# Nature published AlphaFold3 without its code, and it took an open letter and six months to get it released

> When Nature published Google DeepMind's AlphaFold3 in May 2024 it withheld the underlying code, prompting an open letter from hundreds and then more than a thousand researchers; six months later, on 11 November 2024, DeepMind released the AlphaFold3 code on GitHub, with the model weights restricted to noncommercial academic use.

- Verification status: pending
- Case type: deployment
- Provider: Google DeepMind (AlphaFold team) / Springer Nature
- Client: The computational-biology research community and Nature's readers, tech (named)
- Sector: tech / GB / cross
- Canonical URL: https://theinternetninja.com/stories/deepmind-alphafold3-nature-no-code-open-letter-github-noncommercial-release-2024/
- Source: The Internet Ninja (theinternetninja.com), independent verified-proof platform

## Outcomes

| Metric | Before | After |
| --- | --- | --- |
| Signatures on the open letter protesting the missing code | AlphaFold2 (2021) shipped with its full underlying code | 650+ signatures by 14 May 2024, growing to 1,000+ |
| Time from publication-without-code to a public code release | Code withheld at the May 2024 Nature publication | Code released on GitHub (v3.0.0) on 11 November 2024, about six months later; model weights restricted to noncommercial academic use |

## Verification method

Both critical facts are quoted verbatim from independent science journalism fetched live this session via web.archive.org captures (Science, 15 May and 11 November 2024; Genetic Engineering News), cross-checked against Nature's own news desk (22 May 2024 and the November 2024 daily briefing), and anchored by two public primary artifacts captured this session: the open letter's own Zenodo deposit (DOI 10.5281/zenodo.11206103) and the google-deepmind/alphafold3 GitHub release (v3.0.0, 11 November 2024) with its model-weights noncommercial terms of use. Nature is itself a party to the transparency dispute, so Science and GEN are treated as the cleaner independent sources for the protest claim. No confirmation was sought from DeepMind or Nature; only the public record is used. Awaits the checker's audit.

## Full case file

**Verification status: CHECKING — the protest letter and the eventual GitHub release are both quoted verbatim from independent science press (Science, GEN) and cross-checked against Nature's own news desk. Nature is a party to the dispute, so the independent outlets carry the weight. No green badge is claimed, and the story awaits the checker's audit.**

## The problem
AlphaFold reset the terms of an entire field, and openness was part of why. When AlphaFold2 was published in 2021, "the full underlying code was made accessible to all researchers" ([source](https://www.nature.com/articles/d41586-024-01463-0)). AlphaFold3, published in Nature in May 2024, did not follow that path: it "comes with 'pseudocode' — a detailed description of what the code can do and how it works" rather than the code itself ([source](https://www.nature.com/articles/d41586-024-01463-0)). The model is a substantial advance, able to "predict not just the structures of protein complexes, but also when proteins interact with other kinds of molecule, including DNA and RNA" ([source](https://www.nature.com/articles/d41586-024-01463-0)) — which is exactly why the decision to withhold the code drew scrutiny rather than passing quietly.

## What was built
This case file is not about the model's accuracy; it is about the governance of how an AI result entered the scientific record. The published paper was accompanied only by a restricted web server and pseudocode, and researchers could initially use AlphaFold3 only through DeepMind's online portal, "which permitted just 10 (and now 20) requests per day with a restricted set of molecules" ([source](https://www.science.org/content/article/google-deepmind-releases-code-behind-its-most-advanced-protein-prediction-program)). That gap between a headline capability and a reproducible artifact is what the community organised against.

## The outcome
The response was an open letter, posted on 11 May 2024, that accused Nature of "undermining its own rules on code availability" ([source](https://www.science.org/content/article/limits-access-deepmind-s-new-protein-program-trigger-backlash)). The letter is a public primary artifact in its own right: its ten named authors deposited it on Zenodo as "AlphaFold3 Transparency and Reproducibility" (16 May 2024), where its own record states it collected endorsement signatures through a public form and that "This version has endosers as of 7:00 AM PST 5/14/2024" ([source](https://zenodo.org/records/11206103)). Science reported that the letter "had received more than <span class="kpi">650 signatures</span> as of 14 May" ([source](https://www.science.org/content/article/limits-access-deepmind-s-new-protein-program-trigger-backlash)), and Genetic Engineering News reported the protest "culminating in a protest letter signed by more than <span class="kpi">1,000 scientists</span>" ([source](https://www.genengnews.com/topics/artificial-intelligence/alphafold-3-angst-limited-accessibility-stirs-outcry-from-researchers/)). The signatories argued that "using the imprimatur of academic publications without the possibility of reproducing the results, far less building on them, subverts the enterprise" ([source](https://www.science.org/content/article/limits-access-deepmind-s-new-protein-program-trigger-backlash)).

DeepMind then reversed course under the pressure. Its research vice-president Pushmeet Kohli said the team was "working on releasing the AF3 model (incl weights) for academic use" within six months ([source](https://www.genengnews.com/topics/artificial-intelligence/alphafold-3-angst-limited-accessibility-stirs-outcry-from-researchers/)). It made good on that pledge on 11 November 2024: "The computational model itself was made public today on the code repository GitHub with a noncommercial license, while the 'weights' — numbers that help tune how an AI model works — are available to academics who complete a short application form" ([source](https://www.science.org/content/article/google-deepmind-releases-code-behind-its-most-advanced-protein-prediction-program)). Science framed the release plainly: "<span class="kpi">Six months</span> after backlash, AI company fulfills pledge to make AlphaFold3's full computer model available for noncommercial use" ([source](https://www.science.org/content/article/google-deepmind-releases-code-behind-its-most-advanced-protein-prediction-program)), and Nature's own daily briefing confirmed that "the code for AlphaFold3 is finally out in the open" ([source](https://www.nature.com/articles/d41586-024-03728-0)). The repository's own record fixes the date and terms directly: the `google-deepmind/alphafold3` GitHub project's first release, v3.0.0, is stamped "2024-11-11T10:49:52Z" ([source](https://api.github.com/repos/google-deepmind/alphafold3/releases)), and the release primary also sharpens the nuance the press compressed — the code carries an Apache-2.0 license, while the separate noncommercial restriction sits on the model parameters, whose terms of use state that "The AlphaFold 3 model parameters and output are **only** available for non-commercial use by, or on behalf of, non-commercial organizations" ([source](https://raw.githubusercontent.com/google-deepmind/alphafold3/main/WEIGHTS_TERMS_OF_USE.md)).

The weakest point in this record is worth naming where the reader meets it: Nature is not a neutral observer here, because it is the journal the open letter criticised, so its own reporting on the dispute is not fully disinterested ([source](https://www.nature.com/articles/d41586-024-01463-0)). The load-bearing sources for both the protest and the release are therefore the independent outlets — Science and Genetic Engineering News — with Nature's news desk used only as corroboration ([source](https://www.science.org/content/article/limits-access-deepmind-s-new-protein-program-trigger-backlash)).

## How this was verified
Method: every figure and quotation was retrieved live this session from the independent science press and, for the two critical facts, from the primary artifacts themselves, and cross-checked rather than recalled. Live science.org returns HTTP 403 to automated fetchers, so its two articles were read through their web.archive.org captures (dated 2024) and confirmed against the Nature news pages. Both critical facts are now anchored by a public primary in addition to the independent press: the open letter's own Zenodo deposit ("AlphaFold3 Transparency and Reproducibility", DOI 10.5281/zenodo.11206103, 16 May 2024) for the protest, and the `google-deepmind/alphafold3` GitHub repository's release record (v3.0.0, 11 November 2024) plus its model-weights terms of use for the release date and its noncommercial restriction. Each primary was captured to `sources/` and archived to web.archive.org this session. Date of verification: 2 September 2026. No party to the story was contacted, and no signatory or vendor was asked to confirm anything; this is an audit of the public record only.

## Related case files
- [AlphaFold's earlier CASP14 result, whose full code DeepMind did release, sharpens the contrast with AlphaFold3](/stories/deepmind-alphafold-casp14-92-4-gdt-independent-top-rank-2024-nobel-chemistry/)
- [GraphCast, another DeepMind model whose claims were adjudicated in a top journal rather than by the vendor](/stories/google-deepmind-graphcast-outperforms-ecmwf-hres-90pct-1380-targets-2023/)
- [A Nature Medicine study letting the peer-reviewed record, not a vendor, settle a competing AI claim](/stories/nature-medicine-2026-general-purpose-llms-outperform-specialized-clinical-ai-tools-medqa-healthbench-rcq/)

## Sources
1. Science (AAAS) · "Limits on access to DeepMind's new protein program trigger backlash" · 15 May 2024 · https://www.science.org/content/article/limits-access-deepmind-s-new-protein-program-trigger-backlash — **Tier 2** (independent science journalism)
2. Science (AAAS) · "Google DeepMind releases code behind its most advanced protein prediction program" · 11 November 2024 · https://www.science.org/content/article/google-deepmind-releases-code-behind-its-most-advanced-protein-prediction-program — **Tier 2** (independent science journalism)
3. Genetic Engineering News · "AlphaFold 3 Angst: Limited Accessibility Stirs Outcry from Researchers" · 2024 · https://www.genengnews.com/topics/artificial-intelligence/alphafold-3-angst-limited-accessibility-stirs-outcry-from-researchers/ — **Tier 2** (independent trade press)
4. Nature (news) · "AlphaFold3 — why did Nature publish it without its code?" · 22 May 2024 · https://www.nature.com/articles/d41586-024-01463-0 — **Tier 2** (Nature's news desk; note that Nature is itself a party to the dispute, so used as corroboration only)
5. Nature (news) · "Daily briefing: AlphaFold3 is now open source" · November 2024 · https://www.nature.com/articles/d41586-024-03728-0 — **Tier 2** (Nature's news desk, corroboration)
6. Wankowicz, Beltrao, Cravatt, Dunbrack, Gitter, Lindorff-Larsen, Ovchinnikov, Polizzi, Shoichet, Fraser · "AlphaFold3 Transparency and Reproducibility" (the open letter, deposited by its authors) · 16 May 2024 · https://zenodo.org/records/11206103 (DOI 10.5281/zenodo.11206103) — **Tier 1** (the letter's own public deposit; primary artifact for the protest's existence, authorship, date and endorsement snapshot)
7. Google DeepMind · `google-deepmind/alphafold3` GitHub repository — release v3.0.0 record and model-weights Terms of Use · 11 November 2024 · https://api.github.com/repos/google-deepmind/alphafold3/releases and https://raw.githubusercontent.com/google-deepmind/alphafold3/main/WEIGHTS_TERMS_OF_USE.md — **Tier 1** (the vendor's own release primary for the 11 Nov 2024 date, the code's Apache-2.0 license, and the noncommercial restriction on the model parameters)