Molecular dynamics (MD) simulations are a promising tool to guide drug lead optimization. But because these tools are applied prospectively in drug discovery, blind tests provide a key opportunity to validate these for real-world applications. I participated in the Statistical Assessment of Modeling of Proteins and Ligands (SAMPL) 3 blind test (chapter 1) in which I used different force fields to calculate hydration free energies and SAMPL4 (chapter 2) in which I analyzed pose prediction results from different groups using different computational tools. The results from these blind tests improve our understanding of the accuracy of current methods, allowing us to improve these methods.

Large-scale applications of MD simulations to drug discovery have been few, partly because of the difficulty of planning and setting up the simulations. For example, alchemical relative free energy (RFE) calculations have relatively high accuracy in predicting differences in binding between drug lead compounds and new derivatives which are sought to improve binding potency. But setting up RFE calculations for large sets of compounds has required far too much manual intervention to be practical. I helped develop an algorithm, LOMAP (chapter3), to automatically plan and set up these calculations. Resulting applications indicated that it could successfully reduce the time of planning RFE calculations. But in this project, we assumed that relative free energy (RFE) calculations involving ring breaking will introduce substantial error, and we tried to avoid these calculations as much as possible. Later, we quantitatively calculated what these errors would be to confirm this (chapter4).

Beside binding free energy calculations, MD simulations can also be used to predict solubilities. We used free energy calculations (chapter5) to calculate relative solubilities and compared the results with experiment and with results from more empirical chemical engineering methods. We found that our approach is more accurate, despite its straightforward nature.

Long-term, we are working towards developing an automated pipeline to help guide key aspects of drug lead optimization. My work helped with understanding the accuracy of current techniques, improving their automation, and providing a new technique for predicting physical properties like solubilities.