Machine Learning Methods For B-cell Epitope Prediction Based On Side-chain Information Of Protein

B-cell epitope prediction is a major field in the immunology research. B cell epitopes is a group of characteristic sites which are distributed on the surface of antigens and can be recognized by the antigen binding sites of antibodies or B cell receptors.The identifying of B cell epitopes are very useful and meaningful for diagnostics, therapeutics and vaccines and so on. B-cell epitope may be continuous or discontinuous.Linear B cell epitopes are on the surface of peptide chains and form a continuous sequence of amino acid residues, while the discontinuous epitopes comprise atoms from distant residues but close in three-dimensional space. Now we have a lot of results in continuous epitope prediction,but they are very limited for most of the B-cell epitopes are discontinuous. The research of discontinuous B-epitope started late; however develop rapidly in recent years and some algorithms and software platforms are available now.In this paper,we proposed a method about the prediction of discontinuous epitopes based on a patch strategy.We adopt six features of amino acid’s side-chain information,and use support vector machine to do the classification and then to predict candidate epitopes.Our method has been verified on the data set contained161data.With the Non-complex antigen data set,we can get a sensitivity of53.6%,the precision is20.5%and the value of AUC is0.627;with the complex structure data set,the value of AUC is down to0.609,but the value of sensitivity, precision raised a bit.Compared to EPCES,our method get a slight better result.The results show that the combination of machine learning and feature selection methods for epitope prediction is effective.But on the other hand,overall prediction performance is still not satisfactory.To establish more comprehensive epitope data sets, find or mix more efficient epitope features are important to improve the prediction performance.