Research map / Probabilistic and scientific modeling / Molecules and proteins
Learning conformational ensembles of proteins based on backbone geometry
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 BBFlow, a backbone-only SE(3) flow matching model that emulates MD conformational ensembles by conditioning both the flow and a geodesic-interpolated prior on the equilibrium structure, without MSA or folding-model weights.
Problem
Generating MD-emulating protein conformational ensembles currently relies on fine-tuning large folding models (AlphaFold 2, ESMFold) and evolutionary sequence information, which is expensive and biased for de novo proteins and limits applicability.
Approach
Formulate ensemble generation as backbone structure generation conditioned on the equilibrium structure (distance and direction encodings, no residue indices), and use a conditional prior obtained by geodesic interpolation between uniform noise and the equilibrium structure, trained from scratch.
Key results
- On the ATLAS test set, BBFlow achieves inference time 0.8 s per conformation versus 32.6 s for AlphaFlow-T (about 40x faster for 302-residue protein) with comparable accuracy (RMSF MAE 0.42 vs 0.41 for AlphaFlow-T).
- On de novo proteins, BBFlow obtains Pw-RMSD MAE 0.32 versus 7.40 for AlphaFlow (no templates) and 0.38 for AlphaFlow-T.
- BBFlow-light is around 200 times faster than AlphaFlow with 2.5M parameters versus 18.2M for BBFlow.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use backbone-geometry conditioning instead of the AlphaFlow-T template approach | BBFlow vs AlphaFlow-T · ATLAS test set | inference time: 32.6 -> 0.8 | Helped |
| use backbone-geometry conditioning instead of the AlphaFlow-T template approach | BBFlow vs AlphaFlow-T · ATLAS test set | pairwise RMSD MAE: 0.91 -> 0.77 | Helped |
| use backbone-geometry conditioning instead of the strongest baseline | BBFlow vs ESMFlow-T · ATLAS test set | transient contact accuracy: 47 -> 29 | Hurt |
| use BBFlow rather than AlphaFlow-T for de novo proteins | BBFlow vs AlphaFlow-T · de novo proteins | pairwise RMSD MAE: 0.38 -> 0.32 | Helped |
| use BBFlow rather than ESMFlow-T for de novo proteins | BBFlow vs ESMFlow-T · de novo proteins | transient contact accuracy: 56 -> 32 | Hurt |
| add direction encoding (ablation) | BBFlow · ATLAS test set | pairwise RMSD MAE: 1.15 -> 0.77 | Helped |
| add the conditional prior (ablation) | BBFlow · ATLAS test set | pairwise RMSD MAE: 0.90 -> 0.77 | Helped |
| remove residue-index encoding (ablation) | BBFlow · ATLAS test set | pairwise RMSD MAE: 0.82 -> 0.77 | Helped |
| remove amino-acid identity features (ablation) | BBFlow · ATLAS test set | RMSF MAE: 0.42 -> 0.54 | Hurt |
| remove equilibrium distance encoding (ablation) | BBFlow · ATLAS test set | RMSF MAE: 0.42 -> 5.88 | Hurt |
| train with ξ=0.1 rather than ξ=0.2 (ablation) | BBFlow · ATLAS test set | RMSF MAE: 0.42 -> 0.54 | Hurt |
| train with ξ=0.4 rather than ξ=0.2 (ablation) | BBFlow · ATLAS test set | RMSF median: 1.49 -> 1.39 | Mixed |
Surprising findings
- Without amino acid identity input, the model still performs remarkably well, and a backbone-only variant without sequence information is on par with non-template AlphaFlow.
- BBFlow underperforms MSA-based AlphaFlow on transient contact accuracy, suggesting evolutionary information helps predict rare events.
Limitations
- Limited to reproducing MD-derived distributions; cannot predict distant alternative folding states without longer MD training data.
- Underperforms AlphaFlow on transient contact prediction, suggesting evolutionary information is needed for rare events.
- Does not model sidechain conformational ensembles or protein-ligand interactions.
Experimental setup
Datasets: ATLAS de novo proteins generated by FrameFlow and RFdiffusion with MD (50 proteins) five multi-chain systems simulated with MD
Benchmarks: ATLAS test set de novo protein MD dataset multi-chain protein MD systems
Metrics: RMSF Pearson r RMSF MAE pairwise RMSD MAE DCCM Pearson r PCA Wasserstein-2 transient contact Jaccard inference time Ramachandran Wasserstein-2
Baselines: AlphaFlow AlphaFlow-T ESMFlow-T AlphaFlow-T_dist AlphaFlow-T_12L,dist ConfDiff BioEmu MDGen Str2Str ESMDiff
Builds on: FrameFlow GAFL AlphaFlow Riemannian flow matching
Closest papers by meaning
- SE(3)-Stochastic Flow Matching for Protein Backbone Generation (ICLR 2024 · similarity 0.92)
- AlphaFold Meets Flow Matching for Generating Protein Ensembles (ICML 2024 · similarity 0.91)
- Proteina: Scaling Flow-based Protein Structure Generative Models (ICLR 2025 · similarity 0.91)
- Energy-Based Flow Matching for Generating 3D Molecular Structure (ICML 2025 · similarity 0.90)
- Sequence-Augmented SE(3)-Flow Matching For Conditional Protein Generation (NeurIPS 2024 · similarity 0.89)
- Generating Highly Designable Proteins with Geometric Algebra Flow Matching (NeurIPS 2024 · similarity 0.88)
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.