SimBinder-IF: Structure-Aware Antibody Affinity Optimization via Efficient Preference Learning

Area of Science:

• Computational biology
• Protein engineering
• Immunoinformatics

Background:

• Antibody therapeutics require high-affinity target engagement for efficacy.
• Current protein language models (PLMs) and optimization methods are computationally expensive and lack explicit training for high affinity.
• Structure-aware and parameter-efficient antibody optimization methods are needed.

Purpose of the Study:

• To develop a structure-aware antibody optimization model that is parameter-efficient and improves binding affinity.
• To evaluate the performance of the proposed model against existing methods in predicting antibody-antigen interactions.

Main Methods:

• Proposed SimBinder-IF, a structure-aware model freezing the Evolutionary Scale Modeling inverse folding (ESM-IF) structure encoder and fine-tuning only the decoder using Simple Preference Optimization (SimPO).
• Trained the model to prioritize stronger binders.
• Evaluated generalization across seven held-out antibody-antigen complexes.

Main Results:

• SimBinder-IF demonstrated a 22% numerical increase in average Spearman correlation (0.26 to 0.32) compared to vanilla ESM-IF across seven held-out complexes.
• Achieved the highest mean 20-fold improvement@20 (0.453) for top-ranking enrichment.
• In a case study, SimBinder-IF generated variants with significantly lower predicted binding free energy (-68.56 kcal/mol) compared to ESM-IF (-49.77 kcal/mol) for a difficult-to-bind target.
• Updated only 18% of the ESM-IF parameters, showcasing parameter efficiency.

Conclusions:

• SimBinder-IF offers a parameter-efficient and structure-aware approach for antibody affinity optimization.
• The model shows promise in improving antibody-target binding predictions and reducing development costs.
• Further research can explore its application in designing novel antibody therapeutics.