Efficient sampling of slow conformational transitions is a large challenge of molecular dynamics (MD). Here, we describe a hybrid approach that periodically inserts generated conformations proposed by a diffusion model into MD. To preserve ensemble correctness, each proposal is screened with a metropolis criterion before insertion. As a benchmark, we applied the method to alanine dipeptide in implicit solvent. The diffusion model was trained on 2.4 ns of classical MD and used to generate backbone phi/psi proposals every 2 ns during hybrid simulations. We then compared eight 80 ns hybrid replicas to eight classical MD replicas under identical conditions. The hybrid approach was approximately 21 times more likely to yield high efficiency trajectories than classical MD (p = 0.041). Within this regime, effective sampling of the slow phi torsion improved by 147-fold (p = 0.0001). Psi dynamics remained statistically unchanged between regimes of phi sampling. Importantly, thermodynamic and kinetic consistency was maintained. Basin populations were within 1%–5% of classical MD and all free-energy differences were small. The hybrid scheme provides significant acceleration without distorting equilibrium distributions or short-time kinetics. These results show that generated proposals can act as an enhancement to classical MD. The method can be applied to larger biomolecules and integrated into existing enhanced sampling techniques.
Subramani et al. (Sun,) studied this question.