AlphaFold (first version)

What it is

AlphaFold is a protein-structure prediction system built by DeepMind. The version described here is the first one, which the lab entered in the 13th Critical Assessment of protein Structure Prediction, or CASP13, in 2018.^[1][4] The task such a system takes on is to start from a protein's amino-acid sequence and produce a three-dimensional structure. CASP is a community-wide blind test: participating groups submit predictions for proteins whose experimental structures are not yet known, and the predictions are then scored against the experimental results. A system's standing in CASP is therefore treated as evidence of how well it actually predicts structures, rather than how well it fits answers it has already seen.

How it worked

The first AlphaFold did not rely on a contact map alone. A contact map is the field's traditional representation: it records which pairs of residues sit close to each other in space. The system instead estimated a probability distribution of the distances between residues.^[1][3] In effect it turned a contact map into a distance map — rather than stating only that two residues are near each other, it predicted how far apart they are likely to be, with a spread of probabilities around that estimate. Distance information of this kind constrains the geometry of the chain more tightly than a yes-or-no contact, and a structure can then be built around those constraints.

Results at CASP13

At CASP13 the system placed first overall in December 2018.^[1][2][3] Among the 43 hardest targets in the assessment, it gave the best prediction for 25.^[1] On those targets its median GDT score was 58.9, while the next two teams reached 52.5 and 52.4.^[1] GDT measures how closely a predicted structure matches the experimentally determined one, so a higher number means a closer match. Looking back in 2020, DeepMind described the entry as having achieved the highest accuracy among the participants at CASP13.^[4]

Code, limits, and what came next

The code used for the CASP13 entry was released in January 2020, but it could not be used for arbitrary sequences.^[1] For people outside the lab who wanted to apply the method to a protein of their own choosing, that was a real limitation.

The first AlphaFold was superseded by a redesigned system, AlphaFold2, which DeepMind entered at CASP14 in 2020.^[1][4] The CASP14 assessment ran from 30 November to 4 December 2020.^[4] In its announcement of AlphaFold2, DeepMind noted that a median score of around 90 GDT is informally considered competitive with experimental methods.^[4] The first version's median of 58.9 on the hardest targets gives a reference point for reading that informal benchmark.