Prediction Of Binding Affinity Of Human Transporter Associated With Antigen Processing

Processing and presentation of major histocompatibility complex class I antigens are very important for immune surveillance. The generation of CTL epitope from an antigenic sequence is complex and involves a mass of intracellular processes. The endogenous antigen peptides is first degraded into fragments of varying lengths by protease in the cytoplasm, transported into ER by TAP, then inserted into the binding groove of MHC class I molecules, and expressed on the cell surface by cellular efflux system for CD8+ cytotoxic T-lymphocytes receptor to recognize and binds into a triplet, in order to start the immune response. The transporter associated with antigen processing, a transmembrane protein, responsible for the transportation of antigenic peptides into the endoplasmic reticulum (ER), plays a crucial role in the processing and presentation. TAP binding preferences impact significantly on CTL epitope selection. This article has developed a novel method to make quantitative prediction of the binding affinity of nonamer peptides with human TAP. The amino acids sites and physical chemical properties affecting the binding affinity were analyzed and their biological meaning was explained.The main innovations and results in this paper:(1) Binding affinity associates with a large number of physicochemical properties, It chose 15 kinds of properties of 20 amino acids as the basis for the model. Through the machine learning methods, this paper received the more important amino acid positions and the physical and chemistry attributes effecting affinity of human TAP with 9 peptides.(2) For the antigenic 9 peptides, It used the 15 features initial coding scheme. On this basis, it used the machine learning method to choose the top 15 crucial dimensions for binding affinity. Combined comprehensive consideration of the other minor dimensions with statistical methods of principal component analysis, it further constructed three new and different schemes for encoding the input vectors.(3) The data set was divided into training set, validation set and test set. For each coding scheme, using support vector regression machine and artificial neural network as predict engine, experiments were carried out on the independent sets by holdout method and been compared to explain according to test results with training model, optimizing the parameters, independent testing. Among these coding schemes, the third one obtained the best result:the Pearson correlation coefficient test by support vector machine is r=0.9029; the cross validation correlation coefficient q2=0.8068. By artificial neural network r=0.8547,q2=0.6985.(4) 5-folds cross-validation was implemented on the entire data set to cross-training and testing. The optimal parameters were obtained, and test results were analyzed. The support vector regression machine test results of entire date set were:r=0.8225; q2= 0.6697. The artificial neural network test results were:r=0.9417, q2=0.8852. Thus it proved that the prediction technique was reliable and feasibility.(5) According to the model test results, this paper analyzed the corresponding biological meaning and proposed the further research directions.