DCBK: A Fragment-Based Hybrid Simulation and Machine Learning Framework for Predicting Ligand Dissociation Kinetics

Predicting small-molecule binding kinetics is central to drug discovery, but accurately estimating dissociation rates (koff) remains challenging due to high energy barriers and the high computational cost of conventional simulations. Here, we present a hybrid, data-driven computational pipeline that integrates adaptive biased molecular dynamics (ABMD), umbrella sampling (US), fragment-level free energy calculations, and machine learning (ML) correction to achieve scalable, high-precision koff predictions across diverse protein-ligand systems. By leveraging a BRICS-based fragmentation strategy and training ML models on experimental kinetics data, our approach not only reconstructs detailed free energy landscapes but also pinpoints key molecular fragments governing dissociation pathways. In contrast to purely physics-based simulations, this pipeline offers improved accuracy, enhanced transferability across molecular systems, and markedly higher computational efficiency, thereby enabling rigorous quantification of the effect of ligand modifications on residence time. The resulting framework offers a generalizable, interpretable, and reproducible solution for binding kinetics prediction, providing actionable insights for lead optimization and mechanistic studies.