Prediction of B-cell epitopes using diverse class-conditional data selection policies

The rapid and accurate localisation of B cell epitopes on antigenic proteins is a major challenge. However, for diverse class-conditional data, the combination of multiple machine learning algorithms and single algorithm methods seems to have little effect because they have not considered long-tailed distributions of the data. We propose the prediction of continuous B-cell epitopes based on diverse class-conditional data selection policies. The method can generate a set of classifiers with complementary predictive capabilities. In the method, at a certain step, the binary classification is transformed into the four-class classification problem based on the original binary classifier and the training sample set. In our experiments, SVM and partial RNN with a single hidden layer were used to classify the epitope and non-epitope peptides. Four classifiers were trained for six window sizes such as 10, 12, 14, 16, 18 and 20 and obtained similar performance under our selection policy. In the other experiment, the maximum performance achieved by our method is 92.88%, while the other classifiers varied from 57.24% to 68.87%. The results suggest that a group of classifiers where adjacent classifiers have complementary predictive capabilities can improve epitope prediction. In particular, our selection policy of complementary classifiers could not only predict B cell epitopes, but also solve the whole classification problem.