Atlas of the Present Atlas · v0.1

As of 12 September 2026

Dossier 06

Proteins

v1.5-draft · unchecked

1. As of

Date
Version
v1.5-draft
Author / model
Atlas generator
Reviewer
unchecked (Josef)

2. In one sentence

Structure prediction of ordered monomers is within atomic range of experiment on CASP14; ESMFold (Lin, bioRxiv E2) does single-sequence without MSA; RoseTTAFold (Baek Science 2021 AAM E2) is 3-track MSA prediction, behind AF2 on CASP14; Nobel Chemistry 2024 official: Baker design / Hassabis+Jumper prediction (E1); ProteinMPNN inverse folding 52.4% vs 32.9% Rosetta — recovery ≠ binder; de novo design remains a different pipeline; AlphaMissense (Cheng Science 2023, E1) catalogues 71 million missense predictions (not a clinical diagnostic; ACMG-like only terminologically); AF does not replace experiment.

Established now · E1 / E2

3. What works today

Prediction of ordered monomers on CASP14

  1. E1

    AlphaFold2 on CASP14: median backbone 0.96 Å r.m.s.d.95 (n=87 domains) vs. next-best method 2.8 Å. Accuracy falls at median MSA depth < ~30 sequences. Weaker on heterotypic contacts.

    Jumper et al. Highly accurate protein structure prediction with AlphaFold. https://www.nature.com/articles/s41586-021-03819-2. As of 2021-07-15. Checked 2026-08-28. Type: Paper.
  2. E1

    CASP14 organisers: models “rivalling the corresponding experimental ones”; “solution to the classical protein folding problem, at least for single proteins.” Competitive with experimental accuracy for at least 2/3 of the targets. Multimers = “the next barrier”. The phrase is an organisers’ 2021 judgement, not universal physics; MSA dependence and no folding pathway remain objections in the same text.

    Kryshtafovych et al. Critical Assessment of Methods of Protein Structure Prediction (CASP) – Round XIV. https://doi.org/10.1002/prot.26237. As of 2021-12. Checked 2026-08-28. Type: Community overview.
  3. E2

    AlphaFold 1 / CASP13 is a different method (distogram + potential). TM ≥ 0.7 for 24 of 43 FM domains vs. next-best 14 of 43. CASP13 ≠ CASP14.

    Senior et al. Improved protein structure prediction using potentials from deep learning. https://www.nature.com/articles/s41586-019-1923-7. As of 2020-01-15. Checked 2026-08-28. Type: Paper (abstract/extended).

Deployed prediction: coverage ≠ high confidence ≠ experiment

  1. E1

    98.5% of human proteins (20,296 / 20,614) have an AF2 model. 35.7% of all residues pLDDT > 90; 58.0% pLDDT > 70. Long regions with pLDDT < 50 “should not be interpreted as structures”.

    Tunyasuvunakool et al. Highly accurate protein structure prediction for the human proteome. https://www.nature.com/articles/s41586-021-03828-1. As of 2021-07-22. Checked 2026-08-28. Type: Paper.
  2. E2

    AFDB live 2026-08-28: “over 200 million” entries. Complex release May 2026: ~2.2 million homodimeric + ~79,000 heterodimeric high-confidence. Eligible datasets: AF2 or AF-Multimer, not AF3. Do not use 241 million (PDBe v6) — release notes not opened.

    AFDB. AlphaFold Protein Structure Database. https://alphafold.ebi.ac.uk/. As of 2026-08-28. Checked 2026-08-28. Type: Official DB page.
  3. E1

    Terwilliger: AF predictions are hypotheses. Map–model correlation AF 0.56 vs. deposited 0.86. pLDDT > 90: median Cα error 0.6 Å, but ~10% still >2 Å. Sample bias: 86% of residues in this analysis pLDDT >90 vs. 36% of the human proteome.

    Terwilliger et al. AlphaFold predictions are valuable hypotheses and accelerate but do not replace experimental structure determination. https://www.nature.com/articles/s41592-023-02087-4. As of 2023-11-30. Checked 2026-08-28. Type: Paper.

AF3 and complexes: still prediction

  1. E2

    AF3: joint structure of proteins, NA, ligands, ions. PoseBusters: beats Vina without holo pocket (authors’ split). Chirality violation 4.4%. Antibody–antigen: up to 1,000 seeds. Paper limits: clashes, diffusion hallucinations in disordered regions, no solution ensembles (Cereblon closed apo and holo). Code at paper time: not provided.

    Abramson et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. https://www.nature.com/articles/s41586-024-07487-w. As of 2024-05-08. Checked 2026-08-28. Type: Paper.
  2. E2

    AF-Multimer (preprint): heteromeric DockQ ≥ 0.23 in 70%; DockQ ≥ 0.8 in 26%. Authors: “generally not able to predict antibody binding”.

    Evans et al. Protein complex prediction with AlphaFold-Multimer. https://www.biorxiv.org/content/10.1101/2021.10.04.463034v2.full. As of 2022-03-10. Checked 2026-08-28. Type: Preprint.

De novo design is a different task

  1. E1

    RFdiffusion binder wet lab: 95 designs × 5 targets; success = BLI binding ≥50% of the positive-control max at 10 μM. Overall experimental success rate 19%. Hits on all five targets. HA_20 K_D 28 nM; cryo-EM r.m.s.d. to design 0.63 Å. In-silico success does not replace wet lab. Enzymes: no wet lab in this paper.

    Watson et al. De novo design of protein structure and function with RFdiffusion. https://www.nature.com/articles/s41586-023-06415-8.pdf. As of 2023-07-11. Checked 2026-08-28. Type: Paper.
  2. E2

    BindCraft: 12 targets; wet-lab success 10–100%, mean 46.3%. i_pTM does not correlate with affinity; AF2 insensitive to point mutations. Small n, one lab.

    Pacesa et al. One-shot design of functional protein binders with BindCraft. https://www.nature.com/articles/s41586-025-09429-6. As of 2025-08-27. Checked 2026-08-28. Type: Paper.
  3. E1

    AF2/AF3 take a sequence and predict a structure. RFdiffusion/BindCraft generate sequences. AF2 in design pipelines is a filter or loss, not a designer.

    Jumper et al. Highly accurate protein structure prediction with AlphaFold. https://www.nature.com/articles/s41586-021-03819-2. As of 2021-07-15. Checked 2026-08-28. Type: Paper.Watson et al. De novo design of protein structure and function with RFdiffusion. https://www.nature.com/articles/s41586-023-06415-8.pdf. As of 2023-07-11. Checked 2026-08-28. Type: Paper.Pacesa et al. One-shot design of functional protein binders with BindCraft. https://www.nature.com/articles/s41586-025-09429-6. As of 2025-08-27. Checked 2026-08-28. Type: Paper.
  4. E2

    ProtDiff 2022: motif scaffolding in silico, 11.8% designable (scTM>0.5); no wet lab. Not a warrant for RFdiffusion wet lab.

    Trippe et al. Diffusion probabilistic modeling of protein backbones in 3D for the motif-scaffolding problem. https://arxiv.org/html/2206.04119. As of 2022. Checked 2026-08-28. Type: Preprint (ProtDiff, no wet lab).

Inverse folding: backbone → sequence (ProteinMPNN)

  1. E2

    ProteinMPNN (Dauparas Science 2022) is inverse folding: backbone → sequence, not structure prediction. On 402 native monomer backbones, sequence recovery 52.4% vs Rosetta PackRotamersMover 32.9% (1.2 s vs 4.3 min per 100 residues). Median recovery 52 / 55 / 51% on 690 monomers / 732 homomers / 98 heteromers. One lab, one paper.

    Dauparas et al. Robust deep learning–based protein sequence design using ProteinMPNN. https://www.ebi.ac.uk/europepmc/webservices/rest/PMC9997061/fullTextXML. As of 2022-09-15. Checked 2026-08-30. Type: Paper (Science; HHMI author MS PMC9997061).
  2. E2

    Wet lab, same backbones, new sequences: AF hallucinations 73/96 soluble (median 247 mg/L vs 9 mg/L original), 50/96 target oligomeric state (SEC). C5/C6: MPNN 88% soluble and 27.7% correct state vs Rosetta 40% soluble and 0% (SEC-MALS; n unstated). Tetrahedral nanoparticles: 13/76 assemblies ~1 MDa; one crystal 1.2 Å Cα RMSD (T33-27 rescue). Monomer crystal 8CYK, 2.35 Å / 130 residues. Grb2 SH3: BLI binding after MPNN, not after Rosetta — no K_D, no percentage rate in the paper.

    Dauparas et al. Robust deep learning–based protein sequence design using ProteinMPNN. https://www.ebi.ac.uk/europepmc/webservices/rest/PMC9997061/fullTextXML. As of 2022-09-15. Checked 2026-08-30. Type: Paper (Science; HHMI author MS PMC9997061).
  3. E1

    AF2/AF3: sequence → structure. ProteinMPNN: structure → sequence (inverse folding). RFdiffusion/BindCraft: specification → new backbone+sequence. Three tasks; 52.4% recovery is not a binder success rate.

    Jumper et al. Highly accurate protein structure prediction with AlphaFold. https://www.nature.com/articles/s41586-021-03819-2. As of 2021-07-15. Checked 2026-08-28. Type: Paper.Watson et al. De novo design of protein structure and function with RFdiffusion. https://www.nature.com/articles/s41586-023-06415-8.pdf. As of 2023-07-11. Checked 2026-08-28. Type: Paper.Dauparas et al. Robust deep learning–based protein sequence design using ProteinMPNN. https://www.ebi.ac.uk/europepmc/webservices/rest/PMC9997061/fullTextXML. As of 2022-09-15. Checked 2026-08-30. Type: Paper (Science; HHMI author MS PMC9997061).

Single-sequence / ESMFold

  1. E2

    ESMFold (Lin Science 2023 / bioRxiv) is single-sequence structure prediction from ESM-2 (up to 15B parameters): no MSA, no templates. Not ProteinMPNN (inverse folding), not RFdiffusion. Science journal body not opened — warrant is bioRxiv v2 (E2).

    Lin et al. Evolutionary-scale prediction of atomic-level protein structure with a language model. https://www.biorxiv.org/content/10.1101/2022.07.20.500902v2.full.pdf. As of 2022-10-31. Checked 2026-08-31. Type: Paper (bioRxiv preprint; Science 2023 journal body not opened).
  2. E2

    V100, 384 residues: 14.2 s, 6× vs a single AF2 model; shorter sequences up to ∼60× (Fig. S4); practical pipeline up to one–two orders of magnitude without MSA search. Do not collapse 6× (384 aa, forward pass) with ∼60× (shorter sequences).

    Lin et al. Evolutionary-scale prediction of atomic-level protein structure with a language model. https://www.biorxiv.org/content/10.1101/2022.07.20.500902v2.full.pdf. As of 2022-10-31. Checked 2026-08-31. Type: Paper (bioRxiv preprint; Science 2023 journal body not opened).
  3. E2

    TM CAMEO 0.83 (n=194) vs AF2 0.88 / RF 0.82; CASP14 0.68 vs AF2 0.85 (n=51). CAMEO easy/medium/hard: ESMFold 0.90 / 0.79 / 0.45 vs AF2 0.93 / 0.86 / 0.62 (Table S5). Do not average the splits; Fig. 1E LM-only TM 0.71/0.54 is not ESMFold.

    Lin et al. Evolutionary-scale prediction of atomic-level protein structure with a language model. https://www.biorxiv.org/content/10.1101/2022.07.20.500902v2.full.pdf. As of 2022-10-31. Checked 2026-08-31. Type: Paper (bioRxiv preprint; Science 2023 journal body not opened).
  4. E2

    ESM Metagenomic Atlas: >617 million MGnify90 predictions; ∼225 million high confidence (mean pLDDT >0.7 and pTM >0.7); ∼113 million very high (pLDDT >0.9). https://esmatlas.com. Do not collapse 225 million with 113 million.

    Lin et al. Evolutionary-scale prediction of atomic-level protein structure with a language model. https://www.biorxiv.org/content/10.1101/2022.07.20.500902v2.full.pdf. As of 2022-10-31. Checked 2026-08-31. Type: Paper (bioRxiv preprint; Science 2023 journal body not opened).

RoseTTAFold: 3-track with MSA, not single-sequence

  1. E2

    RoseTTAFold (Baek Science 2021, author MS PMC7612213) is 3-track structure prediction: 1D sequence/MSA, 2D distance map, 3D coordinates, with templates. Not ESMFold (single-sequence, no MSA) and not ProteinMPNN (inverse folding). pyRosetta path (all-atom) and end-to-end SE(3) backbone. Science HTML not opened — warrant is the EuropePMC AAM (E2).

    Baek et al. Accurate prediction of protein structures and interactions using a 3-track neural network. https://www.ebi.ac.uk/europepmc/webservices/rest/PMC7612213/fullTextXML. As of 2021-08-20. Checked 2026-09-01. Type: Paper (Science; EuropePMC author MS PMC7612213).
  2. E2

    CASP14 (authors): 3-track beats Zhang-server, BAKER-ROSETTASERVER, BAKER human (rank 2) and their own 2-track model; remains behind AlphaFold2 (Fig. 1B). No mean TM-score in the running text — do not read the bars. CAMEO: since 15 May 2021; 69 medium/hard targets 15 May–19 June 2021, ahead of Robetta, IntFold6-TS, BestSingleTemplate, SWISS-MODEL.

    Baek et al. Accurate prediction of protein structures and interactions using a 3-track neural network. https://www.ebi.ac.uk/europepmc/webservices/rest/PMC7612213/fullTextXML. As of 2021-08-20. Checked 2026-09-01. Type: Paper (Science; EuropePMC author MS PMC7612213).
  3. E2

    After ~1.5 h sequence/template search: end-to-end backbone ~10 min on RTX2080 for proteins <400 residues; pyRosetta 5 min network plus one hour all-atom (15 CPUs). Proteins >400 residues: 8 GB vs 24 GB GPU. Four previously unsolved MR datasets solved with RoseTTAFold; trRosetta models yielded no MR solutions. SLP N-terminus: 95 Cα within 3 Å, Cα RMSD 0.98 Å vs template 4l3a 54 Cα / 1.69 Å. Cryo-EM p101 GBD: Cα RMSD 3.0 Å over the β-sheets.

    Baek et al. Accurate prediction of protein structures and interactions using a 3-track neural network. https://www.ebi.ac.uk/europepmc/webservices/rest/PMC7612213/fullTextXML. As of 2021-08-20. Checked 2026-09-01. Type: Paper (Science; EuropePMC author MS PMC7612213).
  4. E2

    Complexes directly from paired sequences (~30 min, 24G TITAN RTX); in many cases TM-score >0.8 — not a success rate. 693 human disease-related domains; over one-third predicted lDDT >0.8, corresponding to average 2.6 Å Cα RMSD on CASP14 (fig. S8). Code: github.com/RosettaCommons/RoseTTAFold; Robetta server.

    Baek et al. Accurate prediction of protein structures and interactions using a 3-track neural network. https://www.ebi.ac.uk/europepmc/webservices/rest/PMC7612213/fullTextXML. As of 2021-08-20. Checked 2026-09-01. Type: Paper (Science; EuropePMC author MS PMC7612213).

Nobel Prize in Chemistry 2024 — official citation

  1. E1

    Nobel Prize in Chemistry 2024 (press release 9 Oct 2024, nobelprize.org opened): one half to David Baker (University of Washington / HHMI) “for computational protein design”; other half jointly to Demis Hassabis and John Jumper (Google DeepMind) “for protein structure prediction”. Press: AlphaFold2 (2020) predicts structures of virtually all ~200 million identified proteins; Baker designing new proteins since 2003. E1 for the official citation — not for DeepMind marketing figures.

    Nobel Chemistry 2024. Press release: The Nobel Prize in Chemistry 2024. https://www.nobelprize.org/prizes/chemistry/2024/press-release/. As of 2024-10-09. Checked 2026-09-02. Type: Official Nobel press release (HTML opened).

Missense variants: AlphaMissense (catalogue, not diagnosis)

  1. E1

    AlphaMissense (Cheng et al., Science 2023, bronze OA body opened): catalogue of all possible human single amino-acid substitutions — 216 million possible single-AA changes across 19,233 canonical proteins → 71 million missense predictions; classifies 89% as likely benign or likely pathogenic (model bins).

    Cheng et al. AlphaMissense. Accurate proteome-wide missense variant effect prediction with AlphaMissense. https://doi.org/10.1126/science.adg7492. As of 2023-09-19. Checked 2026-09-05. Type: Paper (Science; bronze OA body opened; OCR research/pdfs/Cheng_AlphaMissense_Science2023.txt).
  2. E1

    Of 71 million: 32% (22.8 million) likely pathogenic, 57% (40.9 million) likely benign at score cutoffs with 90% precision on ClinVar (fig. S4A).

    Cheng et al. AlphaMissense. Accurate proteome-wide missense variant effect prediction with AlphaMissense. https://doi.org/10.1126/science.adg7492. As of 2023-09-19. Checked 2026-09-05. Type: Paper (Science; bronze OA body opened; OCR research/pdfs/Cheng_AlphaMissense_Science2023.txt).
  3. E1

    Training: population-frequency weak labels (human/primate); avoids training directly on ClinVar clinical labels. ClinVar used for held-out eval + calibration (2526 early-stop; 90% precision bins). Does not predict mutant structures — pathogenicity as a scalar.

    Cheng et al. AlphaMissense. Accurate proteome-wide missense variant effect prediction with AlphaMissense. https://doi.org/10.1126/science.adg7492. As of 2023-09-19. Checked 2026-09-05. Type: Paper (Science; bronze OA body opened; OCR research/pdfs/Cheng_AlphaMissense_Science2023.txt).
  4. E1

    ClinVar test auROC 0.940 on 18,924 variants (vs EVE 0.911 among methods not trained on ClinVar).

    Cheng et al. AlphaMissense. Accurate proteome-wide missense variant effect prediction with AlphaMissense. https://doi.org/10.1126/science.adg7492. As of 2023-09-19. Checked 2026-09-05. Type: Paper (Science; bronze OA body opened; OCR research/pdfs/Cheng_AlphaMissense_Science2023.txt).

Claimed · E3

4. What is claimed, not shown

Vendor design and impact marketing

  1. E2

    AlphaProteo yeast display: BHRF1 88% (n=94); TNFα 0% (54). Methods not released. SC2RBD/PD-L1/TrkA used in methods development (possible overestimate). Assay difference: RFdiffusion retest in yeast lower than Watson BLI. E3 for generalisability.

    Zambaldi et al. AlphaProteo generates novel proteins for biology and health research. https://arxiv.org/html/2409.08022. As of 2024-09-05. Checked 2026-08-28. Type: Tech report (not journal).
  2. E3

    DeepMind page: “3 million researchers” / “190 countries”; “hundreds of millions of research years saved”. Vendor marketing stays E3 (sources 12/13). Official Nobel citation is E1 as of 2026-09-02 via source 18 — do not upgrade through the vendor page.

    DeepMind AlphaFold. AlphaFold (science/product page). https://deepmind.google/science/alphafold/. As of 2026-08-28. Checked 2026-08-28. Type: Vendor page.DeepMind AlphaProteo Blog. AlphaProteo generates novel proteins for biology and health research (blog). https://deepmind.google/blog/alphaproteo-generates-novel-proteins-for-biology-and-health-research/. As of 2024-09-05. Checked 2026-08-28. Type: Vendor blog.

Constrained · Limit

5. Bottleneck and limit

Prediction, design, and narrative come apart

  1. E1

    AlphaMissense is not a clinical ACMG/AMP classification and not a diagnostic substitute. “Likely pathogenic / likely benign” are model score bins calibrated to 90% ClinVar precision; terminology only “similar” to ACMG (paper).

    Cheng et al. AlphaMissense. Accurate proteome-wide missense variant effect prediction with AlphaMissense. https://doi.org/10.1126/science.adg7492. As of 2023-09-19. Checked 2026-09-05. Type: Paper (Science; bronze OA body opened; OCR research/pdfs/Cheng_AlphaMissense_Science2023.txt).
  2. E1

    Prediction ≠ design. “Success rate” is not a universal scalar: BindCraft 46.3% ≠ RFdiffusion 19% ≠ AlphaProteo 0–88%. Assay, target, n, and filter threshold belong to the figure.

    Jumper et al. Highly accurate protein structure prediction with AlphaFold. https://www.nature.com/articles/s41586-021-03819-2. As of 2021-07-15. Checked 2026-08-28. Type: Paper.Watson et al. De novo design of protein structure and function with RFdiffusion. https://www.nature.com/articles/s41586-023-06415-8.pdf. As of 2023-07-11. Checked 2026-08-28. Type: Paper.Pacesa et al. One-shot design of functional protein binders with BindCraft. https://www.nature.com/articles/s41586-025-09429-6. As of 2025-08-27. Checked 2026-08-28. Type: Paper.
  3. E1

    AF does not replace experiment. 36% of human residues very-high confidence — the rest is uncertain or disordered.

    Terwilliger et al. AlphaFold predictions are valuable hypotheses and accelerate but do not replace experimental structure determination. https://www.nature.com/articles/s41592-023-02087-4. As of 2023-11-30. Checked 2026-08-28. Type: Paper.Tunyasuvunakool et al. Highly accurate protein structure prediction for the human proteome. https://www.nature.com/articles/s41586-021-03828-1. As of 2021-07-22. Checked 2026-08-28. Type: Paper.
  4. E1

    Disorder: AF2 ribbons pLDDT<50 not to be read as structure. AF3 hallucinates compact structures in disordered regions — a different error mode.

    Abramson et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. https://www.nature.com/articles/s41586-024-07487-w. As of 2024-05-08. Checked 2026-08-28. Type: Paper.Tunyasuvunakool et al. Highly accurate protein structure prediction for the human proteome. https://www.nature.com/articles/s41586-021-03828-1. As of 2021-07-22. Checked 2026-08-28. Type: Paper.
  5. E2

    Design fails completely on some targets (AlphaProteo TNFα 0/54). More polar sites in RFdiffusion untested.

    Zambaldi et al. AlphaProteo generates novel proteins for biology and health research. https://arxiv.org/html/2409.08022. As of 2024-09-05. Checked 2026-08-28. Type: Tech report (not journal).Watson et al. De novo design of protein structure and function with RFdiffusion. https://www.nature.com/articles/s41586-023-06415-8.pdf. As of 2023-07-11. Checked 2026-08-28. Type: Paper.
  6. E2

    Native sequence recovery ≠ wet-lab success. Authors: recovery is “not necessarily optimal” for design; may not correlate with folding (a single substitution can block folding). 52.4% is not a deployed binder/drug rate.

    Dauparas et al. Robust deep learning–based protein sequence design using ProteinMPNN. https://www.ebi.ac.uk/europepmc/webservices/rest/PMC9997061/fullTextXML. As of 2022-09-15. Checked 2026-08-30. Type: Paper (Science; HHMI author MS PMC9997061).
  7. E1

    Inverse folding ≠ de novo structure design. RFdiffusion binder 19% BLI (Watson, 95 designs × 5 targets, ≥50% of positive-control max at 10 μM) remains the dossier figure for backbone generation. ProteinMPNN Grb2 is a BLI example without a rate — not 19%, not 52.4%.

    Watson et al. De novo design of protein structure and function with RFdiffusion. https://www.nature.com/articles/s41586-023-06415-8.pdf. As of 2023-07-11. Checked 2026-08-28. Type: Paper.Dauparas et al. Robust deep learning–based protein sequence design using ProteinMPNN. https://www.ebi.ac.uk/europepmc/webservices/rest/PMC9997061/fullTextXML. As of 2022-09-15. Checked 2026-08-30. Type: Paper (Science; HHMI author MS PMC9997061).

Single-sequence / ESMFold: limits

  1. E2

    AF2 with MSA+templates still ahead (CASP14 0.85 vs 0.68; CAMEO hard 0.62 vs 0.45). ESMFold closes the single-sequence gap, does not dominate every split.

    Lin et al. Evolutionary-scale prediction of atomic-level protein structure with a language model. https://www.biorxiv.org/content/10.1101/2022.07.20.500902v2.full.pdf. As of 2022-10-31. Checked 2026-08-31. Type: Paper (bioRxiv preprint; Science 2023 journal body not opened).
  2. E2

    Not trained on complexes; 53.2% same DockQ category vs AF-Multimer on 2,978 complexes.

    Lin et al. Evolutionary-scale prediction of atomic-level protein structure with a language model. https://www.biorxiv.org/content/10.1101/2022.07.20.500902v2.full.pdf. As of 2022-10-31. Checked 2026-08-31. Type: Paper (bioRxiv preprint; Science 2023 journal body not opened).
  3. E2

    RoseTTAFold needs MSA/templates; ESMFold does not. Lin CAMEO “RF 0.82” is Lin’s comparison (bioRxiv), not Baek’s CASP/CAMEO table. TM>0.8 on Baek complexes is “in many cases”, not a rate and not binder wet lab.

    Baek et al. Accurate prediction of protein structures and interactions using a 3-track neural network. https://www.ebi.ac.uk/europepmc/webservices/rest/PMC7612213/fullTextXML. As of 2021-08-20. Checked 2026-09-01. Type: Paper (Science; EuropePMC author MS PMC7612213).Lin et al. Evolutionary-scale prediction of atomic-level protein structure with a language model. https://www.biorxiv.org/content/10.1101/2022.07.20.500902v2.full.pdf. As of 2022-10-31. Checked 2026-08-31. Type: Paper (bioRxiv preprint; Science 2023 journal body not opened).

6. Actors and incentives

Who predicts, who designs, who assesses

  1. E1

    DeepMind supplies AF2; CASP organisers assess independently (“at least for single proteins”). Nobel Chemistry 2024 (source 18): Baker design; Hassabis/Jumper prediction.

    Jumper et al. Highly accurate protein structure prediction with AlphaFold. https://www.nature.com/articles/s41586-021-03819-2. As of 2021-07-15. Checked 2026-08-28. Type: Paper.Kryshtafovych et al. Critical Assessment of Methods of Protein Structure Prediction (CASP) – Round XIV. https://doi.org/10.1002/prot.26237. As of 2021-12. Checked 2026-08-28. Type: Community overview.
  2. E1

    Baker lab/RFdiffusion and Pacesa/BindCraft measure wet-lab binders; code open. Baker lab also ProteinMPNN (Dauparas, source 15) and RoseTTAFold (Baek, source 17). AlphaProteo methods withheld.

    Watson et al. De novo design of protein structure and function with RFdiffusion. https://www.nature.com/articles/s41586-023-06415-8.pdf. As of 2023-07-11. Checked 2026-08-28. Type: Paper.Pacesa et al. One-shot design of functional protein binders with BindCraft. https://www.nature.com/articles/s41586-025-09429-6. As of 2025-08-27. Checked 2026-08-28. Type: Paper.Dauparas et al. Robust deep learning–based protein sequence design using ProteinMPNN. https://www.ebi.ac.uk/europepmc/webservices/rest/PMC9997061/fullTextXML. As of 2022-09-15. Checked 2026-08-30. Type: Paper (Science; HHMI author MS PMC9997061).
  3. E1

    Terwilliger et al. (PHENIX/crystallographic practice): hypotheses, not a replacement.

    Terwilliger et al. AlphaFold predictions are valuable hypotheses and accelerate but do not replace experimental structure determination. https://www.nature.com/articles/s41592-023-02087-4. As of 2023-11-30. Checked 2026-08-28. Type: Paper.

7. State of the dispute

“Folding solved” vs. molecular design

  1. E1

    The CASP “solution” holds for single, ordered proteins in the 2020 blind test, not for the dossier theme of molecular design. Next barrier = complexes.

    Kryshtafovych et al. Critical Assessment of Methods of Protein Structure Prediction (CASP) – Round XIV. https://doi.org/10.1002/prot.26237. As of 2021-12. Checked 2026-08-28. Type: Community overview.
  2. E2

    AlphaProteo vs. RFdiffusion: different assays, different rates. Head-to-head yeast retest is not Watson BLI.

    Zambaldi et al. AlphaProteo generates novel proteins for biology and health research. https://arxiv.org/html/2409.08022. As of 2024-09-05. Checked 2026-08-28. Type: Tech report (not journal).Watson et al. De novo design of protein structure and function with RFdiffusion. https://www.nature.com/articles/s41586-023-06415-8.pdf. As of 2023-07-11. Checked 2026-08-28. Type: Paper.
  3. E0

    Lin et al. ESMFold: bioRxiv body opened (E2, source 16); Science journal HTML/PDF still unopened (2026-09-01 Cloudflare). RoseTTAFold (Baek Science 2021 AAM, source 17) opened. AlphaMissense (Cheng Science 2023, source 19) body opened E1 — catalogue/ClinVar eval, not clinical diagnosis.

    DeepMind AlphaFold. AlphaFold (science/product page). https://deepmind.google/science/alphafold/. As of 2026-08-28. Checked 2026-08-28. Type: Vendor page.

8. Open questions

  1. ESMFold Science/PMC journal body (bioRxiv v2 opened E2; journal still closed 2026-09-01). AlphaMissense body opened 2026-09-05 (source 19, E1).
  2. AFDB v6 release notes + Varadi/Bertoni NAR — do not use 241 million.
  3. AF3 addendum 27 Nov 2024 + full GitHub README.
  4. Cao et al. Nature 2022 RIFdock — <0.1% not an atlas figure.
  5. CASP15/16 protein + RNA assessments.
  6. Wayment-Steele Nature 2024 multi-conformation.
  7. PoseBusters original (Buttenschoen).
  8. AlphaFold Server terms 2026.

9. Changes

  • v1.5-draft2026-09-05: AlphaMissense (Cheng et al. Science 2023, source 19, bronze OA) opened — 216 million/19,233 proteins → 71 million missense; 32%/57% likely path./benign at 90% ClinVar precision; population-frequency training, ClinVar held-out; auROC 0.940; not clinical ACMG diagnosis.
  • v1.4-draft2026-09-02: Nobel Prize in Chemistry 2024 press release (source 18, nobelprize.org) opened — Baker design / Hassabis+Jumper prediction (E1). Vendor marketing (3 million researchers) stays E3; AlphaMissense still E0.
  • v1.3-draft2026-09-01: RoseTTAFold (Baek Science 2021 AAM PMC7612213, source 17) opened — 3-track MSA, CASP14 behind AF2, CAMEO 69 medium/hard, ~10 min RTX2080, MR 4/4. ESMFold journal and AlphaMissense still closed.
  • v1.2-draft2026-08-31: ESMFold (Lin bioRxiv v2, source 16) opened — single-sequence, 14.2 s/6× vs ∼60×, TM CAMEO 0.83 / CASP14 0.68, atlas >617 million / ∼225 million high / ∼113 million very high. Science body still closed. RoseTTAFold/AlphaMissense still E0.
  • v1.1-draft2026-08-30: ProteinMPNN (Dauparas Science 2022, source 15, EuropePMC author MS) opened — inverse folding 52.4% vs 32.9% Rosetta; recovery ≠ binder. ESMFold/RoseTTAFold/AlphaMissense still E0.
  • v1.0-draftFirst version from the 2026-08-28 verification log.

10. Sources

No. Source As of Checked Grade
1 Jumper, J.; Evans, R.; Pritzel, A.; Green, T.; Figurnov, M.; Ronneberger, O.; et al.. Highly accurate protein structure prediction with AlphaFold. https://www.nature.com/articles/s41586-021-03819-2. Type: Paper. E1
2 Abramson, J.; Adler, J.; Dunger, J.; Evans, R.; Green, T.; Pritzel, A.; et al.. Accurate structure prediction of biomolecular interactions with AlphaFold 3. https://www.nature.com/articles/s41586-024-07487-w. Type: Paper. E2
3 Senior, A. W.; Evans, R.; Jumper, J.; Kirkpatrick, J.; Sifre, L.; Green, T.; et al.. Improved protein structure prediction using potentials from deep learning. https://www.nature.com/articles/s41586-019-1923-7. Type: Paper (abstract/extended). E2
4 Tunyasuvunakool, K.; Adler, J.; Wu, Z.; Green, T.; Zielinski, M.; Žídek, A.; et al.. Highly accurate protein structure prediction for the human proteome. https://www.nature.com/articles/s41586-021-03828-1. Type: Paper. E1
5 Terwilliger, T. C.; Liebschner, D.; Croll, T. I.; Williams, C. J.; McCoy, A. J.; et al.. AlphaFold predictions are valuable hypotheses and accelerate but do not replace experimental structure determination. https://www.nature.com/articles/s41592-023-02087-4. Type: Paper. E1
6 Pacesa, M.; Nickel, L.; Schellhaas, C.; Correia, B. E.; et al.. One-shot design of functional protein binders with BindCraft. https://www.nature.com/articles/s41586-025-09429-6. Type: Paper. E2
7 Watson, J. L.; Juergens, D.; Bennett, N. R.; Trippe, B. L.; Yim, J.; Eisenach, H. E.; et al.. De novo design of protein structure and function with RFdiffusion. https://www.nature.com/articles/s41586-023-06415-8.pdf. Type: Paper. E1
8 Zambaldi, V.; La, D.; Chu, A. E.; et al. (Google DeepMind). AlphaProteo generates novel proteins for biology and health research. https://arxiv.org/html/2409.08022. Type: Tech report (not journal). E2
9 Evans, R.; O’Neill, M.; Pritzel, A.; Antropova, N.; Senior, A. W.; et al.. Protein complex prediction with AlphaFold-Multimer. https://www.biorxiv.org/content/10.1101/2021.10.04.463034v2.full. Type: Preprint. E2
10 Kryshtafovych, A.; Schwede, T.; Topf, M.; Fidelis, K.; Moult, J.. Critical Assessment of Methods of Protein Structure Prediction (CASP) – Round XIV. https://doi.org/10.1002/prot.26237. Type: Community overview. E1
11 EMBL-EBI / Google DeepMind. AlphaFold Protein Structure Database. https://alphafold.ebi.ac.uk/. Type: Official DB page. E2
12 Google DeepMind. AlphaFold (science/product page). https://deepmind.google/science/alphafold/. Type: Vendor page. E3
13 Google DeepMind. AlphaProteo generates novel proteins for biology and health research (blog). https://deepmind.google/blog/alphaproteo-generates-novel-proteins-for-biology-and-health-research/. Type: Vendor blog. E3
14 Trippe, B. L.; Yim, J.; Tischer, D.; Baker, D.; et al.. Diffusion probabilistic modeling of protein backbones in 3D for the motif-scaffolding problem. https://arxiv.org/html/2206.04119. Type: Preprint (ProtDiff, no wet lab). E2
15 Dauparas, J.; Anishchenko, I.; Bennett, N.; Bai, H.; Ragotte, R. J.; Milles, L. F.; Wicky, B. I. M.; Courbet, A.; de Haas, R. J.; Bethel, N.; Leung, P. J. Y.; Huddy, T. F.; Pellock, S.; Tischer, D.; Chan, F.; Koepnick, B.; Nguyen, H.; Kang, A.; Sankaran, B.; Bera, A. K.; King, N. P.; Baker, D.. Robust deep learning–based protein sequence design using ProteinMPNN. https://www.ebi.ac.uk/europepmc/webservices/rest/PMC9997061/fullTextXML. Type: Paper (Science; HHMI author MS PMC9997061). E2
16 Lin, Z.; Akin, H.; Rao, R.; Hie, B.; Zhu, Z.; Lu, W.; Smetanin, N.; Verkuil, R.; Kabeli, O.; Shmueli, Y.; dos Santos Costa, A.; Fazel-Zarandi, M.; Sercu, T.; Candido, S.; Rives, A.. Evolutionary-scale prediction of atomic-level protein structure with a language model. https://www.biorxiv.org/content/10.1101/2022.07.20.500902v2.full.pdf. Type: Paper (bioRxiv preprint; Science 2023 journal body not opened). E2
17 Baek, M.; DiMaio, F.; Anishchenko, I.; Dauparas, J.; Ovchinnikov, S.; Lee, G. R.; Wang, J.; Cong, Q.; Kinch, L. N.; Schaeffer, R. D.; Millán, C.; Park, H.; Adams, C.; Glassman, C. R.; DeGiovanni, A.; Pereira, J. H.; Rodrigues, A. V.; van Dijk, A. A.; Ebrecht, A. C.; Opperman, D. J.; Sagmeister, T.; Buhlheller, C.; Pavkov-Keller, T.; Rathinaswamy, M. K.; Dalwadi, U.; Yip, C. K.; Burke, J. E.; Garcia, K. C.; Grishin, N. V.; Adams, P. D.; Read, R. J.; Baker, D.. Accurate prediction of protein structures and interactions using a 3-track neural network. https://www.ebi.ac.uk/europepmc/webservices/rest/PMC7612213/fullTextXML. Type: Paper (Science; EuropePMC author MS PMC7612213). E2
18 The Royal Swedish Academy of Sciences / Nobel Prize Outreach. Press release: The Nobel Prize in Chemistry 2024. https://www.nobelprize.org/prizes/chemistry/2024/press-release/. Type: Official Nobel press release (HTML opened). E1
19 Cheng, J.; Novati, G.; Pan, J.; Bycroft, C.; Žemgulytė, A.; Applebaum, T.; Pritzel, A.; Wong, L. H.; Zielinski, M.; Sargeant, T.; Schneider, R. G.; Senior, A. W.; Jumper, J.; Hassabis, D.; Kohli, P.; Avsec, Ž.. Accurate proteome-wide missense variant effect prediction with AlphaMissense. https://doi.org/10.1126/science.adg7492. Type: Paper (Science; bronze OA body opened; OCR research/pdfs/Cheng_AlphaMissense_Science2023.txt). E1

11. Uncertainty log

Overall uncertainty of this entry, bound to the verification log of 2026-08-28 plus ProteinMPNN 2026-08-30, ESMFold bioRxiv 2026-08-31, RoseTTAFold AAM 2026-09-01 and Nobel press 2026-09-02 and AlphaMissense body 2026-09-05. 19 openings (+1 Cheng AlphaMissense E1; Lin Science body still closed), remaining gaps. Not used as warrant: AlphaMissense as clinical ACMG diagnosis; DeepMind marketing as Nobel citation; Fig. 1B TM with no number in the text; Lin CAMEO RF 0.82 as a Baek figure.

  • Established (layer 1): AF2 CASP14; proteome coverage; Terwilliger hypotheses; RFdiffusion 19% BLI; ProteinMPNN 52.4% recovery; ESMFold single-sequence (bioRxiv E2); RoseTTAFold 3-track MSA (Baek AAM E2); Nobel Chemistry 2024 official citation (E1); AlphaMissense catalogue/ClinVar auROC (E1); BindCraft range; conceptual split prediction / inverse folding / design / missense score.
  • Claimed (layer 2): AlphaProteo 0–88%; DeepMind impact marketing (E3; Nobel citation separately E1); AF3 PoseBusters authors’ comparison.
  • Constrained (layer 3): prediction ≠ design; ProteinMPNN recovery ≠ binder; assay incommensurability; ESMFold vs AF2 with MSA; RoseTTAFold needs MSA (≠ ESMFold); AlphaMissense ≠ clinical ACMG diagnosis.

Not opened (not a warrant)

  • Lin et al. ESMFold Science 2023 journal HTML/PDF + PMC author MS — bioRxiv v2 opened E2 2026-08-31; journal body still closed. Do not cite arXiv 2207.06221.
  • Baek et al. RoseTTAFold Science 2021 AAM opened 2026-09-01 (source 17); science.org HTML still 403.
  • Varadi/Bertoni NAR AFDB; PDBe v6 241,070,489 — search only.
  • Cheng AlphaMissense Science 2023 body opened 2026-09-05 (source 19, E1). Cao 2022 RIFdock; Anishchenko hallucination 2021.
  • AF3 addendum 27 Nov 2024 (Nobel Chemistry 2024 press opened 2026-09-02, source 18).
  • Buttenschoen PoseBusters; ColabFold; RFdiffusion2/3; ESM3.
  • AlphaFold Server terms / GitHub alphafold3 README — landing only, E3 availability.