Research map / Probabilistic and scientific modeling / Molecules and proteins
AlphaFold Meets Flow Matching for Generating Protein 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.
Repurposes AlphaFold and ESMFold as sequence-conditioned flow-matching generative models (AlphaFlow, ESMFlow), fine-tuned on the PDB and then on ATLAS MD ensembles to sample protein conformational ensembles.
Problem
Single-state structure predictors like AlphaFold and ESMFold ignore conformational heterogeneity; existing ensemble methods rely on inference-time MSA subsampling, which cannot apply to PLM-based predictors or learn from MD ensembles.
Approach
Convert regression-style structure predictors into denoising models by adding a noisy-structure input embedding, and train them under a custom flow matching framework on a harmonic prior over C-beta coordinates in the SE(3) quotient space, using a squared FAPE loss to learn the Frechet mean.
Key results
- On PDB ensembles, AlphaFlow traces a superior precision-diversity Pareto frontier relative to MSA subsampling.
- On ATLAS MD test ensembles, AlphaFlow-MD achieves a 0.85 median per-target RMSF Pearson correlation, while no MSA subsampling level meaningfully exceeds AlphaFold (0.52).
- AlphaFlow-MD nearly doubles the success rate of dominant PCA motion modeling (44% vs 23% for AlphaFold with cosine similarity > 0.5).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| fine-tune AlphaFold with flow matching on MD ensembles | AlphaFold · ATLAS test ensembles | Pairwise RMSD r: 0.10 -> 0.48 | Helped |
| fine-tune AlphaFold with flow matching on MD ensembles | AlphaFold · ATLAS test ensembles | Per-target RMSF r: 0.52 -> 0.85 | Helped |
| fine-tune AlphaFold with flow matching on MD ensembles | AlphaFold · ATLAS test ensembles | Root mean W2-distance: 3.58 -> 2.61 | Helped |
| fine-tune AlphaFold with flow matching on MD ensembles | AlphaFold · ATLAS test ensembles | Weak contacts J: 0.27 -> 0.62 | Helped |
| fine-tune AlphaFold with flow matching on MD ensembles | AlphaFold · ATLAS test ensembles | Exposed MI matrix rho: 0.02 -> 0.25 | Helped |
| provide the PDB structure that initialized the simulation as a template | AlphaFlow-MD · ATLAS test ensembles with templates | Root mean W2-distance: 2.61 -> 1.95 | Helped |
| provide the PDB structure that initialized the simulation as a template | AlphaFlow-MD · ATLAS test ensembles with templates | Weak contacts J: 0.62 -> 0.62 | No clear effect |
| distill the generative process into a single forward pass (ablation) | AlphaFlow-MD · ATLAS test ensembles | Root mean W2-distance: 2.61 -> 3.70 | Hurt |
| distill the generative process into a single forward pass (ablation) | AlphaFlow-MD · ATLAS test ensembles | Pairwise RMSD r: 0.48 -> 0.48 | No clear effect |
| distill the generative process into a single forward pass (ablation) | AlphaFlow-MD · ATLAS test ensembles | Weak contacts J: 0.62 -> 0.52 | Hurt |
| use MSA subsampling instead of flow matching | AlphaFold · PDB heterogeneous-protein test set | precision-diversity Pareto frontier | Hurt |
| use ESMFlow instead of deterministic ESMFold | ESMFold · PDB heterogeneous-protein test set | recall | Helped |
Surprising findings
- Neither AlphaFlow nor MSA subsampling meaningfully improves aggregate recall over baseline AlphaFold on PDB ensembles.
- Distilled models converge to lower performance than the full multi-step models but offer greater speedups for short-timescale sampling.
Limitations
- Evaluation limited to phenomena within ATLAS simulation timescales; slower conformational changes not assessed.
- Model is effectively a generative model over C-beta positions only; all-atom output is filled in without iterative refinement.
- Recall is not improved over baseline AlphaFold on PDB ensembles.
Experimental setup
Datasets: PDB ATLAS OpenProteinSet CAMEO
Benchmarks: PDB ensembles test set (100 recent heterogeneous proteins) ATLAS MD test set (82 proteins)
Metrics: precision (lDDT-Calpha) recall (lDDT-Calpha) diversity (1-lDDT-Calpha) pairwise RMSD correlation RMSF correlation root mean Wasserstein distance PCA Wasserstein distance Jaccard similarity of contacts mutual information Spearman correlation
Baselines: AlphaFold with MSA subsampling ESMFold AlphaFold-MD EigenFold replicate MD simulations
Builds on: AlphaFold ESMFold OpenFold flow matching harmonic diffusion
Closest papers by meaning
- SE(3)-Stochastic Flow Matching for Protein Backbone Generation (ICLR 2024 · similarity 0.92)
- Energy-Based Flow Matching for Generating 3D Molecular Structure (ICML 2025 · similarity 0.91)
- Learning conformational ensembles of proteins based on backbone geometry (NeurIPS 2025 · similarity 0.91)
- Sequence-Augmented SE(3)-Flow Matching For Conditional Protein Generation (NeurIPS 2024 · similarity 0.91)
- Inverse problems with experiment-guided AlphaFold (ICML 2025 · similarity 0.90)
- Few-Step Cofolding with All-Atom Flow Maps (NeurIPS 2026 · similarity 0.89)
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.