AbAgKer: A Unified Semi-Supervised Framework for Antigen-Antibody Binding Affinity and Kinetics Prediction

Area of Science:

• Computational biology
• Artificial intelligence in drug discovery
• Protein engineering

Background:

• Generative AI accelerates therapeutic antibody design, but challenges persist.
• Scarcity of high-quality activity data and antibody structural flexibility hinder candidate evaluation.

Purpose of the Study:

• To develop AbAgKer, a novel antibody screening model for predicting antigen-antibody affinity and kinetics.
• To address data scarcity and improve the precision of therapeutic antibody candidate evaluation.

Main Methods:

• A biological prior-guided feature fusion framework integrating epitope knowledge and CDR-specific attention.
• A mixture-of-experts architecture for capturing complex binding landscapes.
• Semi-supervised learning for data self-distillation to enhance affinity prediction.

Main Results:

• AbAgKer significantly improves affinity prediction performance, outperforming baseline models.
• Learned interaction representations are transferable to few-shot learning for predicting dissociation rates.
• The model demonstrates strong generalization capabilities in antibody screening and drug residence time analysis.

Conclusions:

• AbAgKer effectively addresses challenges in antibody screening and drug discovery.
• The model offers a robust solution for predicting antibody binding properties with limited data.
• AbAgKer advances the high-throughput design and evaluation of therapeutic antibody candidates.