AlphaFold Nobel Prize: AI meets the Nobel in Chemistry

The announcement and the divided citation

On 9 October 2024, the Royal Swedish Academy of Sciences awarded the Nobel Prize in Chemistry for work on proteins that had been carried out largely in code. The award was split unevenly. One half went to David Baker "for computational protein design," and the other half was shared by Demis Hassabis and John Jumper "for protein structure prediction." The official prize summary records that Baker took a full half while Hassabis and Jumper each received a quarter, so the citation actually recognizes two related but distinct achievements rather than one. Baker directs the Institute for Protein Design at the University of Washington; Hassabis is chief executive of Google DeepMind and Jumper is a research director there.

The narrow documented claim is precise and worth separating from the broad one. The committee did not certify that any particular predicted structure is correct, nor that AlphaFold has "solved biology." It recognized two research trajectories: designing proteins that do not occur in nature, and predicting the three-dimensional shape of natural proteins from their amino acid sequences. The wider historical judgment, that machine learning had become a first-class instrument of chemical discovery, follows from those trajectories but is a different order of claim. This report keeps the two apart throughout.

The fifty-year problem the prize addressed

The problem behind the structure-prediction half is old and specific. A protein is a linear chain of amino acids that folds into a compact three-dimensional shape, and that shape largely determines what the molecule does. Christian Anfinsen won the 1972 chemistry prize for showing that the folding is driven by the sequence itself, a conclusion he set out in his 1973 review, Principles that Govern the Folding of Protein Chains. If the sequence encodes the structure, then in principle the structure should be computable from the sequence alone.

In practice it was not. Cyrus Levinthal pointed out in 1969 that a modest protein has an astronomically large number of possible conformations, so folding cannot be a blind search and neither can prediction by brute enumeration. For decades the reliable way to obtain a structure was experimental: X-ray crystallography, nuclear magnetic resonance, and later cryo-electron microscopy, each slow and expensive. To measure progress on the computational side, John Moult founded the Critical Assessment of Structure Prediction (CASP) in 1994, a biennial blind trial in which groups predict structures that have been solved experimentally but not yet released. CASP is the reason the field could later make a credible claim of a step change: it is an adversarial benchmark, not a self-reported one.

What CASP14 actually measured

The evidence for the structure-prediction half is concentrated in one CASP round. At CASP14 in 2020, DeepMind's AlphaFold2 produced structures whose accuracy the assessors described as competitive with experiment for a majority of targets. The headline figure, reported in the group's peer-reviewed account, Highly accurate protein structure prediction with AlphaFold, was a median GDT (Global Distance Test) score of 92.4 out of 100 across all targets. On the backbone atoms of its predictions the paper reported a median error of about 0.96 ångström (roughly the width of an atom), against roughly 2.8 ångström for the next best method in the same round.

Two qualifications keep those numbers honest. GDT and root-mean-square deviation are geometric similarity scores against a reference structure, not a percentage of correctness or of biological function; a high GDT means the predicted backbone sits close to the experimental one, nothing more. And the gap over the field, though large, was measured on the CASP target set of a single year, so it describes performance on that distribution of proteins rather than a universal guarantee. What CASP14 established is that a machine-learning system had, on a blind benchmark, closed most of the distance to experimental accuracy for a broad set of single-chain proteins. That is a strong, specific, and independently assessed result.

How AlphaFold2 arrives at a structure

The mechanism matters because it explains both the accuracy and the limits. AlphaFold2 does not simulate the physics of folding. It reads evolutionary information: for a query sequence it builds a multiple sequence alignment of related proteins across species, because positions that mutate together tend to be in contact in the folded structure. This coevolution signal, exploited for years before deep learning, is fed into a neural network the authors call the Evoformer, which reasons jointly over the sequence alignment and a representation of residue pairs, passing information back and forth between the two.

A second module, the structure module, then places every atom in three-dimensional space, treating each residue as a movable frame and refining the whole arrangement over repeated passes, including a "recycling" step that feeds its own output back as input. Crucially, the system emits calibrated confidence estimates alongside the coordinates: a per-residue score called pLDDT and a pairwise "predicted aligned error" that flags which regions and which relative positions are trustworthy. The model was trained on roughly 170,000 experimentally determined structures from the Protein Data Bank, the archive that has collected such structures since 1971. Because the method leans on evolutionary alignments and learned patterns from known structures, it is strongest where those signals are rich and weakest where they are thin, a dependency that reappears in the limitations below.

From a model to a database

A single accurate model is a result; a public resource is what changed daily scientific practice. In partnership with EMBL's European Bioinformatics Institute, DeepMind used AlphaFold2 to predict structures at scale and released them openly through the AlphaFold Protein Structure Database. The infrastructure and its first release are documented in AlphaFold Protein Structure Database: massively expanding the structural coverage of protein-sequence space. Within about a year the database grew from an initial set to over 200 million predicted structures, covering essentially the whole of the UniProt reference collection, which is to say nearly every catalogued protein sequence known to science.

The distinction between a prediction and a fact is the load-bearing one here. Each entry is a hypothesis with a confidence map attached, not an experimentally verified structure, and the interface makes that explicit through the pLDDT colouring. But a hypothesis good enough to guide an experiment is enormously useful when the experiment itself takes months. The database is the reason the CASP14 result propagated so quickly: researchers who would never run the model themselves could look up a molecule of interest, weigh the confidence, and design their next step around it. DeepMind's own account of the milestone, AlphaFold: a solution to a 50-year-old grand challenge in biology, frames the achievement in exactly those terms while stopping short of claiming the folding problem is closed in every case. Judging where a model's biology holds is now paid work in its own right: AI labs pay biologists by the hour to check a model's experimental reasoning, and the biology roles on AITraining.jobs, a jobs board from Chronicle's team, show who is hiring.

Baker's half: designing proteins that never existed

The other half of the prize runs in the opposite direction. Structure prediction asks what shape a given sequence will take; protein design asks which sequence will fold into a shape you specify, and ideally perform a function nature never evolved. David Baker's group built the Rosetta software over decades to attack this inverse problem with a physics-based energy function and search. The proof that de novo design could work at atomic precision came in 2003 with Top7, reported in Design of a Novel Globular Protein Fold with Atomic-Level Accuracy: a protein with a fold not found in nature, designed on a computer and then confirmed by X-ray crystallography to match the intended structure closely.

Baker's more recent work absorbed the same deep-learning wave that produced AlphaFold. His lab's RoseTTAFold, a "three-track" network published in Science in 2021 soon after AlphaFold2, showed the structure-prediction advance was reproducible outside DeepMind and extended it toward protein complexes. For design specifically, the group turned to generative diffusion models with De novo design of protein structure and function with RFdiffusion in 2023, which generates novel backbones for targets such as binders and enzymes that are then validated in the laboratory. The committee's decision to give design a full half, equal to the two structure-prediction laureates combined, signals that inventing new proteins is treated as a peer achievement to predicting existing ones, not a footnote to it.

The competing approaches of its era

AlphaFold2 did not appear in an empty field, and reading the prize as a lone breakthrough understates the surrounding work it depended on and competed with. The coevolution idea it exploits was developed by statistical-genetics and biophysics groups through direct-coupling analysis long before 2020. Physics-based molecular dynamics offered a wholly different route, simulating folding from force fields rather than learning from alignments, powerful for dynamics but generally too costly for routine whole-structure prediction. DeepMind's own first-generation AlphaFold had already won CASP13 in 2018 using a different, distogram-based design, so AlphaFold2 was itself a second act.

The contemporaneous alternatives are the fairest yardstick. RoseTTAFold, from the Baker lab, reached comparable quality with an independent architecture, evidence that the result was a property of the approach rather than one team's secret. A distinct line dispensed with alignments altogether: Meta's Evolutionary-scale prediction of atomic-level protein structure with a language model (ESMFold, 2023) used a protein language model to predict structure from a single sequence, trading some accuracy for speed on very large collections. These are different bets about where the signal lives (in alignments, in physics, or in a learned language of proteins), and the prize landed on the alignment-and-learning approach because that is what cleared the CASP bar first and at scale.

The documented downstream influence

The influence is measurable rather than rhetorical. The open database has been accessed by well over a million researchers, and predicted structures now routinely seed work on enzyme engineering, antibiotic resistance, and drug targets for neglected diseases, where an experimental structure was previously the rate-limiting step. The method also kept advancing: in 2024 DeepMind and Isomorphic Labs published Accurate structure prediction of biomolecular interactions with AlphaFold 3, extending prediction from single proteins to complexes with DNA, RNA, ions, and small-molecule ligands, the interactions that matter most for medicine.

The honest way to state the influence is as a change in the default workflow, not as a replacement of experiment. Cryo-EM, crystallography, and NMR remain the arbiters of truth, and were themselves recognized by the 2017 chemistry prize; what changed is that a fast, openly available prediction now precedes and guides those experiments for a very large fraction of proteins. That is a durable shift in how structural biology is practised, and it is the strongest evidence that the CASP14 result was more than a benchmark victory.

What the prize does not certify

Several caveats are intrinsic to the methods and should temper any grand reading. AlphaFold2 predicts a single static structure, typically the folded native state, and says little about folding pathways, dynamics, or the multiple conformations many proteins adopt as they function. Its accuracy degrades for proteins with few evolutionary relatives, for intrinsically disordered regions that have no fixed shape, and for the effects of individual point mutations, since a single substitution barely changes the input alignment. The grounding note for this event puts it plainly: the award does not validate every prediction or medical application, and design and prediction, though related, are separate contributions.

Two further cautions apply to design and to interpretation. A designed protein that looks correct in silico still has to fold and function in the laboratory, which is why Baker's results are reported with experimental verification rather than model scores alone. And a high confidence score is a statement about geometric reliability, not about biological correctness or safety; the same generative tools that design useful binders raise legitimate biosecurity questions the field is actively governing. Treating a confident prediction as a finished fact is the most common misreading of the whole enterprise, and the confidence metrics exist precisely to prevent it.

How to place a machine-learning result in the Nobel record

The 2024 chemistry prize is best read as the point at which a computational method met the Nobel's traditional standard of demonstrated, checkable impact, rather than as a claim that AI displaced experiment. Its narrow basis is concrete and independently assessed: a CASP14 performance with a median GDT of 92.4, a peer-reviewed method, an open database of more than 200 million structures, and a parallel body of design work verified in the laboratory. Its wider significance is that these were judged by chemists, against a benchmark chemists trust, and found to change practice.

A reader placing the event in AI history can navigate outward with three questions. Which older idea was necessary (Anfinsen's thermodynamic hypothesis, the coevolution signal, decades of accumulated PDB structures)? Which contemporaries showed the result was general rather than singular (RoseTTAFold, and later ESMFold)? And which limitations still demand new work (dynamics, disorder, mutation effects, and the move to complexes that AlphaFold 3 only begins to address)? Held that way, the prize marks a genuine turning point in scientific tooling without collapsing into the myth that one model solved a fifty-year problem outright. It reorganized how structural biology begins; it did not end the questions that follow.

References

• Royal Swedish Academy of Sciences (2024). The Nobel Prize in Chemistry 2024: Summary.
• Jumper et al. (2021). Highly accurate protein structure prediction with AlphaFold.
• Varadi et al. (2022). AlphaFold Protein Structure Database: massively expanding the structural coverage of protein-sequence space.
• Anfinsen (1973). Principles that Govern the Folding of Protein Chains.
• Kuhlman et al. (2003). Design of a Novel Globular Protein Fold with Atomic-Level Accuracy.
• Baek et al. (2021). Accurate prediction of protein structures and interactions using a three-track neural network.
• Watson et al. (2023). De novo design of protein structure and function with RFdiffusion.
• Lin et al. (2023). Evolutionary-scale prediction of atomic-level protein structure with a language model.
• Abramson et al. (2024). Accurate structure prediction of biomolecular interactions with AlphaFold 3.
• DeepMind (2020). AlphaFold: a solution to a 50-year-old grand challenge in biology.
• Protein Structure Prediction Center (2020). CASP14: Critical Assessment of Structure Prediction.
• EMBL-EBI and Google DeepMind (2022). AlphaFold Protein Structure Database.
• RCSB (1971). The Protein Data Bank.