Toward high-throughput predictive modeling of protein binding/unbinding kinetics

One of the unaddressed challenges in drug discovery is that drug potency determined in vitro is not a reliable indicator of drug activity in humans. Accumulated evidences suggest that in vivo activity is more strongly correlated with the binding/unbinding kinetics than the equilibrium thermodynamics of protein-ligand interactions (PLI). However, existing experimental and computational techniques are insufficient in studying the molecular details of kinetics process of PLI. Consequently, we not only have limited mechanistic understanding of the kinetic process but also lack a practical platform for the high-throughput screening and optimization of drug leads based on their kinetic properties. Here we address this unmet need by integrating energetic and conformational dynamic features derived from molecular modeling with multi-task learning. To test our method, HIV-1 protease is used as a model system. Our integrated model provides us with new insights into the molecular determinants of kinetics of PLI. We find that the coherent coupling of conformational dynamics between protein and ligand may play a critical role in determining the kinetic rate constants of PLI. Furthermore, we demonstrate that the relative movement of normal nodes of amino acids upon ligand binding is an important feature to capture conformational dynamics of the binding/unbinding kinetics. Coupled with the multi-task learning, we can predict combined kon and koff accurately with an accuracy of 74.35%. Thus, it is possible to screen and optimize compounds based on their binding/unbinding kinetics. The further development of such computational tools will bridge one of the critical missing links in drug discovery.