Highly Accurate Protein Structure Prediction with AlphaFold
Nature, 2021 · paper · AI and machine learning
Presents AlphaFold, a neural network that predicts a protein's three-dimensional structure from its amino acid sequence. In the CASP14 assessment it is reported to reach accuracy competitive with experimental structures in most cases, well ahead of other methods.
Checked: PDF from nature.com, fetched 4 Oct 2026. Page numbers are those of that PDF.
2 findings
Reference or label "extensively covered in a number of reviews14,40–43" includes refs 42 (Qian & Sejnowski, "Predicting the secondary structure of globular proteins using neural network models", 1988) and 43 (Fariselli et al., "Prediction of contact maps with neural networks and correlated mutations", 2001), which are primary method papers and not reviews.
- p. 6 Related work, first sentence
extensively covered in a number of reviews14,40–43
- p. 7 Reference list, ref. 42
Qian, N. & Sejnowski, T. J. Predicting the secondary structure of globular proteins using
- p. 7 Reference list, ref. 43
Fariselli, P., Olmea, O., Valencia, A. & Casadio, R. Prediction of contact maps with neural
- p. 6 Related work, second sentence
applying neural networks to structure prediction14,42,43
Only the authors can say what is right here.
Judge's note: Superscripts 14,40–43 and 14,42,43 confirmed on the page 6 image; reference entries 40-43 read on page 7. Refs 40 and 41 have review-style titles in review venues; 42 and 43 are, by title and journal, original neural-network predictor papers and are cited appropriately in the next sentence for the history of neural networks in structure prediction. The range probably should end at 41, but only the authors can say what was intended.
Second check: another model agreed
Report a problem with this finding
Reference or label Reference 70 gives a garbled first author: "Ashish, A. M. A. et al. TensorFlow: large-scale machine learning on heterogeneous systems. Preprint at https://arxiv.org/abs/1603.04467 (2015)."; the first author should be Abadi, M.
- p. 10 Reference list, ref. 70
Ashish, A. M. A. et al. TensorFlow: large-scale machine learning on heterogeneous
- p. 10 Reference 70
Ashish, A. M. A. et al. TensorFlow: large-scale machine learning on heterogeneous systems.
Only the authors can say what is right here.
Judge's note: Rests on outside knowledge: that the TensorFlow white paper (arXiv:1603.04467) is by Martín Abadi, Ashish Agarwal et al., so 'Ashish, A. M. A.' is a merge of the first two authors' names. The entry reads as quoted on the page 10 image. The author list could not be checked against an external source here.
Second check: another model agreed
Report a problem with this finding
13 minor (typos and the like; most checked by the judge only)
Typo The same protein (T1064) is written "Orf8 (T1064)" in the Fig. 4 legend and "ORF8 of SARS-CoV-2 (T1064)" in the body text on the same page.
- p. 5 Fig. 4 legend, panel b
targets LmrP (T1024) and Orf8 (T1064)
- p. 5 Interpreting the neural network, body text
ORF8 of SARS-CoV-2 (T1064)
- p. 5 Fig. 4 legend, panel b
LmrP (T1024) and Orf8 (T1064)
Judge's note: Both spellings confirmed on the page 5 crops; reference 84 also uses 'ORF8'. Capitalisation inconsistency only; both forms occur in the literature.
Second check: none (judge only)
Report a problem with this finding
Typo Misspelling in the Discussion: "particular stochiometry, ligand or ion is predictable from the sequence" should read 'stoichiometry' (spelled correctly in the Fig. 5 legend: "without input stoichiometry").
- p. 6 Discussion, second paragraph
particular stochiometry, ligand or ion is predictable
- p. 6 Fig. 5 legend, panel b
correctly predicted without input stoichiometry
- p. 6 Discussion, second paragraph
particular stochiometry, ligand or ion is predictable from the sequence
- p. 6 Discussion, para 2
a particular stochiometry, ligand or ion
Judge's note: The page 6 image shows 'stochiometry' in the body text and 'stoichiometry' in the Fig. 5b legend on the same page. Not an extraction artefact.
Second check: none (judge only)
Report a problem with this finding
Typo Reference 39 is cited twice in one clause: "in a companion paper39, we demonstrate the application of AlphaFold to the entire human proteome39".
- p. 6 Discussion, third paragraph
in a companion paper39, we demonstrate the application of AlphaFold to the entire human proteome39
Judge's note: Both superscripts read 39 on the page 6 image. Redundant duplicate citation; harmless, and either marker alone would do.
Second check: none (judge only)
Report a problem with this finding
Typo Ungrammatical phrase in the Discussion: "AlphaFold is able to handle missing the physical context" (should be 'handle missing physical context' or 'handle the missing physical context').
- p. 6 Discussion, para 2
AlphaFold is able to handle missing the physical context
Judge's note: Wording confirmed on the page 6 image. A stray or misordered 'the'; meaning is recoverable.
Second check: none (judge only)
Report a problem with this finding
Typo Reference 5 reads "Nucleic Acids Res. 47, D520–D528 (2018)"; by the list's own entries (vol. 45 = 2017 in ref. 36, vol. 48 = 2020 in ref. 6) volume 47 is the 2019 volume.
- p. 7 Reference list, ref. 5
structure data. Nucleic Acids Res. 47, D520–D528 (2018)
- p. 7 Reference list, ref. 6
48, D570–D578 (2020)
Judge's note: Entries read on the page 7 and page 10 images. The year is most likely the advance-online date; the same pattern appears in ref. 62 ("Nucleic Acids Res. 49, D480–D489 (2020)", where vol. 49 would be 2021), so the list mixes online year and volume year. The inference is internal to the reference list; the true print year was not checked externally. Harmless inconsistency.
Second check: none (judge only)
Report a problem with this finding
Typo Ref. 62 reads "The UniProt Consortium. UniProt: the universal protein knowledgebase in 2021. Nucleic Acids Res. 49, D480–D489 (2020)": volume 49 is paired with 2020, whereas ref. 6 gives volume 48 as 2020.
- p. 10 Reference list, ref. 62
Acids Res. 49, D480–D489 (2020)
- p. 7 Reference list, ref. 6
48, D570–D578 (2020)
Only the authors can say what is right here.
Judge's note: Rests on outside knowledge: that Nucleic Acids Res. volume 49 is the 2021 volume (the paper was first published online in late 2020). The crop confirms the printed text of ref. 62, and ref. 6 on page 7 reads "48, D570–D578 (2020)", so by the list's own ref. 6 volume 49 would be 2021. However ref. 5 reads "Nucleic Acids Res. 47, D520–D528 (2018)" and ref. 36 reads "45, D170–D176 (2017)", so the list is not consistent in itself: ref. 5 appears to use the online-publication year in the same way as ref. 62. Using the online year for a database-issue paper is a common convention, so this is a minor inconsistency in year style, not a clear error.
Second check: none (judge only)
Report a problem with this finding
Typo Database name written "Mgnify" twice in the ablation paragraph ("removing Mgnify reduced the mean accuracy by 0.7 GDT"; "inclusion of BFD and Mgnify") but "MGnify" elsewhere (Methods, Data and Code availability, ref. 6).
- p. 8 Methods, Inputs and data sources, paragraph on sequence data sources
removing Mgnify reduced the mean accuracy by 0.7 GDT
- p. 8 Methods, Inputs and data sources, same paragraph
effects between inclusion of BFD and Mgnify
- p. 8 Methods, Inputs and data sources, preceding paragraphs
v.2018_08 and MGnify6 v.2018_12
- p. 8 Methods, Inputs and data sources, last paragraph
inclusion of BFD and Mgnify
Judge's note: Page 8 image shows 'Mgnify' in two places in that paragraph (the reviewer said three; the text has two) and 'MGnify' in the preceding paragraphs of the same column. Capitalisation slip, not an extraction artefact.
Second check: none (judge only)
Report a problem with this finding
Typo Database name written "Uniref90" in Methods ("we used Uniref9067 v.2020_01"; "MSA search on Uniref90 and clustered MGnify") but "UniRef90" in Data availability and Code availability and "UniRef" later in the same Methods section and in ref. 67.
- p. 8 Methods, Inputs and data sources
we used Uniref9067 v.2020_01, BFD
- p. 8 Methods, Inputs and data sources
For MSA search on Uniref90 and clustered MGnify
- p. 9 Data availability
UniRef90 v.2020_01
Judge's note: Page 8 image shows 'Uniref90' twice; page 9 image shows 'UniRef90 v.2020_01'; the page 8 ablation paragraph has 'UniRef'. The fused '67' in the text layer is a superscript citation and is ignored.
Second check: none (judge only)
Report a problem with this finding
Typo Tool name written "HHBlits" in the text ("jackhmmer60 and HHBlits61", also Code availability and Author contributions) but "HHblits" in the title of its own reference 61 ("HHblits: lightning-fast iterative protein sequence searching").
- p. 8 Methods, Inputs and data sources
tools including jackhmmer60 and HHBlits61
- p. 10 Reference list, ref. 61
HHblits: lightning-fast iterative protein
Judge's note: Page 8 image shows 'HHBlits'; page 10 image shows ref. 61 titled 'HHblits' and 'HHBlits' again in the author contributions. The article text is internally consistent in using 'HHBlits'; it differs only from the reference title. Capitalisation only.
Second check: none (judge only)
Report a problem with this finding
Typo "The r.m.s.d. of the atoms chosen for the final iterations is the r.m.s.d.95" should read 'final iteration' (singular); the procedure is five iterations and one final set of chosen atoms.
- p. 9 Methods, Metrics, second paragraph
chosen for the final iterations is the r.m.s.d.95
Judge's note: Wording confirmed on the page 9 image. The preceding sentence describes five iterations of align-then-select, so a single r.m.s.d. value can only come from the last iteration. Grammatical slip.
Second check: none (judge only)
Report a problem with this finding
Typo Loose phrasing in Methods: "the 256- and 384-residue inference times are using the memory of a single GPU" (inference times do not use memory; presumably the timings were obtained within a single GPU's memory).
- p. 9 Methods, Inference regimen
the 256- and 384-residue inference times are using the memory of a single GPU
Judge's note: Wording confirmed on the page 9 image. The meaning is recoverable from the surrounding sentences about V100 memory limits; awkward wording only.
Second check: none (judge only)
Report a problem with this finding
Typo Ref. 78 prints "Dunne, M., Ernst, P., Sobieraj, A., Pluckthun, A. & Loessner, M. J." with no umlaut, while other entries keep diacritics ("Wüthrich", "Šali", "Söding").
- p. 10 Reference 78
Pluckthun, A.
Only the authors can say what is right here.
Judge's note: Rests on outside knowledge: that the author's surname is spelled Plückthun. The crop confirms the page prints "Pluckthun, A." and the list elsewhere keeps diacritics (refs 4, 18, 61, 63, 64). Ref. 78 cites a PDB deposition (doi 10.2210/pdb6YJ1/pdb), and PDB author lists are stored without diacritics, so the spelling may have been copied faithfully from the deposition record. Cosmetic either way.
Second check: none (judge only)
Report a problem with this finding
Typo In the Reporting Summary statistics checklist, the row "For Bayesian analysis, information on the choice of priors and Markov chain Monte Carlo settings" is ticked in the "Confirmed" column, not "n/a", although the article describes no Bayesian analysis.
- p. 11 Reporting Summary, Statistics checklist
For Bayesian analysis, information on the choice of priors and Markov chain Monte Carlo settings
Only the authors can say what is right here.
Judge's note: On the full page 11 image the column headers are "n/a" (left) and "Confirmed" (right). The Bayesian row has the right-hand box crossed, the same as the sample-size and statistical-parameters rows, and unlike the covariates, null-hypothesis and hierarchical-design rows, which are crossed in the left n/a box. A text search of the article for Bayes, priors, MCMC and posterior finds only this checklist line (other hits are hidden Markov models and unrelated uses of "prior"). The supplementary information was not provided, so a Bayesian analysis there cannot be excluded, though it is unlikely. This is a form-filling slip in the reporting checklist with no effect on the results, so it is rated cosmetic.
Second check: none (judge only)