Abstract

Background: Proteins are frequently affected by mutations with consequences on their structure’s stability, functionality, and binding affinity towards ligands, with important biological effects. Molecular Dynamics Simulation (MDS) is one of the most effective tools to in silico predict the binding affinity changes upon aminoacidic mutations. A fundamental requirement to conduct this analysis is the availability of the three-dimensional structures of the wild-type (wt) and mutated protein isoforms. Until now, crystallographic models have been mostly used. Recently, with the development of AlphaFold2, protein structure predictions with high model accuracy are possible. In this study, an automated computational workflow to compute the difference in the binding free energy between the wt and the mutated protein using AlphaFold2 is proposed. The workflow consists of generating the AlphaFold models for the wt and mutated proteins and studying their interaction with ligands using MDS to calculate the difference in binding free energy and infer the mutation effect on binding affinity. To evaluate the workflow confidence, two accuracy predictors are proposed: the distance between computed affinities of mutant and wt proteins (i.e. ΔΔG = |Δ G _mutant - Δ G_wt|), and an accuracy evaluation of the wt proteins models predicted with AlphaFold2 based on the analysis of side chains orientation.
Results: We observed how AlphaFold’s models coupled with MDS can be used in binding affinity measurements with an accuracy comparable to that obtained by using crystallographic structures. The workflow is validated through results obtained by using crystallographic structures on a benchmark constituted by proteins in complexes with small ligands, cofactors, and ions, with a high variability of applications. The side chains orientation parameter has been identified as an important workflow confidence estimator because, even though the high accuracy of AlphaFold2 predictions, side chains orientation in predicted wt proteins can significantly differ from the crystallographic models, with considerable effects on binding affinity calculations.
Conclusion: The proposed automated computational workflow successfully integrates AlphaFold2 with MDS for in silico predicting protein-ligand binding affinity. Thus, AlphaFold’s models can be beneficial in studying the effect of mutations on binding affinity even for proteins with unavailable crystallographic structures.