AlphaFold 2 solves protein structure prediction at CASP14

Results from the fourteenth Critical Assessment of Structure Prediction (CASP14) showed that DeepMind’s AlphaFold 2 had predicted the three-dimensional shape of proteins with a median accuracy comparable to laboratory methods. The organisers described the protein-folding problem as, in large part, solved.

CASP is a blind competition: entrants receive amino-acid sequences for proteins whose structures have been determined experimentally but not yet published, and submit predictions. AlphaFold 2 scored a median Global Distance Test score of 92.4 across all targets, against roughly 87 for the hardest category. A score around 90 is generally taken as competitive with experimental determination. The gap to the next-best entrant was large enough that CASP’s organisers said the assessment methods themselves were no longer able to discriminate at the top.

The system combined a transformer-based architecture operating over evolutionary sequence alignments with a structure module that reasoned directly about geometry, trained end to end. The full method was published in Nature in July 2021 alongside open-source code, and DeepMind subsequently released a database of predicted structures covering most catalogued proteins.

The result mattered beyond biology. It was an early, legible demonstration that deep learning could close a scientific problem rather than merely assist with one, and it was cited repeatedly afterwards by both proponents and critics of AI investment as the clearest example of the field delivering something of unambiguous value. Hassabis and Jumper shared the 2024 Nobel Prize in Chemistry for the work.