Deep learning of antibody epitopes using positional permutation vectors

Abstract

BackgroundThe accurate computational prediction of B cell epitopes can vastly reduce the cost and time required for identifying potential epitope candidates for the design of vaccines and immunodiagnostics. However, current computational tools for B cell epitope prediction perform poorly and are not fit-for-purpose, and there remains enormous room for improvement and the need for superior prediction strategies. ResultsHere we propose a novel approach that improves B cell epitope prediction by encoding epitopes as binary positional permutation vectors that represent the position and structural properties of the amino acids within a protein antigen sequence that interact with an antibody. This approach supersedes the traditional method of defining epitopes as scores per amino acid on a protein sequence, where each score reflects each amino acids predicted probability of partaking in a B cell epitope antibody interaction. In addition to defining epitopes as binary positional permutation vectors, the approach also uses the 3D macrostructure features of the unbound protein structures, and in turn uses these features to train another deep learning model on the corresponding antibody-bound protein 3D structures. This enables the algorithm to learn the key structural and physiochemical features of the unbound protein and embedded epitope that initiate the antibody binding process helping to eliminate “induced fit” biases in the training data. We demonstrate that the strategy predicts B cell epitopes with improved accuracy compared to the existing tools. Additionally, we show that this approach reliably identifies the majority of experimentally verified epitopes on the spike protein of SARS-CoV-2 not seen by the model during training and generalizes in a very robust manner on dissimilar data not seen by the model during training. ConclusionsWith the approach described herein, a primary protein sequence and a query positional permutation vector encoding a putative epitope is sufficient to predict B cell epitopes in a reliable manner, potentially advancing the use of computational prediction of B cell epitopes in biomedical research applications.