Can Cavity Prediction Algorithms Help in Docking Experiments?

Blind docking is a method for predicting a binding mode of a ligand with a protein without any prior information about a binding site. Some tools allow this type of docking experiment directly, others, including some established tools, require binding site information being passed as an input. In this latter case, one can use cavity prediction tools and use the results of their prediction as an input in these docking calculations. However, it is still unclear if the results of these predictions can be reliably used in protein-ligand docking and what is the best technical way to pass this information to the docking algorithm. In this study we estimated the applicability of the binding pocket prediction tools in docking experiments to address this gap in knowledge. We use four different computational tools for cavity prediction and use the best predicted cavities represented in different ways to run GOLD docking calculations. Analysis of subsequent use in docking highlights that Fpocket and CAVIAR are the best performing cavity prediction tools in this context. Further analysis shows that accurate binding site input does not guarantee accurate binding pose predictions and, even with the predicted cavities, the more restrained the input is, the more reliable the docking results are.