Integrating Molecular Dynamics and Machine Learning Algorithms to Predict the Functional Profile of Kinase Ligands

The modulation of protein function via designed small molecules is providing new opportunities in chemical biology and medicinal chemistry. While drugs have traditionally been developed to block enzymatic activities through active site occupation, a growing number of strategies now aim to control protein functions in an allosteric fashion, allowing for the tuning of a target's activation or deactivation via the modulation of the populations of conformational ensembles that underlie its function. In the context of the discovery of new active leads, it would be very useful to generate hypotheses for the functional impact of new ligands. Since the discovery and design of allosteric modulators (inhibitors/activators) is still a challenging and often serendipitous target, the development of a rapid and robust approach to predict the functional profile of a new ligand would significantly speed up candidate selection. Herein, we present different machine learning (ML) classifiers to distinguish between potential orthosteric and allosteric binders. Our approach integrates information on the chemical fingerprints of the ligands with descriptors that recapitulate ligand effects on protein functional motions. The latter are derived from molecular dynamics (MD) simulations of the target protein in complex with orthosteric or allosteric ligands. In this framework, we train and test different ML architectures, which are initially probed on the classification of orthosteric versus allosteric ligands for cyclin-dependent kinases (CDKs). The results demonstrate that different ML methods can successfully partition allosteric versus orthosteric effectors (although to different degrees). Next, we further test the models with FDA-approved CDK drugs, not included in the original dataset, as well as ligands that target other kinases, to test the range of applicability of these models outside of the domain on which they were developed. Overall, the results show that enriching the training dataset with chemical physics-based information on the protein–ligand dynamic cross-talk can significantly expand the reach and applicability of approaches for the prediction and classification of the mode of action of small molecules.