Evaluation of AlphaFold 3’s Protein–Protein\nComplexes for Predicting Binding Free Energy Changes upon Mutation

AlphaFold 3 (AF3),\nthe latest version of protein structure prediction\nsoftware, goes beyond its predecessors by predicting protein–protein\ncomplexes. It could revolutionize drug discovery and protein engineering,\nmarking a major step toward comprehensive, automated protein structure\nprediction. However, independent validation of AF3’s predictions\nis necessary. In this work, we evaluate AF3 complex structures using\nthe SKEMPI 2.0 database which involves 317 protein–protein\ncomplexes and 8338 mutations. AF3 complex structures when applied\nto the most advanced TDL model, MT-TopLap (MultiTask-Topological Laplacian),\ngive rise to a very good Pearson correlation coefficient of 0.86 for\npredicting protein–protein binding free energy changes upon\nmutation, which is slightly less than the 0.88 achieved earlier with\nthe Protein Data Bank (PDB) structures. Nonetheless, AF3 complex structures\nled to a 8.6% increase in the prediction RMSE compared to original\nPDB complex structures. Additionally, some of AF3’s complex\nstructures have large errors, which were not captured in its ipTM\nperformance metric. Finally, it is found that AF3’s complex\nstructures are not reliable for intrinsically flexible regions or\ndomains.