The clearest case of AI producing a scientific result is protein structure prediction. AlphaFold2, built by Demis Hassabis and John Jumper at Google DeepMind, predicts the three-dimensional shape of a protein from its amino acid sequence, and in 2024 the work shared the Nobel Prize in Chemistry. The predictions are computed estimates with built-in confidence scores, so they guide experiments but do not replace them.
Key takeaways
- AlphaFold2 turns an amino acid sequence into a predicted 3D structure, a problem researchers had pursued for about 50 years.
- At the 14th Critical Assessment of protein Structure Prediction (CASP14), its median backbone error was 0.96 angstroms, against 2.8 angstroms for the next best method.
- Per-residue confidence is reported as pLDDT, and scores below 50 usually mean a region cannot be trusted as a fixed shape.
- Predicted structures are computational predictions, and the EMBL-EBI database that hosts them says so.
- The 2024 Nobel Prize in Physics, shared with University of Toronto researcher Geoffrey Hinton, recognized the neural network methods behind this kind of work.
The problem AlphaFold addressed
A protein is a chain of amino acids that folds into a specific shape, and the shape largely decides what the protein does. Measuring shapes in a laboratory is slow. The AlphaFold paper in Nature, published July 15, 2021, notes that around 100,000 unique protein structures had been determined experimentally, a small fraction of the billions of known protein sequences, and that each one can take months to years of effort. Predicting the shape from the sequence alone had been an open research problem for more than 50 years.
The Royal Swedish Academy of Sciences announced the 2024 chemistry prize on October 9, 2024. One half went to David Baker of the University of Washington for computational protein design, and the other half jointly to Hassabis and Jumper for protein structure prediction. The Academy says AlphaFold2 was presented in 2020, has been used to predict the structure of virtually all 200 million proteins researchers have identified, and has been used by more than two million people in 190 countries.
How accurate was the breakthrough
The Nature paper describes CASP as a blind test for participating methods and the long-standing gold standard for judging structure prediction accuracy. In CASP14, it reports a median backbone accuracy of 0.96 angstroms (95 percent confidence interval 0.85 to 1.16) for AlphaFold, while the next best method scored 2.8 angstroms. For scale, the paper notes that a carbon atom is about 1.4 angstroms wide. The same paper reports that side chains were highly accurate when the backbone was.
The method is a neural network that works from multiple sequence alignments, meaning the target sequence lined up with related ones, and the authors describe it as building physical and biological knowledge about protein structure into the design of the algorithm. A later system, AlphaFold 3, described in Nature on May 8, 2024, extends prediction to complexes that include nucleic acids, small molecules, ions, and modified residues.
Reading a prediction without overtrusting it
Every predicted structure comes with a score for each residue, called pLDDT, on a scale from 0 to 100. The EMBL-EBI training material on AlphaFold gives the bands below.
| pLDDT score | Label | What EMBL-EBI says it means |
|---|---|---|
| Above 90 | Very high | Backbone and side chains are typically predicted with high accuracy |
| 70 to 90 | Confident | Backbone usually correct, with some side chains misplaced |
| 50 to 70 | Low | Lower confidence; the EMBL-EBI page gives no separate guidance for this band |
| Below 50 | Very low | Often a flexible or disordered region, or one with too little information |
A low score is informative in itself, since many real proteins contain floppy regions with no single fixed shape. The same EMBL-EBI explanation of pLDDT also describes the opposite risk. Some disordered regions fold only when they bind a partner, and for these AlphaFold2 tends to predict the folded state with high scores. The example given is the protein 4E-BP2, which receives a confident helical prediction although it adopts that shape only when bound. A high score therefore does not prove the shape exists in every cell condition.
What a predicted structure is used for
The AlphaFold Protein Structure Database, run by EMBL-EBI with Google DeepMind, holds over 200 million predictions. Its own description calls them predictions of what a protein might look like and presents them as a complement to costly experimental methods. The Nobel committee lists work on antibiotic resistance and enzymes that can break down plastic among the applications. A predicted shape can serve as a starting hypothesis for an experiment. The experiment still decides.
The wider toolkit in research
Protein folding is one example. The Royal Swedish Academy gave the 2024 Nobel Prize in Physics to John Hopfield and Geoffrey Hinton, of Princeton University and the University of Toronto, for foundational discoveries and inventions that enable machine learning with artificial neural networks. The Academy describes Hinton's contribution as a method that can find properties in data on its own, for tasks such as identifying specific elements in pictures. Any such tool in research needs the same checks as above: a test on data the model has not seen, a measure of confidence, and an experiment that can prove it wrong.
FAQ
Did AlphaFold solve the protein folding problem?
The prize citation is for protein structure prediction, meaning the final shape that a sequence adopts, and in the CASP14 test accuracy was competitive with experimental structures in a majority of cases. Regions with low confidence remain uncertain.
Can an AI-predicted structure be used in place of an experiment?
EMBL-EBI describes the database as predictions that complement experimental methods. A prediction can guide a project, but claims about a specific protein usually still need laboratory confirmation.
Who won the Nobel Prize for this work?
Demis Hassabis and John Jumper shared half of the 2024 chemistry prize for protein structure prediction. David Baker received the other half for computational protein design, a related but separate achievement.
Is there a Canadian link to this research?
The 2024 Nobel Prize in Physics went in part to Geoffrey Hinton of the University of Toronto for work that underpins machine learning with neural networks. The chemistry prize recipients are based in the United Kingdom and the United States.
Related reading
- Whether AI can replace scientists
- AI and data science work
- Why the problem domain matters in AI research
- Problems that slow AI research
More in AI Basics.
Last verified: October 2026
Sources
- The Nobel Prize in Chemistry 2024, press release, Royal Swedish Academy of Sciences, October 9, 2024
- The Nobel Prize in Physics 2024, press release, Royal Swedish Academy of Sciences, October 8, 2024
- Jumper J et al., Highly accurate protein structure prediction with AlphaFold, Nature, July 15, 2021
- Abramson J et al., Accurate structure prediction of biomolecular interactions with AlphaFold 3, Nature, May 8, 2024
- AlphaFold: pLDDT, understanding local confidence, EMBL-EBI training, 2026
- About AlphaFold Protein Structure Database, EMBL-EBI and Google DeepMind, 2026