Benchmarking Cross-Docking\nStrategies in Kinase Drug\nDiscovery

In recent years, machine learning has transformed many\naspects\nof the drug discovery process, including small molecule design, for\nwhich the prediction of bioactivity is an integral part. Leveraging\nstructural information about the interactions between a small molecule\nand its protein target has great potential for downstream machine\nlearning scoring approaches but is fundamentally limited by the accuracy\nwith which protein–ligand complex structures can be predicted\nin a reliable and automated fashion. With the goal of finding practical\napproaches to generating useful kinase-inhibitor complex geometries\nfor downstream machine learning scoring approaches, we present a kinase-centric\ndocking benchmark assessing the performance of different classes of\ndocking and pose selection strategies to assess how well experimentally\nobserved binding modes are recapitulated in a realistic cross-docking\nscenario. The assembled benchmark data set focuses on the well-studied\nprotein kinase family and comprises a subset of 589 protein structures\ncocrystallized with 423 ATP-competitive ligands. We find that the\ndocking methods biased by the cocrystallized ligand, utilizing shape\noverlap with or without maximum common substructure matching, are\nmore successful in recovering binding poses than standard physics-based\ndocking alone. Also, docking into multiple structures significantly\nincreases the chance of generating a low root-mean-square deviation\n(RMSD) docking pose. Docking utilizing an approach that combines all\nthree methods (Posit) into structures with the most similar cocrystallized\nligands according to the maximum common substructure (MCS) proved\nto be the most efficient way to reproduce binding poses, achieving\na success rate of 70.4% across all included systems. The studied docking\nand pose selection strategies, which utilize the OpenEye Toolkits,\nwere implemented into pipelines of the KinoML framework,\nallowing automated and reliable protein–ligand complex generation\nfor future downstream machine learning tasks. Although focused on\nprotein kinases, we believe that the general findings can also be\ntransferred to other protein families.