AlphaFold enters the protein-folding contest

Predicting a protein's three-dimensional shape from its sequence of amino acids had been an open problem for fifty years, and since 1994 the community had measured itself every two years at CASP, a blind contest in which structures solved in the laboratory are withheld while computational groups predict them. At CASP13 in December 2018, a team entered for the first time and won by a distance: DeepMind's AlphaFold placed first on 25 of 43 hard targets, against 3 for the next best.

The method used a deep residual network, of the kind that had won ImageNet, to predict the distance between every pair of residues from the evolutionary record of related sequences, and then folded the chain by gradient descent to satisfy the predicted distances. It was an outsider's approach, with little of the physics that structural biologists had built their methods on.

The result was good enough to be startling and not good enough to be useful; most predictions were still far from experimental accuracy. Two years later the same team, having rebuilt the system around attention, returned to CASP14 and closed the gap. The 2018 entry is on this timeline as the moment the laboratories that had beaten humans at games turned to science.