Leveraging diverse data modalities to study kinase inhibitor polypharmacology

Abstract

Kinases constitute a popular class of drug targets and to date over 60 kinase inhibitors have been approved for clinical use by the U.S. Food and Drug Administration (FDA). Despite the clinical efficacy of these inhibitors, characterization of their mechanism of action at a large scale is incomplete, partially because many of them simultaneously interact with multiple kinases and the downstream effects of such multitarget modulation (polypharmacology) remain unclear. This work aims to develop computational pipelines to better elucidate kinase polypharmacology. The introductory chapter of this work reflects on how to leverage the similarities among protein targets to study polypharmacology. Drawing on examples from different protein families, it delineates a workflow that starts with target identification followed by an analysis of cellular networks and discusses how machine learning can accelerate both efforts. The following chapter focuses on kinases and develops a resource to prioritize kinases for research investigation based on their therapeutic potential. We first systematically redefine the kinome by combining sequence, structural and functional annotation. We then assemble a knowledge network of the kinases and identify understudied kinases with disease relevance. Kinases identified to have biological relevance could be prioritized based on the availability of existing compounds, experimental structure, or commercial assays. The third chapter aims to help identify potential kinase targets of a kinase inhibitor. Towards this end, machine learning models are developed to predict the binding affinities between kinases and compounds based on their structure. Since experimental structures of kinase-compound pairs are limited, in silico structures of kinase-compound pairs are generated to leverage as many affinity measurements as possible to infer optimal representations of the kinase-compound pairs for affinity prediction. This approach expands the application domain of the predictive models and led to improved performance compared to training on experimental structures only. Models trained on these docked poses also better recapitulate expected biochemical behavior. The last chapter discusses the application of the kinase resources developed in Chapter 2 and the models developed in Chapter 3. It considers the implications of the results and anticipates future extensions of these modeling projects.