Prediction of Linear B-cell Epitopes Based on Multi-Cluster Feature Selection

The identification of B-cell epitopes is crucial for the design and development of immunodiagnostic kits and vaccines. Existing epitope prediction models require feature preprocessing. This requirement is ”arbitrary” and may reduce the recognition accuracy for the epitope prediction model. In this work, we equip the model training with suitable preprocessing strategy, “Multi-Cluster Feature Selection (MCFS)”, to eliminate the above requirement. The new learning network structure, called BCELM, can improve the linear B-cell epitope classification. On the Lbtope_Fixed_non_redundant(LF) dataset, we demonstrate that MCFS boosts the accuracy of a variety of machine learning architectures despite their different designs. With MCFS’ correlations measuring capabilities, BCELM achieves outstanding classification results using the physico-chemical properties and position entropy of amino acids. In processing test independent antigen proteins, our method is more accurate and effective than the comparison model.