Research map / Image generation / Fast diffusion sampling

Transition Path Sampling with Improved Off-Policy Training of Diffusion Path Samplers

ICLR 2025 · Image generation › Fast diffusion sampling · approach: Diffusion theory and learned samplers · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes TPS-DPS, a collective-variable-free diffusion path sampler trained with off-policy log-variance divergence (learnable control variate, replay buffer, simulated annealing) and a scale-based SE(3)-equivariant bias force, to sample transition paths for double-well, Alanine Dipeptide, and fast-folding proteins.

Problem

Sampling transition paths between meta-stable molecular states is computationally infeasible with unbiased MD due to high energy barriers, and existing ML methods need collective variables or fail on large systems and lack diversity.

Approach

Amortize transition path sampling by minimizing a log-variance divergence between the biased-MD path measure and the target transition path measure, trained off-policy with a learnable scalar control variate, replay buffer and simulated annealing, plus a positive-scale parameterization of the bias force that guarantees distance reduction to the aligned target state.

Key results

• On Alanine Dipeptide, TPS-DPS (P) reaches RMSD 0.16 Å, THP 87.50%, ETS 18.37 kJ/mol versus PIPS (F) RMSD 0.66 Å, THP 43.75%.
• On BBL, TPS-DPS (S) achieves THP 43.75% and RMSD 1.60 Å, while UMD and PIPS hit the target 0%.
• Log-variance loss outperforms reverse KL: TPS-DPS (F) RMSD 0.16 Å and THP 92.19% vs reverse KL (F) RMSD 0.43 Å and THP 53.12% on Alanine Dipeptide.

Measured effects reported in the paper

Surprising findings

• Positive-scale bias parameterization is the only variant that hits the target on Chignolin, Trp-cage, BBA, and BBL, while force and potential parameterizations achieve 0% THP on most fast-folding proteins.
• Simulated annealing is critical: RMSD does not decrease without it even though the loss does.

Limitations

• Only small fast-folding proteins (up to 50 amino acids) evaluated; scalability to proteins with more than 500 residues untested.
• Does not generalize across unseen meta-stable state pairs or different molecular systems.
• Integration with libraries such as PLUMED and DMFF not explored.

Experimental setup

Datasets: ground truth path ensembles from rejection sampling (double-well) unbiased MD trajectories for TICA (fast-folding proteins)

Benchmarks: double-well potential Alanine Dipeptide Chignolin Trp-cage BBA BBL

Metrics: RMSD THP (target hit percentage) ETS (energy of transition state)

Baselines: unbiased MD steered MD PIPS reverse KL divergence

Builds on: path integral path sampling relative trajectory balance log-variance divergence Langevin dynamics Kabsch algorithm

Closest papers by meaning

• Action-Minimization Meets Generative Modeling: Efficient Transition Path Sampling with the Onsager-Machlup Functional (ICML 2025 · similarity 0.91)
• On scalable and efficient training of diffusion samplers (NeurIPS 2025 · similarity 0.90)
• Doob's Lagrangian: A Sample-Efficient Variational Approach to Transition Path Sampling (NeurIPS 2024 · similarity 0.89)
• Consistent Sampling and Simulation: Molecular Dynamics with Energy-Based Diffusion Models (NeurIPS 2025 · similarity 0.87)
• Generative Modeling of Molecular Dynamics Trajectories (NeurIPS 2024 · similarity 0.87)
• Amortized Sampling with Transferable Normalizing Flows (NeurIPS 2025 · similarity 0.86)