Prediction of continuous B-cell epitopes in an antigen using recurrent neural network

Area of Science:

• Bioinformatics
• Immunology
• Computational Biology

Background:

• B-cell epitopes are critical for peptide vaccines, disease diagnostics, and allergy research.
• Experimental epitope characterization is resource-intensive and time-consuming.
• Computational methods can accelerate epitope identification.

Purpose of the Study:

• To evaluate feed-forward neural networks (FNN) and recurrent neural networks (RNN) for B-cell epitope prediction.
• To identify optimal network parameters for accurate epitope prediction.
• To develop a computational tool for B-cell epitope prediction.

Main Methods:

• Training and testing FNN and RNN models on a dataset of 700 B-cell epitopes and 700 non-epitopes.
• Utilizing the Bcipep and Swiss-Prot databases for dataset creation.
• Optimizing network parameters including input window length and hidden units.

Main Results:

• Recurrent neural network (Jordan network) achieved higher prediction accuracy than FNN.
• Optimal performance was observed with an RNN having a single hidden layer, 35 hidden units, and a window length of 16.
• The final RNN model yielded an overall prediction accuracy of 65.93% with sensitivity, specificity, and positive prediction values of 67.14%, 64.71%, and 65.61%, respectively.

Conclusions:

• RNNs are more effective than FNNs for predicting B-cell epitopes from antigenic sequences.
• Peptide length significantly influences prediction accuracy.
• The ABCpred webserver provides a freely accessible tool for B-cell epitope prediction.