Point Transformer-based B-cell Epitope Prediction Study

Immunoinformatics is an interdisciplinary discipline based on informatics and modern immunology established at the beginning of the 21 st century.It uses the laws of immunology and enables prediction,analysis and computer-aided vaccine design containing antigenicity based on relevant experimental results.Epitopes are composed of hydrophilic residues in antigenic proteins that can be specifically recognized by TCR and BCR and are the basis of protein antigenicity.Among them,B-cell epitopes can be specifically recognized by antibodies or BCRs and influence the response process of humoral immunity.Accurate identification of B-cell epitopes is a guide for designing epitope vaccines and establishing immunodiagnostic methods for diseases.The current biological methods for identifying B-cell epitopes include X-ray diffraction methods,bioenzymatic methods,and chemical cleavage methods.These methods are time-consuming,costly,and require professional researchers to operate,which are not conducive to rapid identification of B-cell epitopes.Therefore,the researchers proposed a computational-based B-cell epitope prediction method to assist biological experiments.Computational-based B-cell epitope prediction methods include both antigen sequence-based and structure-based methods.These methods have developed richer strategies to identify B-cell epitopes based on the physicochemical properties such as hydrophobicity and glycosylation of epitopes.However,the existing methods ignore the structural properties of epitopes and do not effectively utilize atomic-level features,resulting in their limited performance and generalization ability.In this paper,we propose a Point Transformer-based method for B-cell epitope prediction.The method first constructs point clouds at the atomic level,takes atomic physicochemical properties,structural information and amino acid evolution information as input features,and uses Point Transformer networks to construct models.Thanks to the incorporation of atomic-level features and the full utilization of structural features,the proposed method has an AP of 0.458 and an AUC of 0.716 on the test set and an AP of 0.276 and an AUC of 0.653 on the independent test set,which are both better than other existing classical B-cell epitope prediction methods.In addition,case studies on H1N1 virus antigenic proteins were also performed in this paper,and the results were also better than other tools.This also demonstrates the effectiveness of the B-cell epitope prediction framework proposed in this paper,which can provide some reference for the design of subsequent vaccines.