AlphaFold: independently ranked the top method in the CASP14 blind assessment (summed z-score 244 vs 91 for second place), scaled to 214M+ predicted structures, and won the 2024 Nobel Prize in Chemistry

Verification status: GRADUATED — publish-ready, awaiting the owner’s sign-off (not yet green). The headline rests on the CASP14 blind assessment, which is genuinely independent of DeepMind, plus a peer-reviewed database paper and the 2024 Nobel citation. The one first-party-only number, the exact median 92.4 GDT, is flagged in the prose below.

The problem

Predicting a protein’s three-dimensional shape from its amino-acid sequence had been, in DeepMind’s framing, a “50-year-old grand challenge in biology” (source). The field’s own yardstick for progress is CASP, the Critical Assessment of protein Structure Prediction, whose organizers “choose protein structures that have only very recently been experimentally determined (some were still awaiting determination at the time of the assessment) to be targets for teams to test their structure prediction methods against; they are not published in advance” (source). That blind design is what makes a strong CASP result an independent test rather than a self-report.

What was built

Google DeepMind built AlphaFold, a deep-learning system that predicts a protein’s 3D structure from its amino-acid sequence, and entered its AlphaFold2 version in the 14th CASP round in 2020 (source). The predictions were later released as the AlphaFold Protein Structure Database, described in the peer-reviewed database paper as “a collaborative project between EMBL-EBI and Google DeepMind” (source).

The outcome

The independent scorers ranked AlphaFold2 first by a wide margin. On the Protein Structure Prediction Center’s official CASP14 results, AlphaFold2 (group 427) posts a summed z-score of 244.0217 at Rank 1, while the second-placed group (BAKER, group 473) reaches a summed z-score of only 90.8241 (source). An independent peer-reviewed review, by University of Sao Paulo authors with no connection to DeepMind, reports the same result: “AlphaFold2 scored 244.0 in summed z-scores compared with 90.8 for the next closest group” (source). In DeepMind’s own account of the same assessment, “our latest AlphaFold system achieves a median score of 92.4 GDT overall across all targets” (source).

The predictions then reached research at scale. The 2023 database paper is titled for providing “structure coverage for over 214 million protein sequences” (source), and EMBL, the database’s host, describes the expanded release as “200 million protein structure predictions” reaching “2 million users in 190 countries” (source). In 2024 the work was recognised at the highest level: the Royal Swedish Academy of Sciences awarded the Nobel Prize in Chemistry “the other half jointly to Demis Hassabis [and] John Jumper” of Google DeepMind “for protein structure prediction” (source).

Weakest load-bearing source, named. The exact figure “92.4 GDT” is stated in DeepMind’s own announcement, which is a first-party source (source). It is not left to stand alone: GDT is computed by the independent CASP organizers, and the same organizers’ summed-z-score ranking places AlphaFold2 first by roughly a factor of 2.7 over the next group, an independent measurement of the same superiority (source). Treat the precise 92.4 value as DeepMind’s number and the top-of-the-field ranking as the independent one.

How this was verified. Method: the CASP14 overall ranking (AlphaFold2 group 427 first at summed z-score 244.02 vs 90.82 for second place) was read from the Protein Structure Prediction Center’s official results page (predictioncenter.org — Tier 1, independent assessor; archived Wayback 20260828155602) and corroborated against an independent peer-reviewed review that reports the identical figures (Bertoline et al., Frontiers in Bioinformatics 2023 — Tier 2; archived Wayback 20260828162038). The median 92.4 GDT figure was quote-matched against DeepMind’s announcement (Tier 3, first-party; archived 20260828155544). The 214-million-sequence coverage is from the peer-reviewed Nucleic Acids Research database paper (Tier 2; archived 20260828155627), and the 200-million / 2-million-users figures from EMBL, the database host (Tier 2; archived 20260828155654). The Nobel citation is verbatim from the Royal Swedish Academy press release (Tier 1; archived 20260828155852). Verified 2026-08-28. No confirmation was sought from DeepMind: asking the subject to confirm its own numbers is a testimonial, not an audit.

Sources

1. Protein Structure Prediction Center · CASP14 official results — summed z-scores (final) · 2020 · https://predictioncenter.org/casp14/zscores_final.cgi — Tier 1 (independent academic assessor’s primary results; AlphaFold2 group 427 ranked 1 at summed z-score 244.0217, second-placed BAKER group 90.8241; archived Wayback 20260828155602, saved to sources/predictioncenter-casp14-zscores.txt).
2. Google DeepMind · AlphaFold: a solution to a 50-year-old grand challenge in biology · 30 November 2020 · https://deepmind.google/discover/blog/alphafold-a-solution-to-a-50-year-old-grand-challenge-in-biology/ — Tier 3 (first-party announcement; source of the exact “median score of 92.4 GDT overall across all targets” and the description of CASP’s blind design; archived Wayback 20260828155544, saved to sources/deepmind-casp14-blog.txt).
3. Varadi M, et al. (EMBL-EBI / Google DeepMind) · AlphaFold Protein Structure Database in 2024: providing structure coverage for over 214 million protein sequences · Nucleic Acids Research, 2 November 2023 (52(D1):D368-D375) · https://pmc.ncbi.nlm.nih.gov/articles/PMC10767828/ — Tier 2 (peer-reviewed database paper; source of the “over 214 million protein sequences” coverage figure; archived Wayback 20260828155627, saved to sources/nar-2024-alphafold-db.txt).
4. EMBL · Case study: AlphaFold uses open data and AI to discover the 3D protein universe · 2023 · https://www.embl.org/news/science/alphafold-using-open-data-and-ai-to-discover-the-3d-protein-universe/ — Tier 2 (independent host of the AlphaFold DB; restates “200 million protein structure predictions” and “2 million users in 190 countries”; archived Wayback 20260828155654, saved to sources/embl-alphafold-casestudy.txt).
5. The Royal Swedish Academy of Sciences · The Nobel Prize in Chemistry 2024 — press release · 9 October 2024 · https://www.nobelprize.org/prizes/chemistry/2024/press-release/ — Tier 1 (authoritative primary; verbatim citation awarding half the prize jointly to Demis Hassabis and John Jumper of Google DeepMind “for protein structure prediction”; archived Wayback 20260828155852, saved to sources/nobel-2024-chemistry.txt).
6. Bertoline LMF, Lima AN, Krieger JE, Teixeira SK · Before and after AlphaFold2: An overview of protein structure prediction · Frontiers in Bioinformatics, 3:1120370, 28 February 2023 · https://www.frontiersin.org/journals/bioinformatics/articles/10.3389/fbinf.2023.1120370/full — Tier 2 (independent, peer-reviewed review by University of Sao Paulo authors unaffiliated with DeepMind; restates the CASP14 outcome verbatim, “AlphaFold2 scored 244.0 in summed z-scores compared with 90.8 for the next closest group”, giving c1 a second independent report; archived Wayback 20260828162038, saved to sources/frontiers-casp14-review.txt).

Related case files

• Insilico Medicine’s AI-discovered drug rentosertib, which reached a Phase 2a readout — the downstream half of the story: AI moving from structure prediction to a molecule that actually entered the clinic.
• The MASAI trial, where AI mammography was measured in a rigorous independent study — the same methodological point that gives AlphaFold its weight: an AI result carries most where an independent trial, not the vendor, does the measuring.
• Waymo’s Swiss Re crash-rate study, an outcome measured by a single source — the contrast: a strong-looking AI result that capped below green precisely because no party independent of the subject re-measured it, which is what CASP does here.

Path to green

The critical claim (c1, the CASP14 ranking) now rests on two independent reports: the primary blind-assessment ranking at predictioncenter.org (Tier 1) and an independent peer-reviewed review that reports the identical 244.0-vs-90.8 result (Bertoline et al., Frontiers in Bioinformatics 2023, Tier 2). The 2024 Nobel citation is a further independent authoritative validation and the scale figure is peer-reviewed. The only remaining first-party-only number is the exact 92.4 median GDT, which is supporting color, not the headline: GDT is scored by the independent CASP organizers, and the independent z-score ranking already carries the top-of-field claim. No confirmation should be sought from DeepMind; the headline stands on independent sources.