Towards automated physics-based absolute drug residence time predictions

The residence time (τ) of a drug bound to a receptor target is increasingly recognized as a key property to control during ligand-optimization campaigns and provides a useful dimension for modulating compound profiles in addition to binding affinity. For this reason, many computational approaches have been developed to predict this quantity. Several methods employ an empirical correlation between the residence time and the measured simulation time for a ligand to escape the binding pocket during biased molecular dynamics (MD), while others rely on more formal approaches that require a substantially larger computational effort and/or setup times often impractical in a fast-paced drug-discovery setting. Here we propose a new scheme to calculate absolute residence times by using two enhanced sampling approaches, consisting of an exploration phase followed by an exploitation phase that estimates the residence time: Random Acceleration Molecular Dynamics (RAMD) to harvest plausible egress pathways, and then Infrequent Metadynamics (iMetaD) to estimate residence time. This protocol caters to drug discovery programs, where a key aspect is the compromise between accuracy, throughput, and ease of use. We benchmark this approach by computing residence times for a congeneric series of ligands binding to several diverse drug targets and show that we can achieve good accuracy (RMSE of 1.22 and R2 of 0.80 in log10(τ)) without manually tuning the enhanced sampling parameters.