AlphaFold2 structure-prediction research appears in Nature

Background

Proteins are chains of amino acids that fold into specific three-dimensional shapes, and the shape largely determines what a protein does inside a cell. For decades, biologists found these shapes mainly through laboratory methods such as X-ray crystallography and cryo-electron microscopy. These methods are slow and costly, so determining a structure for every protein of interest was impractical. Predicting a protein's shape from its amino-acid sequence alone, known as protein structure prediction, was a long-standing problem in biology.

Progress is judged against structures that have been determined experimentally. The Critical Assessment of Structure Prediction (CASP) is a blind competition in which teams predict structures that the organisers already know but have not yet published, so that the entries can be compared fairly.

What happened

On 15 July 2021 the journal Nature published “Highly accurate protein structure prediction with AlphaFold,” a full description of the AlphaFold2 system from DeepMind.^[1] The paper had been received on 11 May 2021 and accepted on 12 July 2021.^[1] Its first author was John Jumper, and the corresponding authors were John Jumper and Demis Hassabis of DeepMind in London.^[1]

About a week after the paper appeared, DeepMind and EMBL-EBI launched the AlphaFold Protein Structure Database, which made predicted structures freely available to anyone.^[2]

Key details

The headline result came from the CASP14 assessment. AlphaFold's median backbone accuracy was 0.96 Å, a measure of how far the predicted backbone sits from the experimental structure. The next-best method reached 2.8 Å. Å (angstrom) is a very small unit of length used for distances between atoms, so lower numbers mean closer agreement.^[1] Measured across all atoms, AlphaFold's accuracy was 1.5 Å, against 3.5 Å for the best alternative.^[1]

DeepMind released the AlphaFold source code, the trained weights and the inference script, which is the program that runs predictions, under an open-source license on GitHub.^[2]

The database started with about 20,000 proteins from the human proteome, the full set of proteins one organism encodes, plus the proteomes of 20 other organisms. It held over 350,000 structures in total.^[2] DeepMind said it planned to expand the database to nearly all sequenced proteins, more than 100 million structures.^[2]

Impact

The clearest later assessment of the work came from the Nobel committee. In October 2024 Demis Hassabis and John Jumper shared half of the Nobel Prize in Chemistry “for protein structure prediction.”^[3] The committee said AlphaFold2, presented in 2020, had been used to predict the structure of virtually all of the 200 million proteins that researchers have identified, and that more than two million people from 190 countries had used it.^[3]

Protein structures help scientists understand how cells work, how diseases develop and how a drug molecule might bind to its target. Before predictions of this kind existed, a researcher studying an uncharacterised protein could spend a long time on experiments to determine a single structure. Because the code and the database were openly available, researchers in many fields could use the predictions without building the system themselves. A predicted structure is still a model's estimate, and many uses still require experimental confirmation.