Analysis of docking for binding affinity prediction

Abstract

New docking algorithmic advances focus on refining the structural fitness of the selected ligand conformation within the protein binding pocket. Parallelly, recent systematic evaluations of novel docking methods focus on the physico-chemical properties of the proposed ligand conformations and how well do the predicted poses correspond with the experimentally determined binding poses. Here, we present an alternative analysis where we study how well do traditional docking scoring functions predict binding affinity prediction when applied to docking poses generated on predicted AlphaFold structures. Our results show that there is no correlation between predicted binding affinity and experimental affinity values across all protein-ligand pairs, but when we consider each protein individually we observe a greater variance with some proteins showing strong correlation. This exploratory analysis suggests that docking scoring function may be useful predictors of binding affinity for certain pocket configurations, but do not generalise properly and we hope it will open a discussion as to how to generalise them to the whole proteome.