What happened?
AlphaFold 2 is an AI system from DeepMind that predicts the 3D shape of a protein from its chain of building blocks, called amino acids DeepMind blog ↗. In the CASP14 blind assessment in 2020, it reached a median score of 92.4 out of 100 on a common accuracy scale, where around 90 is informally seen as competitive with lab methods DeepMind blog ↗.
Proteins are tiny machines inside every living thing, and their shape decides what they do. Finding a shape in the lab, using methods like X-ray crystallography, can take years of work and expensive equipment DeepMind blog ↗. CASP is a contest run every two years since 1994 where teams predict shapes that have been solved in the lab but not yet published, so nobody can peek at the answers DeepMind blog ↗.
The method was published in the peer reviewed journal Nature in July 2021, with open source code Nature paper ↗. DeepMind and the European Bioinformatics Institute (EMBL-EBI) then built a free database that grew to over 200 million predicted structures, covering almost every catalogued protein EMBL news ↗. In 2024, Demis Hassabis and John Jumper of Google DeepMind shared half of the Nobel Prize in Chemistry for this work Science News ↗.
What are the three pieces?
The model
A neural network that reads a protein sequence plus related sequences from other species and outputs a 3D structure Nature paper ↗.
The confidence score
Each part of a prediction comes with a score, called pLDDT, that estimates how reliable it is Nature paper ↗.
The database
A free public collection of over 200 million predictions, run with EMBL-EBI under an open licence EMBL news ↗.
A step that once took a research project can now often start from a free prediction, which lets scientists spend lab time on testing ideas instead of finding shapes.
Leapscope interpretation of the reported result.How did AI help?
Here the AI is the result itself. DeepMind researchers designed the system and trained it on about 170,000 known protein structures from the Protein Data Bank, plus large sequence databases DeepMind blog ↗. The model learned patterns linking sequence to shape, and independent CASP assessors scored its predictions against lab structures that were kept secret during the test DeepMind blog ↗.
Score from DeepMind's CASP14 summary DeepMind blog ↗; database size and usage from EMBL's July 2022 announcement EMBL news ↗.
It is not perfect. The paper reports that accuracy drops a lot when fewer than about 30 related sequences are available, and that it is weaker for parts of proteins that mainly touch other proteins Nature paper ↗. DeepMind also noted that how proteins form complexes and how they bind DNA, RNA or small molecules were still open problems in 2020 DeepMind blog ↗. All predictions in the database are computer predictions, not lab measured structures EMBL news ↗.
Which fields could this affect?
Some uses are happening today, while others are possible future value; the connections below are our assessment.
Structural biology
Researchers use predicted shapes as a starting point when lab structures do not exist. One CASP assessor said it helped his team solve a structure they had been stuck on for nearly a decade DeepMind blog ↗.
Explore scienceBasic biology research
The free database covers almost every organism with a sequenced genome EMBL news ↗. Scientists can look up a protein's likely shape in seconds instead of starting from nothing.
Drug discovery
Knowing a protein's shape can help in designing molecules that fit it. A predicted structure alone does not show that a drug is safe or works.
Explore healthcareEnzymes for industry and environment
DeepMind mentioned finding enzymes that break down industrial waste as a possible use DeepMind blog ↗. This report demonstrates no such enzyme ready for use.
What has been checked?
The evidence includes an independent blind competition (CASP14), a peer reviewed Nature paper with open code, and a large public database. Leapscope reviewed these sources; we did not repeat the experiments.
Shown so far
- Independent CASP14 assessors scored AlphaFold 2 at a median of 92.4, with 87.0 on the hardest targets DeepMind blog ↗.
- The Nature paper reports backbone errors of about 0.96 ångström versus 2.8 for the next best method Nature paper ↗.
- Over 500,000 researchers in more than 190 countries had accessed the database by July 2022 EMBL news ↗.
Still unknown
- How reliable predictions are for proteins with few known relatives, where accuracy is known to drop Nature paper ↗.
- How well a single static shape captures proteins that move or change shape inside cells.
- How many predictions have directly led to approved medicines, which the sources do not show.
Evidence status: External assessment. Stage: In use. Open code and a public database of predicted structures are widely used by researchers.
From predicted shape to real use
This is our suggested way to follow the story, not a promised timetable.
Can I use it today?
Yes. Anyone can browse the free AlphaFold Protein Structure Database, and researchers can download the open source code Nature paper ↗ EMBL news ↗. The predictions are research tools, not lab confirmed structures or medical advice.
A few things you might be wondering
Did AlphaFold 2 solve protein folding completely?
It solved much of the shape prediction problem for single proteins, according to CASP organisers DeepMind blog ↗. It does not fully explain how proteins move, and it is weaker in some cases, such as proteins with few known relatives Nature paper ↗.
Did the AI do this on its own?
No. DeepMind researchers designed and trained it on structures that scientists had measured in labs over decades DeepMind blog ↗. Independent assessors then judged the results DeepMind blog ↗.
Has it led to new medicines?
The sources we reviewed do not show an approved medicine that came from AlphaFold 2. A predicted shape is a starting point and does not establish a drug's safety or effectiveness.
Go straight to the sources
Checked Oct 8, 2026. The first source is the original announcement or research. Later sources add independent context; background pages do not validate the result on their own.
01DeepMind's announcement of its CASP14 results, with scores, training details and outside comments.
The peer reviewed paper describing the method, its accuracy, its confidence score and its limits.
EMBL's announcement expanding the free database to over 200 million predicted structures, with usage figures.
News report on the Nobel Prize awarded to Hassabis and Jumper for AlphaFold and to David Baker for protein design.