Research map / Probabilistic and scientific modeling / Molecules and proteins
4D Diffusion for Dynamic Protein Structure Prediction with Reference and Motion Guidance
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 a 4D score-based diffusion model that jointly generates multi-step MD-like protein trajectories (backbone and side chains) conditioned on a reference structure, using a reference network and a temporal motion alignment module.
Problem
Protein structure prediction has focused on static structures, while dynamic protein conformations over time (MD trajectories) are underexplored; existing learned MD methods mostly handle small systems or coarse-grained representations and predict iteratively.
Approach
A unified SE(3)/torsion-angle score-based diffusion model that generates S time steps simultaneously, with a reference network that encodes the initial structure via self-attention and a temporal self-attention motion alignment module to keep trajectories coherent.
Key results
- On ATLAS (S2L), R32 C-alpha RMSE reduced from 4.60 (FlowMatching) to 2.12 Å.
- On Fast-Folding proteins (S2L), R32 RMSE improved from 5.48 (FlowMatching) to 4.39 Å.
- Motion alignment lowers ATLAS R2 error from 1.40 to 1.30 Å; simultaneous generation beats iterative (R32 2.27 vs 3.22 Å).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Ours (S2L) instead of FlowMatching | 4D Diffusion · ATLAS (R32, S2L) | Cα-RMSE: 4.60 -> 2.12 | Helped |
| Use Ours (S2L) instead of FlowMatching | 4D Diffusion · Fast-Folding Protein (R32, S2L) | Cα-RMSE: 5.48 -> 4.39 | Helped |
| Add motion alignment (ablation) | 4D Diffusion on ATLAS · ATLAS (R2) | Cα-RMSE: 1.40 -> 1.30 | Helped |
| Predict all trajectory steps simultaneously instead of iteratively (ablation) | 4D Diffusion on ATLAS · ATLAS (R32) | Cα-RMSE: 3.22 -> 2.27 | Helped |
| Increase the number of training proteins (ablation) | 4D Diffusion · Protein trajectories; trajectory length held constant | Cα-RMSE | Helped |
| Increase training trajectory length (ablation) | 4D Diffusion · Protein trajectories; number of proteins fixed | Cα-RMSE | Helped |
Limitations
- Accuracy degrades for long-term predictions and large conformational changes.
- Computational cost grows with sequence length; sparse edge features are proposed as future work.
- Evaluated only on 758 ATLAS proteins and six fast-folding proteins.
Experimental setup
Datasets: ATLAS Fast-Folding proteins
Benchmarks: ATLAS Fast-Folding proteins
Metrics: C-alpha RMSE
Baselines: DFF FlowMatching DiffMD iterative prediction
Builds on: AlphaFold2 Invariant Point Attention score-based diffusion model DiffMD GeoFormer
Closest papers by meaning
- Generative Modeling of Full-Atom Protein Conformations using Latent Diffusion on Graph Embeddings (NeurIPS 2025 · similarity 0.88)
- Simultaneous Modeling of Protein Conformation and Dynamics via Autoregression (NeurIPS 2025 · similarity 0.88)
- Energy-Based Flow Matching for Generating 3D Molecular Structure (ICML 2025 · similarity 0.87)
- Predicting directional flexibility in proteins (NeurIPS 2026 · similarity 0.87)
- DPLM-2: A Multimodal Diffusion Protein Language Model (ICLR 2025 · similarity 0.87)
- SE(3)-Stochastic Flow Matching for Protein Backbone Generation (ICLR 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.