Research map / Probabilistic and scientific modeling / Molecules and proteins
Transferable Boltzmann Generators
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 transferable Boltzmann Generators based on flow-matched equivariant CNFs that zero-shot sample Boltzmann distributions of unseen dipeptides, with reweighting, and improves single-molecule alanine dipeptide BGs.
Problem
Boltzmann Generators must be retrained for every molecular system, which is costly; the paper targets generating equilibrium samples and reweighting them for unseen molecules without retraining.
Approach
An EGNN-parameterized continuous normalizing flow trained with conditional flow matching on Cartesian coordinates, using atom-type, amino-acid, and sequence-position embeddings so that one model generalizes across dipeptides, with bond-graph and chirality checks for valid samples.
Key results
- On unseen dipeptides, TBG + full reaches mean ESS 15.29% and 98% correct configurations, versus 0.48% ESS for TBG and 0.58% for TBG + backbone.
- On alanine dipeptide with the classical force field, TBG + full improves NLL from -109.02 to -127.06 and ESS from 1.56% to 6.03% over BG + backbone.
- TBG + full outperforms Timewarp in Wasserstein distance of Ramachandran plots across 30k energy evaluations, 12h and 24h budgets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add amino-acid and sequence-position embeddings to the atom encoding (ablation) | TBG + full · Unseen dipeptides | ESS: 0.58\pm1.04\% -> 15.29\pm9.27\% | Helped |
| Add amino-acid and sequence-position embeddings to the atom encoding (ablation) | TBG + full · Unseen dipeptides | Correct configurations: 17\pm21\% -> 98\pm2\% | Helped |
| Add amino-acid and sequence-position embeddings to the atom encoding (ablation) | TBG + full · Unseen dipeptides | ESS: 0.48\pm0.59\% -> 15.29\pm9.27\% | Helped |
| Add amino-acid and sequence-position embeddings to the atom encoding (ablation) | TBG + full · Unseen dipeptides | Correct configurations: 13\pm18\% -> 98\pm2\% | Helped |
| Train on tenfold shorter trajectories (ablation) | TBG + full · Unseen dipeptides (8 test peptides) | ESS: 15.29\pm9.27\% -> 6.13\pm3.13\% | Hurt |
| Train on tenfold shorter trajectories (ablation) | TBG + full · Unseen dipeptides (8 test peptides) | Correct configurations: 98\pm2\% -> 96\pm3\% | Hurt |
| Train on 100-times shorter trajectories (500 ps) (ablation) | TBG + full · Unseen dipeptides | Capture of metastable states | Hurt |
| Bias training data toward positive φ states (ablation) | TBG + full · Unseen dipeptides | Free-energy landscape accuracy | Helped |
| Bias training data toward positive φ states (ablation) | TBG + full · Unseen dipeptides | ESS: 15.29\pm9.27\% -> 10.24\pm7.14\% | Hurt |
| Use a harmonic prior instead of the standard Gaussian prior (ablation) | Transferable Boltzmann Generator · Molecular systems | Performance | No clear effect |
Surprising findings
- Training on trajectories 100 times shorter (500 ps) still yields usable models for the 2AA dataset, although some metastable states are missed.
- Harmonic prior gave no significant improvement for the transferable model.
Limitations
- Scaling to larger systems remains untested; only dipeptides (2AA) were evaluated.
- CNF inference with Jacobian traces is expensive, taking about four days for 3e4 dipeptide samples.
- Samples with wrong bond graph or chirality must be filtered, and extensive hyperparameter tuning was needed.
Experimental setup
Datasets: 2AA dipeptide dataset (Timewarp) alanine dipeptide trajectories
Benchmarks: alanine dipeptide 2AA dipeptides test set
Metrics: effective sample size negative log likelihood correct configuration rate Wasserstein distance of Ramachandran plots free energy difference
Baselines: BG + backbone TBG TBG + backbone Timewarp
Builds on: Boltzmann Generators conditional flow matching EGNN Timewarp
Closest papers by meaning
- Scalable Equilibrium Sampling with Sequential Boltzmann Generators (ICML 2025 · similarity 0.90)
- Temperature-Annealed Boltzmann Generators (ICML 2025 · similarity 0.89)
- BoltzNCE: Learning likelihoods for Boltzmann Generation with Stochastic Interpolants and Noise Contrastive Estimation (NeurIPS 2025 · similarity 0.89)
- Path Gradients after Flow Matching (NeurIPS 2025 · similarity 0.88)
- Conditional Normalizing Flows for Active Learning of Coarse-Grained Molecular Representations (ICML 2024 · similarity 0.85)
- Few-Step Boltzmann Generators via Scalable Likelihood Flow Maps (NeurIPS 2026 · similarity 0.85)
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.