AlphaFold 3
Era 5 · Generative AI, reasoning, and agents (2022–2026) › AI for science and recognition
- Lineage
- Acting in the world
- Kind
- Implementation
Who
Google DeepMind, Isomorphic Labs
What happened
Google DeepMind and Isomorphic Labs present AlphaFold 3, which predicts the structure of complexes of proteins with DNA, RNA, chemical modifications, ligands, and ions, with an improvement of at least 50 percent on interactions between proteins and other molecules.
Why it mattered
It predicts how proteins fit together with the cell's other molecules, not just their shape. It is the step from the structure of a protein to the structure of its interactions.
The full story
AlphaFold 2 predicted the shape of an isolated protein. But in the cell, proteins act bound to other molecules: to DNA and RNA, to other proteins, or to small molecules such as drugs. Understanding a disease or designing a medicine requires knowing how they fit together.
AlphaFold 3, developed by Google DeepMind with Isomorphic Labs, another Alphabet subsidiary, predicts those complexes: proteins with DNA, RNA, post-translational chemical modifications, ligands, and ions. It improves accuracy on interactions between proteins and other molecules by at least 50 percent over existing methods.
The architecture changed. AlphaFold 2's Evoformer gave way to the Pairformer, a simpler module inspired by the Transformer, and the initial predictions are refined by a diffusion model, the same family of techniques that generates images from text.
The paper was published in Nature on May 8, 2024, with 48 authors. Unlike AlphaFold 2, the code was not released at the time: the model was offered free for noncommercial research through the AlphaFold Server. The code was published on November 11, 2024, and the model weights are made available to researchers who request them, for noncommercial academic use only.
Builds on
AlphaFold 2 (November 2020)
Documented direct influence
AlphaFold 3 replaces AlphaFold 2's Evoformer with the Pairformer and adds a diffusion module; the Nobel honors Hassabis and Jumper for AlphaFold 2
Sources
Original source
- Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature, 630(8016), pp. 493-500 (opens in a new tab). Josh Abramson, Jonas Adler, Jack Dunger et al., 2024. open license: CC BY 4.0
- AlphaFold 3 repository (GitHub, Google DeepMind) (opens in a new tab). Google DeepMind, 2024
Further reading
How to cite
“AlphaFold 3.” The History of AI, revised November 13, 2026. https://thehistoryofai.com/en/milestone/alphafold-3
This entry was last reviewed on November 13, 2026.