henhao
Sign in
BiophysicsNature Methods
Peer-reviewed journal article

AlphaFold as a Prior: Experimental StructureDetermination Conditioned on a Pretrained Neural Network

Alisia FadiniiD, Minhuan LiiD, Airlie J. McCoyiD, Suresh BanjaraiD, Hiroki OkumuraiD, Eve NapieriD, Pietro FontanaiD, Amir R. Khan, Luca JovineiD, Thomas C. TerwilligeriD, Randy J. ReadiD, Doeke R. HekstraiD, Mohammed AlQuraishiiD

Abstract

Advances in machine learning have transformed structural biology, enabling swift and accurate prediction of protein structure from sequence. However, challenges persist in capturing sidechain packing, condition-dependent conformational dynamics, and biomolecular interactions, primarily due to scarcity of high-quality training data. Emerging techniques, including cryo-electron tomography (cryo-ET) and high-throughput crystallography, promise vast new sources of structural data, but translating experimental observations into mechanistically interpretable atomic models remains a key bottleneck. Here, we address these challenges by improving the efficiency of structural analysis through combining experimental measurements with a landmark protein structure prediction method - AlphaFold2. We present an augmentation of AlphaFold2, ROCKET, that refines its predictions using cryo-EM, cryo-ET, and X-ray crystallography data, and demonstrate that this approach captures biologically important structural variation that AlphaFold2 does not. By performing structure optimization in the space of coevolutionary embeddings, rather than Cartesian coordinates, ROCKET automates difficult modeling tasks, such as flips of functional loops and domain rearrangements, that are beyond the scope of current state-of-the-art methods and, in some instances, even manual human modeling. The ability to efficiently sample these barrier-crossing rearrangements unlocks a new horizon for scalable and automated model building. Crucially, ROCKET does not require retraining of AlphaFold2 and is readily adaptable to multimers, ligand-cofolding, and other data modalities. Conversely, our differentiable crystal-lographic and cryo-EM target functions are capable of augmenting other structure prediction methods. ROCKET thus provides an extensible framework for the integration of experimental observables with biomolecular machine learning.

PDFDOI
Full-text digest ↓

02 /Specification & Classification

Click any tag to browse matching papers
Citations
22
FWCI
33.07
Field percentile
top 1%
OpenAlex Topics

In-depth summary

✦ Multi-agent verified full-text synthesis

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 2 (11 Mar 2025) · CC BY 4.0

ROCKET refines AlphaFold2 predictions against crystallographic, cryo-EM and cryo-ET data by gradient descent on MSA cluster profiles, capturing loop flips and domain shifts without retraining.

Refinement in Cartesian coordinates plus pattern-matching model building struggles with large rearrangements and with maps worse than 4 to 5 Å, while ML alternatives either retrain networks (costly, modality-specific) or alternate prediction and rebuilding with poor convergence. ROCKET instead treats pretrained AF2 as an inference-time prior steered by differentiable crystallographic and cryo-EM likelihoods.

Key results

  1. ShownOn 27 high-resolution X-ray datasets, ROCKET brought AF2 predictions closer to PDB-REDO re-refined structures. Its density fit was comparable to human-deposited models, and its R-free was on par with PredictAndBuild plus phenix.refine.

    Resolutions better than 3 Å; Cα RMSD to PDB-REDO improved for all AF2 predictions except one; for the ten cases where AF2 deviated by more than 1 Å, mean RMSD drop 0.47 Å; backbone and sidechain RSCC rose in every case; R-free improved over phenix.refine alone for all but two models, with 10 showing reductions above 3%. · 2.2 Evaluation Dataset and Approach; 2.3 ROCKET Reliably Refines High-Resolution…; Figure legends

  2. SuggestedOptimizing in MSA-profile space let ROCKET make barrier-crossing changes (ligand-induced loop moves, peptide flips, a disulphide switch, domain shifts). PredictAndBuild largely failed to reach the loop and bond rearrangements, although both methods handled domain rearrangements.

    Case studies: activation and P-loops of human c-Abl (PDB 3PYY), L16-loop peptide flip in PTP-1B (1NWL), GroEL cryo-EM subunit domain (9C0B), starch-binding protein domains (7RAW), nucleoside phosphorylase flip (7YEJ), thaumatin disulphide (7AOJ, built the 0.61-occupancy state); the comparison to PredictAndBuild is qualitative (Fig. S7). · 2.4 ROCKET Overcomes Structural Barriers and…; Figure legends

  3. ShownFor latent-state PAI-1, gradient descent from the full MSA got stuck. Scoring AFCluster-style subsampled MSAs by experimental likelihood gave a starting model from which ROCKET reached the latent conformation.

    1.8 Å dataset; RMSD to deposited model (overall / reactive center loop, residues 330–354): AF2 10.77 / 40.4 Å; ROCKET from full MSA 3.2 / 36.8 Å; best subsampled start 4.2 / 13.6 Å; final model 1.9 Å for both. Likelihood did not rise monotonically with pLDDT across subsamples. · 2.5 ROCKET Integrates MSA Subsampling to Escape…; Figure legends

  4. ShownROCKET extracted usable structural information from a 3.82 Å crystallographic dataset and a 9.60 Å cryo-ET subtomogram average, past usual automated-building limits. In places it fit the map better than the deposited models.

    HAI-1 (PDB 5H7V): fit improved in well-resolved regions, AF2 kept where density was poor, and residues 310–330 (apparent register error in deposited model) improved without new geometric outliers. GroEL (8P4P): both ring conformers recovered, map correlation comparable to manual models; chain A residues 250–280 average RSCC 0.5 vs 0.2 for deposited model. · 2.6 ROCKET Extracts Essential Information from…; Figure legends

  5. SuggestedExperimental data add information beyond AF2's own confidence: maximizing pLDDT in MSA-profile space did not bring predictions closer to experiment. Among the embeddings tried, MSA cluster profiles gave the most diverse search.

    pLDDT maximization did not improve Cα RMSD to PDB-REDO models, unlike likelihood maximization (Fig. S6; e.g., g5rp, PDB 7DNT); mean pairwise RMSD ξ over the first 50 phase-1 structures: 1.22 for MSA cluster profile vs 0.48 (pair) and 0.52 (template torsion). · 2.1 Method Overview; 2.3 ROCKET Reliably Refines High-Resolution…; Figure legends

Methods, data and reproducibility

ROCKET extends OpenFold (AF2 model 1 monomer weights) by learning multiplicative and additive biases on the MSA cluster profile, backpropagating differentiable crystallographic (LLGI) or cryo-EM log-likelihood-gain targets after molecular replacement or docking; two-phase descent, optional MSA subsampling and final phenix.refine follow.

  • High-resolution single-chain X-ray datasets (better than 3 Å), solved by SAD and released after AF2 weight training; PDB-REDO models used as references · n = 27 datasets
  • Activator-bound c-Abl kinase crystallographic dataset, two chains in the asymmetric unit · human · PDB 3PYY
  • Ligand-bound protein tyrosine phosphatase 1B (PTP-1B) crystallographic dataset · human · PDB 1NWL
  • Latent-state serpin PAI-1 crystallographic dataset at 1.8 Å · human · PDB 1JL5 (written 1LJ5 in Fig. 4 legend)
  • Single-particle cryo-EM GroEL structure whose subunits differ from the AF2 prediction · E. coli · PDB 9C0B
  • Multidomain protease inhibitor HAI-1 crystallographic dataset at 3.82 Å · human · PDB 5H7V
  • Cryo-ET sub-tomogram average of GroEL at 9.60 Å · E. coli · PDB 8P4P
  • Further crystallographic case studies: starch-binding protein, nucleoside phosphorylase, thaumatin, T4 lysozyme, viral decapping enzymes · E. coli; T. daniellii; bacteriophage T4; viruses (incl. vaccinia) · PDB 7RAW, 7YEJ, 7AOJ, 6LZM, 7SEZ, 7DNT
Methods & tools
OpenFold, AlphaFold2 (model 1 monomer weights), Phasertng (molecular replacement), EM placement (likelihood-based cryo-EM docking), SFCalculator (differentiable structure factors), phenix.refine, PredictAndBuild (comparison method), PDB-REDO, AFCluster-style MSA subsampling with DBSCAN, MolProbity, PyMOL, Moorhen
Reagents & environment
PyTorch 1.12.1; extension of the OpenFold monomer model with AF2 model 1 weights; phenix.refine for final local refinement (version not stated).
Assay conditions & compute
Refinement capped at about 500 residues per protein or domain on an Nvidia 40 GB A100 GPU; memory and runtime in Fig. S10 (values not in text). Defaults: three phase-1 traces of 100 iterations, up to 500 phase-2 iterations.
Reproducibility gaps
Code is not yet public: authors say the ROCKET code and likelihood targets are being prepared for release; no repository or dataset URLs are provided, and Table 1 PDB IDs are not reproduced in the text.
Limitations (6)
  • Backpropagating through OpenFold uses a lot of memory, capping refinement at about 500 residues per protein or domain on a 40 GB A100, and only one chain is refined at a time. Because of the size cap, only the GroEL subunit domain needing most rebuilding was refined, using a 9-glycine loop. · stated by the authors
  • OpenFold has no knowledge of crystal contacts, so ROCKET can miss lattice-dependent conformations (e.g., PDB 7SML) or mishandle a second chain in the asymmetric unit (7DT2). · stated by the authors
  • ROCKET can fail to flip short loops of 3–4 residues that carry bulky sidechains; for the vaccinia D9 decapping enzyme (7SEZ), PredictAndBuild recovered the backbone where ROCKET did not. · stated by the authors
  • Atomic B-factors are estimated heuristically from pLDDT rather than refined. Raw iterative outputs have more bond outliers and clashes than strict protocols allow, so a final phenix.refine pass is required. · stated by the authors
  • Accuracy is judged mainly by fit to experimental maps, and both crystallography and cryo-EM can introduce artifacts, so map agreement alone does not establish physiological states. · stated by the authors
  • The claims for hard cases (barrier crossing, very low resolution, beating human models) rest on single-dataset case studies such as one PAI-1, one HAI-1 and one GroEL tomogram, with no systematic benchmark in that regime. · noted in this summary

Who should read:Read this if you build atomic models from crystallographic, cryo-EM or cryo-ET data and want to use AlphaFold2 as a differentiable prior to automate large rearrangements or low-resolution fitting.

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISHeo '19Edich '22Zhu '25Baek '23Terwilliger '22Qiao '26
73% semantic matchBiophysics
L. Heo, M. Feig·bioRxiv·Aug 10, 2019