效的计算方法用来解决蛋白质功能的自动预测以及功能标注数据中的噪声问题。全文的主要工作概括如下:(1)针对蛋白质功能的自动标注问题,提出了一种新的整合流形学习和多标签学习的蛋白质功能预测框架。首先,利用边介数对蛋白质相互作用网络进行加权处理。然后,利用等度规特征映射(ISOMAP)算法将该加权网络嵌入到低维表示空间中,从而获得蛋白质数据的低维特征表示;最后,将蛋白质功能预测转化成经典地多标签学习问题,并且能够采用多种多标签学习方法进行蛋白质功能的预测与评估工作。实验结果表明,提出的方法能够取得了更加合理的蛋白质低维特征表示,并且相比于其他对比方法取得了更加准确的预测精度。(2)提出一种鲁棒的融合功能相关性的多标签线性回归方法来预测蛋白质的功能。首先,采用基于流形学习的ISOMAP算法将边介数加权的蛋白质相互作用网络嵌入到低维子空间中。然后,根据蛋白质低维数据的分布特点,将线性回归理论扩展到多标签情境中,通过余弦相似性计算蛋白质功能标签之间的相似性,并将其作为规则项加入到多标签线性回归模型的目标函数中。最后,评估了提出的算法在酵母菌数据库上的有效性。实验结果表明,提出的方法相比于其他现有的方法实现了更加令人满意的预测性能。(3)为了解决蛋白质功能标注数据中包含大量噪声的问题,提出了一种基于图规则化l1-范数的主成分分析法(Gl1PCA)进行蛋白质功能优化。首先,该方法通过蛋白质相互作用网络与功能相似性矩阵分别构建了一个蛋白质图与一个功能图。然后,将蛋白质图与功能图经过拉普拉斯变化后以规则项的形式被整合到了l1-范数的主成分分析法(l1PCA)的目标函数之中。最后,给出了该优化模型的一种基于增广拉格朗日乘子法(ALM)的快速解法,并利用理论证明与优化实验验证了提出的算法的正确性。实验结果表明,提出的算法能够有效的优化蛋白质的功能标注数据。
[Abstract]:In the post-genome era, with the rapid development of high-throughput experimental techniques, a large number of protein data have been collected. However, the gap between protein data and functional labeling data is widening. Even in species widely studied, such as yeasts, nearly a quarter of their protein functions remain uncertain. It is one of the most important challenges in bioinformatics to design efficient automatic annotation methods for protein functions from computational perspective. In addition, high-throughput experimental methods or computational prediction methods contain a high proportion of false-positive and false-negative noises, which seriously affect the protein function annotation phase. In this paper, based on the topological structure of protein-protein interaction network, manifold learning method and graph theory, we propose three effective computational methods to solve the problem of automatic prediction of protein function and noise in function labeling data. 1) To solve the problem of automatic annotation of protein function, a new framework for protein function prediction based on integrated manifold learning and multi-label learning is proposed. First, the protein-protein interaction network is weighted by the edge median. Then, the weighted network is embedded into the low-dimensional representation space by the ISOMAP algorithm. Finally, the protein function prediction is transformed into a classical multi-label learning problem, and many multi-label learning methods can be used to predict and evaluate protein function. The experimental results show that the proposed method can achieve a more reasonable protein low-dimensional feature representation. (2) A robust multi-label linear regression method based on functional correlation is proposed to predict protein function. Firstly, an ISOMAP algorithm based on manifold learning is used to embed the edge-median weighted protein-protein interaction network into a low-dimensional subspace. Secondly, according to the distribution characteristics of low-dimensional protein data, the linear regression theory is extended to multi-label situation, and the similarity between protein functional labels is calculated by cosine similarity, which is added as a rule term to the objective function of multi-label linear regression model. Finally, the yeast count of the proposed algorithm is evaluated. Experimental results show that the proposed method achieves better predictive performance than other existing methods. (3) In order to solve the problem of large amount of noise in protein function annotation data, a graph regularized L1-norm based principal component analysis (Gl1PCA) is proposed to optimize protein function. Firstly, a protein graph and a function graph are constructed by the protein interaction network and the function similarity matrix respectively. Then, the protein graph and the function graph are integrated into the objective function of L1-norm principal component analysis (l1PCA) in the form of regular terms after Laplace transformation. A fast algorithm based on augmented Lagrange multiplier (ALM) is proposed to solve the optimization model. The validity of the proposed algorithm is verified by theoretical proof and optimization experiments. The experimental results show that the proposed algorithm can effectively optimize the protein functional annotation data.
【学位授予单位】：安徽大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：Q51;TP181

【相似文献】