Transition Path Sampling with Improved Off-Policy Training of Diffusion Path Samplers
Authors: Kiyoung Seong, Seonghyun Park, Seonghwan Kim, Woo Youn Kim, Sungsoo Ahn
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We extensively evaluate our approach, termed TPSDPS, on a synthetic system, small peptide, and challenging fast-folding proteins, demonstrating that it produces more realistic and diverse transition pathways than existing baselines. |
| Researcher Affiliation | Academia | Kiyoung Seong , Seonghyun Park , Seonghwan Kim, Woo Youn Kim, Sungsoo Ahn KAIST EMAIL |
| Pseudocode | Yes | Algorithm 1 Training 1: Initialize an empty replay buffer ˆD, an policy vθ, a scalar parameter w, the number of rollout I and training per rollout J, and an annealing schedule λstart = λ1 > > λI = λend. 2: for i = 1, . . . , I do |
| Open Source Code | Yes | We provide links to our project page and code. |
| Open Datasets | Yes | We extensively evaluate our method on the synthetic double-well potential with dual channels, Alanine Dipeptide, and four fast-folding proteins: Chignolin, Trp-cage, BBA, and BBL (Lindorff-Larsen et al., 2011). |
| Dataset Splits | No | All metrics are averaged over 1024 paths for the double-well system, and 64 paths for Alanine Dipeptide. All metrics are averaged over 64 paths. |
| Hardware Specification | Yes | RT and RI denote runtime (second) per rollout in training and inference on a single RTX A5000 GPU. |
| Software Dependencies | Yes | All real-world molecular systems are simulated using the Open MM library (Eastman et al., 2023). For the top two TICA components, we use Py EMMA library (Scherer et al., 2015)... |
| Experiment Setup | Yes | We use a 3-layer MLP for the double-well system, and a 6-layer MLP for real-world molecules with Re LU activation functions for neural bias force, potential, and scale. We update the parameters of the neural network with a learning rate of 0.0001, while the scalar parameter w is updated with a learning rate of 0.001. Table 3: Model configurations of TPS-DPS. |