AlphaFold2 at CASP14

Filed as idea in canon/proposals.md, and idea is wrong. The proposal line reads: "alphafold2-2020 — AlphaFold2 (Jumper et al., CASP14 2020, Nature 2021) — structure prediction at experimental accuracy; Nobel 2024. Cite when: the accelerant lens needs a benchmark for what a real AI-accelerated result looks like." That citation occasion — a benchmark for what a real result looks like — is a moment occasion, not an idea occasion, and the proposal contradicts its own filing in the sentence that matters most.

The house definition of idea, set by shannon-chess-1950 and used since by dijkstra-1959 and eliza-1966, is a framing later work is built out of rather than argued about, where the citation is "the thing you are watching descends from this, and the descent explains its shape." AlphaFold2 has a framing — the Evoformer, attention over multiple sequence alignments, the structure module, recycling — and it did descend into AlphaFold3, ESMFold, RoseTTAFold and everything since. But almost nobody who cites AlphaFold2 could describe any of that, and nobody needs to. What is canonical about it is not its architecture. It is that on a stated date, against an opponent it did not choose, under a rule set it did not write, judged by assessors it did not appoint, it produced a number nobody thought was available for another decade. That is moment on the definition eliza-1966 argued out when it talked its way out of the kind, and that deep-blue-1997, watson-jeopardy-2011, clippy-1996, siri-2011, expert-systems-collapse-1987 and alphago-move-37-2016 have used since: a date on which something visibly happened in public.

The entry's first real finding is that the id names the wrong date, and the right one is seven and a half months later. The id says 2020, and 2020 is CASP14 — 30 November 2020, the scoreboard. But CASP14 changed nothing outside a conference. What changed the world is 15 July 2021, when the code went up under Apache 2.0 with the trained parameters under CC-BY-4.0, and 22 July 2021, when the database of predicted structures was given away free. Every measured downstream effect in this entry — all of it — starts on those two days and none of it starts in November 2020. The entry keeps the 2020 id anyway, because the scoreboard is the adjudicated part and adjudication is what this canon is short of; but a reading citing "AlphaFold" for anything about impact is citing July 2021, and should say so.

Why not idea, more precisely. An idea entry here would be about the Nature paper, would be dated 15 July 2021, would be titled Highly accurate protein structure prediction with AlphaFold, and would have the transformer as its parent. That is a real entry and somebody should write it. Its parent id does not exist in canon/ yet — there is no transformer-2017, no alexnet-2012, no scaling-laws-2020 — which is a gap this entry names rather than invents around.

Why not interpretation. The competitor is the phrase "AlphaFold for X," which by 2026 is a standing claim-form applied to materials, weather, chemistry, mathematics and drug design, and which almost always means our field is about to get its miracle. That frame is worth an entry. It is not this one, for the same reason alphago-move-37-2016 gave for declining "a Move 37 moment": the frame is still moving and the event underneath is fixed. Anchor to the fixed thing; grade the frame in section 4.

descends_from holds one id, and the descent is institutional rather than architectural — which DeepMind itself is unusually clear about. In AlphaGo at 10, published 10 March 2026, Hassabis writes of the Seoul match: "It was further proof of what I knew the moment we won the match in Seoul — the technology was ready to be applied to our real goal of accelerating scientific breakthroughs." The same piece names AlphaProof as "the most direct descendant of AlphaGo's architecture" and makes no equivalent claim for AlphaFold. That is exactly right and worth preserving: nothing in the Evoformer is Monte Carlo tree search. What descends from Seoul is not a method but a thesis about where to point a lab — pick a problem with a scoreboard, an independent referee and fifty years of accumulated ground truth, and win it in public. AlphaFold2 is that thesis executed a second time on something that mattered, and the executive who says so is the one who ran both.

Two other ancestors are named here rather than in the header, which is what the job asks for. mycin-1976 is the closest thematic parent in the canon and section 2 leans on it hard: a system that matched or beat experts in a blinded evaluation and was never used on a patient. watson-jeopardy-2011 is the other half of the same series — an impeccable demonstration whose deployment failed in the field. Neither is a genealogy. A reading may cite any of the three; this entry's parent is Seoul.

What it is

The problem, and why "fifty years" is roughly honest

A protein is a chain of amino acids that folds into a specific three-dimensional shape, and the shape is what does the work. In his 1972 Nobel lecture, Christian Anfinsen set out what became Anfinsen's dogma: the sequence determines the structure. If that is true, the structure is in principle computable from the sequence alone. In 1969, Cyrus Levinthal had pointed out why that is not easy — a typical chain has on the order of 10³⁰⁰ possible conformations, so no enumeration will ever finish, and yet real proteins fold in milliseconds.

Between those two facts sits the gap the field lived in for five decades. Determining a structure experimentally — X-ray crystallography, NMR, later cryo-EM — took months to years per protein, and the Nature paper's own abstract puts the resulting coverage bluntly: "Through an enormous experimental effort, the structures of around 100,000 unique proteins have been determined, but this represents a small fraction of the billions of known protein sequences." Hill and Stein's 2026 count of the public archive is roughly 130,000 unique structures; the curated SwissProt subset of UniProt holds 570,000 proteins of which only 8% had an experimental structure before AlphaFold. Sequencing had become cheap and structure had not, and the ratio between them was the bottleneck.

"Fifty years" is therefore a fair description of the structure prediction problem dated from Anfinsen, and it is not a fair description of the folding problem, which is a different question — how the chain gets there — and which nobody claims to have solved. The distinction is the entire content of section 4, and DeepMind's own scientists drew it correctly while DeepMind's marketing did not.

CASP, and why it is the reason this entry exists

CASP — Critical Assessment of Structure Prediction — was founded in 1994 by John Moult and Krzysztof Fidelis and has run biennially since. Its design is the point. Every two years, participants are given the amino-acid sequences of proteins whose structures have already been solved experimentally but not yet released. Predictions are submitted blind against a deadline. The experimental answers are then unsealed and the models scored by independent assessors who are not the predictors.

This canon has spent five entries complaining that AI capability claims are graded by the vendor on a suite the vendor picked. CASP is what the alternative looks like, it predates the deep-learning era by twenty years, it was built by the host field for the host field's own reasons, and no AI lab has any leverage over it. A reading that wants to know what an honest capability evaluation looks like should look at CASP before it looks at anything a lab published this month. That is arguably the most transferable thing in this file, and it has nothing to do with proteins.

What happened at CASP14

Results were presented at the CASP14 meeting and announced publicly on 30 November 2020. DeepMind's entry, group 427, scored a median GDT_TS of 92.4 across all targets and 87.0 on the free-modelling category — the hardest class, where no useful template structure exists. Median error was about 1.6 Å, roughly the width of an atom. DeepMind's own framing of the threshold is worth keeping because it is the one everyone repeated: "A score of around 90 GDT is informally considered to be competitive with results obtained from experimental methods."

The independent numbers are more useful than the vendor's, and they are worse for everyone else. From the Oxford Protein Informatics Group's read of the official CASP14 tables on 1 December 2020: in the assessors' ranking by summed z-score, group 427 scored 244.0 against 90.8 for the second-placed group 473 (BAKER) — David Baker's lab, which would share the 2024 Nobel for other work. AlphaFold2's average z-score was about 2.5 across all targets and 3.8 on the hardest, against roughly zero for the field. 36% of its predictions came in under 2 Å RMSD, against a typical experimental resolution of about 2.5 Å; 86% were under 5 Å, mean 3.8 Å. On one straightforward target, T1046s1, it hit 0.48 Å. The assessor's summary of the difference, quoted in the same write-up, was: "What did AlphaFold 2 get right, that other models did not? The details."

John Moult, who had founded the competition twenty-six years earlier to measure exactly this, said: "We have been stuck on this one problem – how do proteins fold up – for nearly 50 years. To see DeepMind produce a solution for this, having worked personally on this problem for so long and after so many stops and starts, wondering if we'd ever get there, is a very special moment."

The reaction outside the room is best preserved in Andrei Lupas, an evolutionary biologist at the Max Planck Institute in Tübingen, quoted by **Ewen Callaway in Nature on 30 November 2020**: "This will change medicine. It will change research. It will change bioengineering. It will change everything." He was not speaking loosely — AlphaFold had just solved a bacterial protein structure his lab had been stuck on for a decade. Section 3 grades that sentence at five years and nine months. It does not grade it unkindly.

The paper, and what it says about itself

Highly accurate protein structure prediction with AlphaFold, by Jumper, Evans, Pritzel, Green, Figurnov and colleagues, was received 11 May 2021, accepted 12 July 2021, and published in Nature 596, 583–589, on 15 July 2021. Median backbone accuracy 0.96 Å r.m.s.d.95; all-atom 1.5 Å.

The abstract is more careful than the press release, and the gap between them is this entry's spine. It describes the target as "the structure prediction component of the 'protein folding problem'" — the authors carving out precisely the distinction the blog post elided. Its claim is "the first computational method that can regularly predict protein structures with atomic accuracy even in cases in which no similar structure is known," which is a claim about prediction, is checkable, and has held.

The paper also states its own failure modes, which is why they are known: accuracy drops substantially when the multiple sequence alignment is shallower than about 30 sequences, and performance is poor on proteins whose contacts are predominantly heterotypic and cross-chain. Orphan proteins, designed proteins and antibodies were flagged as weak from the start, by the authors, in the paper.

The release, which is the part that mattered

On 15 July 2021 DeepMind published the source code under Apache 2.0 with the trained parameters under CC-BY-4.0. On 22 July 2021, with EMBL-EBI, it launched the AlphaFold Protein Structure Database — around 350,000 structures covering the human proteome and twenty model organisms, free to anyone. [verify] on the exact initial count, which is reported between roughly 350,000 and 365,000 depending on source. By July 2022 the database covered most of UniProt release 2021_04 and now holds over 214 million predicted structures.

Set that against what the same organisation did with the same line of work three years later, in section 3. The trajectory from here is the single cleanest thing in this canon for the concentration lens, and it runs in one direction.

Why a reading would cite it

The admission case: this is the strongest AI-for-science result that exists, which is exactly why it is the right thing to measure every weaker one against. It has now been running long enough to be graded by economists rather than by enthusiasts, and the grade is a ceiling.

Every other entry in this canon's demonstration series has a defect that lets the ambitious reader dismiss the grading. mycin-1976 was never deployed, so we cannot know. watson-jeopardy-2011 was deployed badly, so perhaps that was IBM rather than the technology. deep-blue-1997 was a game. AlphaFold2 has none of those outs. It was adjudicated by a hostile-by-design third party, published in full, given away free, adopted by over 3 million researchers in more than 190 countries on DeepMind's own count, cited in more than 35,000 papers, and awarded a Nobel Prize in Chemistry in 2024. If AI-accelerated science works anywhere, it works here. So what it did in six years is not a floor on what AI can do for a field. It is close to a ceiling, and a reading that treats a 2026 "AlphaFold for X" claim as promising more than this is claiming something the best case does not support.

And the six-year measurement now exists. Ryan Hill (Northwestern) and Carolyn Stein (UC Berkeley), How Artificial Intelligence Shapes Science: Evidence from AlphaFold — first version 11 March 2026, current version 17 July 2026 — exploit the fact that only 8% of curated proteins had a structure before AlphaFold, treat those as controls and the other 92% as treated, and run a difference-in-differences. Their findings, in their words:

• On substitution: "to date, we find that the rate of experimental structure determination has remained almost unchanged." PDB deposits went slightly up. Their survey of working structural biologists finds "the vast majority say they do not trust predicted structures enough to skip experimental validation," about half report pivoting to proteins AlphaFold handles badly, and respondents expect to solve more experimental structures by 2031, not fewer.
• On basic science, the good news, and it is large: research on previously unsolved proteins "increased by 23 to 32% relative to previously solved proteins, with effects as large as 80 to 300% by the last year of our sample." New papers about those proteins disproportionately cite the core AlphaFold papers, which is how they establish the mechanism rather than assuming it. They call it a "floodlight" effect.
• On applied R&D, the finding this project needs: "In contrast with our basic science results, we see small and only marginally significant upticks in patents referencing previously unsolved proteins. Moreover, our relatively tight standard errors allow us to rule out effect sizes larger than a 13% relative increase." Their conclusion: "more basic scientific research needs to occur before scientists can make use of these new protein structures in ways that benefits patients and consumers."

That is the citable shape and it is worth stating in one sentence: the science moved enormously, the medicine has not moved yet, and the bound on how much the medicine moved is 13%.

Three concrete occasions.

One: any 2026 claim of the form "AI has transformed field X." The digest for 2026-08-15-12 records Crouzeix's conjecture acquiring two independently claimed proofs with disclosed model use, neither yet peer-reviewed, and notes that the result reached general attention on 14 August. That is the good shape of an accelerant finding by this project's own test — a checkable artefact, a named model, a disclosed method. This entry is what the next question looks like: five and a half years after the strongest such result in history, the patent effect cannot be distinguished from 13%. A reading should expect the lag between a real scientific acceleration and anything a person outside the field can feel to be measured in many years, and should say so rather than let the excitement stand unqualified.

Two: mycin-1976, updated and inverted. MYCIN beat infectious-disease faculty in a blinded 1979 evaluation and was never used on a patient. Isomorphic Labs — DeepMind's drug-discovery spinout, incorporated 24 February 2021 and announced 4 November 2021, built explicitly on AlphaFold, with Novartis and Eli Lilly partnerships from January 2024, $600 million raised in April 2025 and $2.1 billion more on 12 May 2026 led by Thrive Capital with Alphabet, GV, MGX, Temasek, CapitalG and the UK Sovereign AI Fund — has, as of this entry's writing, dosed no human being. Hassabis had set end-2025 for first-in-human trials and restated it at Davos in January 2026 as end-2026, a one-year slip. [verify] on the Davos restatement, which comes from secondary reporting. The primary quotations from the raise are Hassabis — "Now that we have shown our approach is fundamentally sound, our focus is on scaling our technology to its full potential" — and president Max Jaderberg: "Our drug design engine works, and it's giving us a repeatable way to design new medicines for a wide range of diseases." Both sentences are about the engine. Neither is about a patient. The distance between "the engine works" and "somebody got better" is the thing mycin-1976 exists to hold open, and here it is being held open at a valuation. This is not an accusation of bad faith; drug development is slow for reasons that have nothing to do with AI, which is precisely the point Hill and Stein's bottleneck finding makes quantitatively.

Three, and the one a reading is most likely to need: the same organisation, same line of work, walking from maximally open to maximally closed in five years. Under the concentration lens the trajectory is unusually clean because every step is dated and the actor is constant.

• July 2021, AlphaFold2: code Apache 2.0, weights CC-BY-4.0, database free to the world. This is why the Hill–Stein effect exists at all — an effect that large is not available from an API.
• May 2024, AlphaFold3: published in Nature without code. Ten scientists wrote to the editors: "The amount of disclosure in the AlphaFold3 publication is appropriate for an announcement on a company website" and it "fails to meet the scientific community's standards of being usable, scalable, and transparent." By 28 May 2024 the letter had over 1,000 signatures. DeepMind committed to an academic download within six months and delivered in November 2024. [verify] on the exact May 2024 publication date, which secondary sources give inconsistently.
• 10 February 2026, IsoDDE: Isomorphic Labs' drug design engine, announced in a 27-page technical report. Its claims are large — "IsoDDE more than doubles the accuracy of AlphaFold 3 on a challenging protein-ligand structure prediction generalisation benchmark"; it "outperforms AlphaFold 3 by 2.3x and Boltz-2 by 19.8x in the high-fidelity regime (DockQ > 0.8)"; it surpasses all deep-learning methods on FEP+, OpenFE and the CASP16 blind binding-affinity task. Every one of those numbers is self-reported and the model is not available to anyone outside the company.

Read as a sequence: the result that earned the Nobel was the open one, and the openness is inseparable from the measured impact. What came after is progressively less checkable, and the most recent step is a company grading itself on benchmarks it selected, on a system nobody else can run. A reading meeting a 2026 claim that begins "we have surpassed AlphaFold" should notice that AlphaFold's own number came from CASP and this one did not.

What it got right, and what it got wrong

moment does not require this section. Six dated claims attach and all are due.

Claim 1 — "solved." Made 30 November 2020. Contested by February 2022. Half right, and the wrong half is the famous half.

The claim: DeepMind's announcement was titled "AlphaFold: a solution to a 50-year-old grand challenge in biology," and CASP's own organisers described the problem of protein structure prediction as solved for single chains.

What happened: on 4 February 2022, four of the most senior structural biologists alive — Peter B. Moore (Yale), Wayne A. Hendrickson (Columbia), Richard Henderson (MRC LMB) and Axel T. Brunger (Stanford) — published a letter in Science 375(6580), 507, titled "The protein-folding problem: Not yet solved." Their point is the one the Nature abstract had already conceded: predicting the folded state is not the same as understanding folding. I could not read the letter itself — science.org returned HTTP 403 and the Stanford mirror failed TLS verification — so this entry states its citation, its title and its authorship, and [verify]s its text rather than quoting it second-hand.

The strongest evidence that they are right is experimental. **Carlos Outeiral, Daniel A. Nissley and Charlotte M. Deane, "Current structure predictors are not learning the physics of protein folding," Bioinformatics 38(7), 1881–1887, published 31 January 2022**, compared the folding trajectories implied by AlphaFold2, RoseTTAFold, trRosetta, RaptorX, DMPfold, EVfold, SAINT2 and Rosetta against experimental folding-pathway data across 170 proteins. Their summary, verbatim: "We find evidence that their simulated dynamics capture some information about the folding pathway, but their predictive ability is worse than a trivial classifier using sequence-agnostic features like chain length." On two-state kinetics they measure unsupervised accuracy 0.613 and F1 0.591; the Spearman correlation between predicted folding-event position and log folding rate is −0.23; for seven of nine proteins with known intermediates, Jaccard similarity is about 0.1.

Graded strictly: "structure prediction is solved" is broadly true and holds up. "The protein folding problem is solved" was never claimed by the paper, was claimed by the announcement, and is false. A model can be right about where the chain ends up while carrying no correct account of how it gets there — and it is. That distinction is worth transporting wholesale: a system that predicts an outcome accurately has not thereby acquired the mechanism, and the two get conflated the moment a press release is written.

Claim 2 — Lupas's "it will change everything." Made 30 November 2020. Due, generously, 2026. Right about research, not yet right about medicine.

The claim: "This will change medicine. It will change research. It will change bioengineering. It will change everything."

What happened: on research, he was right and the effect is measured — the 23–32% shift toward previously structureless proteins, rising to 80–300% by the end of Hill and Stein's sample, is a large effect by any standard in the economics of science, and the "floodlight" metaphor is theirs, not a booster's. Concrete instances are easy to name and are not press releases: the cytoplasmic ring of the nuclear pore complex, solved by integrative cryo-EM with AlphaFold-predicted nucleoporins fitted into a medium-resolution map (Science, 2022); the malaria transmission-blocking vaccine candidate Pfs48/45, where an AlphaFold2 model supplied the molecular-replacement solution that let a 2.13 Å dataset be interpreted (Nature Communications, 2022); and Lupas's own decade-old bacterial protein.

On medicine, the honest verdict at five years nine months is not yet, with a measured bound: patents referencing newly structured proteins moved by an amount that cannot be distinguished from zero and cannot exceed 13%; no AlphaFold-lineage molecule has entered a human being. A reading should treat Lupas's sentence as the model of a good overclaim — made by an expert, in good faith, about a real result, and wrong only in its tense. That failure mode is far more common in this material than dishonesty, and much harder to spot at the time.

Claim 3 — that prediction would displace experiment. Made continuously since December 2020. Due now. Wrong.

What happened: nothing. PDB deposits went slightly up. Structural biologists did not stop solving structures; they redirected toward proteins AlphaFold handles badly and used predictions as molecular-replacement templates for the ones it handles well. The Nature Methods framing of this, in a 2023 title that says the whole thing, is that AlphaFold predictions are valuable hypotheses that accelerate but do not replace experimental structure determination. [verify] — I have that paper's title and thrust but did not read it.

This is the canon's cleanest case of a capability that was genuinely superhuman on its benchmark and produced approximately zero labour displacement in the occupation it was supposed to automate, six years on. Under the work and the economy lens that is a finding with a mechanism attached: the output was not trusted enough to be terminal, so it became an input. Whether that generalises is exactly the open question, and this is the best-documented instance either way.

Claim 4 — the benchmark itself. Made 2020, tested at CASP15 and CASP16. The benchmark saturated, which is not the same as the problem being finished.

What happened: at CASP16, held in late 2024 and assessed in 2025 by Yuan, Zhang, Kryshtafovych, Schaeffer, Zhou, Cong and Grishin across 125 human expert groups and 36 server groups, the best monomer models averaged GDT_HA 81.01, against 80.80 at CASP15 — "the progress in monomer modeling from CASP15 to CASP16 was subtle," with no statistically significant difference (p = 0.92). The top strategies were all AlphaFold2 and AlphaFold3 with better multiple sequence alignments and more sampling. AlphaFold3 improved confidence estimation and model ranking rather than raw accuracy. What remains hard: "truncated sequences, irregular secondary structures, and interaction-induced conformational changes."

Recorded plainly: four years after the leap, the benchmark has not moved, and this is ambiguous evidence rather than good or bad news. It is consistent with "the problem is finished" and equally consistent with "the measure stopped resolving the thing." alphago-move-37-2016 records what happened the last time a saturated benchmark was read as a settled question: seven years later an amateur beat superhuman Go engines fourteen games in fifteen by going off the distribution the benchmark lived on. Nobody has run that experiment against AlphaFold, and until somebody does, "solved" means "no longer measured by this."

Claim 5 — the drug-discovery case specifically. Made 30 November 2020. Partially due. Weaker than advertised, and known to be.

The claim: DeepMind's announcement offered "drug design and environmental sustainability" and use "in development of new medicines" — with, to its credit, the caveat "Not every structure we predict will be perfect."

What happened: Masha Karelina, Joseph J. Noh and Ron O. Dror, "How accurately can one predict drug binding modes using AlphaFold models?", eLife 12:RP89386 (2023), tested the specific thing drug discovery needs — docking a small molecule into a predicted pocket. Their finding: ligand poses predicted from AlphaFold2 models were not significantly more accurate than those from traditional homology models, and much less accurate than those from experimentally determined structures. [verify] — I have this from search summaries and the eLife listing, not from the paper read end to end. The structural reason is intuitive once stated: AlphaFold predicts one static apo-like conformation, and binding pockets are exactly where proteins move.

The grade: a structure is upstream of a drug by a distance that a structure does not close. DeepMind's own November 2020 caveat was more honest than the five years of "AI will design drugs" that followed it, and Hill and Stein have now put a number on the shortfall.

Claim 6 — the vendor's own five-year self-assessment. Made 25 November 2025. Gradeable now, and one number is in tension with the independent work.

The claim: DeepMind's AlphaFold: Five Years of Impact reports 3 million+ researchers in 190+ countries, over 1 million users in low- and middle-income countries, 8 million+ folds through AlphaFold Server, 35,000+ citations, 200,000+ papers incorporating AlphaFold2 elements, AlphaFold papers "twice as likely to be cited in clinical articles", "over 30% of AlphaFold-related research is focused on better understanding disease", and — citing independent analysis — "an increase of over 40% in their submission of novel experimental protein structures."

Grading: the adoption figures are the strongest vendor claims in this canon and this entry does not dispute them — the LMIC access number in particular is a real consequence of a genuinely free release and should be cited as such under concentration. The 40% figure sits in visible tension with Hill and Stein's finding that overall experimental structure determination is "almost unchanged." Both can be true if the 40% refers to novel-structure submissions by AlphaFold users specifically rather than to the field's total output, which is the most likely reading, but the vendor page does not say so. [verify]. What should be noticed either way is the shape of the page: everything about usage is quantified, and everything about medicine — "one day solve all diseases", "an era of digital biology" — is aspirational and undated. Five years on, with a Nobel in hand, the vendor's own impact page still cannot name a patient. That is not a scandal. It is the base rate, and it is the reason this entry exists.

Commonly misused as

moment does not require this section; godel-incompleteness-1931 established that the canon writes one wherever a result is repeatedly made to say something it does not.

1. "AlphaFold solved protein folding." It solved structure prediction for a broad class of single chains. The Nature abstract itself says "the structure prediction component of the 'protein folding problem'." The folding process — pathways, intermediates, kinetics — is not predicted, and Outeiral, Nissley and Deane showed experimentally that AlphaFold2's implied dynamics are worse than a classifier using chain length alone. Four senior structural biologists put the correction in Science fifteen months after the announcement. The paper was right and the press release was wrong, and it is the press release that propagated.

2. "AI cracked a 50-year problem, so AI can crack yours." The transferable lesson runs the other way, and it is about preconditions. AlphaFold2 had: a single unambiguous objective expressible as a number; fifty years of curated experimental ground truth in one public archive; and a biennial blind competition with independent assessors, running since 1994, that the vendor did not control. Remove any one and there is no CASP14. Almost no 2026 "AI for science" claim has any of the three — most have no ground-truth archive, no scoreboard, and no referee. When a field is offered its AlphaFold, the question to ask is not how good the model is; it is what the field's PDB is and who its Moult would be. Usually the answer is that neither exists.

3. "AlphaFold means we no longer need experiments." Empirically false and measured false. Deposits rose. Structural biologists report not trusting predictions enough to skip validation and expect to solve more structures by 2031. Prediction became an input to experiment — molecular replacement, cryo-EM map interpretation — rather than a substitute for it. Anyone citing AlphaFold as the template for AI displacing a technical profession is citing the case where that conspicuously did not happen.

4. "The Nobel proves AI is doing science." The 2024 Chemistry prize, announced 9 October 2024, went half to David Baker "for computational protein design" and half jointly to Demis Hassabis and John Jumper "for protein structure prediction." It was awarded to people, for a result, in chemistry. It is evidence that a machine-learning system produced a result the Academy judged worth a chemistry prize — which is a great deal, and is the strongest single credential in this entry — and it is not evidence about autonomy, generality, or anything a language model does. It also arrived unusually fast, four years after CASP14, which Hill and Stein note as "an unusually early indicator of its perceived scientific importance." Early canonisation is a thing to notice, not only a thing to cite.

5. "AlphaFold-scale results are now routine." CASP16's monomer numbers are flat to three significant figures and statistically indistinguishable from CASP15's. The frontier moved to complexes, ligands and affinity, where the leading claims are now made by companies on benchmarks they select, with models nobody outside can run. The one thing that made the 2020 result trustworthy — an adjudicator the vendor did not control — is the thing the field's frontier has been quietly moving away from ever since.

Sources

Primary.

• J. Jumper, R. Evans, A. Pritzel, T. Green, M. Figurnov et al., "Highly accurate protein structure prediction with AlphaFold," Nature 596, 583–589, 15 July 2021, doi:10.1038/s41586-021-03819-2. Read via the PubMed Central deposit; nature.com redirects to an identity provider and could not be read directly. Source of the received/accepted/published dates, the abstract's "structure prediction component" framing verbatim, the 0.96 Å / 1.5 Å r.m.s.d.95 figures, the ~100,000 solved structures figure, and the stated MSA-depth and heterotypic-contact limitations.
• Google DeepMind, "AlphaFold: a solution to a 50-year-old grand challenge in biology," 30 November 2020, read directly. Source of the 92.4 and 87.0 GDT medians, the ~1.6 Å error, the "around 90 GDT… competitive with… experimental methods" framing, John Moult's quotation, the "drug design and environmental sustainability" claim and the "Not every structure we predict will be perfect" caveat that section 3 credits.
• Google DeepMind, "AlphaFold: Five Years of Impact," 25 November 2025, read directly. Source of every figure in claim 6 and of the aspirational medicine language graded there.
• Isomorphic Labs, "The Isomorphic Labs Drug Design Engine unlocks a new frontier," 10 February 2026, read directly. Source of the IsoDDE claims quoted verbatim in section 2, and of the fact that they are self-reported.
• Isomorphic Labs, "Isomorphic Labs secures $2.1 Billion funding to scale its AI drug design engine," PR Newswire, 12 May 2026, read directly. Source of the investor list and of the Hassabis and Jaderberg quotations. Note what it does not contain: no valuation, no programme count, no clinical timeline.
• Google DeepMind / Demis Hassabis, "AlphaGo at 10: How AI Innovation Is Paving the Path to AGI," 10 March 2026, read directly. Source of the "ready to be applied to our real goal of accelerating scientific breakthroughs" line that establishes descends_from, and of the observation that DeepMind names AlphaProof rather than AlphaFold as AlphaGo's architectural descendant.
• R. Hill and C. Stein, "How Artificial Intelligence Shapes Science: Evidence from AlphaFold," Northwestern University and UC Berkeley, first version 11 March 2026, version of 17 July 2026, read directly from the authors' PDF (introduction and results summary; the appendices were not read). Source of every quantitative claim in section 2 — the unchanged rate of experimental determination, the survey findings, the 8% of SwissProt, the ~130,000 structures, the 23–32% and 80–300% basic-research effects, the 13% patent upper bound and the bottleneck conclusion — and of the Lupas quotation, which it reproduces from Callaway.
• C. Outeiral, D. A. Nissley and C. M. Deane, "Current structure predictors are not learning the physics of protein folding," Bioinformatics 38(7), 1881–1887, published 31 January 2022. Abstract read directly and quoted verbatim; the 0.613 / 0.591 / −0.23 / ~0.1 figures are from the same fetch.
• R. Yuan, J. Zhang, A. Kryshtafovych, R. D. Schaeffer, J. Zhou, Q. Cong and N. V. Grishin, "CASP16 protein monomer structure prediction assessment," bioRxiv 2 June 2025 (v2, 30 July 2025), subsequently in Proteins. Read via the PubMed Central deposit. Source of the 81.01 / 80.80 GDT_HA figures, the p = 0.92, the group counts and the "subtle" and remaining-challenges quotations.

Secondary, and used as such.

• Oxford Protein Informatics Group (blopig), "CASP14: what Google DeepMind's AlphaFold 2 really achieved, and what it means for protein folding, biology and bioinformatics," December 2020, read directly. This is an academic group blog, not a peer-reviewed assessment, and is cited as expert commentary. It is the entry's source for the independent CASP14 numbers — group 427's summed z-score of 244.0 against 90.8 for group 473 (BAKER), the 2.5 and 3.8 average z-scores, the 36%-under-2 Å and 86%-under-5 Å distribution, the 0.48 Å on T1046s1 and the assessor's "the details" line — and for the contemporaneous statement of the limitations (dynamics, complexes, orphan proteins, the easy-to-crystallise bias) that sections 3 and 4 grade. Preferred over the vendor's own figures wherever the two overlap.
• P. B. Moore, W. A. Hendrickson, R. Henderson and A. T. Brunger, "The protein-folding problem: Not yet solved," Science 375(6580), 507, 4 February 2022 (ADS bibcode 2022Sci...375..507M). Not read at first hand — science.org returned HTTP 403 and the Stanford mirror failed certificate verification. Cited for its existence, title, authorship and date; its text is deliberately not quoted. [verify].
• E. Callaway, "'It will change everything': DeepMind's AI makes gigantic leap in solving protein structures," Nature 588, 30 November 2020. Not read at first hand for the same nature.com reason; the Lupas quotation is taken from Hill and Stein's direct quotation of it, and the decade-stuck-protein detail from consistent secondary accounts. [verify] on the latter.
• M. Karelina, J. J. Noh and R. O. Dror, "How accurately can one predict drug binding modes using AlphaFold models?", eLife 12:RP89386, 2023. Read via the eLife listing and consistent secondary summaries rather than the paper end to end. [verify].
• T. C. Terwilliger et al., "AlphaFold predictions are valuable hypotheses and accelerate but do not replace experimental structure determination," Nature Methods, 2023. Title and thrust only; not read. Cited in claim 3 as corroboration, not as the basis of it — the basis is Hill and Stein.
• Nature news coverage of the AlphaFold3 code dispute (May and November 2024), read via search summaries: the ten-scientist letter to the editors, its verbatim "appropriate for an announcement on a company website" and "fails to meet the scientific community's standards" phrasing, the 1,000+ signature count as of 28 May 2024, DeepMind's six-month commitment and the November 2024 academic release. [verify] on the AlphaFold3 publication date, which secondary sources give as both 8 and 22 May 2024.
• EMBL-EBI and EMBL press material on the AlphaFold Protein Structure Database launch of 22 July 2021 and the July 2022 expansion, and the AFDB database papers in Nucleic Acids Research (2022, 2024, 2025) for the 214-million figure. [verify] on the initial release count, reported between roughly 350,000 and 365,000.
• NobelPrize.org, "The Nobel Prize in Chemistry 2024," announced 9 October 2024, for the division of the prize and the citation wording quoted in section 4.
• Forbes (13 May 2026), Tech.eu and Drug Discovery World on the $2.1 billion raise, and secondary reporting of Hassabis's January 2026 World Economic Forum restatement of the clinical timeline. The one-year slip from end-2025 to end-2026 is [verify].
• Reported accounts of the AlphaFold-enabled results named in claim 2: the cytoplasmic ring of the nuclear pore complex (Science, 2022) and the Pfs48/45 malaria vaccine candidate (Nature Communications, 2022, with the 2.13 Å dataset phased by molecular replacement from an AlphaFold2 model). Both read via abstracts and institutional summaries rather than in full. [verify].
• CASP founding facts — 1994, John Moult and Krzysztof Fidelis, biennial, blind, independent assessors — from the CASP project's own description and consistent secondary sources.
• Anfinsen's 1972 Nobel lecture and Levinthal's 1969 paradox are recorded here as the standard framing of the fifty-year claim, from general reference sources rather than the originals.

Attempted and failed, so that nothing above silently depends on it: nature.com redirects all article URLs to an identity provider, so neither the 2021 AlphaFold paper nor Callaway's news piece could be read there — the paper was read from its PubMed Central deposit and the news piece was not read at all; science.org returned HTTP 403 for the Moore et al. letter and the Stanford mirror of it failed TLS verification, which is why that letter is cited but not quoted; pnas.org returned HTTP 403, so "AlphaFold two years on: Validation and impact" is not used; tandfonline.com returned HTTP 403 for the BioTechniques piece asking the same question this entry asks; statnews.com's January 2025 report on CASP after AlphaFold is behind STAT+ and only its date (7 January 2025) and author (Brittany Trang) could be established; nobelprize.org returned HTTP 403 for Hassabis's Nobel lecture PDF, which is why the descends_from argument rests on DeepMind's own March 2026 blog instead; and one PubMed Central article served a reCAPTCHA rather than the CASP14 round assessment, which is why the CASP14 figures here come from DeepMind and from blopig rather than from the official assessment paper.

Claims circulating that this entry deliberately does not use: any figure for the cost or duration of determining a single protein structure experimentally, which is quoted everywhere as "years" and "$100,000" with no traceable source; any count of drugs "discovered with AlphaFold," a category with no agreed definition and no registry; and any of the several 2026 trade-press and SEO-generated pages on IsoDDE and Isomorphic's pipeline, which agree with each other closely enough to be suspicious and trace back to the company's own February 2026 report in every case where they trace anywhere.

The 2026 citation occasions — the Crouzeix's conjecture proofs and the digest's own quiet-places caveat — are as recorded in this project's digests/2026-08-15-12.md, which holds the primary links. They are named here as occasions to cite this entry, not as evidence for anything in it. Nothing in this file is evidence, nothing in it is deposited in the ledger, and nothing in it touches the needle.