Custom scoring Aurora-A with SMINA

In 2002, Sir Philip Cohen predicted that protein kinases would become ‘the drug targets of the 21st century’. So far, kinases have lived up to this expectation, in fact protein kinases have emerged as one of the most successful families of drug targets. Actually, Virtual Screening (VS) studies in the kinase field, are hampered by the fact that many specific inhibitors are not fully selective for a single target, due to the high structural similarity of the ATP-binding site of kinases. The aim of this work is to explore strategies to improve the effectiveness of VS on kinase proteins. As a paradigm, we focused on Aurora A kinase. The Aurora kinases are implicated in a variety of hematological and solid cancers, because of the essential role in mitosis and cell cycle regulation. In recent years, Aurora kinases have proved popular cancer targets and many inhibitors have been developed. The majority of these clinical candidates are multi-targeted, rendering them inappropriate as tools for studying Aurora kinase mediated signaling. Our starting point was a structural analysis of 24 Aurora-A crystal structures complexed with different ligands. Starting from this dataset and using data derived from PubChem, we benchmarked the performance of the default scoring function (SF) of the popular Vina docking tool to identify true binders out of a set of decoys, obtained by DUDE. These decoys are selected to be chemically dissimilar from the provided active compounds. Thus, we used the RDkit library to generate 3D conformations from the 2D SMILES obtained by DUDE. The docking performance was analyzed using the receiver operating characteristic (ROC) curves, which represent the plot of false positive rate versus the true positive one. By analyzing the ROC curves for docking performance against the various Aurora-A structures, we were able to simultaneously evaluate our docking protocol and the most appropriate receptor structure to be used for VS purposes. Moreover, in addition to the default scoring of VINA, we developed a custom SF that was parametrized using the previously described benchmark. To this end, we made use of logistical regression with backward variable selection, to fit docking scores to activity data. We show that the obtained custom SF that was tailored on Aurora-A outperformed the original VINA custom scoring. Finally, an ad-hoc set of heuristic rules on receptor’s side-chains flexibility was conceived during the analysis to further improve the docking performance of Aurora-A inhibitors. The protocol used to derive this custom SF and the obtained heuristic rules can be easily adopted and made generalizable to whole human kinome.