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

Abstract

Predicting small-molecule dissociation kinetics remains challenging due to the high computational cost of simulating rare unbinding events and the limited scalability of existing approaches. Here, we introduce DCBK (Divide-and-Compute Binding Kinetics), a physics-informed, fragment-resolved framework for predicting ligand dissociation across diverse protein-ligand systems. DCBK integrates steered molecular dynamics and umbrella sampling to reconstruct unbinding pathways and free energy profiles, followed by BRICS-based ligand fragmentation. Fragment-level energetic and structural features are combined with machine learning trained on experimental kinetics data, enabling efficient prediction of dissociation rate constant (koff) values. Validation across five protein families demonstrates strong agreement with experiments, with ablation and unseen-target analyses highlighting the contribution of fragment-level features and the frameworks transferability. Importantly, DCBK allows dissociation kinetics to be interpreted in terms of fragment-level contributions, offering mechanistic insight and guiding ligand modification to optimize residence time. This interpretable and extensible framework links molecular structure, energetics, and kinetics, providing actionable insights for lead optimization and rational drug design.