Research map / Probabilistic and scientific modeling / Molecules and proteins
JAMUN: Bridging Smoothed Molecular Dynamics and Score-Based Learning for Conformational Ensemble Generation
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 JAMUN, an SE(3)-equivariant walk-jump sampler that runs Langevin MD in a single-noise-level smoothed space to generate all-atom peptide conformational ensembles faster than MD and transferably to unseen and longer peptides.
Problem
Generating conformational ensembles of molecules with MD is computationally expensive due to femtosecond timesteps, while many ML ensemble generators fail to transfer to systems outside their training data.
Approach
Perform Langevin dynamics in a Gaussian-smoothed (noised) all-atom coordinate space, with a single learned denoiser at one fixed noise level (sigma=0.4 Å) providing both the score for the walk step and the jump back to clean data; the denoiser is an SE(3)-equivariant GNN with normalization adjusted for relative vectors and no sequence-index features.
Key results
- JAMUN outperforms TBG (un-reweighted) on Timewarp 2AA-Large at equal time, and is only slightly worse than TBG run 20x longer.
- JAMUN passes PoseBusters checks at ~94.7% overall on 20 unseen MDGen 4AA-Explicit peptides.
- Sampling time per sample ~2-8 ms for JAMUN versus ~40 minutes to 258 hours of MD-based baselines in the reported tables; JAMUN roughly 5x faster than Boltz-1 and ~60x faster than BioEmu with H-Packer.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use JAMUN instead of reference MD (10× shorter) | Capped 2AA sampling · Capped 2AA | Jensen-Shannon distance, all torsions: 0.424 ± 0.056 -> 0.304 ± 0.112 | Helped |
| Use JAMUN instead of TBG run for the same sampling time | Timewarp 2AA-Large sampling · Timewarp 2AA-Large | Jensen-Shannon distance, all torsions: 0.105 ± 0.038 -> 0.165 ± 0.030 | Hurt |
| Use JAMUN instead of TBG run 20× shorter | Timewarp 2AA-Large sampling · Timewarp 2AA-Large | Jensen-Shannon distance, all torsions: 0.225 ± 0.071 -> 0.165 ± 0.030 | Helped |
| Use JAMUN instead of MDGen | MDGen 4AA-Explicit sampling · MDGen 4AA-Explicit | Jensen-Shannon distance, all torsions: 0.107 ± 0.028 -> 0.187 ± 0.054 | Hurt |
| Use JAMUN instead of reference MD shortened 10× | MDGen 4AA-Explicit sampling · MDGen 4AA-Explicit | Jensen-Shannon distance, all torsions: 0.095 ± 0.025 -> 0.187 ± 0.054 | Hurt |
| Train JAMUN on 4AA peptides and apply it to 5AA peptides | JAMUN · Uncapped 5AA | Jensen-Shannon distance, all torsions: 0.197 ± 0.010 vs 0.411 ± 0.029 (Boltz-1); 0.420 ± 0.018 (BioEmu) | Helped |
| Use JAMUN instead of MD sampling | Capped 2AA sampling · Capped 2AA | sampling time per sample: 40 -> 2 ms | Helped |
| Use JAMUN instead of Boltz-1 | Uncapped 5AA sampling · Uncapped 5AA | sampling time per sample: 360 -> 8 ms | Helped |
| Use JAMUN with H-Packer side-chain reconstruction instead of BioEmu | Uncapped 5AA sampling · Uncapped 5AA | sampling time per sample: 4320 -> 8 ms | Helped |
| Use JAMUN instead of full diffusion sampling with the same architecture (ablation) | JAMUN architecture · Timewarp 2AA | JSD-Backbone Torsions: 0.1501 vs 0.1363 (3149 samples); 0.0496 vs 0.0460 (200000 samples) | Mixed |
| Increase noise level from 0.4 Å to 0.8 Å (ablation) | JAMUN · Timewarp 4AA-Large | topology validity | Hurt |
| Decrease noise level from 0.4 Å to 0.2 Å (ablation) | JAMUN · Timewarp 4AA-Large | sampling efficiency | Hurt |
Surprising findings
- A JAMUN model trained only on 4-residue peptides accurately predicts ensembles for 5-residue peptides, while training on 2AA data does not transfer.
- E(3)-equivariant denoisers produced symmetric Ramachandran plots due to parity constraints, motivating the use of SE(3)-equivariant networks.
Limitations
- No reweighting or convergence guarantees; samples are approximate Boltzmann emulation.
- Does not generalize well to much longer peptides (e.g., 10AA) since 4AA training lacks secondary structure.
- Samples look more diffusive than ground truth for macrocycles because CREST data keeps only local minima.
Experimental setup
Datasets: Timewarp 2AA-Large Timewarp 4AA-Large MDGen 4AA-Explicit Capped 2AA Uncapped 5AA CREMP 4AA CREMP 6AA
Benchmarks: Timewarp 2AA-Large Timewarp 4AA-Large MDGen 4AA-Explicit Capped 2AA Uncapped 5AA CREMP 4AA CREMP 6AA
Metrics: Jensen-Shannon distance on backbone/sidechain torsions TICA projections JSD metastable state probabilities decorrelation time speedup wall-clock sampling time PoseBusters pass rate
Baselines: TBG MDGen Boltz-1 BioEmu RINGER Timewarp
Builds on: walk-jump sampling EDM preconditioning NequIP BAOAB integrator Timewarp
Closest papers by meaning
- Amortized Sampling with Transferable Normalizing Flows (NeurIPS 2025 · similarity 0.88)
- Aligning Protein Conformation Ensemble Generation with Physical Feedback (ICML 2025 · similarity 0.87)
- Simultaneous Modeling of Protein Conformation and Dynamics via Autoregression (NeurIPS 2025 · similarity 0.87)
- AlphaFold Meets Flow Matching for Generating Protein Ensembles (ICML 2024 · similarity 0.87)
- TEMPO: Temporal Multi-scale Autoregressive Generation of Protein Conformational Ensembles (NeurIPS 2025 · similarity 0.87)
- Swallowing the Bitter Pill: Simplified Scalable Conformer Generation (ICML 2024 · 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.