Protein Complexes Identification Based On The Protein-Protein Interaction Networks

Since the beginning of System Biology, IT has become an indispensable means of biological research. The researchers major in Computer Science try to solve the biological problems with a different angles using data mining, machine learning and other computational methods. Specially, the research of Protein-Protein Interaction Networks (PPINs) is becoming the hotspot. Recearches try to predict protein function, identify protein complexes and detect functional modules from PPINs by computational methods.Starting from the analysis about the effect of realiability in PPINs on protein complexes identification, this study presented efficient protein complexes detection methods for biologists’demand. The main orinigal works include:For providing the basis for the rational design of protein complexes algorithm, this paper first quatify the reliablity of Protein-Protein Interactions(PPI) through utilizing the gene co-expression data, protein domain, gene ontology annotation and some other rich resources in bioinformatics. Then study the effect of noise data in PPINs on protein complexes identification by comparing the difference of fuzzy Naive Bayes model and Naive Bayes model in classifying protein complexes task.When having a known PPINs, but the PPINs are unchangeable, this paper proposed a semantic similarity method based on Gene Ontology for weigthing the edge in network, and designed a protein complexes detection method (OIIP) for weighted PPINs. Results shows that utilizing the Gene Ontology to weight PPINs reduces the effect of false positives. Comparing with the existing methods, OIIP got the highest F measure.When having a known PPINs, and the PPINs are unchangeable, this paper propsed a method incorporating rich bioinformatics resources to reconstruct PPINs based on machine learning. Results shows that the reconstructed PPINs has higher proportion of reliable PPI, and the perfromances of existing protein complexes identification method are all improved, particularly reflected in Precision and Recall.When having only a few known PPIs, this paper proposed a multi-layer network conversion method to construct PPINs for protein complexes identification. Results shows that it achieves higher performances than the state-of-the-art methods, particularly reflected in Recall and F measure, and moreover, the new constructed PPINs has higher propotion of interactions with biology significance. The performances of existing protein complexes detection method are also improved on the new constructed PPINs.In summary, this paper proposed effective protein complexes detection methods by solving the unreliability of PPINs under different conditions. Some new predicted protein complexes are all proved by biological significance from a statistical sense. They probably will inspire the biological experiments. Furthermore, the proposed PPINs constructed algorithms are also helpful for link-prediction in complex networks. This will be investigated in the future work.