Research map / Probabilistic and scientific modeling / Molecules and proteins
Inverse problems with experiment-guided AlphaFold
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 experiment-guided AlphaFold, which guides AlphaFold's diffusion sampling with electron density and NOE likelihoods via non-i.i.d. joint ensemble sampling, followed by force-field relaxation and matching-pursuit ensemble selection.
Problem
Current structure predictors output a single conformation and miss the conformational heterogeneity revealed by experiments such as X-ray crystallography (electron density, altlocs) and NMR (NOE restraints, dynamics).
Approach
Treat AlphaFold as a sequence-conditioned structural prior and cast ensemble modeling as posterior inference given experimental data; guidance gradients of the likelihood of the whole ensemble (non-i.i.d.) are added to the diffusion SDE, followed by force-field relaxation and matching-pursuit selection.
Key results
- On ubiquitin, non-i.i.d. NOE guidance raises correlation with experimental N-H S2 from r=0.47 (AlphaFold) to r=0.72.
- On NMR peptides 1DEC, density of NOE violations drops from 31.2% (AlphaFold) to 15.0% with guidance, matching the NMR-deposited 11.4% closely.
- Guided ensembles take about 1-5 minutes versus hours for MD-based NMR structure determination (e.g., 73-324 s for crystallographic guidance).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Guide AlphaFold with electron density | AlphaFold · ORF8 7F5F:A residues 102-112 and 7JX6:A residues 102-112 | cosine similarity: 0.562 -> 0.699; 0.561 -> 0.696 | Helped |
| Guide AlphaFold with electron density | AlphaFold · ORF8 7F5F:A and 7JX6:A residues 38-41 | cosine similarity: 0.645 -> 0.785; 0.579 -> 0.736 | Helped |
| Guide AlphaFold with NOE restraints | AlphaFold · 1DEC and 2LI3 | NOE restraint violations: 31.2% -> 15.0%; 29.3% -> 5.3% | Helped |
| Guide AlphaFold with NOE restraints | AlphaFold · 3BBG | NOE restraint violations: 33.3% -> 21.7% | Helped |
| Guide AlphaFold with NOE restraints | AlphaFold · 3BBG | backbone flexibility agreement (ρ-RMSF): 0.63 -> 0.49 | Hurt |
| Use non-i.i.d. rather than i.i.d. ensemble sampling (ablation) | NOE-guided AlphaFold · Ubiquitin | correlation with experimental N-H S²: r=0.52 -> r=0.72 | Helped |
Surprising findings
- Guided ensembles outperform PDB-deposited NMR structures in NOE restraint compliance for several NMRDb proteins.
Limitations
- Guided sampling overfits to experimental noise, creating broken bonds and clashes that need filtering and relaxation.
- Runtime grows linearly with ensemble size and guidance is not separable across ensemble members.
- Limited to monomers; complexes, ligands and cryo-EM not yet addressed.
Experimental setup
Datasets: PDB crystallographic electron density maps NMR NOE restraints from PDB NMR structures 100 NMR spectra database (NMRDb) ubiquitin N-H S2 order parameters
Benchmarks: PDB electron density cosine similarity (ORF8, NBR1 altlocs, homologous pairs) NOE violation percentage on NMRDb and peptide benchmarks ubiquitin N-H S2 correlation
Metrics: cosine similarity between observed and calculated density NOE violation percentage and median violation rho-RMSF Pearson correlation with S2 runtime
Baselines: AlphaFold2 AlphaFlow ESMFlow Chroma Guided Chroma PDB-deposited structures NMR ensembles from PDB
Builds on: AlphaFold2 AlphaFlow Chroma AMBER
Closest papers by meaning
- AlphaFold Meets Flow Matching for Generating Protein Ensembles (ICML 2024 · similarity 0.90)
- AlphaFold Database Debiasing for Robust Inverse Folding (NeurIPS 2025 · similarity 0.88)
- Learning conformational ensembles of proteins based on backbone geometry (NeurIPS 2025 · similarity 0.87)
- Unlocking hidden biomolecular conformational landscapes in diffusion models at inference time (NeurIPS 2025 · similarity 0.87)
- Aligning Protein Conformation Ensemble Generation with Physical Feedback (ICML 2025 · similarity 0.87)
- Simultaneous Modeling of Protein Conformation and Dynamics via Autoregression (NeurIPS 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.