Memory-resolved prediction of ligand residence times on the minutes timescale
π arXiv:2610.20271 Β· π₯ PDF Β· 2026-10-14 Β· cond-mat.stat-mech
Authors: Hossein Batebi [arXiv Β· scholar] , Lucas Tepper [arXiv Β· scholar] , Roland R. Netz [arXiv Β· scholar]
π Abstract
AI-enabled computational drug discovery is advancing rapidly in predicting molecular structures, binding poses and affinities, but ligand dissociation kinetics remain harder to access. Bound-state residence time, important for biological response and kinetic selectivity, is not fixed by equilibrium affinity but depends on the dissociation free-energy landscape and dissipative friction effects. Here, we show that restrained umbrella-sampling trajectories can be used to calculate residence times on the minutes timescale. Time-resolved coordinate fluctuations yield friction profiles that incorporate non-Markovian effects on dissociation kinetics. Combining them with the free-energy profile, we predict a residence time of $Ο_{\rm res}$ = 275 s for the adenosine A$_{2A}$ receptor with ZM241385 (comparable to the experimental value of 125 s) and reproduce the experimental ranking of wild type and two receptor mutants. For HSP90 with compound 18 and RNase S with the S15 peptide, we predict $Ο_{\rm res}$ = 2,169 s and 2,259 s (experimental values are 833 s and 2,439 s). Neglecting memory effects on residence kinetics overestimates $Ο_{\rm res}$ by factors of 2.6 to 29.