Molecular dynamics emulator generating time-coarsened trajectories for small molecules, peptides, and proteins from one shared atomic representation.
No providers recorded yet. Browse all providers
Classical molecular dynamics resolves atomic motion by numerically integrating equations of motion at timesteps around a femtosecond, which makes reaching biologically interesting timescales prohibitively expensive. Machine-learned surrogates address this by predicting a system's state directly some long interval into the future — time-coarsened dynamics — but almost all of them are fitted to a single chemical domain. A model trained on small organic molecules transfers poorly to peptides, and one trained on peptides transfers poorly to folded proteins, even though the underlying interatomic physics is the same in each case.
UniSim, from researchers at Tsinghua University and Renmin University of China, attacks that fragmentation directly. It first learns one atomic representation model by multi-head pretraining across a deliberately heterogeneous mixture of quantum-chemistry, peptide, protein-structure, and solvated-fragment datasets, then reuses that representation as the encoder of a generative model that pushes a molecular state forward over a long coarsened timestep. A separate force guidance module lets the same trajectory model adapt to different chemical environments — different solvent conditions or force-field parameters — without retraining the generator. The work was presented as a poster at ICML 2025.
Unlike equilibrium ensemble generators such as AlphaFlow and BioEmu-1, which sample conformations without ordering them in time, UniSim produces successive states of a trajectory, so kinetic quantities and transition pathways remain accessible.
The atomic representation model is a 4-layer SO(3)-equivariant TorchMD-NET graph network with a hidden dimension of 256, 8 attention heads, and a 5 Å interaction cutoff. Pretraining draws on PCQM4Mv2 (roughly 3M DFT-optimized small molecules), ANI-1x (5M DFT calculations of energies and forces), Solvated Protein Fragments, a PDB-derived set of protein monomers, the PepMD peptide trajectory set (283 peptides of 3-10 residues simulated in OpenMM), and ATLAS all-atom protein simulations, filtered to 834 entries and clustered at 30% sequence identity into 790 training and 14 test proteins. The dynamics model builds on the stochastic interpolant framework, reinitializing the pretrained encoder to accept a continuous diffusion time and attaching GVP output heads with vector layer normalization for the drift and noise terms; sampling solves an SDE across the diffusion time. The force guidance kernel uses a second TorchMD-NET encoder residually coupled to the frozen representation.
On the 14 held-out PepMD peptides, UniSim outperforms FBM, Timewarp, ITO, and Score Dynamics on nearly all reported metrics — Cα validity plus Jensen-Shannon distances over pairwise distances, radius of gyration, and TIC projections — and ablations show that both pretraining and force guidance contribute. Transfer is demonstrated by training a guidance kernel on MD17 and evaluating on five larger MD22 molecules, and by evaluating protein dynamics on the ATLAS test split.
UniSim targets computational chemists and structural biologists who need trajectory-level sampling faster than explicit integration allows: exploring peptide conformational landscapes, generating ensembles for flexible protein targets, or estimating how a small molecule behaves across solvent conditions. Because a single checkpoint covers small molecules, peptides, and proteins, it suits workflows that would otherwise stitch together separate domain-specific surrogates, and the guidance mechanism offers a cheap route to re-targeting an existing model at a new force field.
UniSim's contribution is the demonstration that cross-domain pretraining on chemically diverse data measurably improves a learned dynamics model, rather than merely broadening its input range. Its limits are stated plainly by the authors: the released version has not undergone engineering optimization and is not yet suitable for real-world deployment, with data and model scaling still in progress. Evaluation is in silico throughout, on modest test sets of 14 peptides and 14 proteins, and the coarsened timestep is fixed per domain at training time rather than chosen at inference. The open MIT-licensed code, downloadable checkpoints, and published data splits nevertheless make it a usable starting point for work on unified molecular simulators.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.