Abstract
Epitope prediction is a key challenge in immunology and therapeutic antibody design. Existing computational methods rely on residue-level graph representations that fail to capture fine-grained atomic-level geometric information essential for antibody-antigen recognition. Considering that protein structure files (e.g. Protein Data Bank (PDB) files) inherently contain 3D atomic coordinates, we model proteins as atomic-level point clouds to directly preserve high-resolution spatial features . Building on this representation, we propose Point cloud-based Epitope Prediction Network(PEPNet), a two-stage point cloud framework for epitope prediction. Inspired by the natural atom-to-residue hierarchy in proteins, PEPNet employs a residue-aware hierarchical embedding module to aggregate atomic features into residue-level embeddings. To capture sequential dependencies absent in unordered point clouds, we integrate rotary positional encoding. Additionally, PEPNet leverages a BERT-style pretraining strategy with data augmentation to mitigate data scarcity, and a cross-attention decoder to explicitly model antigen-antibody interactions. Experimental results show that PEPNet achieves the best overall performance (MCC = 0.401, AUC = 0.765). Even when evaluated on AlphaFold3-predicted structures, PEPNet maintains strong robustness (MCC = 0.346), still outperforming WALLE (MCC = 0.305). These results underscore PEPNet's potential for real-world antibody-antigen analysis and design.