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

Abstract

The availability of AlphaFold2 has led to great excitement in the scientific communityparticularly among drug huntersdue to the ability of the algorithm to predict protein structures with high accuracy. However, beyond globally accurate protein structure prediction, it remains to be determined whether ligand binding sites are predicted with sufficient accuracy in these structures to be useful in supporting computationally driven drug discovery programs. We explored this question by performing free-energy perturbation (FEP) calculations on a set of well-studied protein–ligand complexes, where AlphaFold2 predictions were performed by removing all templates with >30% identity to the target protein from the training set. We observed that in most cases, the ΔΔG values for ligand transformations calculated with FEP, using these prospective AlphaFold2 structures, were comparable in accuracy to the corresponding calculations previously carried out using crystal structures. We conclude that under the right circumstances, AlphaFold2-modeled structures are accurate enough to be used by physics-based methods such as FEP in typical lead optimization stages of a drug discovery program.

Recent Advances in Alchemical Binding Free Energy Calculations for Drug Discovery

“…For protein models built on template structures of 30–50% sequence identity, FEP results were comparable to those obtained with protein crystal structures in an impressive 86% of the cases. Likewise, AlphaFold2 models for 14 proteins structures built on templates with <30% sequence identity yielded FEP results comparable to those generated with crystal structures suggesting suitability for lead optimization purposes . These examples support the notion that RBFE calculations can succeed when using protein models.…”

“…Likewise, AlphaFold2 models for 14 proteins structures built on templates with <30% sequence identity yielded FEP results comparable to those generated with crystal structures suggesting suitability for lead optimization purposes. 20 These examples support the notion that RBFE calculations can succeed when using protein models. However, we believe that caution is still warranted because unsuccessful attempts to use model structures, as experienced on multiple occasions by us (unpublished) and others, 2 are rarely advertised.…”

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

“…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”.…”

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

“…Our AF2 customization has been described in detail in our recent FEP study [23]. Briefly we systematically removed all template structures above 30% sequence identity from the database used to build the models.…”

“…For example, in a recent publication we showed how a state of the art implementation of FEP (FEP+ from Schrödinger), when applied to AF2 structures, could produce analogous results to those when using crystal structures [23]. In that study, in order to impose more realistic prospective conditions on our benchmark experiment, we developed a custom AF2 version, named AF2 30 , where we eliminated all structural templates with >30% identity to the target protein from the training set.…”