Names, versions and owners change fast. Checked October 2026.
In plain words
AlphaFold is not a chatbot. It is a scientific tool built by Google DeepMind to solve one hard problem in biology: working out what shape a protein folds into.
Proteins do most of the work in living things: carrying oxygen, fighting infections, digesting food. Each one starts as a long chain of smaller pieces called amino acids, which then folds into a precise 3D shape, and that shape decides what the protein does. Finding the shape in a lab can take months or years. AlphaFold predicts it from the chain alone, often in minutes.
In 2020 its second version reached accuracy close to lab methods in a major scientific competition. Predicted shapes for over 200 million proteins were then shared free in a public database. In 2024 two of its creators, Demis Hassabis and John Jumper, shared the Nobel Prize in Chemistry for this work.
Everyday example
Guessing the shape from the string
Imagine a long string of beads of 20 colours, given to you straight. Drop it, and it always crumples into the same knot, because each colour pulls on the others in its own way. Biologists know the bead order of millions of proteins but not the knot. AlphaFold learned, from proteins whose knots were already known, to look at a bead order and predict the knot.
Where the example breaks: Real proteins are not fixed knots: many wiggle, change shape or only fold properly next to partners. AlphaFold predicts a likely shape and says how confident it is, but some predictions still need checking in the lab.
Latestas of October 2026
- Newest public version
- AlphaFold 3, announced 8 May 2024; code and weights for academic use since 11 November 2024
- AlphaFold Server
- Free for non-commercial research since May 2024
- Database
- Version 6, released 21 October 2025, holds about 241 million predicted structures
These details change often. The rest of this page explains what it is, which stays true.
How it actually works
AlphaFold learned the patterns linking a protein's sequence to its shape from known examples.
Learning from known shapes
It was trained on protein structures that scientists had measured in labs over decades, collected in the public Protein Data Bank. This is supervised learning: the known shape is the answer to learn.
Clues from evolution
It compares a protein's sequence with related proteins in other species. Pieces that change together across evolution are probably touching in the folded shape, a strong hint about the 3D layout.
Attention over pairs
A deep learning network repeatedly updates its guess of how every pair of amino acids sits relative to the others, using attention, the same mechanism behind modern language models.
A shape with a confidence score
The output is a 3D structure plus a confidence score for each part, so scientists know which regions to trust. AlphaFold 3 extended this to how proteins fit with DNA, RNA and small drug-like molecules.
- Made by
- Google DeepMind; AlphaFold 3 with Isomorphic Labs
- Breakthrough
- AlphaFold 2 at the CASP14 competition, November 2020, with a median error under 1 ångström
- Open database
- With EMBL-EBI from July 2021, expanded to over 200 million predicted structures in July 2022
- Nobel Prize
- Chemistry 2024: half to Demis Hassabis and John Jumper for protein structure prediction, half to David Baker
AlphaFold is a different system from AlphaGo. Both come from DeepMind, but AlphaGo plays a board game; AlphaFold predicts molecules.
Questions people ask
What is protein folding, in simple words?
A protein is made as a long chain of building blocks called amino acids. The chain then curls and folds into a precise 3D shape, like a ribbon folding into a knot. The shape decides what the protein can do, such as grab a virus or carry oxygen.
Is AlphaFold free to use?
Scientists can browse predicted structures for free in the AlphaFold Protein Structure Database, run with EMBL-EBI. AlphaFold Server lets researchers run AlphaFold 3 predictions free for non-commercial research, and the AlphaFold 3 code and weights are available for academic use.
Why did AlphaFold win a Nobel Prize?
Predicting a protein's shape from its sequence had been a major unsolved problem for about 50 years. AlphaFold 2 made those predictions accurate enough to be useful, at scale. The 2024 Nobel Prize in Chemistry recognised Demis Hassabis and John Jumper for it.
Does AlphaFold make new medicines?
Not by itself. It predicts shapes, which helps scientists understand diseases and design drugs faster. Turning that into a medicine still needs lab work and clinical trials. DeepMind's sister company Isomorphic Labs uses AlphaFold-style models for drug discovery.