How Good is AlphaFold3 at Ranking Drug Binding Affinities?

Xin Hong, Bowen Gao, Yinjun Jia, Wenyu Zhu, Qixuan Chen, Xiaohe Tian, Zhenyi Zhong, Jianhui Wang, Yanyan Lan

Keywords: AlphaFold3, lead optimization, affinity prediction, FEP+, ranking, AlphaRank

TL;DR: We show that AlphaFold3-like models, even without fine-tuning, outperform traditional methods in ranking ligand affinities, and our fine-tuned version, AlphaRank, rivals FEP+ accuracy with far lower computational cost.

Abstract: Accurate affinity ranking of small molecules is pivotal for drug discovery. We investigate whether structure prediction models like AlphaFold3, pretrained on protein-ligand interactions, can address this task. Zero-shot evaluation of Protenix (an AlphaFold3-like model) demonstrates superior prioritization of active compounds over conventional scoring functions and state-of-the-art deep learning models. By further fine-tuning Protenix on structure-agnostic protein-ligand bioactivity data from ChEMBL and BindingDB, we develop AlphaRank that predicts pairwise affinity relationships. AlphaRank achieves prediction accuracy comparable to computationally intensive free energy perturbation (FEP+) workflows on standard benchmarks, while requiring substantially less computational resources. Our findings highlight the emergent potential of AlphaFold3-derived models in affinity ranking tasks and emphasize the necessity for targeted methodological exploration to fully harness their capabilities in drug discovery applications.

Submission Number: 65