henhao
Sign in
BiophysicseLife
Peer-reviewed journal article

Large scale prospective evaluation of co-folding across hundreds of Mac1-ligand complexes and three virtual screens

Jongbin KimiD, G.J. CorreyiD, Brendan W. HalliD, Moira M. RachmaniD, Olivier MailhotiD, Takaya TogoiD, Ryan L. GonciarziD, Priyadarshini JaishankariD, R. Jeffrey NeitziD, Eric R. HantziD, Yagmur Umay DorukiD, Maisie GV StevensiD, Morgan E. DiolaitiiD, Rashad L. ReidiD, Saumya GopalkrishnaniD, Nevan J. KroganiD, Adam R. RensloiD, Alan AshworthiD, Brian K. ShoichetiD, James S. FraseriD

Abstract

Accurate prediction of ligand-bound protein complexes and ranking them by affinity are central problems in drug discovery. While deep learning co-folding methods can help address these challenges, their evaluation has been hampered by the difficulties in assessing independence from training data and insufficiently large test sets. Here we test the ability of co-folding methods to predict the structures of 551 ligands, 489 of which are of sufficient quality for evaluation, bound to the SARS-CoV-2 NSP3 macrodomain (Mac1) that were determined after the training cut-off dates. AlphaFold3 (AF3), Boltz-2, and Chai-1 each reproduced >50% of the Mac1 ligand poses to better than 2 [A] RMSD of experiment. Despite the potential for co-folding to describe protein conformational changes that stabilize ligand binding, we did not find that common conformational rearrangements, including peptide flip and a large loop opening, were recapitulated by the co-folding prediction. For AF3 and Chai-1, ligand pose prediction confidence weakly, but significantly, tracked experimental potency, while DOCK3.7 energies were only weakly correlated. Boltz-2 affinity predictions showed the strongest correlation with measured potency and, after calibration, achieved lower mean absolute error than a baseline predictor. We next assessed whether co-folding scores could rescore docking hit-lists to distinguish true ligands from non-binders among hundreds of molecules prospectively experimentally tested against AmpC {beta}-lactamase, the dopamine D4 and the {sigma}2 receptors. AF3 ligand pose confidence values did not separate true ligands from high-scoring false-positives as effectively as docking scores or Boltz-2 affinity predictions did. Taken together, the modest, but independent correlations of docking score and co-folding confidence or affinity suggests that integrating physics-based and deep-learning approaches may help with hit prioritization and subsequent optimization in structure-based ligand discovery.

PDFDOICodeData
Full-text digest ↓

02 /Specification & Classification

Click any tag to browse matching papers
Citations
7
FWCI
25.41
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 4 (21 Sept 2026) · CC BY 4.0

On 489 evaluable new Mac1 complexes, AlphaFold3, Chai-1 and Boltz-2 placed over half of ligands within 2 Å, but on docking hit-lists AF3 and Boltz-2 added little over docking.

Co-folding methods have lacked large prospective tests, and earlier studies pointed to structural memorization and implausible high-confidence poses. This work tests three models on 551 Mac1 ligand complexes not previously disclosed and on hit-lists from three prospective docking screens.

Key results

  1. ShownAF3 and Chai-1 placed most ligands within 2 Å RMSD of the crystal pose, with Boltz-2 at 53% and DOCK3.7 at 40%. AF3 pose error showed no meaningful correlation with similarity to pre-cutoff training ligands.

    489 evaluable Mac1 complexes. Ligand centre of mass within 2.5 Å: 449 AF3 (92%), 440 Chai-1 (90%), 444 Boltz-2 (93%), 392 DOCK3.7 (80%). Boltz-2 scored lower despite its later training cutoff. · Co-folding can accurately reproduce poses of…

  2. ShownCo-folding did not recapitulate rare protein conformational changes in the binding site, although ligands were often still posed correctly.

    AF3 correctly predicted 4/19 twisted (Phe156 backbone flip) states and 0/20 open (~8 Å loop) states, and predicted 8 false-positive twisted states. Global Cα RMSD was 0.1 to 0.4 Å. · Co-folding can accurately reproduce poses of…

  3. ShownBoltz-2 predicted affinity tracked measured Mac1 potency best and, after linear calibration, beat a mean-value baseline by a modest margin. AF3 confidence correlated slightly better than DOCK3.7, but not significantly so.

    202 ligands with IC50. Pearson r: Boltz-2 pIC50 0.6, AF3 L-pLDDT 0.314, DOCK3.7 -0.225; Boltz-2 vs DOCK p = 0.0057. Boltz-2 MAE 1.19 pIC (~1.6 kcal) before and 0.54 pIC (~0.7 kcal) after linear fit, vs a mean-value baseline of 0.68 pIC (0.93 kcal). · Model confidence and energy scores reflect pose…; Figure legends

  4. ShownOn hit-lists from three prospective docking screens, AF3 confidence separated actives from docking false positives poorly. DOCK3.7 and Boltz-2 pIC50 did better, and co-folding showed little advantage over docking.

    AUC for actives vs false positives: σ2 DOCK 78.8, AF3 56.1, Boltz-2 83.8; AmpC 76.5, 60.5, 68.1; D4 Boltz-2 and DOCK ~71, AF3 46.4. In σ2 with random false-positive potencies, Boltz-2 MAE was 1.66 kcal vs 1.84 for the mean baseline. · Co-folding struggles to reliably re-rank Docking…

  5. SuggestedInternal confidence scores flagged correct Mac1 poses. AF3 and Chai-1 pose errors were more correlated with each other than with DOCK3.7 or Boltz-2, which the authors read as complementarity.

    AUC for RMSD < 2 Å: AF3 L-pLDDT 75.6, Chai-1 ipTM 74.6, Boltz-2 ligand ipTM 75.2, DOCK3.7 67.1, Boltz-2 pIC50 53.8. RMSD correlation: AF3 vs Chai-1 r = 0.72, vs DOCK 0.45, vs Boltz-2 0.52. · Model confidence and energy scores reflect pose…

Methods, data and reproducibility

Three co-folding models and DOCK3.7 were run from SMILES and sequence on Mac1 ligands with new X-ray structures (RDKit heavy-atom RMSD, ROC and correlation analyses). AF3 and Boltz-2 were then used to rescore experimentally tested hit-lists from three docking campaigns.

  • Mac1 (SARS-CoV-2 NSP3 macrodomain) ligand complexes solved by X-ray crystallography; 489 evaluable; 202 with HTRF IC50 · SARS-CoV-2 (viral protein) · n = 551 ligands (489 evaluable, 202 with IC50)
  • Mac1 structure used as DOCK3.7 receptor · SARS-CoV-2 (viral protein) · n = 1 structure · PDB 5SQW
  • Experimentally tested docking hit-list, σ2 receptor (490-million molecule screen) · n = 506 tested (201 actives, 305 inactive)
  • Experimentally tested docking hit-list, AmpC β-lactamase (1.7-billion molecule screen) · n = 1,293 tested (247 actives)
  • Experimentally tested docking hit-list, dopamine D4 receptor (138-million molecule screen) · n = 541 tested (205 actives, 336 inactive); Methods give n = 549
Methods & tools
AlphaFold3, Chai-1, Boltz-2, DOCK3.7, RDKit, PyMOL, PanDDA, phenix.elbow, HTRF peptide displacement assay, Large-Scale Docking database (lsd.docking.org)
Code
https://github.com/jongbin99/Cofoldinghttps://github.com/google-deepmind/alphafold3https://github.com/chaidiscovery/chai-lab.githttps://github.com/jwohlwend/boltz
Data & accessions
https://lsd.docking.org/
Experimental conditions
Co-folding run on an array of NVIDIA A40 GPUs; DOCK3.7 sampled over 290 million complexes for the 551 Mac1 compounds.
Reproducibility gaps
The 468 new PDB codes, poses and scores are in supplementary files rather than printed in the text; software versions are not stated.
Limitations (5)
  • The Mac1 set falls into four MCS clusters and has many fragment structures of the protein in the training data, which may inflate pose recovery. · stated by the authors
  • The docking hit-lists cover only three targets and were pre-selected by docking, which the authors say arguably favours docking over co-folding. · stated by the authors
  • Only 202 of 551 Mac1 ligands have dose-response IC50 values, skewed toward weak binders, and Mac1 is a single system with a limited set of pharmacophores. · stated by the authors
  • Active/non-binder labels for σ2 and D4 use an adjusted 25% displacement cutoff rather than the original 50%, and Chai-1 was not run on the hit-lists. · noted in this summary
  • Hit-list affinity errors for Boltz-2 depend on assumed potencies for false positives, so the MAE comparisons rest on author-chosen assumptions. · noted in this summary

Who should read:Read this if you are deciding whether co-folding models can replace or complement docking for pose prediction, potency ranking or hit prioritization.

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
THISMenon '26Röhrig '26Škrinjar '26Wu '25Xu '26Akabari '26
79% semantic matchPharmacology & Toxicology
Kartikeya M. Menon, Aakash Davasam, Guo Chen+8 more·bioRxiv·Jun 12, 2026· 10 citations