What Is AlphaFold and Why Does It Matter?

AlphaFold is an artificial intelligence system built by Google DeepMind. It predicts the three-dimensional shape of a protein from the sequence of amino acids it is made of. It matters because a protein's shape determines what it does. Working out a single shape in the lab could take months or years. AlphaFold can produce a prediction in minutes. Its predictions for more than 200 million proteins are free to use. Two of its creators shared half of the 2024 Nobel Prize in Chemistry.

What problem does AlphaFold solve?

Proteins are the working molecules of life. Each is a chain, generally built from 20 kinds of amino acids, that folds into a particular shape. That shape governs whether the protein acts as an enzyme, an antibody, a hormone or a building block of tissue. In 1961, the American scientist Christian Anfinsen concluded that the sequence alone determines the shape. Predicting one from the other became a central challenge of biochemistry, according to the Royal Swedish Academy of Sciences. The academy says it stayed unsolved for about 50 years.

The alternative was experiment, chiefly X-ray crystallography. The AlphaFold team wrote in 2021 that laboratories had determined the structures of roughly 100,000 unique proteins, against billions of known protein sequences.

How accurate is AlphaFold?

The field measures progress with a blind test called CASP, run every two years since 1994. In it, teams predict structures that have been solved in the lab but kept secret. DeepMind's first AlphaFold won in 2018 with an accuracy of almost 60%, the academy's account says. That was up from a previous best of about 40%, it says. A redesigned version, AlphaFold2, went much further in 2020.

In the Nature paper describing that system, the DeepMind team reported a median error of 0.96 angstroms for the protein backbone in the 2020 test. The next best method had a median error of 2.8 angstroms, the team reported. An angstrom is one ten-billionth of a meter. A carbon atom is about 1.4 angstroms wide. Each prediction carries a confidence estimate for every part of the structure. The estimate shows which regions to trust.

On October 9, 2024, the Royal Swedish Academy of Sciences awarded half of the chemistry prize for this work. That half went to Demis Hassabis and John Jumper of Google DeepMind. The other half went to David Baker of the University of Washington for designing entirely new proteins by computer.

What is AlphaFold used for?

Researchers use it to work out what a protein does, how it interacts with other molecules and what goes wrong in disease. The Nobel announcement cited studies of antibiotic resistance and of enzymes that can break down plastic.

DeepMind published the AlphaFold2 code. With the European Bioinformatics Institute (EMBL-EBI), it runs the AlphaFold Protein Structure Database. DeepMind says the database grew to more than 200 million structures in July 2022, covering nearly every catalogued protein. The database's home page cited more than 260 million predictions in October 2026. The data are free for academic and commercial use under a CC-BY-4.0 license. DeepMind said in November 2025 that more than 3 million researchers in over 190 countries were using AlphaFold.

A newer version goes beyond proteins. AlphaFold 3 was published in Nature in May 2024 by researchers at Google DeepMind and Isomorphic Labs. Isomorphic Labs is a drug design company also led by Hassabis. AlphaFold 3 predicts how proteins fit together with DNA, RNA, small molecules and ions. That is relevant to drug design, where the question is often how a small molecule attaches to a protein. The authors reported that it positioned small molecules on proteins more accurately than standard docking programs.

The model was first offered through AlphaFold Server, a free website for non-commercial research. The paper said code would not be provided. An addendum published in November 2024 said DeepMind had since released the underlying inference code. The company says it offers the code and model weights for academic use.

What are AlphaFold's limitations?

A predicted structure is a model, not a measurement, and it is not a medicine. Documented limits include:

• Still images. AlphaFold predicts static structures. The AlphaFold 3 authors say it does not capture how molecules move in solution.
• Mutations. EMBL-EBI's training material says AlphaFold2 is not sensitive to changes in a single amino acid, which limits its use for judging a specific mutation.
• Sparse data. The same guide says AlphaFold2 struggles with proteins that have few close relatives. It says the model is less accurate for highly variable molecules such as antibodies.
• Plausible-looking errors. The AlphaFold 3 authors say the model can produce orderly structure in regions that are actually disordered. On one benchmark, they say, it violated a molecule's handedness, or chirality, in 4.4% of cases.

The bottom line

AlphaFold made accurate structure prediction practical for a large share of known proteins. Researchers had pursued that goal for half a century. The results are freely available. Its value so far is as a research tool that guides experiments. Whether that leads to approved medicines depends on steps AlphaFold does not perform. These include designing a safe molecule and testing it in people.