Research On Protein Complex Detection Algorithms Based On Dynamic Protein-Protein Interaction Network

With the development of bioinformatics, computers are becoming indispensable tools inbiology research. By using the knowledge of machine learning and data mining computerprofessionals designed a series of related algorithms which greatly promoted the developmentof biology. Research shows that proteins rarely function in isolation in vivo, instead theytends to combined with other proteins and form protein complexes. Identification of proteincomplexes is of great importance in understanding of cellular organization and functions, andis becoming a hot research topic.Protein complex is a group of two or more function associated polypeptide chainscombined through disulfide bond or other proteins, and they participate in a number ofimportant activities like DNA transcription, mRNA translation, signal transduction, reactioncatalyzing and so on. Through immunization experiments biologist can achieve higher proteincomplex recognition accuracy, however, such tests are time consuming and have certainblindness. Therefore, at present, many scholars introduce using computer algorithms to reducecosts and improve accuracy.Recently, study of protein complex prediction has been greatly prompted, and manyclassical algorithms have been proposed, such as MCODE and G-N. Many of thosealgorithms detect protein complexes mainly based on the topology of protein-proteininteraction (PPI) networks but seldom considering biological properties of related proteins.Interactions between proteins are extremely complex and rapidly changing, simply based onthe static network is difficult to accurately identify the protein complex.In this paper, firstly, through reading related protein complex prediction research papers,we summarized the major categories and characteristics and currently used algorithms as wellas their advantages and disadvantages. Secondly, we proposed to construct dynamic PPInetworks, and use GO-Slim to annotate protein functions. Finally, based on weighted PPInetworks we proposed a novel protein complex prediction algorithm GECluster.Dynamic PPI network is constructed from static PPI network which is constructed byremoving unexpressed proteins using gene expression datasets. GO-Slim is a trim-downversion of GO, which can give three aspects annotation of a protein including molecule function, biological process and cellular component. By using GO-Slim we can pre-classifiedall the proteins. Weighted PPI network is a combination between dynamic PPI network andGO-Slim annotation results.In order to verify the prediction accuracy of GECluster algorithm, both training datasetsand test datasets were used. GECluster was applied to a variety of datasets and it predictedmore credible complexes than peer methods. The results indicate that using weighted PPInetwork can efficiently improve protein complex prediction accuracy.In this paper, we introduced dynamic PPI networks and weighted PPI networks. Indynamic PPI networks, only proteins translated are kept, others all deleted. Weighted PPInetwork constructed based on dynamic PPI network and GO-Slim annotations of proteins. Weuse GO Slims to annotate each protein in the dynamic PPI networks. By doing this, proteinswith similarly functions are almost fall into the same group. As a result, our method canachieve much higher accuracy. However, relying solely on the analysis of gene expressiondata to determine whether or not the proteins are expressed is not enough. In fact proteinstructure plays very important roles in protein complex formation, therefore, continuing toconduct the thorough research to further improve the prediction accuracy and provideresearchers with more reliable protein complexes is the focus of our future work.