Assessing the Performance of AF2 and AF3-Implementations on Antibody-Antigen Complexes

This study evaluates the performance of AF2, Boltz-1, Boltz-1x, Boltz-2, Chai-1, Protenix and ESMFold, on a curated dataset of 200 novel antibody-antigen (Ab-Ag) complexes. Our findings reveal that AF2 remains a strong predictor against newer AlphaFold3 (AF3)-inspired methods in Ab-Ag complex prediction. Chai-1 consistently ranks as the second-best performer across multiple success metrics. We observed diverse effects of recycling iterations, with AF2, Chai-1 and Protenix benefiting from increased cycles, unlike Boltz variants. We analyzed various model confidence scores, noting high precision from pDockQ2 and high recall from pTM. By integrating these two scores, we developed AntiConf, a novel metric that achieves superior performance for all methods in terms of precision and unique target recall. This work addresses current limitations in AI-based Ab-Ag complex prediction, showcasing the potential of AntiConf for future predictions and assessments of these complexes, and providing a guideline for improving the accuracy of Ab-Ag complex prediction.