Protein Structural Model Selection Informed by Comparison of Predicted Ligand Binding Poses

Recent advances in protein structure prediction have highlighted the importance of a longstanding problem: given multiple structural models of a protein, how does one select the best model to use when predicting interactions between that protein and candidate drug molecules? Here we demonstrate the value of a previously unutilized source of information in addressing this problem. We show that given multiple ligands known to bind the protein, one can perform effective model selection by comparing the predicted binding poses of multiple ligands at each model. We introduce a method, RevBind, that exploits this information, leveraging the statistical tendency of different ligands to form similar chemical interactions with a protein's binding pocket. RevBind can be used, for example, to select among variants of AlphaFold models, identifying those that are most useful for molecular docking. Our findings pave the way for the development of even better model selection methods that draw simultaneously on the information used by RevBind and the information used by previous methods.