Computation Alanalysis Of Protein Interaction Networks

Protein participates and controls almost all the activities inside a cell. It is afundamental problem to understand the functions of proteins for further and insightfulinterpretations of biological systems. With the development of Information technology,computational analysis of Protein-Protein Interaction (PPI) networks provides vitalmeans of detecting and identifying protein functions. The computational methodsprovide the fundamental knowledge about the living system in a cell, and offerassociate information for biologists and biomedicians. This work focuses on twoaspects, including the module detection in static PPI networks and the critical proteinsidentification in dynamic PPI networks.Functional module detection in static PPI networksFunctional module detection in static PPI networks has faced with the stubbornproblems of incomplete information and high noise. The recent development of high-throughput experimental techniques provides a variety of biological data fordenoising PPI data and improving the accuracy of functional analysis. Based on theresearch of traditional algorithms of protein functional module detection, this workproposes to solve these problems in two different ways:1. A graph based full-information-representation model for functional moduledetection is developed on the output level of multiple data source integration. Intraditional data fusion methods on the output level, the correlational information in theoriginal data is dropped in the ensembling model, which induces inaccurate finalclustering results. To address this problem, a new method based on afull-information-representation graph model is proposed, which comprehensivelyrepresents the relationships between the objects in this problem, including the proteinsand the base clusters from the basic cluster methods. The graph partition model in thismethod is easily explained in a probabilistic fashion by which the overlappingproteins are well addressed. Extensive experimental results show that our method issuperior to the baseline methods; further analysis is addressed to discuss the benefitsof integrating multiple biological information sources and diverse clustering methods.2. A collective Non-negative Matrix Factorization method for finding thefunctional modules in static PPI network is proposed which is built on a modelintegration level. In this method, the integration problem is reduced to optimizingapproximations of multi-view data by the productions of their common matrix factorwith basis matrices. As a result, the common matrix factor provides an intuitiveinterpretation of soft clustering. Extensive experiments show that the proposedmethod outperforms most of the baseline methods listed in the paper and is effectivein extracting functional modules in PPI networks.Dynamic PPI network characteristicsThe detection of critical proteins and critical protein modules in dynamic PPI networks are essential for lots of biomedical industry like disease diagnosis, medicinedesign, and health care. However, the traditional methods of static PPI networkanalysis is not appropriate for dynamic PPI network analysis; just a small group ofproteins, rather than the majority, play more essential roles at crucial points ofbiological processes; and the complexity of dynamic PPI networks are exacerbated bythe changing PPIs. In the study of dynamic PPI network, the main research contentsand innovations are:1. Critical protein detection in dynamic PPI networksIn this thesis, a comprehensive way of modeling the dynamic PPIs is presentedwhich simultaneously analyzes the activity of proteins and assembles the dynamicco-regulation correlation between proteins at each time point. By comparing withtraditional algorithms, the proposed algorithm gives out a more accurate descriptionof dynamic PPIs. Based on the constructed dynamic PPI networks, a novel criticalprotein detection method is proposed, which models a common representation ofmultiple PPI networks using a multiple-source based Deep Belief Network framework.By analyzing the reconstruction errors, the critical proteins with severe variabilitiesacross the time courses are extracted. Experiments were implemented on data of yeastcell cycles. The ranking results of critical proteins in msiDBN were compared withthe results from the baseline methods and the results of comparison showed thatmsiDBN had better reconstruction rate and identified more proteins of critical value toyeast cell cycle process.2. Co-regulated protein functional modules with varying activities in dynamicPPI networksBased on the dynamic PPI networks constructed by the above method, thedynamic active modules are detected using a method based on the Bayesian graphicalmodel and the modules with the most varied dispersion of clustering coefficients,which could be responsible for the dynamic mechanism of the cell cycle, areidentified. Comparison of results of our proposed method with the state-of-artfunctional module detection methods and validations of the ranking of activities offunctional modules using GO annotations demonstrate the efficacy of our method fornarrowing the scope of possible essential responding modules that will providemultiple targets for biologists to further experimentally validate.