GCAT-BCE: A hybrid GCN-GAT framework for enhanced conformational B-cell epitope prediction

The prediction of conformational B-cell epitopes (BCEs) is crucial for vaccine development and therapeutic antibody design. However, reliable identification of BCEs remains challenging because epitope residues are spatially discontinuous and represent only a small fraction of antigen surface residues, leading to severe class imbalance and high false-positive rates. In this study, we propose GCAT-BCE, a hybrid graph neural network that integrates graph convolutional networks (GCN) and graph attention network (GAT) for conformational BCE prediction. To reduce prediction noise, buried residues are first removed through a relative solvent accessibility (RSA)-guided filtering strategy prior to graph construction. Then, GCAT-BCE leverages multi-modal residue-level features (including amino acid types, secondary structure, relative solvent accessibility, and epitope propensity) combined with three stacked GCN layers with residual connections to capture local spatial interactions, followed by a GAT layer to refine long-range residue dependencies. Comprehensive evaluations on two independent benchmark test sets comprising 15 and 45 antigens demonstrated that GCAT-BCE consistently outperformed state-of-the-art sequence-based and structure-based models. Notably, GCAT-BCE achieved the highest AUC-PR and AUCPR10% values, indicating superior capability in identifying true epitope residues among highly imbalanced samples. Furthermore, we evaluated the GCAT-BCE model on two newly curated independent test sets with 215 non-redundant antigens. The results demonstrated that GCAT-BCE consistently outperformed BepiPred-3.0, CALIBER, BIDpred, and CLBTope with particularly pronounced improvements in AUC-PR. On the RoBep_187 test set, GCAT-BCE achieved an AUC-PR of 0.392, approximately twice that of the second-ranked model, while on the PDB2526_28 test set it maintained the highest AUC-PR and AUCPR10% performance, highlighting its robust predictive capability for minority-class epitope residues.