Technology & Medicine › Internet, genomes & AI › Deep learning (modern AI)
AlphaFold protein structure prediction
November 2020 (CASP14)· Invention or discovery· Germs, drugs & genetics
Well documentedSolid records, physical evidence or scientific consensus.
DeepMind's AI system that predicts a protein's 3D shape from its chemical sequence, often as accurately as lab experiments.
See it on the Technology & Medicine map →How it works
Proteins are chains of amino acids that fold into precise 3D shapes, and shape determines what they do. Solving one structure by experiment could take a lab years. DeepMind's AlphaFold 2, trained on about 170,000 known structures, predicted shapes from the sequence alone; at the CASP14 blind contest in November 2020 it matched experimental accuracy for about two-thirds of targets. By 2022 DeepMind had released predictions for over 200 million proteins.
Why it mattered
It largely solved a 50-year-old grand challenge in biology, giving researchers free structures for nearly every known protein and speeding work on drugs, enzymes and vaccines.
Think of it like
Like reconstructing a building's 3D form from its parts list alone, after studying thousands of finished buildings.
How we know
Independent blind scoring at CASP14; DeepMind's 2021 Nature papers and open code; the 2024 Nobel Prize in Chemistry to Demis Hassabis and John Jumper, shared with David Baker.
Still unknown or debated
It predicts static shapes well but struggles with proteins that move or change shape, and how much it will speed up drug discovery is still unclear.
Connections (3)
▸X-rays (Röntgen) supplied, via crystallography, the training data of AlphaFold protein structure predictionWell documented
The connection
X-ray wavelengths are about the size of atoms, so beams passing through a crystal scatter into patterns that reveal how atoms are arranged (X-ray crystallography). Over decades scientists used it to solve most of the 170,000-plus protein structures in the public Protein Data Bank. AlphaFold learned to predict shapes by studying those structures.
When and how
Max von Laue and the Braggs, 1912–13; Kendrew and Perutz's first protein structures at Cambridge, 1958–60; DeepMind trained AlphaFold in London on Protein Data Bank entries before 2020.
How we know
DeepMind's papers state AlphaFold was trained on Protein Data Bank structures, the large majority determined by X-ray crystallography.
Debates and open questions
Because AlphaFold learned largely from crystals, it may favor rigid, crystal-friendly shapes and be less reliable for flexible proteins.
▸Deep learning (modern AI) led to AlphaFold protein structure predictionWell documented
The connection
AlphaFold 2 is a deep neural network that takes a protein's amino-acid sequence, plus related sequences from other species, and outputs the 3D position of every atom. It used attention, the mechanism behind transformers, to relate distant parts of the chain, and learned from known structures much as image models learn from labeled photos.
When and how
DeepMind in London; AlphaFold 1 won the CASP13 contest in 2018, and AlphaFold 2, led by John Jumper, dominated CASP14 in November 2020.
How we know
DeepMind's 2021 Nature paper and open code document the design; independent teams reproduced it (OpenFold) or built similar systems (RoseTTAFold).
Debates and open questions
Whether AlphaFold learned the real physics of folding or clever pattern matching is debated; it reveals little about how proteins fold step by step.
▸AlphaGo beats Lee Sedol was followed by the same lab's AlphaFold protein structure predictionWell documented
The connection
DeepMind's leaders said that after AlphaGo they wanted to turn the same approach, deep learning plus search and self-improvement, to science. Demis Hassabis started the protein-folding project around 2016, shortly after the Seoul match. AlphaFold reused the team's know-how in large-scale training, though its design differs: it does not play games against itself.
When and how
DeepMind in London; protein work began around 2016; first CASP entry in 2018; Hassabis and John Jumper shared the 2024 Nobel Prize in Chemistry.
How we know
Hassabis has described the path from games to proteins in interviews and in his 2024 Nobel lecture.
Debates and open questions
How much AlphaFold owed to AlphaGo's specific methods, as opposed to the lab's money, talent and computing power, is unclear; the designs are quite different.