Transition Path Sampling with Diffusion Path Samplers

Diffusion path sampler

We propose Diffusion Path Sampler (DPS) which is diffusion SDEs to sample from a target path measure of Transition Path Sampling (TPS) problem. Our method, called TPS-DPS, introduces a neural bias force into unbiased MD without relying on collective variables (CVs). TPS-DPS minimizes the log-variance divergence between path measure induced by the biased MD and target path measures. The log-variance divergence with off-policy training improves accuracy and diversity of sampled paths.

Double-well synthetic system

The bias force of TPS-DPS accelerates crossing energy barriers while unbiased MD does not escape the initial meta-stable state.

Unbiased MD

TPS-DPS

Fast folding proteins

TPS-DPS scales to fast-folding proteins, such as Chignolin, Trp-cage, BBA, and BBL. Our scale-based equivariant parameterization of the bias force enables reaching the target state in large molecules while maintaining realistic intermediate states.

• Spin button: toggles the spin of the molecule.
• Run/pause MD button: runs or pauses the molecular dynamics simulation.
• Drag (left click): rotates the molecule.
• Scroll (up/down): zooms in/out the molecule.