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

Abstract

Machine learning-based protein structure prediction algorithms, such as RosettaFold and AlphaFold2, have greatly impacted the structural biology field, arousing a fair amount of discussion around their potential role in drug discovery. While there are few 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 model based on low prior structural information. In order to address this, we have developed an AlphaFold2 version where we exclude all structural templates with more than 30% sequence identity from the modelbuilding process. In a previous study, we used those models in conjunction with state-of-the-art free energy perturbation methods and demonstrated that it is possible to obtain quantitatively accurate results. In this work, we focus on using these structures in rigid receptor-ligand docking studies. Our results indicate that using out-of-the-box Alphafold2 models is not an ideal scenario for virtual screening campaigns; in fact, we strongly recommend to include some post-processing modeling to drive the binding site into a more realistic holo model.

AlphaFold2 in Molecular Discovery

“…7−9 This result was explained by the observation that AF2 structures usually present a significant collapse of the binding site in comparison to holo structures, being more closely related to the apo structures. 8 Zhang and colleagues 7 observed that AF2 structures showed similar enrichment when compared to apo structures. Interestingly, the authors proposed a protocol to refine AF2 structures, involving energy-guided structure refinement and molecular dynamics simulations, which led to structures that provide enrichment similar to the one provided by holo structures.…”

“…Refinement of AF2 structures before docking is in line with the results and recommendations from Di ́az-Rovira and colleagues. 8 Moreover, Kersten and colleagues 9 observed that accuracy in binder-decoy discrimination using AF2 structures was better than the accuracy obtained using structures obtained by homology modeling. This is a surprising result, since many expected homology models to show more accurate predictions for side chain positions and, therefore, better results in docking.…”

“…The authors discuss challenges to be tackled in the future to make AF2 even more helpful in drug discovery, such as improving predictions of different conformation states, structures with post-translational modifications, multidomain structures, and protein–ligand complexes. Different authors performed virtual screening using AF2 structures and docking, − comparing the results to the ones obtained using experimental structures or structures from homology modeling. AF2 structures showed worse enrichment of known active compounds when compared to holo structures. − This result was explained by the observation that AF2 structures usually present a significant collapse of the binding site in comparison to holo structures, being more closely related to the apo structures .…”

“…Different authors performed virtual screening using AF2 structures and docking, − comparing the results to the ones obtained using experimental structures or structures from homology modeling. AF2 structures showed worse enrichment of known active compounds when compared to holo structures. − This result was explained by the observation that AF2 structures usually present a significant collapse of the binding site in comparison to holo structures, being more closely related to the apo structures . Zhang and colleagues observed that AF2 structures showed similar enrichment when compared to apo structures.…”

“…Interestingly, the authors proposed a protocol to refine AF2 structures, involving energy-guided structure refinement and molecular dynamics simulations, which led to structures that provide enrichment similar to the one provided by holo structures. Refinement of AF2 structures before docking is in line with the results and recommendations from Díaz-Rovira and colleagues . Moreover, Kersten and colleagues observed that accuracy in binder-decoy discrimination using AF2 structures was better than the accuracy obtained using structures obtained by homology modeling.…”