| Proteins play key roles in various aspects of life by physically interacting with other proteins.Protein-protein interactions(PPIs)are the molecular basis for many biological processes,such as signal transduction,transport,metabolism,gene expression,growth and proliferation of cells.Protein-protein interaction sites are critical domains for selective recognition of molecules and for the formation of complexes.Identification of protein-protein interaction sites is pivotal for understanding protein function,elucidating signal transduction networks and drug design.Experimental methods such as NMR and X-ray crystallography have been used to characterize PPI sites.But these techniques are expensive and time-consuming.With the fast expansion of resolved sequences and structural data of proteins,several kinds of computational methods such as molecular dynamics methods and machine learning methods were proposed to predict PPI sites.However,currently there are few computational methods for PPI sites prediction,and the results obtained by these methods are not ideal.Exploring better methods for PPI sites prediction can play a better role in promoting better understanding of protein interaction and drug design.The main work of this thesis is as follows:1.PPI sites prediction is a binary classification problem,we use the protein-protein docking benchmark data set(DBD,version 5.0),and analyze the limitations of the definition of existing data sets,then we propose an improved data sets method based on co-evolution and statistical methods,we use the relative abundance of interacting residues(RAIR)to improve the original data set.2.We extract features of the improved sample data sets,and map the feature vectors to the high-dimensional space,then input them into the convolutional neural network for training.We evaluate our model through the area under receiver operating characteristic curv(AUC)area and the first rank of the first positive prediction(RFPP)by using leave-one-complex-out-cross validation and five-fold cross validation.We also compare our model with other PPI sites predictors,our approach has made significant improvement in AUC score(0.912)and some of the RFPP values(RFPP(100)=580).The model was applied to the complex between hemopexin and hemopexin binding protein,the AUC achieved 0.87,which reflect that our model has better robustness.The goal of this thesis is to predict PPI sites.We proposed a novel statistics-based method for judging the binding propensity of amino acids and apply it to the partitioning of samples.The convolutional neural network is used to train the improved sample data set.Compared with other five predictors,our model obtains better prediction results,and provides a reference for promoting computational methods to predict protein interaction sites and drug design. |
