Generating Minimum Free Energy Paths With Denoising Diffusion Probabilistic Models

A method combining denoising diffusion probabilistic models (DDPMs) with the string method is presented to generate minimum free energy paths between metastable states in molecular systems. It has been demonstrated in recent work that DDPMs at low noise levels can approximate the gradient of the potential of mean force, allowing efficient sampling of high-dimensional configurational spaces. Building on this insight, it is shown here that DDPM-derived force fields accurately generate transition pathways for the analytical Muller-Brown potential and for the alanine dipeptide system at some range of noise levels for DDPMs, recovering the transition path and implicitly capturing solvent effects in the case of alanine dipeptide.