Sequence and Functional Annotations-Based Prediction of Protein-Protein Interactions

摘要

Protein-protein interactions (PPIs) are the basis of most cellular processes and biological functions, and many computational methods complementary to experimental methods have been proposed for predicting PPIs. However, two kinds of typical computational methods, namely, sequence-based and prior-knowledge-driven ones, are mainly studied separately. Consequently, the latent trait that links and complements protein sequence and prior knowledge is overlooked. Most methods also suitably perform in small genomes like S. cerevisiae but perform less suitably in large genomes such as Homo sapiens. These models have poor generalization capability as well. Therefore, a novel, effective, robust, and highly reliable method of predicting PPIs must be developed. In this paper, a protein functional annotation-based method (FNM) was presented for the first time. The method incorporated protein essentiality, domain-domain interactions, gene ontology annotations, and sequence annotations. Different fusion strategies were then used to establish three methods (JFM, ELM, and WELM) that combined protein sequence features with functional annotation features in FNM. On both datasets of the small-genome S. cerevisiae and the large-genome Homo sapiens, fusion methods demonstrated a high success rate of classification than other sequence-based methods. Finally, a yeast and human protein interaction network was obtained, and cluster analysis showed that the PPI networks were biologically meaningful, indicating that they can be used for network mining research, such as detecting protein complexes and hub proteins, searching for disease-causing proteins and drug targets, etc.