A deep learning-based model for linear B-cell epitope prediction

Accurate prediction of linear B-cell epitopes plays a key role in vaccine design, diagnosis and antibody production. Here we provide a series of models DLBEpitopex for linear B-cell epitopes prediction, in which x=11, 12, …, and 50 are corresponding to the potential epitope lengths 11, 12, …, and 50, respectively. Each model is composed of 11 classifiers developed using deep learning methods. The training datasets and test datasets were extracted from the IEDB database. The following figure demonstrates the relationship between potential epitope lengths and AUC values on the independent test datasets, which shows that our models have a much improvement in AUC values. When apply one model, e.g. DLBEpitope38, to predict a potential peptide, all 11 classifiers will be used and 11 values (one for epitope and zero for non-epitope) will be obtained. The sum of the 11 values will be taken as an index to indicate whether the potential peptide as an epitope or not. Obviously, if the sum is 11, the peptide has the strongest signal to be an epitope.

This work has been published in BioData Mining on April 17, 2020.