Research map / Probabilistic and scientific modeling / Molecules and proteins
Generative Modeling of Molecular Dynamics Trajectories
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 MDGen, a flow-matching (stochastic interpolant + SiT/Hyena transformer) generative model over SE(3)-invariant key-frame-relative tokens of MD trajectories, supporting forward simulation, transition path sampling, upsampling, and inpainting design on tetrapeptides and proteins.
Problem
Existing deep learning surrogates of molecular dynamics learn only transition densities or equilibrium distributions, so they cannot readily address trajectory-level inverse problems such as transition path sampling, upsampling, or dynamics-conditioned design.
Approach
Model entire MD trajectories as time-series of SE(3)-invariant tokens (roto-translation offsets relative to conditioning key frames plus torsion angles) with a stochastic-interpolant flow model using alternating residue and time attention; different conditioning masks yield forward simulation, interpolation, upsampling, and inpainting from one framework.
Key results
- Forward simulation achieves JSD on TICA-0 of 0.230 versus 0.323 for 10 ns replicate MD, with ~60 GPU-seconds per 100 ns trajectory versus ~3 GPU-hours for MD.
- Transition path sampling yields higher path likelihoods than MSMs from replicate MD shorter than 100 ns, with best JSD and valid-path fraction.
- Inpainting recovers ground-truth residues substantially more often than DynMPNN; protein ensembles fall between AlphaFlow and MSA subsampling in accuracy while being much faster.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MDGen forward simulation instead of a 10 ns replicate MD simulation | Test peptide forward simulation · Test peptides, all torsions | JSD: .125 -> .109 | Helped |
| Use MDGen forward simulation instead of a 10 ns replicate MD simulation | Test peptide forward simulation · Test peptides, TICA-0 | JSD: .323 -> .230 | Helped |
| Use MDGen forward simulation instead of a 10 ns replicate MD simulation | Test peptide forward simulation · Test peptides, MSM states | JSD: .363 -> .235 | Helped |
| Use MDGen forward simulation instead of a 100 ns replicate MD simulation | Test peptide forward simulation · Test peptides, all torsions | JSD: .076 -> .109 | Hurt |
| Use MDGen forward simulation instead of replicate MD | Test peptide forward simulation · 100 ns-equivalent trajectories | Runtime: 3h -> 60s | Helped |
| Generate transition paths with MDGen rather than replicate MD MSMs shorter than 100 ns | Transition path sampling · Test peptides, rare transitions | Path likelihood under reference MSM | Helped |
| Generate transition paths with MDGen rather than replicate MD MSMs | Transition path sampling · Test peptides, rare transitions | JSD of visited-state distribution | Helped |
| Generate transition paths with MDGen rather than replicate MD MSMs | Transition path sampling · Test peptides, rare transitions | Valid path fraction | Helped |
| Upsample 10 ps trajectories with MDGen | Trajectory upsampling · Test peptides, torsion autocorrelation | Recovery of fast dynamics | Helped |
| Condition peptide inpainting on a residue trajectory rather than static endpoint frames | MDGen inpainting versus DynMPNN · Test peptides, high-flux and random paths | Sequence recovery | Helped |
| Use MDGen instead of MSA subsampling with AlphaFold | Protein forward simulation · ATLAS proteins | Ensemble statistical properties | Helped |
Surprising findings
- Withholding amino acid identities and torsions of flanking residues in inpainting was necessary to prevent overfitting and improve generalization.
Limitations
- Reliance on key frames means no unconditional generation or inpainting of residue roto-translations.
- Weaker performance on proteins than on peptides, suggesting the architecture may not handle larger motions.
- Tokenization limited to peptides/proteins; does not yet cover ligands, materials, or explicit solvent.
Experimental setup
Datasets: ATLAS
Benchmarks: tetrapeptide MD simulations ATLAS
Metrics: Jensen-Shannon divergence decorrelation time flux matrix Spearman correlation valid path rate path likelihood sequence recovery runtime
Baselines: replicate MD simulations DynMPNN S-MPNN AlphaFlow MSA subsampling
Builds on: Scalable Interpolant Transformer (SiT) stochastic interpolants Dirichlet flow matching Hyena DiT
Closest papers by meaning
- Action-Minimization Meets Generative Modeling: Efficient Transition Path Sampling with the Onsager-Machlup Functional (ICML 2025 · similarity 0.91)
- Atomic Trajectory Modeling with State Space Models for Biomolecular Dynamics (NeurIPS 2026 · similarity 0.89)
- Simultaneous Modeling of Protein Conformation and Dynamics via Autoregression (NeurIPS 2025 · similarity 0.89)
- Looking Under the Streetlight: Evaluation in Generative Molecular Dynamics (NeurIPS 2026 · similarity 0.88)
- TEMPO: Temporal Multi-scale Autoregressive Generation of Protein Conformational Ensembles (NeurIPS 2025 · similarity 0.88)
- FlashMD: long-stride, universal prediction of molecular dynamics (NeurIPS 2025 · similarity 0.87)
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.