MMOOZZGG.com

Glossary

What is AlphaFold?

In short

AlphaFold is a Google DeepMind neural network that predicts the 3D shape of a protein from its chain of amino acids, with accuracy close to experiments.

History of AI · AlphaFold

MMOOZZGG.com editorsChecked: 1 min read

For a child

Imagine a string with colored knots that folds itself into a shape, like origami. Inside us, tiny helpers called proteins do the work, and each one is folded in its own way. How it is folded decides what it does. Scientists used to spend months finding the shape of one such helper. Now a computer guesses it from the order of the knots very quickly, and doctors find it easier to look for new medicines.

In plain words

A protein is built from links, amino acids, like a chain. The finished chain folds into a complex 3D shape, and this shape decides what the protein does in the cell. Finding the shape by experiment is hard: one structure takes months or years, so about 100,000 proteins have been studied, while billions of chains are known. AlphaFold solves the task by calculation. The model looks at the sequence and at similar chains in other organisms and gives the coordinates of the atoms with a confidence score. A scientist gets a model of the protein without a long experiment and sees at once which parts can be trusted.

For business

For pharmaceuticals, biotech and agrochemistry, AlphaFold makes the early stages cheaper: teams look at the structure of the target protein before lab tests and run fewer of them. Google DeepMind and the European Bioinformatics Institute EMBL-EBI maintain the AlphaFold DB database, with more than 200 million predicted structures, and the data may be used in commercial projects too under the CC BY 4.0 license. The AlphaFold 2 code is open under Apache 2.0 and the model parameters under CC BY 4.0, so you can run it on your own servers and keep sequences inside the company. You need graphics cards and a specialist to set up the pipeline. A prediction remains a hypothesis, and an experiment confirms it.

For a specialist

Jumper and co-authors described AlphaFold 2 in the paper “Highly accurate protein structure prediction with AlphaFold” (Nature, online 15 July 2021); in the CASP14 competition the model in most cases reached accuracy comparable to experimental structures. The trunk of the network consists of Evoformer blocks with attention-based components: they exchange information between the multiple sequence alignment (MSA) representation and the representation of residue pairs. Then the structure module sets a rotation and a shift for each residue and builds the atom coordinates. Confidence is expressed by the pLDDT metric. A week later, a paper by Tunyasuvunakool and co-authors came out: the model was applied to almost the whole human proteome, 98.5% of proteins. The 2024 Nobel Prize in Chemistry recognized protein structure prediction; according to the press release, more than 2 million people from 190 countries had used AlphaFold2 by then.

Example

A lab is looking for an enzyme to process waste. Its staff download the predicted structures of candidates from AlphaFold DB and send to costly experimental testing only those whose active site looks right.

Related terms

Questions and answers

What does AlphaFold do? It predicts the 3D structure of a protein from its amino acid sequence and shows how confident the model is in each part. This helps you understand how the protein works and find a drug for it.
Can AlphaFold be used in commercial projects? The AlphaFold 2 code is released under the Apache 2.0 license, and the model parameters and the prediction database under CC BY 4.0, which allows commercial use with attribution. Check the terms of other versions separately before the project starts.

Sources

Where companies use it

Not sure whether your company needs this?

AI solutions map
Contact us