An Automated Workflow for Diagnosing Sampling Issues Caused by Slow Torsional Motions in Molecular Simulations

Physics-based methods, such as protein-ligand binding free energy calculations, are increasingly used in early-stage drug discovery to prioritize compounds. Accurate free energy estimates require adequate sampling of all relevant protein-ligand conformations, including ligand and protein side chain's rotatable bonds. Sampling issues can arise from slow torsion conformation changes and may manifest as high statistical errors or variability between repeated calculations. However, apparent convergence does not guarantee sufficient sampling, and identifying the underlying causes of slow convergence can be difficult. Here, instead of simply focusing on convergence of free energy estimates, we assess the sampling of specific structural degrees of freedom to identify potential sampling problems. Particularly, we develop an automated method for diagnosing sampling issues caused by slow torsional rotation events in the protein or ligand during binding free energy calculations. Here, our focus is on postsimulation analysis. Our method analyzes torsions in the ligand and in residues near the protein's binding site to define the dihedral angle states for each torsion. We then flag potential sampling issues when there are low transitions in and out of each dihedral angle state. We find that our method automatically detects sampling issues caused by slow torsional rotations that otherwise would have gone unnoticed and may have noticeable impacts on the calculated free energy values.