Assessment of Alphafold Protein Models for Small-Molecule Ligand Docking versus Co-Folding

Molecular docking is a powerful computational tool for predicting protein-ligand interactions, widely employed in drug discovery. However, its effectiveness is often constrained by the availability of experimentally determined, high-resolution protein structures, a process that is both time-consuming and resource-intensive. AlphaFold (AF), a machine learning (ML)-based method, offers an efficient alternative by predicting high-accuracy 3D protein structures directly from amino acid sequences. This study assesses the utility of AF-generated protein models for fragment and larger-ligand docking with Glide, which is a widely used docking approach. The docking workflow is evaluated in an unbiased manner by carrying out binding site identification with FTMap, a binding hot spot prediction software. We show that fragment docking to AF models outperforms docking to the respective unbound protein crystal structures and performs comparably to docking to the corresponding ligand-bound structures when using an unbiased approach. Leveraging the computational efficiency of AF model generation, we also employ ensembles of AF models to incorporate protein flexibility. Results show that docking to AF ensembles improves larger-ligand docking compared to docking to singular AF models and outperforms docking to unbound structures. In addition, we compare docking to AF ensembles to co-folding with AF3 and Boltz-2. The results provide insights into the effectiveness of integrating AF protein models into docking procedures, highlighting their potential for streamlining computational drug discovery processes.