Docking-Informed\nMachine Learning for Kinome-wide\nAffinity Prediction

Kinase inhibitors are an important class of anticancer\ndrugs, with\n80 inhibitors clinically approved and >100 in active clinical testing.\nMost bind competitively in the ATP-binding site, leading to challenges\nwith selectivity for a specific kinase, resulting in risks for toxicity\nand general off-target effects. Assessing the binding of an inhibitor\nfor the entire kinome is experimentally possible but expensive. A\nreliable and interpretable computational prediction of kinase selectivity\nwould greatly benefit the inhibitor discovery and optimization process.\nHere, we use machine learning on docked poses to address this need.\nTo this end, we aggregated all known inhibitor-kinase affinities and\ngenerated the complete accompanying 3D interactome by docking all\ninhibitors to the respective high-quality X-ray structures. We then\nused this resource to train a neural network as a kinase-specific\nscoring function, which achieved an overall performance (<i>R</i><sup>2</sup>) of 0.63–0.74 on unseen inhibitors across the\nkinome. The entire pipeline from molecule to 3D-based affinity prediction\nhas been fully automated and wrapped in a freely available package.\nThis has a graphical user interface that is tightly integrated with\nPyMOL to allow immediate adoption in the medicinal chemistry practice.