Answers are grounded in this paper and cite their sources.
Rare-event conformational changes, such as cryptic pocket opening and allosteric rearrangements, often occur on timescales that are impractical for classical molecular dynamics (MD). Here, we introduce a hybrid molecular dynamics scheme that intermittently proposes backbone torsion updates from a score-based diffusion model and accepts or rejects them with a Metropolis–Hastings criterion, guaranteeing sampling from the canonical equilibrium distribution. We benchmark this method on alanine dipeptide. The diffusion model was trained on 2.4 nanoseconds of trajectories and evaluated in five 80 nanosecond replicas each of classical and hybrid MD. After accounting for training and data generation time, hybrid runs achieved a 68-fold improvement (p≈0.00755) in mean effective sample size per GPU-hour for the slow ϕ dihedral compared to classical MD. Equilibrium basin populations agreed within ≤5% across three low free energy basins. This work shows that the method can achieve large gains in sampling efficiency without distorting thermodynamic properties, serving as a proof of concept for the integration of diffusion-based proposals into MD with applications to biologically relevant protein-ligand systems and potential for integration into other enhanced sampling techniques.
No takes yet. Share an insight, caveat, or question.
Subramani et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: