A systematic evaluation framework for universal antibody-antigen binding affinity prediction and candidate...

Area of Science:

• Computational biology
• Immunology
• Drug discovery

Background:

• Accurate prediction of antigen-antibody binding is vital for drug discovery and protein engineering.
• Current binding predictors struggle with generalization, especially for novel antigens or datasets lacking non-binders.

Purpose of the Study:

• To establish a standardized benchmarking framework for evaluating universal antibody-antigen binding affinity prediction.
• To compare the generalization capabilities of sequence-based and structure-based prediction methods.

Main Methods:

• Developed a benchmarking framework using pairwise accuracy and retrieval metrics.
• Proposed MochiBind, a sequence-only pairwise binding affinity predictor.
• Compared MochiBind against structure-based models (Boltz-2, GeoDock, Graphinity).

Main Results:

• MochiBind achieved performance comparable or superior to structure-based methods in pairwise accuracy and retrieval.
• Sequence-only models demonstrate strong generalization capabilities across diverse antigens.
• The proposed benchmark facilitates fair comparison of different prediction approaches.

Conclusions:

• Sequence-based methods can match or surpass structure-based models in generalizing antibody-antigen binding affinity prediction.
• The developed benchmark and MochiBind model offer scalable, sequence-driven solutions for in silico screening in drug discovery.
• Future development can leverage this framework for more robust and generalizable binding prediction tools.