Are Deep Learning Structural Models Sufficiently Accurate for Virtual Screening? Application of Docking Algorithms to AlphaFold2 Predicted Structures

Abstract

Machine learning protein structure prediction, such as RosettaFold and AlphaFold2, have impacted the structural biology field, raising a fair amount of discussion around its potential role in drug discovery. While we find some preliminary studies addressing the usage of these models in virtual screening, none of them focus on the prospect of hit-finding in a real-world virtual screen with a target with low sequence identity. In order to address this, we have developed an AlphaFiold2 version where we exclude all structural templates with more than 30% sequence identity. In a previous study, we used those models in conjunction with state of the art free energy perturbation methods. In this work we focus on using them in rigid receptor ligand docking. Our results indicate that using out-of-the-box Alphafold2 models is not an ideal scenario; one might think in including some post processing modeling to drive the binding site into a more realistic holo target model.

Are Deep Learning Structural Models Sufficiently Accurate for Free-Energy Calculations? Application of FEP+ to AlphaFold2-Predicted Structures

“…While all data produced here were obtained without any user intervention, and as such, it is likely that model quality can be improved through customization of sequence alignment and model building parameters, there is clearly a large gap between AF2 and the methods compared to here (as seen already in the recent CASP competitions) . Still, many studies have shown limitations when modeling structures with AF2, including erroneous predictions for highly disordered structures, bias toward a particular conformational state, or the tendency to collapse potential binding sites. , …”

Discovery of a cryptic pocket in the AI-predicted structure of PPM1D phosphatase explains the binding site and potency of its allosteric inhibitors

“…(Saldanõ et al, 2022) Docking efforts against AlphaFold structures show lower performance than against holo structures available on the PDB. (Díaz-Rovira et al, 2022; Wong et al, 2022) Here, we show that this can be mitigated by considering conformational heterogeneity using MSMs. Using a highly flexible system, we can sample conformations and identify cryptic pockets that can be successfully used in downstream virtual screening applications.…”

Hic Sunt Dracones: Molecular Docking in Uncharted Territories with Structures from AlphaFold2 and RoseTTAfold

“…52 More sophisticated methods, like free energy perturbation (FEP) combined with AlphaFold2 homology models, were demonstrated to achieve accurate results recently. 53 However, the requirement of a correct binding mode for those calculations was also achieved by aligned ligand poses from crystal structures and refinement to resolve "significant clashes in some cases". We hence conclude that for a prospective virtual screening by molecular docking, multiple programs and homology models and surrogate structures should be considered and validated using known ligands if these are available.…”

“…However, even though docking with experimental structures (“control”) showed slightly better results than homology models, observed ROC AUCs (average 0.67/67, median 0.74/69 for FlexX/FRED) and adjusted log AUCs (average 0.13/0.15, median 0.14/0.13 for FlexX/FRED) are not ideal, and current molecular docking software might not yet be able to take advantage of the potentially high accuracy of, for example, MD-refined and ligand-steered AlphaFold2 homology models . More sophisticated methods, like free energy perturbation (FEP) combined with AlphaFold2 homology models, were demonstrated to achieve accurate results recently . However, the requirement of a correct binding mode for those calculations was also achieved by aligned ligand poses from crystal structures and refinement to resolve “significant clashes in some cases”.…”