Biology · 15 July 2021
Atomic-accuracy protein structures from sequence alone
AlphaFold predicted protein structures with atomic accuracy, even when no similar structure was known, and reached experimental quality in a majority of cases in the CASP14 assessment.
- Field
- Biology
- Date
- 15 July 2021
- AI system
- AlphaFold
- Institution
- Google DeepMind
- Published in
- Nature
- Identifier
- 10.1038/s41586-021-03819-2
The three-dimensional shape of a protein follows from its amino-acid sequence, but predicting that shape had been an open problem for fifty years. Experimental structure determination is slow and expensive; by 2021 it had covered only a small fraction of known protein sequences.
In a paper published in Nature on 15 July 2021, John Jumper, Demis Hassabis and colleagues at DeepMind described AlphaFold, a neural-network method they present as the first computational approach that can regularly predict protein structures with atomic accuracy even when no similar structure is known. In the 14th Critical Assessment of protein Structure Prediction (CASP14), AlphaFold structures had a median backbone accuracy of 0.96 Å r.m.s.d.95, compared with 2.8 Å for the next-best method. The all-atom accuracy was 1.5 Å r.m.s.d.95.
The structures are predictions, later released at scale for a large fraction of known proteins. They are in this record because the model produced scientifically usable three-dimensional structures — a solution, in the sense recognized by the CASP community, to a core problem in molecular biology.