Sequence-based Structural B-cell Epitope Prediction by Using Two Layer SVM Model and Association Rule Features

Background: Immune reaction is the most important defense mechanism for destroying invading pathogens in our body, and the epitope is the position of the antigen–antibody interaction on pathogenic proteins. Objective: The majority of epitopes are structural; however, the existing sequence-based predicting websites still have several methods to improve the predicting performance. Therefore, in this study, we used SVM as a machine learning tool to predict the epitope-based on protein sequences. Methods: Firstly, we built five SVM models in the first layer according to five features, including binary composition, position-specific scoring matrix, secondary structure, accessible surface area, and association rule, and then chose the patterns that exhibited the best performance in each model. Secondly, using the confidence score of the first-layer models as the input value for the SVM model in the second layer, that SVM model was integrated into the first-layer SVM models for improving the predicting accuracy. Results: The final prediction model was able to achieve up to 63% accuracy in predicting epitope results, and the predicting performance was better than that achieved by the existing predicting websites. Conclusion: Finally, a case study using a two-subunit cytochrome c oxidase of Paracoccus denitrificans was tested, achieving an accuracy of up to 66%.