ABCpred: Prediction of Continuous B-Cell Epitopes in an Antigen Using Recurrent Neural Network

Dataset for Continuous B-Cell Epitope Prediction

Overview

This repository contains the curated dataset used in our study on the prediction of continuous B-cell epitopes using machine learning approaches, specifically recurrent neural networks (RNNs).

The dataset was originally developed for the identification and prediction of linear (continuous) B-cell epitopes in antigenic protein sequences. This resource may be useful for researchers working in:

Immunoinformatics

Vaccine design

Antibody epitope prediction

Computational immunology

Machine learning-based peptide classification

Reference

Saha S, Raghava GP. Prediction of continuous B-cell epitopes in an antigen using recurrent neural network. Proteins. 2006 Oct 1;65(1):40-8. doi: 10.1002/prot.21078. PMID: 16894596. href="https://onlinelibrary.wiley.com/doi/10.1002/prot.21078"

Dataset Description

The dataset consists of experimentally validated B-cell epitopes and negative peptide samples.

Positive Dataset

Contains 700 non-redundant experimentally validated continuous B-cell epitopes

Collected from the Bcipep database

Only epitopes of length ≤ 20 amino acids were considered

Redundant sequences were removed to reduce bias

Negative Dataset

Contains 700 non-epitope peptide sequences

Randomly generated from Swiss-Prot proteins

Any sequence identical to known epitopes was removed

Thus, the final benchmark dataset contains:

Dataset Type
Number of Sequences

B-cell Epitopes
700

Non-Epitopes
700

Total
1400

Data Processing

To create fixed-length patterns suitable for neural network training:

Variable-length epitopes were normalized to fixed window lengths

Neighboring residues from the parent antigen sequence were added when needed

Multiple window sizes (10, 12, 14, 16, 18, and 20 residues) were evaluated

The best performance was achieved with:

Window Length: 16 residues

Model: Recurrent Neural Network (Jordan Network)

Applications

This dataset can be used for:

Training machine learning/deep learning models

Benchmarking epitope prediction tools

Feature engineering on peptide sequences

Comparative studies with modern protein language models