AlphaFold 2 solves protein structure prediction

On 30 November 2020 the organisers of CASP14 announced that DeepMind's second AlphaFold had predicted the structures of the competition's proteins with a median accuracy comparable to experiment, and John Moult, who had run the contest since 1994, said the problem was in some sense solved. Two-thirds of its predictions were within the error of the laboratory methods. The previous best, AlphaFold's own 2018 entry, had been nowhere near.

The system had been rebuilt around attention. A transformer-like module reasoned jointly over the sequence, its evolutionary relatives and a matrix of pairwise relationships, passing information between them repeatedly, and a structure module then produced three-dimensional coordinates directly, refining them through the same network several times. The paper appeared in Nature in July 2021 with the code, and by 2022 DeepMind had released predicted structures for essentially every known protein, some 200 million.

AlphaFold 2 is the strongest case that the methods on this timeline advance science rather than merely imitate it. John Jumper and Demis Hassabis received half the 2024 Nobel Prize in Chemistry for it, three years after the paper, the fastest award in the prize's recent history.