AlphaFold is a family of protein-structure prediction models developed by Google DeepMind. AlphaFold 2 (Nature paper July 15, 2021; CASP14 result December 2020) predicts a protein's 3D structure from its amino-acid sequence; AlphaFold 3 (Nature paper May 8, 2024) extends prediction to complexes of proteins, nucleic acids, small molecules, and ions. The work was recognized with one half of the 2024 Nobel Prize in Chemistry. It is among the most cited examples of AI-for-science and, to date, the only AI system to underpin a Nobel Prize.
| Field | Value | |
|---|---|---|
| Developer | [[google-deepmind | Google DeepMind]] |
| AlphaFold 2 | July 15, 2021 (Nature paper); CASP14 breakthrough December 2020 | |
| AlphaFold 3 | May 8, 2024 (Nature paper) | |
| Recognition | 2024 Nobel Prize in Chemistry — Demis Hassabis and John Jumper (DeepMind); David Baker (UW, for protein design) | |
| Access | AlphaFold Protein Structure Database (free, 200M+ predictions); AlphaFold 3 via AlphaFold Server (rate-limited free + commercial tier via Isomorphic Labs) |
AlphaFold 2 (2021)
AlphaFold 2 predicts the 3D structure of a protein from its amino-acid sequence. At CASP14 in December 2020 it reached a median GDT-TS of 92.4, within experimental error of crystallography for most targets, a result described as the solution to the 50-year-old protein folding problem. Its architecture combines an Evoformer with a structure module, trained end-to-end and differentiably on Protein Data Bank (PDB) structures and multiple-sequence-alignment (MSA) inputs.
The model was released in July 2021, with code on GitHub under the Apache 2.0 license and public weights. The accompanying AlphaFold Protein Structure Database, launched with EMBL-EBI, now holds more than 200 million predicted structures covering nearly all catalogued proteins; the database is released under CC-BY-4.0. AlphaFold 2's public database is the single most downloaded scientific-data artifact attributable to an AI model.
AlphaFold 3 (2024)
AlphaFold 3 predicts the joint structure of complexes involving proteins, nucleic acids (DNA, RNA), small molecules (ligands), and ions. It replaced the Evoformer with a Pairformer and added a generative diffusion module for the final coordinate prediction. It was the first model to predict drug-protein binding structures at scale, relevant to pharmaceutical development.
AlphaFold 3 was not released fully open. Weights were initially withheld and inference was restricted to DeepMind's AlphaFold Server with usage limits, with commercial use routed through Isomorphic Labs. This departed from the AlphaFold 2 approach and drew criticism from the Nature Methods editorial board and open-science advocates. After that criticism, in May 2024 DeepMind committed to releasing inference code and weights within six months; the code was released under a non-commercial license in November 2024. The progression from AlphaFold 2's full open release to AlphaFold 3's restricted access is cited in Open-Source AI / Open-Weight Models debates as a non-LLM case of frontier capability moving toward closed access.
Downstream: IsoDDE (2026)
In early 2026, Isomorphic Labs announced IsoDDE, a proprietary closed-weight drug-design engine built on top of AlphaFold 3. Isomorphic's technical paper claims IsoDDE "more than doubles the accuracy of AlphaFold 3" on its drug-design tasks. (Source: zenodo.org) President Max Jaderberg said at WIRED Health on April 16, 2026 that the company is "gearing up to go into the clinic" with AI-designed molecules across oncology and immunology, which would be the first clinical translation of an AlphaFold-lineage system. (Source: wired.com)
Nobel Prize (2024)
The 2024 Nobel Prize in Chemistry was split, with one half for protein design (David Baker, UW) and one half for protein-structure prediction (Demis Hassabis and John Jumper, DeepMind). It was the first Nobel Prize in which AI methods were the primary recognized contribution, and it coincided with the 2024 Nobel Prize in Physics to Geoffrey Hinton and John Hopfield for neural-network foundations. The award was cited extensively in 2024–2025 policy discourse as evidence that AI is producing Nobel-grade scientific contributions.
Policy and safety relevance
In policy discourse focused on LLM risk, AlphaFold is frequently invoked as a benefit case, including in frameworks such as the Bletchley declaration. It is also cited in AI-in-healthcare and drug-discovery policy discussions, including references to AI in biology in EO 14110.
AlphaFold 3's multi-molecule prediction (protein–DNA, protein–ligand, protein–protein) sits closer to the bio-capability threshold around which bioweapon-uplift evaluations are framed, including those in Anthropic's RSP, the OpenAI Preparedness Framework, and Biological Weapons Convention discussions referenced in ICRC positions. AlphaFold itself has not been demonstrated to provide meaningful bioweapon uplift beyond public PDB data, but it serves as the reference point against which uplift claims in LLMs are compared.
Related models and pages
AlphaFold is DeepMind's primary research showcase following AlphaGo, with Demis Hassabis as the Nobel-linked principal alongside John Jumper. Connected pages include AI for Science, Open-Source AI, John Jumper, Isomorphic Labs, and IsoDDE.
Relationships
- instance-of: General-Purpose AI (GPAI) (debated — AlphaFold is narrow-domain)
- related: Google DeepMind, Demis Hassabis, John Jumper, Isomorphic Labs, IsoDDE, AI for Science, Open-Source AI / Open-Weight Models, Anthropic's Responsible Scaling Policy (Version 3.1), OpenAI Preparedness Framework V.2, The Bletchley Declaration (AI Safety Summit, 1–2 November 2023), ICRC Position on Autonomous Weapon Systems (source summary)