ARID-sf: A Physics-Informed Deep Learning Scoring Function to Improve Antibody-Antigen Docking Model Ranking

Area of Science:

• Computational biology
• Structural biology
• Immunoinformatics

Background:

• Accurate prediction of antibody-antigen (Ab-Ag) complexation is vital for immunology, diagnostics, and therapeutic antibody development.
• Current molecular docking scoring functions often fail to identify near-native poses for Ab-Ag interactions.

Purpose of the Study:

• To develop and validate a novel scoring function, ARID-sf (Antibody-antigen Residue Interface Docking scoring function), for improved Ab-Ag complexation prediction.
• To enhance the accuracy and generalization capabilities of computational methods in predicting antibody-antigen interactions.

Main Methods:

• ARID-sf combines classical force fields, structural features, and protein language model embeddings using a self-attention neural network.
• The model was trained on over 1.5 million docking models and rigorously evaluated on four independent test sets (806 cases, >700,000 models).

Main Results:

• ARID-sf demonstrated superior performance compared to existing scoring functions across various docking scenarios.
• The function maintained high accuracy even with diverse sequence identities and significant conformational changes from unbound states.
• ARID-sf is highly parallelizable, processing thousands of models per minute.

Conclusions:

• ARID-sf offers a robust and efficient solution for predicting antibody-antigen complexation, outperforming current methods.
• Its generalization capabilities make it suitable for computational antibody engineering and drug discovery pipelines.
• The freely available code and pipeline facilitate broader adoption and further research.