Recurrent Neural Networks for Linear B-Epitope Prediction in Antigens

Abstract

Experimental methods used for characterizing epitopes that play a vital role in the development of peptide vaccines, in diagnosis of diseases, and also for allergy research are time consuming and need huge resources. There are many online epitope prediction tools that can help experimenters in short listing the candidate peptides. To predict B-cell epitopes in an antigenic sequence, Jordan recurrent neural network (JRNN) are found to be more successful. To train and test neural networks, 262.583 B epitopes are retrieved from IEDB database. 99.9% of these epitopes have lengths in the interval 6-25 amino acids. For each of these lengths, committees of 11 expert recurrent neural networks are trained. To train these experts alongside epitopes, non-epitopes are needed. Non-epitopes are created as random sequences of amino acids of the same length followed by a filtering process. To distinguish epitopes and non-epitopes, the votes of eleven experts are aggregated by majority vote. An overall accuracy of 97.23% is achieved. Then these experts are used to predict the linear b epitopes of antigen, ESAT6 (Tuberculosis).