AlphaFold2 achieves near-experimental accuracy at CASP14

Background

Working out the three-dimensional shape of a protein from its amino acid sequence had stood as one of biology's hardest open problems, and DeepMind describes it as a 50-year-old grand challenge ^[1]. Shape matters because what a protein does in a cell is bound up with how its chain folds up in space; a reliable computational route from sequence to structure would therefore touch almost every part of the life sciences.

What happened

On 30 November 2020, DeepMind announced that the organisers of CASP14 had recognised its latest AlphaFold system as a solution to that problem ^[1]. CASP is a community assessment in which participating groups submit blind predictions that are then judged against structures determined experimentally; the verdict therefore came from outside the company.

How accurate it was

The headline measure at CASP is the GDT score, which summarises how closely a predicted structure agrees with the experimental one. Higher is better, and a score of around 90 is informally treated as competitive with experimental methods ^[1]. Across all targets, AlphaFold reached a median of 92.4 GDT ^[1]. On the hardest targets, the free-modelling cases where no close template is available, it still managed a median of 87.0 ^[1]. Measured as distance, its predictions were off by about 1.6 ångströms on average, roughly the width of a single atom ^[1].

How it worked

The system was built on an attention-based neural network, a design in which the network learns to weigh how strongly different parts of its input relate to one another, rather than treating the sequence as a flat string. It was trained on about 170,000 protein structures from the Protein Data Bank, the public archive of experimentally determined structures, and the training run used roughly 16 TPUv3 chips over a few weeks ^[1].

Stated limitations

DeepMind was explicit that the result was not the end of the story. It noted that the method did not yet handle protein complexes, interactions with DNA, RNA or small molecules, or the precise positions of side chains ^[1]. These are exactly the situations in which many biologically interesting questions arise, so the announcement described a large step rather than a finished tool.

Why it mattered

The Nobel committee later cited AlphaFold2, the system presented in 2020, when it awarded Demis Hassabis and John Jumper half of the 2024 Nobel Prize in Chemistry ^[2]. For a result first announced as a benchmark score in a community assessment, that is an unusually direct route from a technical evaluation to the field's highest recognition.