Chemoinformatics-Driven Approaches for Kinase Drug Discovery

Abstract

Given their importance for the majority of cell physiology processes, protein kinases are among the most extensively studied protein targets in drug discovery. Inappropriate regulation of their basal levels results in pathophysiological disorders. In this regard, small-molecule inhibitors of human kinome have been developed to treat these conditions effectively and improve the survival rates and life quality of patients. In recent years, kinase-related data has become increasingly available in the public domain. These large amounts of data provide a rich knowledge source for the computational studies of kinase drug discovery concepts.<br /> This thesis aims to systematically explore properties of kinase inhibitors on the basis of publicly available data. Hence, an established "selectivity versus promiscuity" conundrum of kinase inhibitors is evaluated, close structural analogs with diverging promiscuity levels are analyzed, and machine learning is employed to classify different kinase inhibitor binding modes. In the first study, kinase inhibitor selectivity trends are explored on the kinase pair level where kinase structural features and phylogenetic relationships are used to explain the obtained selectivity information. Next, selectivity of clinical kinase inhibitors is inspected on the basis of cell-based profiling campaign results to consolidate the previous findings. Further, clinical candidates are mapped to medicinal chemistry sources and promiscuity levels of different inhibitor subsets are estimated, including designated chemical probes. Additionally, chemical probe analysis is extended to expert-curated representatives to correlate the views established by scientific community and evaluate their potential for chemical biology applications. Then, large-scale promiscuity analysis of kinase inhibitor data combining several public repositories is performed to subsequently explore promiscuity cliffs (PCs) and PC pathways and study structure-promiscuity relationships. Furthermore, an automated extraction protocol prioritizing the most informative pathways is proposed with focus on those containing promiscuity hubs. In addition, the generated promiscuity data structures including cliffs, pathways, and hubs are discussed for their potential in experimental and computational follow-ups and subsequently made publicly available. Finally, machine learning methods are used to develop classification models of kinase inhibitors with distinct experimental binding modes and their potential for the development of novel therapeutics is assessed.