AlphaFold 2
Era 4 · Deep learning and foundation models (2012–2021) › AI for science
- Lineage
- Acting in the world
- Kind
- Application
Who
DeepMind
What happened
AlphaFold 2 wins CASP14 with the best prediction for 88 of the 97 targets and a median GDT score of 92.4 out of 100, an accuracy that in many cases approaches that of experimental methods.
Why it mattered
An end-to-end model built on attention predicts the shape of proteins with an accuracy that transforms structural biology.
The full story
AlphaFold 1 depended on a two-stage process: the network predicted distances, and another program then searched for a structure consistent with them. For CASP14 the team replaced all of that with a system of connected subnetworks forming a single differentiable model, trained end to end to produce the structure directly.
The model uses attention modules that refine, over several passes, the information about each pair of amino acids and about related sequences of the same protein in other species. The network learns at the same time which amino acids are close to each other and how they are arranged in space.
In November 2020 AlphaFold 2 gave the best prediction for 88 of the 97 CASP14 targets, with a median GDT score of 92.4 out of 100. The structural biologist and Nobel laureate Venki Ramakrishnan called it "a stunning advance on the protein folding problem."
It did not solve the problem entirely. It does worse on disordered regions, which have no fixed shape, and it struggles with proteins that take on several shapes or form complexes. But within four years its creators would receive the Nobel Prize in Chemistry: in 2024 Demis Hassabis and John Jumper shared half the prize for protein structure prediction, and the other half went to David Baker for computational protein design.
Videos
- AlphaFold: The making of a scientific breakthrough (opens in a new tab) (Google DeepMind)
Builds on
The Transformer (June 12, 2017)
Inherited, with no proven direct link
AlphaFold 2's Evoformer blocks are built on attention and its structure module uses an "equivariant transformer"; the paper does not present its design as a direct adaptation of the Transformer
AlphaFold 1 (December 2018)
Inherited, with no proven direct link
The same DeepMind team; the AlphaFold 2 paper describes it as a completely different model from the CASP13 system
Leads to
AlphaFold 3 (May 8, 2024), Nobel Prize in Chemistry for Hassabis, Jumper, and Baker (October 9, 2024)
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
RoseTTAFold (July 15, 2021)
Independent convergence
Baker's lab reached, on its own, performance close to what DeepMind had shown at CASP14, before the method was published; the two papers came out on the same day
Threads that pass through here
- From protein to Nobel (step 3 of 4)
Sources
Original source
- Highly accurate protein structure prediction with AlphaFold. Nature, 596(7873), pp. 583-589 (opens in a new tab). John Jumper, Richard Evans, Alexander Pritzel et al., 2021. open license: CC BY 4.0
Scholarship and context
- The Nobel Prize in Chemistry 2024 (nobelprize.org) (opens in a new tab). Nobel Foundation, 2024
Further reading
How to cite
“AlphaFold 2.” The History of AI, revised November 13, 2026. https://thehistoryofai.com/en/milestone/alphafold-2
This entry was last reviewed on November 13, 2026.