DeepMind's AlphaFold 2 Achieves Highest-Accuracy Results at CASP14 Protein Structure Prediction Competition
In November–December 2020, DeepMind's AlphaFold 2 system achieved a median Global Distance Test score of approximately 92.4 across all CASP14 targets, far surpassing the next-best group, in a result that computational biologists described as largely solving the 50-year-old protein-folding problem for single-chain proteins.
Background
Every protein in a living cell is a chain of amino acids that folds into a precise three-dimensional shape. That shape determines what the protein does: whether it speeds up a chemical reaction, carries a signal across a membrane, or holds a structure together. Biologists had understood this since the 1950s, and for just as long they had known that predicting the folded shape from the amino acid sequence alone was extraordinarily hard. The number of possible arrangements a chain can take is so vast that random search is hopeless, and the physical forces involved interact in ways that resist simple calculation.
The main tool for measuring progress on this problem was CASP, the Critical Assessment of Protein Structure Prediction, a biennial competition run since 1994. Participating groups receive protein sequences whose experimentally determined structures are already known but not yet published, submit their predictions, and are scored against the real answer. The scoring unit used to rank groups is the Global Distance Test, or GDT, which measures how many predicted atom positions fall within a few ångströms of the true positions. A GDT score of 100 would mean a perfect match. For most of the competition’s history, the best groups scored somewhere in the 40s or 50s on hard targets. Progress was slow, and genuine cases of a model matching experimental accuracy were rare.
DeepMind had entered a system called AlphaFold at CASP13 in 2018, where it outperformed other groups by a clear margin. That version used deep learning to predict distances between pairs of amino acids, which then guided a separate search for a plausible structure. It was a strong result, but it still fell well short of what X-ray crystallography or cryo-electron microscopy could tell you. The team went back and rebuilt the approach from the ground up.
What happened
The system that competed at CASP14 was a different architecture, referred to as AlphaFold 2. It was built by a large team at DeepMind led by John Jumper, with Demis Hassabis as senior author alongside Pushmeet Kohli, Koray Kavukcuoglu, Oriol Vinyals and David Silver, among dozens of others. At its core was a new neural network module called the Evoformer, which processed two kinds of information together: a multiple sequence alignment (a comparison of the target sequence against related proteins from other species, showing which positions tend to vary and which stay fixed) and a representation of which pairs of residues were likely to be close in space. The Evoformer let these two representations update each other through many rounds, passing information back and forth rather than treating sequence and geometry as separate problems.
The results were announced at the CASP14 conference, held 30 November to 4 December 2020. Across all CASP14 targets, AlphaFold 2 achieved a median GDT score of approximately 92.4. The gap to the next-best group was large enough that the organisers and outside observers described it as unlike anything the competition had produced before. On the hardest category of targets, where other methods had historically struggled most, AlphaFold 2 still returned predictions close enough to the experimental structures to be genuinely useful for biological work. Several targets that AlphaFold 2 predicted were later confirmed by crystallography, matching the prediction at near-atomic resolution.
The Nature paper describing the method in full was published in July 2021, with Jumper and Richard Evans as joint first authors, giving the scientific community the technical detail needed to understand and build on what had been done.
Why it mattered
AlphaFold 2's performance at CASP14 demonstrated that a deep-learning system trained on known protein sequences and structures could predict three-dimensional protein conformations with accuracy approaching experimental methods such as X-ray crystallography, for a large fraction of targets. This effectively transformed structural biology: researchers could now obtain reliable structural models for proteins that had resisted experimental determination for decades, accelerating drug discovery, enzyme engineering, and fundamental biological research. The result prompted a broad reappraisal of what machine learning could achieve in the natural sciences beyond pattern recognition on human-generated data.
People
John Michael Jumper, Richard Evans, Alexander Pritzel, Tim Green, Michael Figurnov, Olaf Ronneberger, Kathryn Tunyasuvunakool, Russ Bates, Augustin Zidek, Anna Potapenko, Andrew Bridgland, Clemens Meyer, Simon Kohl, Andrew Ballard, Andrew Cowie, Bernardino Romera Paredes, Stanislav Nikolov, Rishub Jain, Jonas Adler, Trevor Back, Stig Petersen, David Reiman, Ellen Clancy, Michal Zielinski, Martin Steinegger, Michalina Pacholska, Tamas Berghammer, Sebastian Bodenstein, David Silver, Oriol Vinyals, Andrew Senior, Koray Kavukcuoglu, Pushmeet Kohli, Demis Hassabis
Organisations
Deepmind, Critical Assessment of protein Structure Prediction
Sources
- Highly accurate protein structure prediction with AlphaFold.Nature.Primary source
- 'It will change everything': DeepMind's AI makes gigantic leap in solving protein structures.Nature News.Secondary
- CASP14 Press Release.Critical Assessment of Protein Structure Prediction.Official
Cite this page
AI Achievements. (2020). DeepMind's AlphaFold 2 Achieves Highest-Accuracy Results at CASP14 Protein Structure Prediction Competition. Retrieved 2026-08-22, from https://achievements.ai/milestone/deepminds-alphafold-won-casp-protein-contest
@misc{achievements_deepminds_alphafold_won_casp_protein_contest,
title = {DeepMind's AlphaFold 2 Achieves Highest-Accuracy Results at CASP14 Protein Structure Prediction Competition},
author = {{AI Achievements}},
year = {2020},
url = {https://achievements.ai/milestone/deepminds-alphafold-won-casp-protein-contest}
}