AlphaFold and the Protein Folding Problem

For half a century the protein folding problem was the most famous unsolved question in molecular biology. The question was simple to state and impossible to answer. Given the linear sequence of amino acids that makes up a protein chain, how does that chain fold into a unique three-dimensional shape that lets the protein do its work. The chains are long. The behaviour is fast. The shapes are intricate. The answer, written into the chemistry itself, must be there. The problem was to extract it from a number whose size dwarfs any quantity that human intuition can manage. Decades of work produced a generation of skilled laboratories, a mountain of small datasets, and three useful principles. They did not, until very recently, produce a general solution. Then a British artificial-intelligence laboratory called DeepMind, originally a London start-up founded in 2010 and acquired by Google in 2014, did something that the protein-folding community had spent years assuming was impossible. In two CASP competitions, in 2018 and 2020, their system went from world-class to essentially perfect. Six years later, in 2024, the work earned the Nobel Prize in Chemistry. The story of how that happened is the story of a fifty-year scientific problem and a particular kind of artificial neural network called a transformer applied to one of biology most intractable computational problems.

The problem and a thirty-year competition

The protein folding problem is older than computing. It had been suspected since the 1930s, and demonstrated by protein x-ray crystallography in the 1950s and 60s, that proteins adopt unique three-dimensional structures. The full statement of the problem came in the early 1960s through the work of two American biochemists, Christian Anfinsen and Cyrus Levinthal, working from independent observations. Anfinsen demonstrated that a polypeptide chain, when left to itself, will fold back into its native shape. He won the 1972 Nobel Prize for that observation. Levinthal attacked the problem from the other direction. He computed the number of possible conformations that a single polypeptide chain could in principle occupy, a number that grows exponentially with the chain length, and observed that the chain folds into the right one within microseconds. A protein with one hundred amino acids might, by an exhaustive enumeration, take longer than the age of the universe to find the right conformation by trial and error. Yet real proteins fold in a fraction of a second. The contradiction was named Levinthal paradox. The eventual solution would require either an empirical model of the kinetics of folding or a statistical sampling of the energetics of folding. The first path led to folding simulations. The second path led to structural genomics.

The community chose to organise itself around a competition. The Critical Assessment of protein Structure Prediction, or CASP, was set up in 1994 by John Moult, a structural biologist then at the University of Maryland. The format was unusual for biology. A small committee would collect the experimental structures that other groups had just finished solving, kept them secret, and feed the sequences of those structures to prediction groups around the world. The predictors would send back their best guesses. Six months later, the experimental structures would be released, and the predictions would be compared against them. CASP became the Olympic games of protein structure prediction. Every two years, between fifty and one hundred groups would enter. The early competitions rewarded careful physics-based modelling, carefully tuned force fields, and a great deal of computational patience. None of the methods coming out of CASP were reliable for proteins whose closest relative had not already been solved experimentally.

The 1990s and 2000s produced a steady refinement of the available prediction methods. The dominant techniques combined, by the early 2010s, fragment-based assembly from the database of known protein structures with energy minimisation under a force field. A representative tool of that era was Rosetta, developed in David Baker laboratory at the University of Washington. Rosetta and similar tools could, given careful tuning and a high-quality multiple sequence alignment, often produce useful predictions for small proteins whose close structural relatives were in the database. For arbitrary proteins, the predictions were unreliable. The CASP competitions chronicled that limitation year after year. The median accuracy, measured on the GDT-TS scale that runs from zero to one hundred, hovered between forty and fifty through the 2000s. A score of fifty meant that roughly half of the predicted backbone atoms were within a few angstroms of their experimentally correct positions. For many applications, that was not good enough.

DeepMind enters CASP

DeepMind was founded in 2010 by Demis Hassabis, Shane Legg, and Mustafa Suleyman. Its early work was in reinforcement learning for board games and Atari video games. The protein folding effort began in the lab around 2016. John Jumper, the biochemist who would lead the AlphaFold 2 technical work, joined DeepMind in 2017 from the D. E. Shaw Research group where he had been working on a molecular dynamics approach to the same problem. The first paper was modest in ambition. It described a convolutional residual network trained on the Protein Data Bank to predict inter-residue distances and torsion angles. The submission to CASP13 in late 2018 was a system called AlphaFold. It came first in the competition. Its median GDT-TS score, on the held-out test set of ninety-eight proteins, was 58.9, where the next best method from any other group reached 41.7. The improvement was large. It was not a solution. The improvement was large enough, however, that the CASP organisers invited the DeepMind team back for the next round.

What changed between AlphaFold 1 and AlphaFold 2 is what mattered. AlphaFold 1 had used convolutional networks, hand-engineered distance potentials, and a final optimisation pass that used gradient descent in Cartesian coordinates to relax the predicted structure into a chemically reasonable one. AlphaFold 2 threw most of that away. The new system was an end-to-end attention-based neural network that produced, from the multiple sequence alignment of a protein chain and a template-free representation of the available structural information, a per-residue pair representation that could be turned directly into a three-dimensional structure. The architecture was driven by a module called the Evoformer, an evolution-aware transformer that the team described as the heart of the model. The model also used iterative refinement. Prediction and structure passed back and forth through the network for three rounds, each round refining a learned representation of the structure before producing a final prediction. The system was trained on a curated subset of the Protein Data Bank, plus synthetic samples produced by distillation from earlier versions of the same network. Training took several weeks on a few hundred tensor processing units. Inference, on a single protein, took a few seconds to a few minutes depending on length.

CASP14, in late 2020, settled the question. AlphaFold 2 produced predictions with a median GDT-TS of 92.4 on the ninety-seven protein targets. Where AlphaFold 1 had matched the best conventional methods at fifty-eight GDT, AlphaFold 2 reached ninety-two. A score of ninety corresponds to predictions whose backbone atoms are within roughly one angstrom of their correct positions for most residues. The result was so far ahead of the rest of the field that the CASP14 assessor, Andriy Kryshtafovych, wrote that the problem had, for practical purposes, been solved. The headline in Nature that month was simple. The protein folding problem had been solved, after fifty years. (The phrasing was, characteristically, contested in some scientific circles, but the underlying result was unprecedented.) The paper describing AlphaFold 2, led by John Jumper and his colleagues, appeared in Nature in July 2021. The author list ran to over thirty names. DeepMind moved the code and the model weights over the summer of 2021 to a managed release for non-commercial use and to an open-source license for academic use, paired with an open database of predicted structures.

The open database and the Nobel

Between 2021 and 2024 the AlphaFold predictions moved from a single demonstration into the infrastructure of biology. DeepMind partnered with the European Bioinformatics Institute (EMBL-EBI) at Hinxton, near Cambridge, and began populating the AlphaFold Database with structural predictions for every protein in the reviewed eukaryotic, bacterial, and archaeal proteomes of the UniProt reference set. The first release, in July 2021, contained 350,000 structures. The second release, expanded coverage in late 2022, contained over 200 million structures, one for almost every protein sequence in the databases of the world. The database was free for any non-commercial use. By late 2023 the database held over 250 million structures, with new entries added through automated pipelines rather than batched releases. The accompanying AlphaFold Server, a hosted inference platform, let researchers submit a sequence and receive a prediction in seconds. A separate code release, AlphaFold-Multimer, extended the prediction to protein-protein complexes, the essential pair-to-pair interactions that drive so much of cellular biology. The infrastructure scale of the work had, by 2022, ceased to be a scientific problem and become a logistics problem.

The 2024 Nobel Prize in Chemistry was awarded jointly to Demis Hassabis and John Jumper, for the AlphaFold computational work, and to David Baker of the University of Washington, for the parallel development of a complementary system, RoseTTAFold, that emerged out of the same CASP14 cycle and that the Baker laboratory had published in Science in July 2021, three weeks after the Jumper paper in Nature. The parallel histories are an instructive detail. Baker and his laboratory had been working on protein structure prediction for nearly twenty years before AlphaFold 2. RoseTTAFold reached nearly the same level of accuracy as AlphaFold 2 using a different neural architecture (a three-track network combining one-dimensional sequence information, two-dimensional distance information, and three-dimensional coordinate information). Baker lab also released the code and weights for RoseTTAFold on the same timeframe as the AlphaFold 2 release. The Baker laboratory, in the years that followed, would go on to design entirely new proteins from scratch using their own generative network RFdiffusion, and would design miniproteins that neutralise the SARS-CoV-2 spike protein. AlphaFold 2 and RoseTTAFold were, by 2024, two different keys to the same door, plus a third (ESMFold, released by Meta in late 2022 from the EvolutionaryScale team). The award recognised both. The Nobel assembly is not given for a single method. It is given for a body of work. The 2024 prize was given for a body of work, distributed across the two camps, that had changed the way structural biology is done.

What AlphaFold has changed

The list of things AlphaFold has enabled is long enough that I will compress it. The first category is drug discovery. Knowing the three-dimensional structure of a protein target is a starting point for designing a small molecule that binds to it. Industrial pharmacology had, before 2021, typically paid to crystallise or freeze-dry their targets in the laboratory, a process that can take years per target. By 2024 pharmaceutical companies were using AlphaFold predictions to triage targets, to construct initial binding hypotheses, and to skip the experimental work for targets whose structure is clearly predictable. The pharma company that worked most publicly with AlphaFold was Isomorphic Labs, founded by Demis Hassabis and its then-president Colin Murdoch, which uses AlphaFold and related systems for drug design work. The second category is vaccine science. The Baker laboratory had designed miniproteins that neutralise the spike proteins of SARS-CoV-2. The Malaria vaccine RTS,S was developed without AlphaFold and the structure-based design of the next-generation malaria vaccines used AlphaFold-style structural prediction as a starting point. The third category is enzyme engineering. Industrial enzymes are now routinely designed by predicting the structure of a candidate sequence and screening for substrate complementarity. The fourth category is rare-disease biology. The structure of a missense variant of a protein can, in many cases, be predicted. Predicted impacts on folding, stability, and binding give an inference from a single genomic observation that previously would have required experimental structure determination.

There are limits that the technique does not yet cover. The AlphaFold prediction is essentially a single static snapshot. It does not, in the standard release, capture the dynamic conformations that proteins adopt in the cell. Where a protein operates by switching between conformations (the G-protein-coupled receptors, the ribosome, the spliceosome, many of the immune-system receptors), the predicted structure gives the dominant state but not the regulatory motions. AlphaFold-Multimer handles pairwise complexes but not the larger assemblies that drive most cellular functions. The intrinsically disordered proteins, which make up roughly a third of eukaryotic proteomes, are not really proteins in the AlphaFold sense at all; AlphaFold will produce a structure where the protein actually has no stable structure in solution. The community understands these limitations well and uses the predictions accordingly. The AlphaFold Server released in 2024 added automated confidence metrics that flag low-confidence regions, and that flag, for any given submission, the cases where the prediction is not reliable.

The second-order effects of AlphaFold on biology are still being worked out. The first-order effect, the speed with which structural biology can now produce protein models, has changed the day-to-day practice of structural prediction in most academic laboratories. The second-order effect on biological discovery, on the kinds of question that biologists are now asking because a model is always available, is still in progress. By 2024, in many research laboratories, computing a protein structure was no longer a research bottleneck. It was a routine step in hypothesis generation. The bottleneck had moved to other things, including the experimental validation of structural predictions and the integration of structural information into the broader cellular context. None of these changes would have been possible without the engine behind the AlphaFold predictions, the most aggressive application to date of large-scale transformer architectures to a scientific problem that had defeated every previous computational approach.

The history of the protein folding problem, as a scientific problem, can now be told in three phases. The first phase, from the 1950s to the early 1990s, was the experimental phase. The second phase, from the early 1990s to the mid 2010s, was the comparative-modelling phase. The third phase, from the mid 2010s onwards, has been the deep-learning phase. The transition from phase two to phase three did not, technically, finish the protein folding problem. The problem in its full form includes the question of how the dynamic motions of a protein emerge from its sequence. That part of the problem is still open. What AlphaFold 2 did was finish the static-structure problem, which had been the part of the problem that the structural-biology community had tried, in vain, to solve for thirty years. The static-structure problem is, for most biological purposes, the more important of the two. The dynamic problem is also, by 2024, becoming a target for an active new generation of machine-learning approaches that go under the names of AlphaFold-Multimer, ESM-3, RFdiffusion, and others. None of those approaches have the same accuracy AlphaFold 2 has on the static problem. The protein folding field is not solved. It is, for the first time in fifty years, tractable.

The story has not ended. The CASP competitions continue, with new categories for protein-protein complexes, disordered proteins, and protein-ligand interactions. The databases continue to expand. The 2024 Nobel did not end the field. It closed a long chapter and opened several new ones. Anyone working in biology in 2025 can, with a single web browser, submit a sequence and receive a structural prediction that would have taken the combined effort of a major laboratory a year to produce in 1995. That change is permanent. The protein folding problem will continue to be discussed for a long time to come. The expectation that a single structural prediction will cost a few pence worth of computer time and arrive in seconds will, from now on, be a baseline expectation, not a research achievement.

References

1. Anfinsen, Christian B. “Principles That Govern the Folding of Protein Chains.” Science 181, no. 4096 (1973): 223-230.
2. Levinthal, Cyrus. “How to Fold Graciously.” In Mossbauer Spectroscopy in Biological Systems: Proceedings, edited by J. T. Edsall, 22-24. Urbana: University of Illinois Press, 1969.
3. John Moult, F. and colleagues, eds. “CASP: Critical Assessment of Techniques for Protein Structure Prediction.” Proteins: Structure, Function, and Bioinformatics, supplement issues from 1995 onwards.
4. Senior, Andrew W. and others. “Improved Protein Structure Prediction Using Potentials From Deep Learning.” Nature 577 (2020): 706-710.
5. Jumper, John and others. “Highly Accurate Protein Structure Prediction With AlphaFold.” Nature 596 (2021): 583-589.
6. Baek, Minkyung and others. “Accurate Prediction of Protein Structures and Interactions Using a Three-Track Neural Network.” Science 373, no. 6557 (2021): 871-876.
7. Varadi, Mihaly and others. “AlphaFold Protein Structure Database: Massively Expanding the Structural Coverage of Protein Sequences at the Proteome Scale.” Nucleic Acids Research 50 (2022): D439-D444.
8. Evans, Richard and others. “Protein Complex Prediction with AlphaFold-Multimer.” bioRxiv preprint, March 2022, doi 10.1101/2021.10.04.463034.
9. Lin, Zeming and others. “Evolutionary-Scale Prediction of Atomic-Level Protein Structure With a Language Model.” Science 379, no. 6637 (2023): 1123-1130.
10. Watson, James L. and others. “Broadly Applicable and Generalised Protein Design With De Novo Diffusion.” Nature 620 (2023): 1089-1100.
11. Kryshtafovych, Andriy and others. “Critical Assessment of Methods of Protein Structure Prediction (CASP) Round XV.” Proteins: Structure, Function, and Bioinformatics 91 (2023): 1539-1549.
12. Royal Swedish Academy of Sciences. Press Release: The Nobel Prize in Chemistry 2024. Stockholm, 9 October 2024.
13. DeepMind. “AlphaFold Server.” https://alphafold.ebi.ac.uk/.
14. EMBL-EBI. “AlphaFold Database.” https://www.alphafold.ebi.ac.uk/.
15. Protein Data Bank. “PDB Statistics.” https://www.rcsb.org/stats/.
16. Pereira, Joicymara and others. “High-Throughput Protein Structure Prediction at Scale.” Nature Biotechnology 41 (2023): 1610-1615.