AlphaFold

AlphaFold is an artificial intelligence program made by DeepMind, part of Alphabet. It uses special learning ways called deep learning to figure out the shapes of proteins, tiny parts in our bodies.

In 2018, the first version, AlphaFold 1, did very well in a big competition called the 13th Critical Assessment of Structure Prediction. It was good at guessing hard shapes when there were no similar examples before.

In 2020, AlphaFold 2 did even better in another competition. It guessed protein shapes so well that two out of three guesses were almost perfect.

In May 2024, AlphaFold 3 was announced. This new version can guess not just protein shapes, but also how proteins work with DNA, RNA, and other tiny parts in our bodies.

Because of this great work, the leaders of AlphaFold, Demis Hassabis and John Jumper, along with David Baker, received important awards, including the 2024 Nobel Prize in Chemistry, for helping scientists learn more about proteins.

Background

See also: Protein structure prediction and De novo protein structure prediction

Proteins are made of chains of amino acids that fold into special 3-D shapes. These shapes help us learn what proteins do in our bodies.

Finding these shapes usually needs special experiments like X-ray crystallography, cryo-electron microscopy and nuclear magnetic resonance (NMR). These experiments take a lot of time and money. Over many years, scientists found shapes for many proteins, but there are still millions more to discover.

Because of this, scientists tried to use computers to guess protein shapes from their building blocks, the amino acids. A competition called CASP, started in 1994, tests how well these guesses work. By 2016, the best guesses were not perfect. Then, AlphaFold used artificial intelligence and deep learning to make much better guesses in 2018.

Algorithm

DeepMind trained this program on many protein structures from the Protein Data Bank. The program uses a kind of deep learning called an attention network. This helps the AI solve big problems by breaking them into smaller parts.

AlphaFold 1, released in 2018, was the first version and did well in a competition for predicting protein shapes. AlphaFold 2, released in 2020, changed a lot. It uses connected parts that work together to learn patterns and get better at its guesses. By May 2024, AlphaFold 3 was announced. This new version can predict not just proteins, but also how proteins connect with DNA, RNA, and other tiny parts inside cells. It uses a method called the "Pairformer" to make its predictions more accurate.

Competitions

CASP13

In December 2018, DeepMind's AlphaFold placed first in a big contest called the 13th Critical Assessment of Techniques for Protein Structure Prediction (CASP).

The program did very well at guessing the shapes of hard-to-predict proteins, especially when there were no similar shapes already known. AlphaFold gave the best guess for many of these hard proteins.

CASP14

In November 2020, DeepMind's new version, AlphaFold 2, won CASP14. It gave the best guess for most of the proteins.

AlphaFold 2 was extremely accurate, with many of its guesses being almost perfect. It was as good as some lab experiments at finding the right shapes.

CASP15

In 2022, DeepMind did not join CASP15, but many teams used AlphaFold or tools that included AlphaFold to help with their guesses.

Reception

AlphaFold 2 did very well in a test called the global distance test. It scored more than 90, which is a big success in computational biology. Venki Ramakrishnan, a Nobel Prize winner and structural biologist, called the result "a stunning advance on the protein folding problem." He said it happened much sooner than many expected and will change biological research in many important ways.

AlphaFold 2 got a lot of attention from news outlets like Nature, Science, MIT Technology Review, and New Scientist. People talked about how this technology can help us understand diseases and create new medicines by predicting how proteins are shaped. In 2023, Demis Hassabis and John Jumper won important awards for their work on the AlphaFold project, and in 2024, they won the Nobel Prize in Chemistry together with David Baker of the University of Washington.

Source code

DeepMind shared the source code for AlphaFold in 2022 after scientists asked for it. The code for AlphaFold 3 was shared in November 2024 and became publicly available in February 2025, but only for non-commercial use.

Many copies of AlphaFold have been made by different groups, often with flexible sharing terms. Examples include Protenix by ByteDance, OpenFold-3 by the AlQuraishi Laboratory, and Boltz-1/2. Tools like ColabFold help make AlphaFold work faster on platforms such as Google Colab.

Database of protein models generated by AlphaFold

The AlphaFold Protein Structure Database (AFDB), a joint project between AlphaFold and EMBL-EBI, was launched on July 22, 2021. It contains AlphaFold-predicted models for many proteins from humans and other organisms. The database does not include very small or very large proteins.

In July 2021, UniProt-KB and InterPro was updated to show AlphaFold predictions when available.

In July 2022, the team added structures for around 200 million proteins from many species. As of 2024, the database has 214 million structures.

As of 2025, the AFDB uses AlphaFold 2 for its predictions. Foldseek helps search these structures quickly.

Derived databases

AlphaFill adds parts to AlphaFold models when needed.

TmAlphaFold places AlphaFold models in biological membranes.

AFTM uses AlphaFold models to find areas in human proteins that cross membranes.

ChannelsDB 2.0 uses models to study how molecules move to reach enzymes or cross membranes.

AlphaSync keeps the AFDB updated with changes from UniProt.

The Encyclopedia of Domains (TED) studies domains in AFDB structures.

The Evolutionary Classification of Protein Domains database (ECOD) classifies proteins in AFDB.

Unrelated AlphaFold-based databases

isoform.io is a database of AlphaFold2-generated structures of proposed splice isoforms in the human genome.

Performance, validations and limitations

AlphaFold has some limits in what it can do.

AlphaFold DB gives models of single protein chains, not their more complex forms. Many parts of proteins are predicted with less confidence, especially parts that don’t have a fixed shape.

AlphaFold 3 can predict how small molecules bind to proteins better than some older methods. It can also predict structures of protein complexes, but many human proteins are missing important sugar attachments.

AlphaFold sometimes has trouble with proteins it hasn’t seen before, especially those used in drug discovery. It may also rely too much on what it learned during training instead of understanding chemistry properly.

The model sometimes creates shapes that aren’t possible in real life because it doesn’t keep the protein chain intact. It works best with proteins that are similar to those it was trained on and may not do as well with completely new or synthetic proteins.

AlphaFold also has a hard time showing the different shapes a protein can take, which is important for understanding how proteins work.

Applications

See also: Earth BioGenome Project

AlphaFold has helped scientists learn about the shapes of proteins from the SARS-CoV-2 virus, which causes COVID-19. In early 2020, these protein shapes were hard to study in labs, so researchers used AlphaFold to predict them. The predictions were checked by experts at the Francis Crick Institute in the United Kingdom before sharing them with scientists around the world. The team compared their predictions with real experiments, especially for the spike protein, which they found in the Protein Data Bank. Learning about these shapes helps everyone understand the virus better. For example, AlphaFold 2 predicted the shape of the ORF3a protein, which matches what scientists at the University of California, Berkeley found using special microscopes.

Published works

Here are some important papers and presentations about AlphaFold:

• Andrew W. Senior et al. (December 2019), "Protein structure prediction using multiple deep neural networks in the 13th Critical Assessment of Protein Structure Prediction (CASP13)", Proteins: Structure, Function, Bioinformatics 87(12) 1141–1148 doi:10.1002/prot.25834
• Andrew W. Senior et al. (15 January 2020), "Improved protein structure prediction using potentials from deep learning", Nature 577 706–710 doi:10.1038/s41586-019-1923-7
• John Jumper et al. (December 2020), "High Accuracy Protein Structure Prediction Using Deep Learning", in Fourteenth Critical Assessment of Techniques for Protein Structure Prediction (Abstract Book), pp. 22–24
• John Jumper et al. (December 2020), "AlphaFold 2". Presentation given at CASP 14.
• Abramson, J., Adler, J., Dunger, J. et al. (May 2024), "Accurate structure prediction of biomolecular interactions with AlphaFold 3", Nature 630, 493–500 (2024)