Research map / Image generation / Autoregressive and latent generation
NMA-tune: Generating Highly Designable and Dynamics Aware Protein Backbones
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.
Introduces NMA-tune, a small trainable GVP-GNN conditioner added to frozen RFdiffusion that steers backbone generation toward a target lowest normal mode (hinge motion) while keeping designability.
Problem
Generating protein backbones that are both designable and exhibit a user-specified large-scale collective motion (lowest normal mode), since prior NMA-guidance trades off designability and conditioning efficiency.
Approach
Replace the hand-set loss-guidance weighting with a trainable correction network that takes the gradient of an NMA loss w.r.t. the predicted clean structure (ignoring the Jacobian) and learns a noise correction added to frozen RFdiffusion's noise prediction, trained with noise-matching, NMA, chain-distance and radius-of-gyration losses.
Key results
- On 1exr with eta=0.0, NMA-tune achieves 9.4% designable samples with sc-cossim>0.9 vs 7.6% for NMA-guidance and 1.2% unconditioned (110 samples per setting, 3 seeds).
- NMA-tune increases sampling time by about 8% versus about 75% for NMA-guidance.
- For 1hhp at eta=0.0, MD PCA-cossim for dynamics-conditioned designs reaches 0.86 (sc-cossim 0.92) vs 0.23 for unconditioned designs during 20-100 ns.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the trainable conditioner for normal-mode dynamics conditioning rather than using NMA-guidance | RFdiffusion joint motif and dynamics conditioning · 1hhp, 1exr, 1hhp assembly, and 4tgl | % designable and sc-cossim > 0.8: 1hhp: 15.5 -> 20.3; 1exr: 20.9 -> 26.1; 1hhp assembly: 2.1 -> 7.0; 4tgl: 0.0 -> 0.9 | Helped |
| Add joint dynamics conditioning rather than structure conditioning only | RFdiffusion · 1hhp, 1exr, and 1hhp assembly (η = 0.0) | % designable and sc-cossim > 0.8: 1hhp: 7.3 -> 20.3; 1exr: 4.8 -> 26.1; 1hhp assembly: 0.6 -> 7.0 | Helped |
| Add joint dynamics conditioning rather than structure conditioning only | RFdiffusion · 4tgl (η = 0.0) | % designable and sc-cossim > 0.8: 0.0 -> 0.9 | Helped |
| Use NMA-tune rather than NMA-guidance | RFdiffusion motif and dynamics conditioning · 1hhp, 1exr, 1hhp assembly, and 4tgl (η = 0.0) | % designable: 1hhp: 67.3 -> 58.2; 1exr: 68.8 -> 53.3; 1hhp assembly: 47.7 -> 20.3; 4tgl: 13.2 -> 11.5 | Hurt |
| Use joint dynamics conditioning rather than structure conditioning only | RFdiffusion · 1hhp and 1exr (η = 0.0) | % designable: 1hhp: 72.7 -> 58.2; 1exr: 81.8 -> 53.3 | Hurt |
| Increase noise scale η from 0.0 to 1.0 (ablation) | NMA-tune · 1hhp, 1exr, 1hhp assembly, and 4tgl | % designable: 1hhp: 58.2 -> 42.7; 1exr: 53.3 -> 46.4; 1hhp assembly: 20.3 -> 15.8; 4tgl: 11.5 -> 8.2 | Hurt |
| Increase noise scale η from 0.0 to 1.0 (ablation) | NMA-tune · 1hhp, 1exr, 1hhp assembly, and 4tgl | % designable and sc-cossim > 0.8: 1hhp: 20.3 -> 13.9; 1exr: 26.1 -> 17.3; 1hhp assembly: 7.0 -> 2.4; 4tgl: 0.9 -> 0.3 | Mixed |
| Add the trainable conditioner instead of leaving dynamics unconditioned | RFdiffusion-generated 1hhp designs · 1hhp, MD (20–100 ns) | PCA-cossim: 0.23 -> 0.86; 0.34 -> 0.58 | Helped |
| Add the trainable conditioner instead of leaving dynamics unconditioned | RFdiffusion-generated 1hhp designs · 1hhp, MD (tailored stable interval) | PCA-cossim: 0.51 -> 0.50; 0.80 -> 0.48 | Mixed |
| Replace loss-guidance with a trainable correction network | RFdiffusion dynamics conditioning · Protein sampling | sampling time: NMA-tune increases the sampling time only by 8%; NMA-guidance ... increase by about 75% | Helped |
Surprising findings
- Both NMA-guidance and NMA-tune reduce designability relative to structure-only generation, and NMA-tune's designability is lower than unconditioned RFdiffusion on most targets.
- Structure-conditioned samples already match the dynamics target to some degree because the force-field Hessian depends on pairwise distances.
Limitations
- Targeted mode does not consistently persist in alternative conformations during long MD simulations.
- Conditioning only controls displacement directions, not absolute displacement amplitudes.
- Designability decreases with dynamics conditioning, and some targets were not designable even with structure-only conditioning, requiring relaxed thresholds.
Experimental setup
Datasets: SCOPe PDB 4tgl PDB 1exr PDB 1hhp
Benchmarks: 4tgl 1exr 1hhp 1hhp assembly
Metrics: designability (scRMSD, sc-motif-RMSD, pLDDT, pAE) sc-cossim orig-cossim PCA-cossim success rate fraction
Baselines: NMA-guidance RFdiffusion unconditional (structure-only)
Builds on: RFdiffusion NMA-guidance loss-guided diffusion ProteinMPNN
Closest papers by meaning
- Proteina: Scaling Flow-based Protein Structure Generative Models (ICLR 2025 · similarity 0.89)
- TEMPO: Temporal Multi-scale Autoregressive Generation of Protein Conformational Ensembles (NeurIPS 2025 · similarity 0.86)
- Predicting directional flexibility in proteins (NeurIPS 2026 · similarity 0.86)
- SE(3)-Stochastic Flow Matching for Protein Backbone Generation (ICLR 2024 · similarity 0.86)
- 4D Diffusion for Dynamic Protein Structure Prediction with Reference and Motion Guidance (AAAI 2025 · similarity 0.85)
- CarbonNovo: Joint Design of Protein Structure and Sequence Using a Unified Energy-based Model (ICML 2024 · 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.