Bridging Scales in Drug Binding Kinetics: Residence Time Estimation via All-Atom and Coarse-Grained Simulations

The equilibrium binding affinity has traditionally guided the drug discovery. Despite offering insight into the extent of binding, it is occasionally inaccurate in predicting biological response. Emerging evidence suggests that the lifetime of binary complex, known as residence time (RT), is more directly correlated with pharmacological activity. Eventually, the RT model has gained significant traction in recent years. Several computational approaches have recently been emerged to estimate experimentally observed RT using all-atom molecular dynamics simulations. While some methods rely on establishing empirical correlations between computed relative RTs and experimental values, others aim to calculate absolute RTs through more rigorous modeling. Despite these efforts, achieving a scalable solution that balances the accuracy-computation cost trade-off remains a challenge – especially for drug-like molecules with longer RT. In this work, we propose a workflow for calculating RT, applicable across both all-atom (AA) and coarse-grained (CG) simulation scales, enabling estimations from nanoseconds to thousands of seconds. This workflow integrates three key components: (a) reaction coordinate learning using Deep-LDA, a deep-learning based protocol that extracts meaningful coordinates from metastable state information; (b) guiding the simulations using Frequency Adaptive Metadynamics, a modified infrequent metadynamics strategy, and (c) classification of multiple unbinding pathways using dynamic time-warping algorithm to compute RT corresponding to each pathway cluster. Here, we demonstrate that this approach not only recovers experimentally observed RT using AA simulations, but also correctly ranks the relative RTs of congeneric ligand series at a significantly reduced computational cost using CG simulations.