Atlas record: https://www.aitimeline.dev/en/events/evt-0239/
Benchmark · November 2020
AlphaFold 2 at CASP14
The system predicted the three-dimensional structure of a protein from its amino acid sequence at accuracy comparable to experiment.
Why it matters
A problem biology had worked at for fifty years was solved, the first time AI produced a result another science needed rather than one for itself.
The assessment was blind, on structures the experimentalists had not yet published, so fitting was ruled out. Median accuracy passed the threshold at which a structure counts as usable for most biological purposes. DeepMind later released predictions for two hundred million proteins. It is the main argument against the view that large models are useful only for text and pictures.
Event record
- Event date
- November 2020
- Timeline date
- Event date
- Verification
- Sources gathered automatically · September 17, 2026
- Lines
- ID
- evt-0239
The result was announced in November 2020 from CASP14; the Nature paper appeared in July 2021.
Related events
Earlier
- Related A network finds a new antibiotic 2020links to this record
In the same year, 2020, deep learning gave biology two results checked in the laboratory: protein structures and a new antibiotic. The record does not claim a connection between the two.
- Builds on Transformer 2017
The attention mechanism is applied to pairs of amino acid residues.
- Related Residual connections (ResNet) 2015
The architecture rests on deep networks with residual connections and attention.
Later
- Related AlphaDev 2023links to this record
A second problem outside AI where a model produced a result another field needed.
- Related GNoME and new materials 2023links to this record
The same move: a model predicts structures that experiment then checks.
- Extends AlphaFold 3 2024links to this record
The third generation of the same system, now not only for proteins.
- Builds on The Nobel Prize in Chemistry for protein structure prediction 2024links to this record
Half the prize was awarded for protein structure prediction by this system.