Research map / Probabilistic and scientific modeling / Molecules and proteins
Str2Str: A Score-based Framework for Zero-shot Protein Conformation Sampling
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 Str2Str, a score-based SE(3)^n forward-backward (perturb-then-anneal) structure-to-structure translation model trained on PDB crystal structures that samples protein conformations zero-shot without simulation data.
Problem
Protein conformation sampling is expensive with MC/MD and learning-based samplers need target-specific simulation data that is costly to acquire and does not generalize to unseen proteins.
Approach
Treat conformation sampling as a roto-translation-equivariant structure-to-structure translation on backbone frames in SE(3)^n, using a simulated-annealing-inspired forward perturbation plus score-based backward annealing, with an amortized denoising score matching objective trained on general PDB structures.
Key results
- On the fast folding benchmark, Str2Str (SDE) reaches JS-PwD 0.348 and JS-Rg 0.365 versus 0.480 and 0.661 for idpGAN, and MAE-TM 0.133 versus 0.189 for idpGAN.
- Str2Str (PF) achieves Val-Bond 0.992 and Val-Clash 0.963 while idpGAN has Val-Bond 0.032.
- On the WW domain, Str2Str (PF) takes 510 GPU seconds versus more than 160 GPU days for 100us MD, with comparable JS metrics (JS-PwD 0.379 vs 0.399).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Str2Str instead of zero-shot baselines | Fast-folding protein sampling · Fast-folding proteins | JS-PwD: 0.480 -> 0.348 | Helped |
| Use Str2Str instead of zero-shot baselines | Fast-folding protein sampling · Fast-folding proteins | MAE-RMSD: 0.592 -> 0.184 | Helped |
| Use Str2Str instead of MSA subsampling | Fast-folding protein sampling · Fast-folding proteins | Val-Clash: 0.999 -> 0.977 | Hurt |
| Use Str2Str (PF) instead of 100us MD | WW domain sampling · WW domain | sampling time: $>$160 GPU days -> 510 GPU secs | Helped |
| Use coil cutoff r_c = 0.75 instead of r_c = 0.5 during training (ablation) | Str2Str (PF) · Fast-folding proteins | JS-TIC: 0.397 -> 0.373 | Helped |
| Use coil cutoff r_c = 0.75 instead of r_c = 0.5 during training (ablation) | Str2Str (PF) · Fast-folding proteins | Val-Clash: 0.963 -> 0.953 | Hurt |
| Replace Gaussian forward perturbation with 3σ-uniform noise (ablation) | Str2Str (PF) · Fast-folding proteins | JS-PwD: 0.375 -> 0.360 | No clear effect |
| Replace Gaussian forward perturbation with 3σ-uniform noise (ablation) | Str2Str (SDE) · Fast-folding proteins | JS-PwD: 0.348 -> 0.343 | No clear effect |
| Initialize sampling with AlphaFold2 instead of ESMFold (ablation) | Str2Str (PF) · Fast-folding proteins | JS-TIC: 0.397 -> 0.374 | Helped |
| Initialize sampling with AlphaFold2 instead of ESMFold (ablation) | Str2Str (SDE) · Fast-folding proteins | JS-PwD: 0.348 -> 0.336 | Helped |
| Remove rotation modeling and use input rotations (ablation) | Str2Str (PF) · Fast-folding proteins | JS-PwD: 0.375 -> 0.675 | Hurt |
| Remove rotation modeling and use random rotations (ablation) | Str2Str (PF) · Fast-folding proteins | Val-Clash: 0.963 -> 0.002 | Hurt |
Surprising findings
- Removing rotation modeling causes validity to collapse with random rotations (Val-Clash 0.002), showing orientation modeling is necessary.
- Str2Str still underperforms 100us MD on distributional metrics in general.
Limitations
- Isotropic perturbation kernels may be biased toward less efficient subspaces relative to collective variables.
- Still underperforms long MD (e.g., 100us) on distributional fidelity metrics.
- Only sampling from a single identical input structure is evaluated; sequential use as an MCMC proposal is left to future work.
Experimental setup
Datasets: Protein Data Bank (PDB) crystal structures fast folding protein MD reference trajectories (12 targets, up to 1 ms) BPTI long MD trajectory (1.01 ms)
Benchmarks: fast folding proteins benchmark (12 targets, MD reference) BPTI TICA case study
Metrics: Val-Clash Val-Bond JS-PwD JS-TIC JS-Rg MAE-TM MAE-RMSD
Baselines: MSA subsampling EigenFold idpGAN Boltzmann generator
Builds on: denoising score matching Riemannian diffusion models FrameDiff Invariant Point Attention (AlphaFold2 structure module) simulated annealing
Closest papers by meaning
- Action-Minimization Meets Generative Modeling: Efficient Transition Path Sampling with the Onsager-Machlup Functional (ICML 2025 · similarity 0.88)
- Few-Step Cofolding with All-Atom Flow Maps (NeurIPS 2026 · similarity 0.86)
- Protein Conformation Generation via Force-Guided SE(3) Diffusion Models (ICML 2024 · similarity 0.86)
- Swallowing the Bitter Pill: Simplified Scalable Conformer Generation (ICML 2024 · similarity 0.86)
- Aligning Protein Conformation Ensemble Generation with Physical Feedback (ICML 2025 · similarity 0.86)
- SE(3)-Stochastic Flow Matching for Protein Backbone Generation (ICLR 2024 · 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.