DeepRank-Ab: a scoring function for antibody-antigen complexes based on geometric deep learning.

Area of Science:

• Structural biology
• Immunology
• Computational chemistry

Background:

• Accurate modeling of antibody-antigen interactions is crucial for understanding immune responses and designing therapeutics.
• Current computational methods, including AI-based approaches like AlphaFold3, face challenges in generating and ranking near-native conformations of these complexes.

Purpose of the Study:

• To develop a novel, highly accurate scoring function for antibody-antigen complex modeling.
• To address the limitations of existing methods in predicting the structural accuracy of antibody-antigen interfaces.

Main Methods:

• Developed DeepRank-Ab, a geometric deep learning-based scoring function specifically designed for antibody-antigen interfaces.
• Created a large, diverse benchmark dataset of approximately 2.3 million decoys from 1,442 complexes for training and evaluation.
• Systematically evaluated various graph representations, structural/energetic features, and sampling strategies, identifying atom-level representations with Voronoi-based surface decomposition and antibody-specific features as optimal.

Main Results:

• DeepRank-Ab consistently outperformed AlphaFold3, HADDOCK, and other state-of-the-art scoring functions across multiple independent test sets.
• Achieved a 35.5% increase in AlphaFold3's Top1 success rate and more than doubled the mean Top1 DockQ score.
• Demonstrated strong generalization capabilities, achieving a 100% Top5 success rate on external antibody-antigen CAPRI targets.

Conclusions:

• DeepRank-Ab represents a significant advancement in scoring antibody-antigen complex models.
• The method substantially improves the identification of near-native antibody-antigen conformations, offering a powerful tool for structural biology and drug discovery.