AlphaFold 2 nears laboratory accuracy at CASP14

A protein is a chain of amino acids that folds into a particular three-dimensional shape, and the shape largely determines what the protein does. Reading the chain is cheap and routine. Determining the shape has meant X-ray crystallography, nuclear magnetic resonance or electron microscopy: months of laboratory work, and often considerable expense, for a single structure. Whether the shape could instead be computed from the sequence alone had been an open question for about fifty years.

CASP was built to answer it. Every two years the Critical Assessment of protein Structure Prediction gives entrants the amino-acid sequences of proteins whose structures have been determined in the laboratory but not yet published, collects the predictions, and grades them against the experimental answers afterward. The measure is the competition’s own and is applied blind to every entrant alike.

At the fourteenth round, organizers said on 30 November 2020 that DeepMind’s AlphaFold 2 had predicted around two-thirds of the target proteins to an accuracy comparable to the experimental methods themselves, working from the sequence alone.

The result was announced rather than published. The description of the system, and the predictions themselves, followed the next year, when DeepMind and EMBL’s European Bioinformatics Institute opened a public database of predicted structures covering the human proteome and twenty other organisms.

The remaining third mattered too. Accuracy was uneven across target types, and a predicted structure is a hypothesis rather than a measurement, which DeepMind had said plainly when it released predictions for coronavirus proteins earlier in the year. What changed at CASP14 was the default. For a large share of proteins, computation became a reasonable first step and the laboratory a way of checking it.