BIGGER PICTURE Scientific discovery
A protein’s shape became something any lab could look up
A genetic sequence lists the ingredients; a three-dimensional structure helps explain what the protein can do. AlphaFold made predicted structures available at a scale that changed the starting point of biological research.
Source illustration- Open database milestone
- >200 million predictions in 2022
- Access
- Free; CC BY 4.0 structures
- What the colours show
- Local prediction confidence
How the idea works
Compare related sequences
Evolutionary patterns suggest spatial relationships.
Predict a folded structure
A neural network refines a three-dimensional hypothesis.
Guide experiments with confidence
Researchers combine the model and its uncertainty with real measurements.
From a sequence to a spatial hypothesis
Proteins are chains of amino acids, but their folded arrangement matters to how they work. AlphaFold2 compares related sequences across organisms, using evolutionary patterns to infer which parts may lie near one another. Its neural network refines a three-dimensional structural hypothesis. This gives a researcher more than a list of letters to begin investigating.
Sources: [2]
A shared resource instead of a private bottleneck
In July 2022, DeepMind and EMBL-EBI expanded the public database to more than 200 million predicted structures. Scientists could inspect and download them without first building their own prediction infrastructure. The release covered proteins from plants, bacteria, animals and other sequenced organisms, making a broad structural starting point available beyond well-funded specialist laboratories.
Sources: [1]
Help assemble a molecular gateway
EMBL-EBI describes how predicted structures can complement experimental images and measurements. One example is the human nuclear pore complex, the enormous protein assembly lining gateways into the cell nucleus. Work incorporating AlphaFold helped resolve much of that structure. The benefit is better-informed experiments: scientists can ask how pieces fit, where interactions occur and which hypotheses deserve testing.
Sources: [4]
Read the uncertainty as carefully as the shape
A prediction is not an experimentally established structure. Confidence varies along the same protein; a poorly predicted region may be flexible or simply insufficiently understood by the model. AlphaFold supplies local confidence scores so users can identify those boundaries. The 200 million figure counts predictions, not 200 million verified functions, medicines or solved biological problems.
Sources: [3]
The original sources
- EMBL-EBIAlphaFold predicts structures across the protein universe
- EMBL-EBI TrainingAlphaFold2: A high-level overview
- EMBL-EBI TrainingpLDDT: Understanding local confidence
- EMBL-EBI TrainingHow is AlphaFold2 used by scientists?