While a plethora of different protein–ligand docking protocols have been developed over the past twenty years, their performances greatly depend on the provided input protein–ligand pair. In this study, we developed a machine-learning model that uses a combination of convolutional and fully connected neural networks for the task of predicting the performance of several popular docking protocols given a protein structure and a small compound. We also rigorously evaluated the performance of our model using a widely available database of protein–ligand complexes and different types of data splits. We further open-source all code related to this study so that potential users can make informed selections on which protocol is best suited for their particular protein–ligand pair.

The number of entries in the Protein Data Bank (PDB) has doubled in the last decade, and it has increased tenfold in the last twenty years. The availability of an ever-growing number of structures is having a huge impact on the Structure-Based Drug Discovery (SBDD), allowing investigation of new targets and giving the possibility to have multiple structures of the same macromolecule in a complex with different ligands. Such a large resource often implies the choice of the most suitable complex for molecular docking calculation, and this task is complicated by the plethora of possible posing and scoring function algorithms available, which may influence the quality of the outcomes. Here, we report a large benchmark performed on the PDBbind database containing more than four thousand entries and seventeen popular docking protocols. We found that, even in protein families wherein docking protocols generally showed acceptable results, certain ligand-protein complexes are poorly reproduced in the self-docking procedure. Such a trend in certain protein families is more pronounced, and this underlines the importance in identification of a suitable protein–ligand conformation coupled to a well-performing docking protocol.