Research map / Probabilistic and scientific modeling / Molecules and proteins
TEMPO: Temporal Multi-scale Autoregressive Generation of Protein Conformational Ensembles
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.
TEMPO hierarchically generates protein backbone MD-like trajectories with a slow 20 ns autoregressive model conditioning a fast 1 ns model, reporting gains over BioEMU, AlphaFlow-MD, ESMFlow-MD and MDGen on mdCATH and ATLAS.
Problem
Generating realistic, temporally coherent protein conformational trajectories (MD-like dynamics) is costly with MD and existing ensemble generators produce static ensembles without modeling temporal evolution.
Approach
A two-scale autoregressive SDE-inspired model: a low-resolution model generates 20 ns-spaced slow collective conformations, and a high-resolution model fills in 1 ns local fluctuations conditioned on these anchors, using an IPA plus attention plus GRU encoder with a steric clash penalty.
Key results
- On ATLAS, TEMPO reaches Pairwise RMSD Pearson r 0.91 and Global RMSF r 0.89, versus 0.48 and 0.50 for MDGen.
- On mdCATH, TEMPO achieves RMSD error 1.78 versus MDGen's 3.76, but MDGen has lower root mean W2 (3.36 vs 4.21).
- TEMPO generates 400-frame trajectories in about 0.006 hours per protein versus 4.5 to 4.7 hours for AlphaFlow-MD and ESMFlow-MD.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TEMPO rather than MDGen | protein trajectory generation · mdCATH test | RMSD Error: 3.76 -> 1.78 | Helped |
| Use TEMPO rather than AlphaFlow-MD | protein ensemble generation · ATLAS test | MD PCA W2: 1.52 -> 0.60 | Helped |
| Use TEMPO rather than AlphaFlow-MD | protein dynamics generation · ATLAS test | Pairwise RMSD r: 0.48 -> 0.91 | Helped |
| Use TEMPO rather than ESMFlow-MD | protein ensemble generation · mdCATH test | Weak contacts J: 0.51 -> 0.43 | Hurt |
| Add hierarchical multi-scale guidance instead of direct single-scale generation (ablation) | TEMPO · mdCATH test | Pairwise RMSD: 7.46 -> 2.78 | Helped |
| Add hierarchical multi-scale guidance instead of direct single-scale generation (ablation) | TEMPO · mdCATH test | Pairwise RMSD r: 0.14 -> 0.77 | Helped |
| Add hierarchical multi-scale guidance instead of direct single-scale generation (ablation) | TEMPO · mdCATH test | All-atom RMSF: 4.27 -> 1.60 | Helped |
| Add hierarchical multi-scale guidance instead of direct single-scale generation (ablation) | TEMPO · mdCATH test | Global RMSF r: 0.15 -> 0.67 | Helped |
| Add hierarchical multi-scale guidance instead of direct single-scale generation (ablation) | TEMPO · mdCATH test | Root mean W2: 8.27 -> 4.21 | Helped |
| Add hierarchical multi-scale guidance instead of direct single-scale generation (ablation) | TEMPO · mdCATH test | MD PCA W2: 2.53 -> 2.33 | Helped |
| Add hierarchical multi-scale guidance instead of direct single-scale generation (ablation) | TEMPO · mdCATH test | % PC-sim > 0.5: 3.12 -> 7.81 | Helped |
| Add hierarchical multi-scale guidance instead of direct single-scale generation (ablation) | TEMPO · mdCATH test | Weak contacts J: 0.23 -> 0.43 | Helped |
Surprising findings
- ESMFlow achieves broader free energy surface coverage than TEMPO, which the authors attribute to TEMPO's trajectory-focused sampling.
- On mdCATH, MDGen has a lower clash ratio (0.42%) than TEMPO (4.75%) and a lower root mean W2 than TEMPO.
Limitations
- Limited generalization to unseen proteins, especially for large conformational changes.
- Restricted to backbone dynamics without side chains or multi-molecular interactions.
- Performance gap between training and test sets on mdCATH (PC-sim >0.5: 46.7% vs 7.81%).
Experimental setup
Datasets: mdCATH ATLAS
Benchmarks: mdCATH ATLAS
Metrics: pairwise RMSD RMSF root mean Wasserstein distance MD PCA W2 PC cosine similarity contact Jaccard clash ratio RMSD error inference time
Baselines: BioEMU AlphaFlow-MD ESMFlow-MD MDGen single-scale TEMPO
Builds on: AlphaFlow MDGen Invariant Point Attention BioEMU
Closest papers by meaning
- Simultaneous Modeling of Protein Conformation and Dynamics via Autoregression (NeurIPS 2025 · similarity 0.92)
- Atomic Trajectory Modeling with State Space Models for Biomolecular Dynamics (NeurIPS 2026 · similarity 0.89)
- Action-Minimization Meets Generative Modeling: Efficient Transition Path Sampling with the Onsager-Machlup Functional (ICML 2025 · similarity 0.89)
- AlphaFold Meets Flow Matching for Generating Protein Ensembles (ICML 2024 · similarity 0.88)
- Generative Modeling of Molecular Dynamics Trajectories (NeurIPS 2024 · similarity 0.88)
- Aligning Protein Conformation Ensemble Generation with Physical Feedback (ICML 2025 · 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.