AF3Score: A Score-Only Adaptation of AlphaFold3 for Biomolecular Structure Evaluation

Abstract

Scoring biomolecular complexes remains central to structural modeling efforts. Recent studies suggest that AlphaFold (AF) - a revolutionary deep learning model for biomolecular structure prediction - has implicitly learned an approximate biophysical energy function. While many researchers highly rely on AF-derived scores for structure evaluation, existing AlphaFold2-based implementations require iterative refinement of the input structure, leading to biased scoring. To address this limitation, we adapted AlphaFold3 into a score-only model, AF3Score, by directly feeding input coordinates into the confidence head while bypassing the diffusion-based structure module. AF3Score demonstrates robust performance in structural quality assessment across diverse systems, including monomeric proteins, protein-protein complexes, de novo designed binders, fold-switching proteins, and protein-ligand complexes. In benchmarking designed binder screening, AF3Score outperformed state-of-the-art methods for 8 out of 10 targets. Moreover, combining AF3Score with AlphaFold2-derived methods significantly improved the enrichment of experimentally validated binders, increasing the success rate from 15.2 to 31.6%. Additionally, AF3Score effectively identified stable conformations in fold-switching proteins, whereas AlphaFold predominantly predicted only the dominant fold. These findings highlight the broad applicability of AF3Score, from high-throughput screening in de novo binder design to filtering docking-generated poses and molecular dynamics (MD) trajectories.