Benchmarking Antibody-Antigen Affinity Prediction Across Sequence-, Structure-, & Physics-Based Methods
- Structure-aware approaches outperform sequence-based models, highlighting the importance of explicit 3D epitope–paratope interactions
- PLM and LLM methods underperform overall, indicating limits of sequence-only representations for binding affinity prediction
- Reliable prediction remains driven by structural and physicochemical information, with hybrid approaches showing the most promise