PEPNet: a two-stage point cloud framework with hierarchical embedding and antigen-antibody interaction modeling for...

Area of Science:

• Computational immunology
• Structural bioinformatics
• Machine learning in drug discovery

Background:

• Epitope prediction is crucial for immunology and antibody design.
• Current methods using residue-level graphs miss atomic geometric details vital for antibody-antigen recognition.

Purpose of the Study:

• To develop a novel computational framework for epitope prediction using atomic-level protein representations.
• To enhance the accuracy and robustness of epitope prediction models by incorporating fine-grained structural information.

Main Methods:

• Modeled proteins as atomic-level point clouds to preserve high-resolution spatial features.
• Developed PEPNet, a two-stage point cloud framework with residue-aware hierarchical embedding and rotary positional encoding.
• Employed BERT-style pretraining with data augmentation and a cross-attention decoder for antigen-antibody interaction modeling.

Main Results:

• PEPNet achieved superior performance with MCC = 0.401 and AUC = 0.765.
• Demonstrated strong robustness on AlphaFold3-predicted structures (MCC = 0.346), outperforming existing methods like WALLE (MCC = 0.305).

Conclusions:

• PEPNet's atomic-level point cloud approach significantly advances epitope prediction accuracy.
• The model shows potential for practical applications in antibody-antigen analysis and therapeutic antibody design.