Empirical methods to predict free energies of ligand binding often err in their predictions by more than 3 kcal/mol. In our work, we use a simple model of ligand-protein complexes to simulate possible artifacts and errors in empirical modelling that might contribute to its large error bars. We study the magnitudes, and the physical and statistical bases, of several types of error; we illustrate ways to remedy the errors; ad we make recommendations for statistical procedures not currently in common use in the field. Quantitative results from our model suggest that if drug designers correct the errors we study, they can improve the accuracy of free energy predictions by several kcal/mol. We use Computer Graphics Laboratory resources to produce illustrations of both the lattice model and all-atom molecular examples where artifacts may contribute to errors in free energy prediction. The color and graphics capabilities of the CGL machines are vital to our project, because they allow us to produce pictures that are clear and easy to understand.
Showing the most recent 10 out of 508 publications
