(405a) Identifying Drug Binding Locations and Poses Using Hamiltonian Replica Exchange Molecular Dynamics

Fast and accurate identification of protein-ligand binding locations is crucial in drug discovery. While docking can quickly identify the binding locations, its accuracy is limited by the effectiveness of scoring function, lack of solvent influence and other known disadvantages.  In this research, we present results of a study to identify drug binding locations using GPU-accelerated Hamiltonian replica exchange molecular dynamics simulations.

If run sufficiently long with an accurate force field, MD simulations of a protein with a known ligand will eventually converge on the true bound structure. However, such simulations can get stuck in local minima which leads to insufficient sampling.  We use a variety of methods to greatly accelerate this process and improve sampling. We use a flat-bottom potential to restrain the ligand in the vicinity of the protein and let the ligand sample all available conformational space. We also useHamiltonian replica exchange method to escape local minima. Replicas of fully interacting systems are coupled with less strongly interacting systems, allowing sampling at one thermodynamic state to be shared with that performed at other states. The use of multiple fully coupled states and fully coupled states are investigated and applied. Finally, molecular dynamics are GPU accelerated with the OpenMM toolkit and Monte Carlo moves are added to improve sampling. Because of the rigorous nature of the statistical mechanical sampling, we can also extract binding free energy estimates at all putative binding sites and the overall protein-ligand binding free energy estimates. We develop the methodology on the T4 lysozyme model system, including rigorous convergence measures from independent starting configurations. The results for the T4 lysozyme model system show that with the above mentioned sampling-improvement methods, the binding site is successfully identified 70X faster than similar codes on standard processors with both the accurate location and binding free energy compared to the literature. Further results for the 85-protein Astex Diverse Set will provide us insights on the validity of the methodology and to what extent our methodology can improve the current docking methods.