Research map / Probabilistic and scientific modeling / Dynamical systems
UniSim: A Unified Simulator for Time-Coarsened Dynamics of Biomolecules
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 UniSim, a pretrained unified full-atom model with stochastic-interpolant vector field and force guidance kernel for time-coarsened MD simulation across small molecules, peptides, and proteins.
Problem
Classical MD requires tiny timesteps; existing deep learning time-coarsened dynamics models are restricted to single molecular domains and hand-crafted representations, lacking transferability across small molecules, peptides and proteins.
Approach
Pretrain a unified SO(3)-equivariant atomic representation (TorchMD-NET) with multi-head denoising/force objectives, gradient-environment subgraphs and atomic embedding expansion; then learn long-timestep transitions with stochastic interpolants and adapt to new force fields via a force guidance kernel trained with frozen backbones.
Key results
- On PepMD, UniSim achieves VAL-CA 0.575 versus 0.539 for FBM and best JS distances on PWD (0.328) and RG (0.330).
- On ATLAS, validity (VAL-CA) rises to 0.079 from 0.052 for ITO, the best prior baseline reported.
- UniSim reports about 25x higher ESS/s than MD with OpenMM on the PepMD test set.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add unified representation pretraining (ablation) | UniSim vector field model · PepMD test set | VAL-CA: 0.332 -> 0.505 | Helped |
| add force guidance (ablation) | UniSim/g · PepMD test set | VAL-CA: 0.505 -> 0.575 | Helped |
| add force guidance (ablation) | UniSim/g · MD22 test set | TIC JS distance: 0.408/0.111 -> 0.368/0.132 | Helped |
| add force guidance (ablation) | UniSim/g · ATLAS test set | CONTACT: 0.171/0.031 -> 0.173/0.031 | Hurt |
| remove atomic embedding expansion A_e (ablation) | UniSim/g · PepMD test set | VAL-CA: 0.505/0.112 -> 0.397/0.132 | Hurt |
| increase guidance strength from α=0.05 to α=0.07 at T=15 (ablation) | UniSim · PepMD test set | VAL-CA: 0.575/0.139 -> 0.607/0.138 | Helped |
| increase guidance strength from α=0.05 to α=0.07 at T=15 (ablation) | UniSim · PepMD test set | P W D JS distance: 0.328/0.149 -> 0.349/0.144 | Hurt |
| increase SDE steps from T=15 to T=25 at α=0.05 (ablation) | UniSim · PepMD test set | VAL-CA: 0.575/0.139 -> 0.441/0.117 | Hurt |
| increase SDE steps from T=15 to T=25 at α=0.05 (ablation) | UniSim · PepMD test set | TIC JS distance: 0.510/0.124 -> 0.507/0.134 | No clear effect |
| use UniSim instead of OpenMM molecular dynamics | UniSim · PepMD test set | ESS/s: approximately 25 times | Helped |
Surprising findings
- Ablating atomic embedding expansion degrades nearly all metrics on PepMD, e.g., VAL-CA drops from 0.505 to 0.397 for UniSim/g.
Limitations
- Validity of generated protein conformations is unreliable due to cumulative autoregressive prediction errors.
- Generated trajectories evaluated are relatively short, limiting discovery of metastable states.
Experimental setup
Datasets: PCQM4Mv2 ANI-1x PepMD PDB protein monomers ATLAS Solvated Protein Fragments MD17 MD22
Benchmarks: PepMD MD22 ATLAS
Metrics: JS distance (PWD, RG, TIC, TIC-2D) VAL-CA CONTACT ESS/s
Baselines: FBM Timewarp ITO Score Dynamics (SD)
Builds on: stochastic interpolants FBM TorchMD-NET DPA-2 Timewarp
Closest papers by meaning
- FlashMD: long-stride, universal prediction of molecular dynamics (NeurIPS 2025 · similarity 0.88)
- UMA: A Family of Universal Models for Atoms (NeurIPS 2025 · similarity 0.88)
- Generative Modeling of Molecular Dynamics Trajectories (NeurIPS 2024 · similarity 0.87)
- UniMoMo: Unified Generative Modeling of 3D Molecules for De Novo Binder Design (ICML 2025 · similarity 0.86)
- Stein Kernelized Molecular Dynamics for Active Learning of Interatomic Potentials (NeurIPS 2026 · similarity 0.86)
- Few-Step Cofolding with All-Atom Flow Maps (NeurIPS 2026 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.