Research Of Complex Detection Algorithms In Protein Interaction Network

As abundant of experiments in biomedical field have been taken, vast amount of protein interaction data are produced. All these interactions are collected and founded as some protein interaction databases. For the vast interactions can be used to make a complex network, researching on the network has been a hot topic in biomedical field, where the complex prediction and module structures detection are the most important issues. Protein complex is made up by two or more proteins, and these proteins work with each other, by which they can realize some biologic function or impact the biologic process. Therefore, the research on complex detection has most significance for the understanding of cellular organization and biological function. Moreover, the interaction network making by vast interaction data can provide a good basis for the auto complex detection algorithm.At first, the related complex detection technologies are been introduced, and the classical complex detection algorithms are showed briefly. Meanwhile, the insufficient of the existed algorithms is pointed, which will be solved in following contents. We analyze the structure and properties of the protein interaction network. For the modularity in the network, it proves that the complex detection in the network is viable. Aiming at the problems in the classical methods, we present two solutions, one is based on combining different cores from different protein interaction network and the other is based on the supervised learning method.The algorithm based on combining different cores is aiming at the problem that most complex detection algorithms are operated on only one network. Here, we adopt different algorithms of core detection in three different networks. Then, the cores are combined and filtered. After that, the attachment proteins are been checked and added to the cores. So, we will get the perfect predicted complexes.Finally, we provide an algorithm based on supervised learning, which solved the problem that unsupervised method couldn't use the full information of the complexes. In this method, different information is used as the features, including gene ontology information, weighted clustering coefficient and so on. Totally, there are eighteen features in the feature set finally. Three-classification Regression is used in our method. After training the train set, we get the supervised model, which can be used to judge the sub-graph a complex or not in the complex detection algorithm. Many comparison experiments have been taken, including the parameters comparison, the model comparison, the feature set comparison and compared with other methods, and all the results show that our method is effective in complex detection. In conclusion, two methods of complex detection are provided, which are algorithm based on the different cores combining and algorithm based on the supervised learning. The experiments have showed that the supervised method is effective in complex detection, and solved the complex detection problem perfectly.